Science 13.11.2020

RESEARCH | RESEARCH ARTICLE

66. M. T. Maurano et al., Systematic localization of common 88. A. Bery, B. Martynoga, F. Guillemot, J.-S. Joly, S. Rétaux, 111. S. T. Sherry et al., dbSNP: The NCBI database of genetic
disease-associated variation in regulatory DNA. Science 337, Characterization of enhancers active in the mouse embryonic variation. Nucleic Acids Res. 29, 308–311 (2001).
1190–1195 (2012). doi: 10.1126/science.1222794; cerebral cortex suggests Sox/Pou cis-regulatory logics and doi: 10.1093/nar/29.1.308; pmid: 11125122
pmid: 22955828 heterogeneity of cortical progenitors. Cereb. Cortex 24,
2822–2834 (2014). doi: 10.1093/cercor/bht126; 112. Y. Benjamini, Y. Hochberg, Controlling the false discovery
67. H. K. Finucane et al., Partitioning heritability by functional pmid: 23720416 rate: A practical and powerful approach to multiple testing.
annotation using genome-wide association summary J. R. Stat. Soc. B 57, 289–300 (1995). doi: 10.1111/
statistics. Nat. Genet. 47, 1228–1235 (2015). doi: 10.1038/ 89. S. E. McCandless, J. W. Brunger, S. B. Cassidy, The burden of j.2517-6161.1995.tb02031.x
ng.3404; pmid: 26414678 genetic disease on inpatient care in a children’s hospital.
Am. J. Hum. Genet. 74, 121–127 (2004). doi: 10.1086/ 113. W. J. Kent et al., The human genome browser at UCSC.
68. C. T. Harvey et al., QuASAR: Quantitative allele-specific 381053; pmid: 14681831 Genome Res. 12, 996–1006 (2002). doi: 10.1101/
analysis of reads. Bioinformatics 31, 1235–1242 (2015). gr.229102; pmid: 12045153
doi: 10.1093/bioinformatics/btu802; pmid: 25480375 90. M. H. Wojcik et al., Genetic disorders and mortality in infancy
and early childhood: Delayed diagnoses and missed 114. T. Stuart, timoast/signac. GitHub (2020); https://github.
69. A. Buniello et al., The NHGRI-EBI GWAS Catalog of published opportunities. Genet. Med. 20, 1396–1404 (2018). com/timoast/signac.
genome-wide association studies, targeted arrays and summary doi: 10.1038/gim.2018.17; pmid: 29790870
statistics 2019. Nucleic Acids Res. 47 (D1), D1005–D1012 (2019). 115. Broad Institute, Picard Tools; http://broadinstitute.github.io/picard.
doi: 10.1093/nar/gky1120; pmid: 30445434 91. W. Meuleman et al., Index and biological spectrum of human 116. J. A. Castro-Mondragon, S. Jaeger, D. Thieffry,
DNase I hypersensitive sites. Nature 584, 244–251 (2020).
70. G. Kichaev et al., Leveraging polygenic functional enrichment doi: 10.1038/s41586-020-2559-3; pmid: 32728217 M. Thomas-Chollier, J. van Helden, RSAT matrix-clustering:
to improve GWAS power. Am. J. Hum. Genet. 104, 65–75 Dynamic exploration and redundancy reduction of
(2019). doi: 10.1016/j.ajhg.2018.11.008; pmid: 30595370 92. M. R. Corces et al., Lineage-specific and single-cell chromatin transcription factor binding motif collections. Nucleic Acids
accessibility charts human hematopoiesis and leukemia Res. 45, e119 (2017). doi: 10.1093/nar/gkx314;
71. N. Tanimura et al., GATA/heme multi-omics reveals a trace evolution. Nat. Genet. 48, 1193–1203 (2016). doi: 10.1038/ pmid: 28591841
metal-dependent cellular differentiation mechanism. Dev. Cell ng.3646; pmid: 27526324 117. L. Navarro-Núñez, S. A. Langan, G. B. Nash, S. P. Watson,
46, 581–594.e4 (2018). doi: 10.1016/j.devcel.2018.07.022; The physiological and pathophysiological roles of platelet
pmid: 30122630 93. G. J. Pruijn, W. van Driel, P. C. van der Vliet, Nuclear factor III, CLEC-2. Thromb. Haemost. 109, 991–998 (2013).
a novel sequence-specific DNA-binding protein from doi: 10.1160/TH13-01-0060; pmid: 23572154
72. B. Zhao et al., Genome-wide association analysis of 19,629 HeLa cells stimulating adenovirus DNA replication. Nature 322,
individuals identifies variants influencing regional brain 656–659 (1986). doi: 10.1038/322656a0; pmid: 3748145 ACKNOWLEDGMENTS
volumes and refines their genetic co-architecture with
cognitive and mental health traits. Nat. Genet. 51, 1637–1644 94. D. Tantin, C. Schild-Poulter, V. Wang, R. J. G. Haché, We thank past and present members of the Shendure, Trapnell,
(2019). doi: 10.1038/s41588-019-0516-6; pmid: 31676860 P. A. Sharp, The octamer binding transcription factor Oct-1 is and Cusanovich laboratories; A. Adey; the Brotman Baty Institute
a stress sensor. Cancer Res. 65, 10750–10758 (2005). for Precision Medicine (BBI) Advanced Technology Laboratory;
73. F. Tissir, A. M. Goffinet, Reelin and brain development. Nat. doi: 10.1158/0008-5472.CAN-05-2399; pmid: 16322220 and the University of Washington Birth Defects Research
Rev. Neurosci. 4, 496–505 (2003). doi: 10.1038/nrn1113; Laboratory. We thank C. Spurrell, D. Ahrendsen, and R. Blecher
pmid: 12778121 95. V. E. H. Wang, T. Schmidt, J. Chen, P. A. Sharp, D. Tantin, for reviewing the sci-ATAC-seq3 protocol. We thank C. Cenik and
Embryonic lethality, decreased erythropoiesis, and defective Z. Ma from M. Snyder’s laboratory for providing the CH12.LX
74. A. N. Schep, B. Wu, J. D. Buenrostro, W. J. Greenleaf, octamer-dependent promoter activation in Oct-1-deficient cell line. We thank R. Pique-Regi for advice on implementing QuASAR.
chromVAR: Inferring transcription-factor-associated accessibility mice. Mol. Cell. Biol. 24, 1022–1032 (2004). doi: 10.1128/ We acknowledge the ENCODE Consortium for access to their data.
from single-cell epigenomic data. Nat. Methods 14, 975–978 MCB.24.3.1022-1032.2004; pmid: 14729950 Funding: S.D. was supported by an EMBO Long-Term Fellowship
(2017). doi: 10.1038/nmeth.4401; pmid: 28825706 and a HFSP Long-Term Fellowship. D.A.C. was supported in part
96. D. Lee et al., A method to predict the impact of regulatory by an NHLBI fellowship (T32HL007828). Aspects of this work were
75. M. W. Dorrity, L. M. Saunders, C. Queitsch, S. Fields, variants from DNA sequence. Nat. Genet. 47, 955–961 supported by funding from the BBI, the Paul G. Allen Frontiers
C. Trapnell, Dimensionality reduction by UMAP to visualize (2015). doi: 10.1038/ng.3331; pmid: 26075791 Foundation (Allen Discovery Center grant to J.S. and C.T.), The
physical and genetic interactions. Nat. Commun. 11, 1537 Chan Zuckerberg Initiative (to C.T.) and the NIH (HD000836 to
(2020). doi: 10.1038/s41467-020-15351-4; pmid: 32210240 97. S. Domcke, A. J. Hill, R. M. Daza, C. Trapnell, I.A.G.). J.S. is an investigator of the Howard Hughes Medical Institute.
D. A. Cusanovich, J. Shendure, sci-ATAC-seq3. protocols.io Authors contributions: R.M.D., S.D., and D.A.C. developed
76. E. Eisenberg, E. Y. Levanon, Human housekeeping genes, (2020). doi: 10.17504/protocols.io.be8mjhu6 techniques and performed sci-ATAC-seq3 experiments with
revisited. Trends Genet. 29, 569–574 (2013). doi: 10.1016/ assistance from F.Z., D.P., F.J.S., and J.H.M.; D.R.O. performed
j.tig.2013.05.010; pmid: 23810203 98. M. R. Corces et al., An improved ATAC-seq protocol reduces tissue collection and nuclei extraction under supervision of D.D.
background and enables interrogation of frozen tissues. and I.A.G.; S.D. and A.J.H. analyzed the data with assistance from
77. J. R. Dixon et al., Topological domains in mammalian Nat. Methods 14, 959–962 (2017). doi: 10.1038/nmeth.4396; D.A.C. and H.A.P.; J.C. provided sci-RNA-seq data; K.A.A.
genomes identified by analysis of chromatin interactions. pmid: 28846090 performed genotyping; M.A.Z. developed the website, with
Nature 485, 376–380 (2012). doi: 10.1038/nature11082; assistance from S.D.; S.D., D.A.C., and J.S. wrote the manuscript
pmid: 22495300 99. A. Saunders, L. J. Core, C. Sutcliffe, J. T. Lis, H. L. Ashe, Extensive with input from all co-authors; J.S., D.A.C., and C.T. supervised
polymerase pausing during Drosophila axis patterning enables the project. Competing interests: D.P., F.Z., and F.J.S. declare
78. S. S. P. Rao et al., A 3D map of the human genome at high-level and pliable transcription. Genes Dev. 27, 1146–1158 competing financial interests in the form of stock ownership and
kilobase resolution reveals principles of chromatin looping. (2013). doi: 10.1101/gad.215459.113; pmid: 23699410 paid employment by Illumina. J.S. has competing financial
Cell 159, 1665–1680 (2014). doi: 10.1016/j.cell.2014.11.021; interests (paid consulting and/or equity) with Guardant Health,
pmid: 25497547 100. S. Domcke, A. Hill, shendurelab/human-atac: First release of Maze Therapeutics, Camp4 Therapeutics, Nanostring, Phase
sci-ATAC-seq3 demultiplexing. Zenodo (2020). doi: 10.5281/ Genomics, Adaptive Biotechnologies, and Stratos Genomics. One
79. T. H. Kim et al., Analysis of the vertebrate insulator protein zenodo.4012524 or more embodiments of one or more patents and patent
CTCF-binding sites in the human genome. Cell 128, applications filed by Illumina and the University of Washington may
1231–1245 (2007). doi: 10.1016/j.cell.2006.12.048; 101. Y. Zhang et al., Model-based analysis of ChIP-Seq (MACS). encompass the methods, reagents, and data disclosed in this
pmid: 17382889 Genome Biol. 9, R137 (2008). doi: 10.1186/gb-2008-9-9-r137; manuscript. Data and materials availability: In contrast with
pmid: 18798982 previous versions of sci-ATAC-seq, custom Tn5 is not required
80. R. M. Gittelman et al., Comprehensive identification and for sci-ATAC-seq3. A detailed version of the sci-ATAC-seq3
analysis of human accelerated regulatory DNA. Genome Res. 102. S. L. Wolock, R. Lopez, A. M. Klein, Scrublet: Computational protocol is available on protocols.io (97). A demultiplexing script
25, 1245–1255 (2015). doi: 10.1101/gr.192591.115; identification of cell doublets in single-cell transcriptomic and tutorial are provided on Zenodo at (100). Supplementary
pmid: 26104583 data. Cell Syst. 8, 281–291.e9 (2019). doi: 10.1016/ data files S1 to S7 are provided at GEO (accession no. GSE149683).
j.cels.2018.11.005; pmid: 30954476 Raw data are provided at dbGaP (accession no. phs002003.v1.p1).
81. A. Visel, S. Minovitsky, I. Dubchak, L. A. Pennacchio, VISTA Processed data are available at GEO (accession no. GSE149683
Enhancer Browser—A database of tissue-specific human 103. C. Bravo González-Blas et al., cisTopic: Cis-regulatory topic (both unfiltered and peak module-based filtered dataset)). An
enhancers. Nucleic Acids Res. 35 (Database), D88–D92 modeling on single-cell ATAC-seq data. Nat. Methods 16, interactive version of the peak module–based filtered dataset
(2007). doi: 10.1093/nar/gkl822; pmid: 17130149 397–400 (2019). doi: 10.1038/s41592-019-0367-1; facilitates the exploration of these data by tissue, cell type, locus,
pmid: 30962623 or motif at descartes.brotmanbaty.org (15).
82. D. C. King et al., Finding cis-regulatory elements using
comparative genomics: Some lessons from ENCODE data. 104. R. Fang et al., Fast and accurate clustering of single cell SUPPLEMENTARY MATERIALS
Genome Res. 17, 775–786 (2007). doi: 10.1101/gr.5592107; epigenomes reveals cis-regulatory elements in rare cell
pmid: 17567996 types. bioRxiv 615179 [Preprint] 13 May 2019). science.sciencemag.org/content/370/6518/eaba7612/suppl/DC1
Materials and Methods
83. E. Montecino-Rodriguez, K. Dorshkind, Formation of B-1 B 105. A. Butler, P. Hoffman, P. Smibert, E. Papalexi, R. Satija, Figs. S1 to S19
cells from neonatal B-1 transitional cells exhibits NF-kB Integrating single-cell transcriptomic data across different References (118–124)
redundancy. J. Immunol. 187, 5712–5719 (2011). conditions, technologies, and species. Nat. Biotechnol. 36, Tables S1 to S7
doi: 10.4049/jimmunol.1102416; pmid: 22031760 411–420 (2018). doi: 10.1038/nbt.4096; pmid: 29608179
View/request a protocol for this paper from Bio-protocol.
84. A. T. Satpathy et al., Massively parallel single-cell chromatin 106. O. Fornes et al., JASPAR 2020: Update of the open-access
landscapes of human immune cell development and database of transcription factor binding profiles. Nucleic 3 January 2020; accepted 10 September 2020
intratumoral T cell exhaustion. Nat. Biotechnol. 37, 925–936 Acids Res. 48, D87–D92 (2020). doi: 10.1093/nar/gkz1001; 10.1126/science.aba7612
(2019). doi: 10.1038/s41587-019-0206-z; pmid: 31375813 pmid: 31701148

85. C. Luo et al., Single nucleus multi-omics links human cortical 107. A. Zeisel et al., Molecular architecture of the mouse nervous
cell regulatory genome diversity to disease risk variants. system. Cell 174, 999–1014.e22 (2018). doi: 10.1016/
bioRxiv 873398 [Preprint] 12 December 2019; .doi: 10.1101/ j.cell.2018.06.021; pmid: 30096314
2019.12.11.873398
108. https://data.broadinstitute.org/alkesgroup/UKBB/UKBB_409K
86. Cell Types Database, RNA-Seq Data—brain-map.org; 109. D. Kim, J. M. Paggi, C. Park, C. Bennett, S. L. Salzberg,
https://portal.brain-map.org/atlases-and-data/
rnaseq#Human_Cortex. Graph-based genome alignment and genotyping with HISAT2
and HISAT-genotype. Nat. Biotechnol. 37, 907–915 (2019).
87. D. Polioudakis et al., A single-cell transcriptomic atlas of doi: 10.1038/s41587-019-0201-4; pmid: 31375807
human neocortical development during mid-gestation. 110. H. Li et al., The Sequence Alignment/Map format and
Neuron 103, 785–801.e8 (2019). doi: 10.1016/ SAMtools. Bioinformatics 25, 2078–2079 (2009).
j.neuron.2019.06.011; pmid: 31303374 doi: 10.1093/bioinformatics/btp352; pmid: 19505943

Domcke et al., Science 370, eaba7612 (2020) 13 November 2020 15 of 15

RESEARCH

◥ plex and found that a large portion of the target
genes are expressed in, and are even restricted
RESEARCH ARTICLE SUMMARY to, the outermost root hair cells. This suggests
a possible involvement of the TMO5/LHW com-
PLANT SCIENCE plex in root hair development. In line with this
hypothesis, misexpression of TMO5 and LHW
Vascular transcription factors guide plant epidermal in the entire root meristem was found to in-
responses to limiting phosphate conditions crease root hair density, a phenotype resembling
wild-type roots grown under phosphate-limiting
Jos R. Wendrich*, BaoJun Yang*, Niels Vandamme*, Kevin Verstaen*, Wouter Smet, conditions. Moreover, the response of epider-
Celien Van de Velde, Max Minne, Brecht Wybouw, Eliana Mor, Helena E. Arents, Jonah Nolf, mal cells to phosphate deficit was shown to be
Julie Van Duyse, Gert Van Isterdael, Steven Maere, Yvan Saeys†, Bert De Rybel† TMO5-dependent. Phosphate-limiting condi-
tions were found to induce TMO5 expression,
INTRODUCTION: In plants, vascular tissues serve RATIONALE: By intersecting a high-resolution and genetically increasing TMO5 expression
a dual function by providing both structural single-cell gene expression atlas with TMO5/ only in the vascular bundle is sufficient to trig-
support and transport of water, nutrients, hor- LHW target genes, we probed the tissue-specific ger epidermal responses. These results indicate
mones, and other signaling molecules through- distribution of transcriptional responses upon that the TMO5-dependent increase in root hair
out the plant body. Proliferation of vascular induction of this transcription factor complex density under low-phosphate conditions occurs
tissues is, in part, controlled by the TARGET in the Arabidopsis root meristem. We next used through a cell nonautonomous effect.
OF MONOPTEROS 5/LONESOME HIGHWAY a combination of genetic and molecular tools
(TMO5/LHW) transcription factor complex and to validate the existence and importance of a TMO5/LHW-dependent vascular cytokinin
involves intricate cell communication through cytokinin-dependent signaling mechanism biosynthesis was found to be sufficient to increase
mobile factors. The activity of this heterodimer bridging the inner vascular tissues and the epidermal root hair density because mutants
complex is limited to young xylem cells and is outer epidermis cell layers. with reduced cytokinin levels are less sensitive to
required and sufficient to control vascular cell the effects of phosphate-limiting conditions on
proliferation by inducing expression of rate- RESULTS: We generated a single-cell gene ex- root hair density. Furthermore, phosphate-limiting
limiting enzymes in the cytokinin biosynthetic pression atlas that represents all known cell conditions increase cytokinin signaling in epi-
pathway. Because cytokinin is mobile, the mo- types and sub–cell types of the Arabidopsis dermal cells of wild-type roots, and treatment
lecular and developmental responses to this root apical meristem. Cell-type annotations with exogenous cytokinin can restore root hair
hormone signaling cascade are likely to have and developmental trajectories were validated densities back to wild-type levels in the absence
an effect in various cell types surrounding through expression of known marker genes, of TMO5 activity. This suggests that cytokinin
the xylem and perhaps even outside the vas- newly generated promoter reporter lines, and is the mobile signal linking TMO5/LHW activ-
cular bundle. To what extent this is the case analysis of the endoreduplication status in each ity in vascular cells to root hair density changes
and how this controls developmental pro- cell cluster. We next intersected this expression in response to phosphate deficit. The TMO5
cesses other than vascular proliferation re- atlas with a set of known target genes of the and cytokinin-dependent root hair density in-
main unknown. vascular TMO5/LHW transcription factor com- crease under low-phosphate conditions was
found to originate not only from changes in
scRNA-seq Developmental High Pi Low Pi epidermal cell length but also from changes in
trajectory inference cell identity because both phosphate-limiting
conditions and exogenous cytokinin treatment
were shown to scramble identities of the other-
wise highly organized epidermal cells.

in vivo validation High-resolution CONCLUSION: Through its effect on cytokinin
root tip atlas biosynthesis, the vascular TMO5/LHW hetero-
dimer complex controls epidermal root hair
Arabidopsis reporter lines density by modifying cell length and cell identity.
We hypothesize that phosphate-limiting condi-
Single-cell Whole-genome Root hair tions may trigger increased auxin signaling in
deconvolution transcriptomics density xylem cells, inducing activity of the TMO5/LHW
pathway and downstream local cytokinin bio-
Differentially mock DEX synthesis. Cytokinin may then diffuse outward
expressed from the vasculature to direct both length and
genes are Induction of fate of the outer trichoblast cells. This hor-
mapped to TMO5/LHW mone signaling cascade spans multiple tissue
distinct cell
populations ▪layers, allowing roots to efficiently forage the

Xylem Trichoblast soil for phosphate.

Cytokinin The list of author affiliations is available in the full article online.
gradient *These authors contributed equally to this work.
†Corresponding author. Email: [email protected]
Low-phosphate Root hair density (B.D.R.); [email protected] (Y.S.)
perception Cite this article as J. R. Wendrich et al., Science 370,
eaay4970 (2020). DOI: 10.1126/science.aay4970
Vascular transcription factors guide epidermal responses. A validated high-resolution single-cell gene
expression atlas of the Arabidopsis root was intersected with TMO5/LHW target genes, uncovering an enrichment of READ THE FULL ARTICLE AT
epidermal-restricted expression patterns. By activating local cytokinin biosynthesis, the vascular TMO5/LHW https://doi.org/10.1126/science.aay4970
complex was shown to regulate epidermal root hair density in response to the availability of phosphate. scRNA-seq,
single-cell RNA-sequencing; mock, untreated condition; DEX, dexamethasone; Pi, inorganic phosphate.

Wendrich et al., Science 370, 810 (2020) 13 November 2020 1 of 1

RESEARCH

◥ tated dataset representing all major cell types
in the root, including quiescent center cells
RESEARCH ARTICLE (Fig. 1A and fig. S1F). The annotations were
confirmed for all cell identities by the ob-
PLANT SCIENCE served expression of key signature marker
genes and in vivo expression of 41 newly gen-
Vascular transcription factors guide plant epidermal erated promoter reporter lines (Fig. 1B, figs.
responses to limiting phosphate conditions S2 to S12, and table S1). Moreover, cells un-
dergoing division were found in two specific
Jos R. Wendrich1,2*, BaoJun Yang1,2*, Niels Vandamme3,4*, Kevin Verstaen3,4*, Wouter Smet1,2, subclusters (Fig. 1), indicating that their tran-
Celien Van de Velde1,2, Max Minne1,2, Brecht Wybouw1,2, Eliana Mor1,2, Helena E. Arents1,2, scriptomes are more similar to each other
Jonah Nolf1,2, Julie Van Duyse5,6, Gert Van Isterdael5,6, Steven Maere1,2, than to the actual cell identity determinants
Yvan Saeys3,4†, Bert De Rybel1,2† of the transcriptome. The scRNA-seq dataset
contains clusters of all cell types of the root
Optimal plant growth is hampered by deficiency of the essential macronutrient phosphate in most soils. meristem, which were identified according
Plant roots can, however, increase their root hair density to efficiently forage the soil for this immobile to predicted in vivo expression patterns and
nutrient. By generating and exploiting a high-resolution single-cell gene expression atlas of Arabidopsis roots, validated using a set of newly generated re-
we show an enrichment of TARGET OF MONOPTEROS 5/LONESOME HIGHWAY (TMO5/LHW) target gene porter lines.
responses in root hair cells. The TMO5/LHW heterodimer triggers biosynthesis of mobile cytokinin
in vascular cells and increases root hair density during low-phosphate conditions by modifying both the Trajectory analysis establishes a blueprint
length and cell fate of epidermal cells. Moreover, root hair responses in phosphate-deprived conditions of cell lineages
are TMO5- and cytokinin-dependent. Cytokinin signaling links root hair responses in the epidermis to
perception of phosphate depletion in vascular cells. We next used trajectory analyses to refine
identification of cell types and developmen-
V ascular cell proliferation in plant roots ing output in Arabidopsis root meristems tal transitions within each cell identity clus-
is, in part, controlled by the hetero- and found that this vascular heterodimer com- ter. We first analyzed xylem (251 cells, 5%; fig.
dimer complex formed by TARGET OF plex is required for the root hair responses S3A) and phloem (388 cells, 8%; fig. S4A) cell
MONOPTEROS 5 and LONESOME HIGH- to phosphate-deficit conditions. We show how lineages, because these undergo identity changes
WAY (TMO5/LHW) (1–7). This complex cytokinin signaling links vascular perception throughout development. Xylem initial cell
is required and sufficient to control cell pro- of limiting phosphate to epidermal responses, lineages branch into proto- and metaxylem
liferation by inducing expression of the direct which allows plants to efficiently forage the identities, which differ in their subsequent
downstream target LONELY GUY4 (LOG4) soil for this immobile macronutrient. differentiation processes, including secondary
and its close homolog LOG3 (1, 6), which en- cell wall generation (18). Phloem cell initials
code rate-limiting enzymes in the final step of Single-cell RNA-sequencing analysis undergo several oriented divisions, generat-
the conversion of the phytohormone cytokinin ing lineages that branch to generate phloem
into its bioactive form (8, 9). The TMO5/LHW We generated a high-resolution scRNA-seq atlas procambium, sieve elements, and companion
complex is limited to xylem cells, which produce of the wild-type Arabidopsis root tip (12–16), cells (19, 20). The complexity of both these
cytokinin but are themselves insensitive to making use of the 10x Genomics Chromium cell types was captured in the inferred trajec-
cytokinin. The xylem-produced cytokinin dif- technology (fig. S1A). After protoplast isola- tories, and gene expression patterns of the
fuses to neighboring procambium cells, where tion, sorted cells were collected and processed reporter lines validated inferred developmen-
it promotes cell proliferation by induction of for single-cell transcriptomics (see supplemen- tal trajectories and subcluster identities (figs.
DOF-type transcription factors (10, 11). Be- tary materials for details). A total population S3 and S4; see supplementary materials for
cause the TMO5/LHW pathway induces pro- of 15,918 cells was recovered across three re- details). Similar analyses validated trajecto-
duction of cytokinin as a mobile intermediate plicates and then filtered to retain 5145 high- ries for the procambium, pericycle, endodermis,
that functions in neighboring cells, the target quality cells with unique molecular identifier cortex, epidermis, lateral root cap, and colu-
genes (11) in this hormone signaling cascade (UMI) counts >17,290 (fig. S1B). Taking into ac- mella clusters (Fig. 1A and figs. S5 to S11).
are likely to be expressed in various cell types count only these high-quality cells, a total of
surrounding the xylem and perhaps even out- 21,492 genes were detected in our root meristem Because many root cell types in Arabidopsis
side of the vascular bundle. Here, we used dataset, covering nearly 80% of the genome, increase in ploidy with development (21), the
single-cell RNA-sequencing (scRNA-seq) to with a median expression of 6781 genes per developmental trajectories for these cell types
probe the tissue-specific TMO5/LHW signal- cell and a mean of 208,937 reads per cell (Fig. should correspond to trajectories of increasing
1A). Unsupervised clustering and t-distributed cellular endoreplication levels. Thus, to further
1Department of Plant Biotechnology and Bioinformatics, stochastic neighbor embedding (tSNE) projec- validate our trajectories, we predicted the
Ghent University, Ghent, Belgium. 2VIB Center for Plant tions were performed on the 5145 high-quality endoreplication state of each cell on the basis
Systems Biology, Ghent, Belgium. 3Data Mining and single cells, recovering distinct clusters of cells of the expression of a validated set of endo-
Modelling for Biomedicine, VIB Center for Inflammation (Fig. 1A). replication markers (21) (see supplementary
Research, Ghent, Belgium. 4Department of Applied materials for details) and superimposed the
Mathematics, Computer Science and Statistics, Ghent After quality control (see supplementary predicted cell ploidy on the tSNE plot (Fig. 1C).
University, Ghent, Belgium. 5VIB Flow Core, VIB Center for materials and fig. S1, C to E), cell-type anno- Developmental trajectories of cortex, endo-
Inflammation Research, Ghent, Belgium. 6Department of tation and cluster identification were per- dermis, pericycle, epidermis and atrichoblast,
Biomedical Molecular Biology, Ghent University, Ghent, formed by mapping the top 20 differentially lateral root cap, and xylem clusters exhibit
Belgium. expressed genes (DEGs) for each cluster (com- clear ploidy transitions (Fig. 1C), validating
*These authors contributed equally to this work. pared with the rest of the dataset) on a pub- these trajectories and their orientation. For
†Corresponding author. Email: [email protected] (B.D.R.); licly available bulk RNA-seq dataset (17) of the the procambium cell cluster, correspondence
[email protected] (Y.S.) Arabidopsis root. This resulted in an anno- between developmental trajectories and ploidy
was less evident. For the phloem cell lineage,

Wendrich et al., Science 370, eaay4970 (2020) 13 November 2020 1 of 9

RESEARCH | RESEARCH ARTICLE

A pressed in root hair (trichoblast) cells, of
which 47 (17%) were expressed only in tri-
C choblast cells (Fig. 2A and fig. S13). Such
B expression patterns were not expected given
the literature on TMO5/LHW function in vas-
Fig. 1. Identification of Arabidopsis root meristem cell types using scRNA-seq. (A) Color-coded tSNE cular proliferation (1, 2, 5, 6) and the over-
plot (right) showing the classification of 5145 high-quality (UMI count >17,290) cells into distinct cell lapping expression domain in the young xylem
identities corresponding to the schematic representation of the root meristem on the left. Gray dots cells (2). We confirmed the induction and ex-
represent predicted doublet cells. All inferred and validated developmental trajectories are projected onto pression pattern of a subset of these genes by
the tSNE plot as black lines. Cells within the dashed line are initials. QC, quiescent center; ppc, phloem quantitative real-time fluorescence polymer-
procambium; se, sieve element; cc, companion cell; px, protoxylem; mx, metaxylem. (B) Dot plot showing ase chain reaction (QPCR) and promoter–green
the expression of known tissue-specific reporter genes in the scRNA-seq dataset, validating the annotation fluorescent protein (GFP) fusions, respectively
of tissue-specific clusters. The size of the circles represents the percentage of cells with expression (Fig. 2B and fig. S14A). The trichoblast-specific
(pct. exp.), whereas the color indicates the scaled average expression (avg. exp. scale). Dashed boxes expression patterns of these target genes thus
represent major tissue types and cellular stages present in the root. (C) Projection of predicted ploidy levels suggest a putative role for TMO5/LHW in
of each cell onto the tSNE plot. the regulation of root hair development or
patterning. Although homozygous tmo5 single
which undergoes continuous divisions as it a developmental blueprint for all root cell and tmo5 tmo5-like1 double mutants showed
passes through the meristem, no correlation lineages, including progenitor populations normal root hair densities under standard
was found. Cells in the quiescent center and for several cell identities. growth conditions (high phosphate), misex-
hair cell (trichoblast) clusters mostly contained pression of TMO5 and LHW in all cells of the
2C and 16C cells, respectively (Fig. 1C). TMO5/LHW targets are enriched root meristem (pRPS5A::TMO5-GR x pRPS5A::
in root hair cells LHW-GR or dGR) resulted in a strong increase
Both newly generated reporter lines and in root hair density (Fig. 2, C and D), whereas
ploidy analysis confirm the inferred devel- We next intersected our scRNA-seq root misexpression of unrelated basic helix-loop-
opmental trajectories of all main cell identi- dataset with the set of 273 genes identified by helix (bHLH) factors did not result in this root
ties. Our results thus allow the identification bulk transcriptome analysis to be target genes hair density increase (fig. S15). A strong in-
of distinct subclusters linked to the develop- of the TMO5/LHW complex (11). About 80 crease in root hair density can also be observed
mental stage of each cell identity, establishing target genes (29%) were predicted to be ex- in wild-type roots grown under phosphate-
limiting conditions [Fig. 2, C and D, and table
S2; see fig. S16 and supplementary materials
for a detailed description of three-dimensional
(3D) root hair quantifications] (22–24). Auxin
biosynthesis, transport, and signaling are all
required for this root hair response to phosphate-
limiting conditions (22): Under low-phosphate
conditions, auxin signaling is induced in the
columella and lateral root cap region and in
xylem cells, where the TMO5/LHW dimer is
active (2, 4). Auxin-dependent TMO5 func-
tion is required for the root hair response to
low phosphate, because tmo5 tmo5-like1 dou-
ble mutants were less sensitive to these lim-
iting conditions (Fig. 2, C and D, and table S2).
Phosphate starvation genes (25) were, however,
still induced in this mutant background (fig.
S14B), suggesting that perception was un-
affected. Moreover, induction of several root
hair–specific TMO5/LHW target genes was
tmo5 tmo5-like1–dependent (fig. S14C). The
TMO5 homologs seem redundantly required
for this response, because we found no signi-
ficant difference in root hair density between
wild type and the tmo5 single mutant in low-
phosphate conditions (Fig. 2, C and D) (26).
Taken together, these results show that the
increase in root hair density under low-
phosphate conditions specifically requires
TMO5 activity. Because low-phosphate con-
ditions have also been associated with changes
in root hair length (27), we quantified this pa-
rameter in our mutant lines and treatments
and found similar responses (fig. S14D; see fig.
S16 and supplementary materials for details

Wendrich et al., Science 370, eaay4970 (2020) 13 November 2020 2 of 9

RESEARCH | RESEARCH ARTICLE

Fig. 2. TMO5 activity is required for root hair responses to low-phosphate conditions. (A) Number of tions (Fig. 2E). Similar results were obtained
TMO5/LHW target genes expressed in each of the tissue types of the Arabidopsis root meristem. Note the upon dGR induction (Fig. 2E). These results
high number of trichoblast-expressed genes. (B) Predicted (left) and validated (right) expression of root suggest that a reduction of epidermal cell length
hair–specific target genes in the root hair. Arrowheads indicate nuclear expression. Images on the right in the root hair zone drives the low-phosphate
are confocal laser scanning micrographs. (C) Stereo microscope images showing root hair phenotype of dGR response, leading to an increase in root hair
[induced or noninduced with dexamethasone (dex)] and wild type, tmo5 single mutant, and tmo5 tmo5like1 double density.
mutants grown under control conditions (high phosphate, HP) or phosphate-limiting conditions (low phosphate,
LP). (D and E) Quantification of the root hair density (D) and epidermal cell length (E) of the lines depicted in TMO5/LHW-dependent cytokinin controls root
(C). “Mock” indicates that the growth media contained no additives. Lowercase letters on top of box plots indicate hair responses to low phosphate
significantly different groups as determined by one-way analysis of variance (ANOVA) with post-hoc Tukey honest
significant difference (HSD) testing (p < 0.001); the number of individuals is shown at the bottom of the plot, To understand how TMO5/LHW in the xylem
and biological repeats are indicated using different symbols. Boxes represent the 1st and 3rd quartiles, and the might be involved in the low-phosphate re-
center line represents the median. dmso, dimethyl sulfoxide. sponse of root hairs in the epidermis, we
analyzed expression of the transcriptional
on quantification). Thus, activity of the TMO5/ length in control conditions compared with pTMO5::n3GFP reporter line (28). TMO5 expres-
LHW complex is required for the complete root wild type, this did not result in altered root sion increased in response to low-phosphate
hair response to low-phosphate conditions. hair densities (Fig. 2D and fig. S14, E and F). conditions (29) (Fig. 3, A and B), consistent
This suggests that root and meristem length with the reported increase in auxin signaling
To understand the cellular basis of the root are not directly contributing to changes in under these limiting conditions (22) and the
hair density increase in response to low- root hair density. Fitting with the observed auxin-inducibility of TMO5 (2). No ectopic ex-
phosphate conditions, we quantified several changes in root hair density, epidermal cell pression of TMO5 was observed in the tricho-
parameters that could contribute to this ef- length was not different in wild type and blast cells (Fig. 3A). Furthermore, increasing
fect (including root length, meristem length, tmo5 tmo5-like1 under control conditions but TMO5 levels only in the xylem axis or in the
and epidermal cell length; Fig. 2E and figs. was decreased in wild type and was signif- vascular bundle [using pTMO5::TMO5:GR (1)
S14, E and F, and S16) in wild-type and tmo5 icantly less decreased in the tmo5 tmo5-like1 or using newly generated pSHR::TMO5:GR
tmo5-like1 roots. Although the tmo5 tmo5-like1 double mutant under low-phosphate condi- and pWOL::XVE>>TMO5 lines] was sufficient
mutant showed reduced root and meristem to increase root hair density (Fig. 3, C to E, and
fig. S17, A to C). This suggests that the effect
of TMO5 on the low-phosphate-induced root
hair density increase is cell nonautonomous.

The TMO5/LHW complex binds the LOG4
promoter, thus promoting cytokinin biosyn-
thesis. Cytokinin can then diffuse to neighboring
cells, where its perception induces cell prolif-
eration (1, 6). To investigate if low-phosphate
conditions might lead to an increased induc-
tion of the cytokinin signaling pathway in epi-
dermal cells, we analyzed the pTCSn::ntdTomato
cytokinin-signaling reporter (11, 30). Under
low-phosphate conditions, the TCSn reporter
was induced in the epidermal cells of the
root meristem (Fig. 4A). Additionally, A-type
ARABIDOPSIS RESPONSE REGULATORS (ARRs)
(including ARR4, 5, 6, 8, 9, 12 and 15) that our
scRNA-seq dataset showed to be expressed
in trichoblast cells were up-regulated upon
TMO5/LHW induction (11) (figs. S13 and S14A).
Additionally, expression of three trichoblast-
restricted TMO5/LHW target genes was found
to be induced by exogenous cytokinin treat-
ment (fig. S18A), fitting with published data
(31). These results suggest an increased induc-
tion of the cytokinin signaling pathway in the
epidermis under low-phosphate conditions
and fit with the published effect of cytokinin on
cell length (32). To show that increased cyto-
kinin levels and/or signaling might lead to an
increase in root hair density owing to a reduc-
tion in epidermal cell lengths, we next treated
wild-type roots with 0.1 mM 6-benzylaminopurine
(BAP), a synthetic cytokinin. Indeed, treatment
with cytokinin under high-phosphate conditions
mimicked the root hair density–promoting
effect of low-phosphate conditions (Fig. 4B)
and a reduction in epidermal cell length (fig.

Wendrich et al., Science 370, eaay4970 (2020) 13 November 2020 3 of 9

RESEARCH | RESEARCH ARTICLE

Fig. 3. Vascular TMO5/LHW expression increases root hair density. (A and B) Expression (A) and line that expresses increased levels of the
quantification (B) of the pTMO5::n3GFP reporter line in the root meristem under high- and low-phosphate CKX3 cytokinin conjugating enzyme (35). Both
conditions by confocal microscopy. (C to E) Root hair phenotype (by stereo microscopy) and quantification genetic tools to reduce cytokinin levels re-
of wild-type, pTMO5::TMO5:GR, pSHR::TMO5:GR, and pWOL::XVE>>TMO5:YFP roots grown under high- sulted in an inhibition of the low-phosphate
phosphate conditions or induced by dexamethasone or estradiol (est) (see Fig. 2 for wild-type control). response (Fig. 4C and fig. S18C). To strengthen
Lowercase letters on top of box plots indicate significantly different groups as determined by one-way ANOVA the hypothesis that vascular-derived cytokinin
with post-hoc Tukey HSD testing [p < 0.001 in (C) and (E), and p < 0.01 in (D)]; the number of individuals is responsible for the root hair response to low-
is shown at the bottom of the plot, and biological repeats are indicated using different symbols. Boxes phosphate conditions, we generated a vascular-
represent the 1st and 3rd quartiles, and the center line represents the median. specific estradiol-inducible LOG4 transgenic
line (pWOL::XVE>>LOG4). Upon induction
S18B). Additionally, cytokinin was sufficient uniformly altered upon TMO5/LHW induction of the cytokinin biosynthetic gene LOG4 only
to rescue the root hair density and epidermal (fig. S19B). Given that the upstream ein2 recep- in the vascular domain, an increase in root hair
cell length effects to low-phosphate–like re- tor single mutant does show some resistance density was observed (Fig. 4D). Additionally,
sponses in the tmo5 tmo5-like1 double mutant to cytokinin but not low-phosphate treatment we complemented the log 1234578 heptuple
(Fig. 4B and fig. S18B). To investigate the pos- (fig. S19A), we cannot rule out involvement of mutant (9), which has very low levels of active
sible role of ethylene in this response (33), we complex cytokinin-ethylene hormonal cross- cytokinin, with LOG4 expressed only in the
next analyzed the responses of the ein3 eil1 talk during this developmental process. TMO5 domain (1). Under phosphate-limiting
double mutant in downstream ethylene sig- conditions, pTMO5::LOG4 expression was suf-
naling on low-phosphate medium and upon To provide additional genetic support for ficient to restore a wild type–like response (fig.
cytokinin treatment and found no difference the hypothesis that cytokinin signaling in tri- S20). Taken together, these experiments show
compared with wild-type plants (fig. S19A), choblast cells drives the root hair response to that vascular-derived cytokinin is capable of
fitting with published reports that cytokinin low-phosphate conditions in a TMO5-dependent triggering responses in the trichoblast cells.
effects on root hair length are not dependent manner, we first examined the effect of reduced Thus, vascular-derived cytokinin can drive the
on ethylene signaling (34) and the fact that cytokinin levels by analyzing the log347 triple root hair response to low-phosphate condi-
ethylene response markers (33) were not biosynthesis mutant (8) and by analyzing a tions in a TMO5-dependent manner.
newly generated pRPS5A::CKX3 transgenic
Previously, prolonged low-phosphate condi-
tions were shown to increase the number of
cortex cell files and modify epidermal cell fates
(36). To understand if these parameters might
contribute to our observed increase in root
hair density, we first analyzed the number of
cells in radial sections of cortex and epidermal
cell files because these determine the number
of root hairs (36). Although prolonged growth
under phosphate-deprived conditions leads to
an increase in the number of cortex cells (36)
(fig. S21), this was not observed after 10 days,
a time point used in all of our experiments
(table S2), suggesting that the effects observed
in our experiments are not due to additional
cortex cells. We next examined the possible
change in cell fate by examination of epi-
dermal cell identity using markers for hair
(pCOBL9::GFP) (37) and nonhair (pGL2::GFP)
(38) cell files. Epidermal cell identities were
mixed upon cytokinin treatment, similar to
the effect observed under low-phosphate con-
ditions (Fig. 5, A to C) (36). This cytokinin-
dependent effect most likely feeds into the
known pathways determining epidermal cell
identity, because cytokinin treatment was not
able to induce hair formation in the cpc try
double mutant (fig. S22) (39). Moreover, sig-
nificantly more root hairs were formed in
nonhair positions upon dGR induction and
exogenous cytokinin treatment and under
low-phosphate conditions (Fig. 5D). This low-
phosphate effect was absent in the tmo5 tmo5-
like1 mutant and in plants with reduced
cytokinin signaling levels (Fig. 5, D and E).
These results suggest that alterations in the
epidermal cell identity contribute to the ob-
served increase in root hair density under
low-phosphate conditions.

Wendrich et al., Science 370, eaay4970 (2020) 13 November 2020 4 of 9

RESEARCH | RESEARCH ARTICLE

Fig. 4. TMO5/LHW-dependent cytokinin triggers root hair responses. (A) Expression and quantification sucrose, and stratified at 4°C for 2 days be-
of pTCSn::ntdTomato in the root meristem under high- and low-phosphate conditions by confocal fore they were grown at 22°C in continuous
microscopy. Asterisks indicate epidermal cell layer. (B) Root hair phenotype (by stereo microscopy) and light conditions. Ten-day-old seedlings were
quantification of wild-type and tmo5 tmo5like1 roots grown under high-phosphate conditions or induced transferred to soil and grown in greenhouse
by cytokinin (BAP). (C) Root hair phenotype (by stereo microscopy) and quantification of wild-type, conditions. For phosphate-limiting experi-
log347, and pRPS5A::CKX3 roots grown under high- or low-phosphate conditions. (D) Root hair phenotype ments, plants were grown for 10 days on plates
(by stereo microscopy) and quantification of pWOL::XVE>>LOG4:YFP roots grown under high-phosphate containing ½ modified MS salts (M407, Phyto
conditions or induced by estradiol (see Fig. 3 for wild-type control). Lowercase letters on top of box Technology Laboratories), 0.55 mM myo-inositol,
plots indicate significantly different groups as determined by one-way ANOVA with post-hoc Tukey HSD 20.6 mM NH4NO3, 18.8 mM KNO3, 0.8% washed
testing (p < 0.001); the number of individuals is shown at the bottom of the plot, and biological repeats agar, and 1.25 mM or 100 mM KH2PO4 [for high-
are indicated using different symbols. Boxes represent the 1st and 3rd quartiles, and the center line phosphate (HP) or low-phosphate (LP), re-
represents the median. spectively]. Dexamethasone (dex) treatment
was performed by either germinating seeds
Outlook trichoblast cells. As such, this hormone sig- on 10 mM dex–supplemented medium or by
We show that the vascular bHLH heterodimer naling cascade spans multiple tissue layers in transferring plants from ½ MS to 10 mM dex–
TMO5/LHW controls root hair density by the meristem to regulate roots that are for- supplemented medium and continuing growth
modifying epidermal cell length and cell fates. aging for phosphate. for the indicated time. BAP and b-estradiol
Phosphate deficit may trigger increased auxin (est) treatments were performed by germi-
signaling in xylem cells, inducing the TMO5/ Materials and Methods nation on plates supplemented with 0.1 mM
LHW pathway and downstream local cyto- Plant material and growth conditions BAP or 10 mM est, respectively. BAP, dex, and
kinin biosynthesis. Cytokinin may then diffuse est were dissolved in dimethyl sulfoxide as
outward to direct the length and fates of outer All seeds were surface-sterilized, sown on solid solvent (DMSO, 10 mM stock solutions each).
½ Murashige and Skoog (MS) plates without Addition of DMSO to the medium was found
to have an effect on root hair length, largely
canceling out the effect of limiting phosphate
conditions on hair length. This effect was lim-
ited to hair length, because a large phosphate-
dependent effect could still be observed at the
level of root hair density (fig. S23). Media con-
taining no additives is referred to as “mock.”
The Arabidopsis thaliana (L.) Heynh. Col-0
ecotype served as wild-type control in all ex-
periments, unless otherwise stated. Plant lines
used in this study: pRPS5A::TMO5:GR-pRPS5A::
LHW:GR or dGR (11), pTMO5::n3GFP (28),
pTCSn::ntdTomato (11), tmo5 (2), tmo5 tmo5l1
(2), log3 log4 log7 (8), log 1234578 heptuple mu-
tant (9), pTMO5::LOG4 in log 1234578 heptuple
mutant (1), pCOBL9::GFP (37), pGL2::GFP (38),
pWOL::erGFP (40), ein2 (41), ein3 eil1 (42), pGL2::
GUS in Ws ecotype (39), pGL2::GUS in cpc try
double mutant (39), p35S::IBH1 (43), p35S::
IBL1 (43), p35S::MYC2 (44), and TRANSPLANTA
lines TPT_1.01260.2C, TPT_4.09180.1C, TPT_
1.01260.2I, TPT_1.05805.1C, TPT_1.05805.1E,
TPT_1.05805.1H, TPT_1.62975.1A, TPT_1.62975.1B,
TPT_1.62975.1D, TPT_1.62975.1F, TPT_1.62975.1H,
TPT_1.62975.1I, TPT_1.68920.2A, TPT_1.68920.2F,
TPT_2.18300.1B, TPT_2.18300.1C, TPT_2.18300.1I,
TPT_2.22750.2G, TPT_4.09180.1B, and TPT_
1.01260.2G (45).

Protoplasting conditions and
fluorescence-activated cell sorting

Protoplasting was performed as described pre-
viously (46) with minor modifications. Briefly,
root tips of 6-day-old seedlings were cut and
incubated in protoplasting Solution B {1.5%
(wt/vol) cellulysin and 0.1% (wt/vol) pecto-
lyase in Solution A [600 mM mannitol, 2 mM
MgCl2, 0.1% (wt/vol) bovine serum albumin
(BSA), 2 mM CaCl2, 2 mM MES, 10mM KCl,
pH 5.5]} for about 1 hour at room temperature.
Conditions were optimized to ensure that the

Wendrich et al., Science 370, eaay4970 (2020) 13 November 2020 5 of 9

RESEARCH | RESEARCH ARTICLE

Fig. 5. Cytokinin scrambles epidermal cell identities. (A) Expression of pCOBL9::GFP and pGL2::GFP after demultiplexing were used as the input
in roots grown under high-phosphate, high-phosphate cytokinin (HP BAP), or low-phosphate conditions. for “cellranger count,” which aligns the reads
Arrowheads indicate cells with ectopic expression of the marker line. (B and C) Quantification of number to the Arabidopsis thaliana reference genome
of cells per millimeter of root in nonhair (NH) or hair (H) positions with GFP expression. (D and E) (Ensemble TAIR10.40) using STAR and col-
Quantification of the number of cells in nonhair positions that form root hairs, along 1 mm of root from lapses them to UMI counts. The result is a large
dGR, wild type, tmo5, tmo5 tmo5l1, log347, or pRPS5A::CKX3 grown under the indicated conditions. digital expression matrix with cell barcodes
Lowercase letters on top of the box plots indicate significantly different groups as determined by one-way as rows and gene identities as columns. The
ANOVA with post-hoc Tukey HSD testing (p < 0.001); the number of individuals is shown at the bottom aggregation of the replicate samples was done
of the plot, and biological repeats are indicated using different symbols. Boxes represent the 1st and using “cellranger aggr.” Initial filtering in
3rd quartiles, and the center line represents the median. CellRanger (gene count >200), recovered
15,918 cells. To ensure that we only used high-
inner vascular cell types were dissociated well concentration of 1000 cells/ml, before being quality cells for further analysis, we used the
(even if this was to some extent at the expense subdivided to generate three replicates that filtered data provided by CellRanger. This cor-
of the outer cell files) by microscopic analysis have gone through the rest of the pipeline as responds to a minimum UMI count of 17,290
of the protoplasting at regular intervals. Cells independent samples. Cellular suspensions in the aggregated data, recovering a final
were filtered through a 70-mm cell strainer and were loaded on a GemCode Single Cell 3′ Gel 5145 cells.
spun down at 200g for 6 min, resuspended Bead and Library Kit (V2 chemistry, 10X Ge-
in Solution A, filtered through a 40-mm cell nomics) according to the manufacturer’s in- Data analysis (clustering, identity assignment,
strainer, and stained for viability using 4′,6- structions. Sequencing libraries were loaded DEGs, and quality control)
diamidino-2-phenylindole (DAPI) at 14 mM final on an Illumina HiSeq4000 and sequenced fol-
concentration. Cells were sorted on a BD Aria II, lowing the recommendations of 10X Genomics All analyses were performed in R (version 3.5.2).
and protoplasts not containing the DAPI sig- at the VIB Nucleomics Core (VIB, Leuven). Preprocessing of the data was done by the scran
nal were selected for further analysis. (version 1.10.1) and scater (version 1.10.1) pack-
Raw data processing and generation of gene ages according to the workflow proposed by
10X genomics sample preparation, library expression matrix the Marioni lab (47). Outlier cells were iden-
construction, and sequencing tified based on two metrics (library size and
Demultiplexing of the raw sequencing data number of expressed genes) and were tagged
Sorted cells were centrifuged at 4°C at 400g was done by the 10x CellRanger (version 2.0.2) as outliers when they were three median ab-
and resuspended in phosphate buffered saline software “cellranger mkfastq,” which wraps solute deviation (MADs) away from the median
(PBS) with 0.04% BSA to yield an estimated Illumina's bcl2fastq. The fastq files obtained value of these metrics across all cells. Low-
abundance genes were removed using the
“calcAverage” function and the proposed work-
flow by Lun and colleagues (48). The raw counts
were normalized and log2 transformed by first
calculating “size factors” that represent the
extent to which counts should be scaled in each
library. Detecting highly variable genes, find-
ing clusters, and creating tSNE plots was done
using the Seurat pipeline (version 2.3.4). Dif-
ferential expression analysis for marker gene
identification per subpopulation was based on
the nonparametric Wilcoxon rank sum test
implemented within the Seurat pipeline. Clus-
ters with the same cell annotation based on
gene expression analysis were combined to
generate a more comprehensible dataset. The
relative contribution of each replicate sample
to the entire dataset and each individual clus-
ter ranged between 28 and 38%, indicating
reproducibility and limited batch effects (fig.
S1C). Potential doublets were identified using
the DoubletFinder algorithm (version 2.0.0)
(49). The number of high confidence doublets
(indicated in gray on the tSNE plot in Fig. 1A)
was below 1% (50 out of 5145 cells). To assess
whether a capturing bias of meristem cells
was introduced during protoplast isolation,
Arabidopsis root meristems were imaged
using 3D confocal microscopy (see Plant imag-
ing, image processing, and data analysis section
below). After segmentation, cells were quanti-
fied according to their individual identities,
and the in vivo proportions were compared
with the proportions within the scRNA-seq
dataset. Out of the total number of cells, we

Wendrich et al., Science 370, eaay4970 (2020) 13 November 2020 6 of 9

RESEARCH | RESEARCH ARTICLE

observed a 15% overrepresentation of the inner subcluster. All newly generated reporter lines of a given cell was then taken to be the ploidy
vascular cell types and a 13% underrepresentation confirmed the predicted pseudotime and tissue- level for which the PCC between the cortex
of more outer cell layers (fig. S1D). This was specific expression patterns for the xylem tra- and cell data was maximal.
expected, because we optimized the protocol jectory (fig. S3B).
to distinguish inner vascular cells from other Gene selection, cloning, and plant transformation
cells. No other selective capturing was appar- For the phloem cell lineage, the GNG trajec-
ent. The effect of protoplasting on gene expres- tory algorithm ordered the cells across five cel- Genes from the top 20 differentially expressed
sion (50) was limited to a maximum 8% of the lular states of phloem development with more genes per cell type were selected on the basis
top 100 DEGs and only observed in the most elaborate branching (fig. S4A). Known reporter of predicted unique expression after manual
outer tissue types of the root (fig. S1E) genes [including APL for non-procambium curation of the feature plots. Transcriptional
(57), JUL1 for procambium (58), the PEAR reporter lines were obtained by amplifying the
Data comparison gene DOF5.6 for sieve elements (10), and upstream region of the transcriptional start
SUC2 for companion cells (57)] confirmed pre- and cloning this into the pDONRP41R entry
Comparison of different available Arabidopsis dicted subcluster identities within the phloem vector. The final expression clones were gen-
root scRNA-seq datasets (12–15) was done, re- developmental trajectories (fig. S4A). Also for erated by cloning the promoter regions into
vealing similarities and differences between this more complex trajectory, all newly gen- the pBGWFS7 destination vector using Gateway
these datasets, as shown in Table 1. erated reporter lines confirmed the predicted cloning. Only lines with stable expression pat-
spatiotemporal expression patterns (fig. S4B). terns in at least 10 independent transformants
Pseudotime trajectory analysis were kept, resulting in a set of 41 new tran-
A similar analysis and validation were scriptional reporter lines. pTMO5::TMO5:GR
Pseudotime trajectory analysis was performed performed on the predicted developmental line was generated by first generating a ligation-
with different methods compiled in the Dyno trajectories and subcluster identities for the independent cloning (LIC)–compatible vec-
R package (version 0.1.1) (51). The optimal procambium, pericycle, endodermis, cortex, epi- tor containing TMO5 promoter, after which
method for each cell identity was selected dermis, lateral root cap, and columella clusters TMO5:GR was amplified from pRPS5A::TMO5:
based on prior biological knowledge and the (figs. S5 to S11) GR template (1) and combined using LIC (59).
guidelines in the Dyno package. The consen- The pRPS5A::CKX3 line was generated by clon-
sus trajectory was fixed after iterative and Online tool for visualization of data ing the coding sequence (CDS) into pPLV28
reproducible inference calculation. The GNG using LIC. pSHR::nYFP was generated by clon-
algorithm was used to infer trajectories for To allow optimal accessibility of this scRNA-seq ing the promoter fragment (60) together with
the following cell identities: phloem (seed set dataset by the scientific community, cell type– nuclear-localized yellow fluorescent protein
at 4), xylem (seed set at 2), cortex (seed set at specific expression patterns of individual genes, (nYFP) into pH7m34GW destination vector
1), lateral root cap and columella (seed set at 1), expression levels along developmental trajec- using Gateway cloning. pSHR::TMO5:GR was
and pericycle (seed set at 13); Slingshot was tories, and bulk downloads can be accessed generated by combining the SHR promoter
used for endodermis and cambium (52). through a freely accessible on-line browser fragment, TMO5 CDS, and glucocorticoid recep-
tool (http://bioit3.irc.ugent.be/plant-sc-atlas/), tor (GR) tag into pH7m34GW, using Gateway
For trajectory inference on the xylem sub- which includes a warning flag for protoplast- cloning. pWOL::XVE>>TMO5:YFP and pWOL::
populations, cells were ordered along a bifur- induced genes. XVE>>LOG4:YFP were generated by subcloing
cating trajectory from xylem initials toward the respective CDS with pDONRP4p1r-pWOL-
protoxylem and metaxylem (fig. S3A). The Cell ploidy predictions XVE (60) and YFP into pH7m34GW using Gate-
identification and validation of the subpop- way cloning. All constructs were verified by
ulations and branching points is based on For predicting the ploidy of individual cells, Sanger sequencing and were transformed into
gene expression of known regulators of xylem we correlated the scRNA-seq dataset with a wild type using simplified floral dipping (61).
development such as ACL5 and ATHB15 for dataset of ploidy-specific gene expression lev- All primer sequences used for cloning and se-
early metaxylem (53, 54), TMO5 for early proto- els in Arabidopsis root cortex tissue (21). The quencing can be found in table S3.
and metaxylem (2), VND2 for intermediate expression values of 332 previously established
xylem (55), and VND7 and CESA4 for late ploidy marker genes [supplemental table 5 in Plant imaging, image processing,
protoxylem (55, 56) (fig. S3A). Although our (21)] were standardized in both datasets sepa- and data analysis
method depends on generating protoplasts rately (zero mean, unit variance), and Pearson
and is not efficient at retrieving differentiat- correlation coefficients (PCCs) were calculated Marker lines and different genotypes were either
ing cells, we acquired many genes involved in between the resulting marker gene expression fixed with ClearSee, stained with Calcofluor
secondary cell wall formation in the proto- profile in each cell and the marker gene ex- White and mounted on sides with ClearSee so-
xylem subcluster but less so in the metaxylem pression profiles of 2C, 4C, 8C, and 16C cells in lution (62), or directly mounted on slides con-
the cortex dataset. The predicted ploidy level taining 0.1 mg/ml propidium iodide (PI) for cell
wall staining. Confocal microscopy on marker
Table 1. Comparison of different available Arabidopsis root scRNA-seq datasets. The “–” lines was performed on a Leica SP8X using a
indicates data that were not reported. 40× lens (water-corrected objective lens with
NA 1.2). For 3D imaging and segmentation
Dataset Technology Number Average reads Median genes analysis, 6-day-old wild-type roots were stained
used of cells per cell per cell using the modified pseudo-Schiff–PI (mPS-PI)
staining method (63). 3D segmentation of the
Denyer et al., 2019 (13) 10X v2 4727 87,000 4276 primary root meristem was performed on 3D
..................................................................................................................................................................................................................... z-stacks using the MorphoGraphX software
(64), and cell numbers per cell type were scored
Jean-Baptiste et al., 2019 (12) 10X 3121 – 2445 in two individual roots (averages used in plot).
..................................................................................................................................................................................................................... For root hair analysis, 10-day-old seedling
roots were imaged on the plate with a Leica
Zhang et al., 2019 (14) 10X v2 7695 39,667 1875
.....................................................................................................................................................................................................................

Ryu et al., 2019 (15) 10X v2 7522 75,000 5000
.....................................................................................................................................................................................................................

This study 10X v2 5145 208,937 6781
.....................................................................................................................................................................................................................

Wendrich et al., Science 370, eaay4970 (2020) 13 November 2020 7 of 9

RESEARCH | RESEARCH ARTICLE

stereomicroscope at 32× magnification. Addi- summary of all phenotypic measurements re- 16. C. N. Shulse et al., High-throughput single-cell transcriptome
tionally, confocal imaging was applied to gen- lated to cell numbers and root hairs, on respec- profiling of plant cell types. Cell Rep. 27, 2241–2247.e4
erate a 3D stack of a full 1 mm of mature tive genotypes and treatments, is presented (2019). doi: 10.1016/j.celrep.2019.04.054; pmid: 31091459
primary root on a Leica SP8X using a 10× dry in table S2.
lens on samples fixed with ClearSee and stained 17. S. Li, M. Yamada, X. Han, U. Ohler, P. N. Benfey, High-resolution
with Calcofluor White. This stack was gen- Relative expression analysis expression map of the Arabidopsis root reveals alternative
erated by imaging in the XZY direction, taking splicing and lincRNA regulation. Dev. Cell 39, 508–522 (2016).
150 images at ~6.7-mm intervals. For a descrip- Total RNA was isolated using an RNeasy kit doi: 10.1016/j.devcel.2016.10.012; pmid: 27840108
tion of the phenotypic parameters that were (QIAGEN). cDNA was synthesized using 1 mg
measured, see fig. S16 and the materials and of total RNA and a qScript cDNA SuperMix 18. R. Ruonala, D. Ko, Y. Helariutta, Genetic networks in plant
methods Phenotyping section. Calcofluor White, (QUANTABIO) according to the manufacturer’s vascular development. Annu. Rev. Genet. 51, 335–359 (2017).
GFP, YFP, tdTomato, and PI samples were im- instructions. Relative expression was measured doi: 10.1146/annurev-genet-120116-024525; pmid: 28892639
aged at excitations of 405, 488, 514, 561, and in triplicate by QPCR on a LightCycler480 II
514 nm, respectively. Calcofluor White, GFP, apparatus (Roche) with SYBR GREEN I Master 19. A. Rodriguez-Villalon et al., Molecular genetic framework
YFP, tdTomato, and PI were visualized at emis- kit (Roche) according to the manufacturer’s for protophloem formation. Proc. Natl. Acad. Sci. U.S.A.
sions of 425 to 475 nm, 500 to 535nm, 520 to instructions. Data were analyzed using qBase+ 111, 11551–11556 (2014). doi: 10.1073/pnas.1407337111;
550nm, 580 to 630nm, and 600 to 700 nm, software package (Biogazelle). Expression lev- pmid: 25049386
respectively. b-Glucuronidase (GUS) staining els were normalized to those of EEF1a4 and
was performed as described previously (65). UBC21. All primers used for QPCR analysis 20. A. P. Mähönen et al., A novel two-component hybrid molecule
For aesthetic reasons, some sets of images were can be found in table S3. regulates vascular morphogenesis of the Arabidopsis root.
globally adjusted in brightness and contrast, Genes Dev. 14, 2938–2943 (2000). doi: 10.1101/gad.189200;
rotated, and displayed on a matching back- REFERENCES AND NOTES pmid: 11114883
ground. Plots were generated using R package
“ggplot2” (66). In all box plots, boxes represent 1. B. De Rybel et al., Integration of growth and patterning during 21. R. Bhosale et al., A spatiotemporal DNA endoploidy map of the
the 1st and 3rd quartiles, and the center line vascular tissue formation in Arabidopsis. Science 345, 1255215 Arabidopsis root reveals roles for the endocycle in root
represents the median. Data-point shapes rep- (2014). doi: 10.1126/science.1255215; pmid: 25104393 development and stress adaptation. Plant Cell 30, 2330–2351
resent biological replicates, and the numbers (2018). doi: 10.1105/tpc.17.00983; pmid: 30115738
of measurements are shown above the x axis. 2. B. De Rybel et al., A bHLH complex controls embryonic
Statistical analyses were performed using either vascular tissue establishment and indeterminate growth in 22. R. Bhosale et al., A mechanistic framework for auxin dependent
one-way analysis of variance (ANOVA) or chi- Arabidopsis. Dev. Cell 24, 426–437 (2013). doi: 10.1016/ Arabidopsis root hair elongation to low external phosphate.
square tests, and p values below 0.001 were con- j.devcel.2012.12.013; pmid: 23415953 Nat. Commun. 9, 1409 (2018). doi: 10.1038/s41467-018-
sidered significant, unless otherwise stated. All 03851-3; pmid: 29651114
figures were compiled using Adobe Illustrator. 3. K. Ohashi-Ito, D. C. Bergmann, Regulation of the Arabidopsis
root vascular initial population by LONESOME HIGHWAY. 23. J. E. Salazar-Henao, I. C. Vélez-Bermúdez, W. Schmidt,
Phenotyping Development 134, 2959–2968 (2007). doi: 10.1242/ The regulation and plasticity of root hair patterning and
dev.006296; pmid: 17626058 morphogenesis. Development 143, 1848–1858 (2016).
The approaches used for analysis of the range doi: 10.1242/dev.132845; pmid: 27246711
of phenotypic parameters described in fig. 4. K. Ohashi-Ito, M. Matsukawa, H. Fukuda, An atypical bHLH
S16 are explained below in more detail. From transcription factor regulates early xylem development 24. M. Grebe et al., Cell polarity signaling in Arabidopsis
scanned plates, total primary root lengths were downstream of auxin. Plant Cell Physiol. 54, 398–405 (2013). involves a BFA-sensitive auxin influx pathway. Curr. Biol. 12,
measured using Fiji software. From stereo- doi: 10.1093/pcp/pct013; pmid: 23359424 329–334 (2002). doi: 10.1016/S0960-9822(02)00654-1;
microscope images, both the number of root pmid: 11864575
hairs in 1 mm of root and the root hair lengths 5. K. Ohashi-Ito, M. Oguchi, M. Kojima, H. Sakakibara, H. Fukuda,
were measured in the maturation zone of the Auxin-associated initiation of vascular cell differentiation by 25. H. Rouached, A. B. Arpat, Y. Poirier, Regulation of phosphate
primary root using Fiji software. From 3D- LONESOME HIGHWAY. Development 140, 765–769 (2013). starvation responses in plants: Signaling players and cross-
confocal stacks, the number of cortex and doi: 10.1242/dev.087924; pmid: 23362345 talks. Mol. Plant 3, 288–299 (2010). doi: 10.1093/mp/ssp120;
number of epidermal cells (hair versus non- pmid: 20142416
hair position) were counted in one represen- 6. K. Ohashi-Ito et al., A bHLH complex activates vascular cell
tative cross section. Additionally, the total division via cytokinin action in root apical meristem. 26. Z. H. Chen, G. A. Nimmo, G. I. Jenkins, H. G. Nimmo, BHLH32
number of hairs and the number of cells in Curr. Biol. 24, 2053–2058 (2014). doi: 10.1016/ modulates several biochemical and morphological processes
nonhair position that form root hairs were j.cub.2014.07.050; pmid: 25131670 that respond to Pi starvation in Arabidopsis. Biochem. J. 405,
counted along the full stack. Similar imag- 191–198 (2007). doi: 10.1042/BJ20070102; pmid: 17376028
ing and phenotyping were performed on 7. F. Vera-Sirera et al., A bHLH-based feedback loop restricts vascular
pCOBL9::GFP (hair marker) and pGL2::GFP cell proliferation in plants. Dev. Cell 35, 432–443 (2015). 27. Z. Ma, D. G. Bielenberg, K. M. Brown, J. P. Lynch, Regulation
(nonhair marker) lines. Here, the number of doi: 10.1016/j.devcel.2015.10.022; pmid: 26609958 of root hair density by phosphorus availability in Arabidopsis
cells, ectopically expressing the marker, were thaliana. Plant Cell Environ. 24, 459–467 (2001). doi: 10.1046/
counted along the entire stack. Meristem 8. T. Kuroha et al., Functional analyses of LONELY GUY cytokinin- j.1365-3040.2001.00695.x
length was determined from longitudinal con- activating enzymes reveal the importance of the direct
focal sections of the primary root and mea- activation pathway in Arabidopsis. Plant Cell 21, 3152–3169 28. A. Schlereth et al., MONOPTEROS controls embryonic root
sured as the distance from quiescent center (2009). doi: 10.1105/tpc.109.068676; pmid: 19837870 initiation by regulating a mobile transcription factor. Nature 464,
to the first elongated cortical cell. Epidermal 913–916 (2010). doi: 10.1038/nature08836; pmid: 20220754
cell length was determined by measuring the 9. H. Tokunaga et al., Arabidopsis lonely guy (LOG) multiple
length of 10 trichoblast cells on a longitudinal mutants reveal a central role of the LOG-dependent pathway in 29. P. Wu et al., Phosphate starvation triggers distinct alterations
confocal section from the same 1-mm region cytokinin activation. Plant J. 69, 355–365 (2012). doi: 10.1111/ of genome expression in Arabidopsis roots and leaves.
of the root where root hairs were counted. A j.1365-313X.2011.04795.x; pmid: 22059596 Plant Physiol. 132, 1260–1271 (2003). doi: 10.1104/
pp.103.021022; pmid: 12857808
10. S. Miyashima et al., Mobile PEAR transcription factors integrate
positional cues to prime cambial growth. Nature 565, 490–494 30. E. Zürcher et al., A robust and sensitive synthetic sensor
(2019). doi: 10.1038/s41586-018-0839-y; pmid: 30626969 to monitor the transcriptional output of the cytokinin signaling
network in planta. Plant Physiol. 161, 1066–1075 (2013).
11. W. Smet et al., DOF2.1 controls cytokinin-dependent doi: 10.1104/pp.112.211763; pmid: 23355633
vascular cell proliferation downstream of TMO5/LHW.
Curr. Biol. 29, 520–529.e6 (2019). doi: 10.1016/ 31. K. C. Potter, J. Wang, G. E. Schaller, J. J. Kieber, Cytokinin
j.cub.2018.12.041; pmid: 30686737 modulates context-dependent chromatin accessibility through
the type-B response regulators. Nat. Plants 4, 1102–1111
12. K. Jean-Baptiste et al., Dynamics of gene expression in single (2018). doi: 10.1038/s41477-018-0290-y; pmid: 30420712
root cells of Arabidopsis thaliana. Plant Cell 31, 993–1011
(2019). doi: 10.1105/tpc.18.00785; pmid: 30923229 32. I. H. Street et al., Cytokinin acts through the auxin influx carrier
AUX1 to regulate cell elongation in the root. Development 143,
13. T. Denyer et al., Spatiotemporal developmental trajectories 3982–3993 (2016). doi: 10.1242/dev.132035; pmid: 27697901
in the Arabidopsis root revealed using high-throughput
single-cell RNA sequencing. Dev. Cell 48, 840–852.e5 (2019). 33. L. Song et al., The molecular mechanism of ethylene-mediated
doi: 10.1016/j.devcel.2019.02.022; pmid: 30913408 root hair development induced by phosphate starvation.
PLOS Genet. 12, e1006194 (2016). doi: 10.1371/
14. T. Q. Zhang, Z. G. Xu, G. D. Shang, J. W. Wang, A single-cell journal.pgen.1006194; pmid: 27427911
RNA sequencing profiles the developmental landscape of
Arabidopsis root. Mol. Plant 12, 648–660 (2019). doi: 10.1016/ 34. S. Zhang et al., Multiple phytohormones promote root hair
j.molp.2019.04.004; pmid: 31004836 elongation by regulating a similar set of genes in the root
epidermis in Arabidopsis. J. Exp. Bot. 67, 6363–6372 (2016).
15. K. H. Ryu, L. Huang, H. M. Kang, J. Schiefelbein, Single-cell doi: 10.1093/jxb/erw400; pmid: 27799284
RNA sequencing resolves molecular relationships among
individual plant cells. Plant Physiol. 179, 1444–1456 (2019). 35. T. Schmülling, T. Werner, M. Riefler, E. Krupková,
doi: 10.1104/pp.18.01482; pmid: 30718350 I. Bartrina y Manns, Structure and function of cytokinin
oxidase/dehydrogenase genes of maize, rice, Arabidopsis
and other species. J. Plant Res. 116, 241–252 (2003).
doi: 10.1007/s10265-003-0096-4; pmid: 12721786

36. G. Janes et al., Cellular patterning of Arabidopsis roots under
low phosphate conditions. Front Plant Sci 9, 735 (2018).
doi: 10.3389/fpls.2018.00735; pmid: 29922313

37. S. M. Brady, S. Song, K. S. Dhugga, J. A. Rafalski, P. N. Benfey,
Combining expression and comparative evolutionary analysis.
The COBRA gene family. Plant Physiol. 143, 172–187 (2007).
doi: 10.1104/pp.106.087262; pmid: 17098858

Wendrich et al., Science 370, eaay4970 (2020) 13 November 2020 8 of 9

RESEARCH | RESEARCH ARTICLE

38. Y. Lin, J. Schiefelbein, Embryonic control of epidermal cell 50. K. Birnbaum et al., A gene expression map of the Arabidopsis 64. P. Barbier de Reuille et al., MorphoGraphX: A platform for
patterning in the root and hypocotyl of Arabidopsis. root. Science 302, 1956–1960 (2003). doi: 10.1126/ quantifying morphogenesis in 4D. eLife 4, e05864 (2015).
Development 128, 3697–3705 (2001). pmid: 11585796 science.1090022; pmid: 14671301 doi: 10.7554/eLife.05864; pmid: 25946108

39. M. Simon, M. M. Lee, Y. Lin, L. Gish, J. Schiefelbein, Distinct 51. W. Saelens, R. Cannoodt, H. Todorov, Y. Saeys, A comparison 65. B. De Rybel et al., A novel aux/IAA28 signaling cascade
and overlapping roles of single-repeat MYB genes in root of single-cell trajectory inference methods. Nat. Biotechnol. activates GATA23-dependent specification of lateral root
epidermal patterning. Dev. Biol. 311, 566–578 (2007). 37, 547–554 (2019). doi: 10.1038/s41587-019-0071-9; founder cell identity. Curr. Biol. 20, 1697–1706 (2010).
doi: 10.1016/j.ydbio.2007.09.001; pmid: 17931617 pmid: 30936559 doi: 10.1016/j.cub.2010.09.007; pmid: 20888232

40. M. Bonke, S. Thitamadee, A. P. Mähönen, M. T. Hauser, 52. K. Street et al., Slingshot: Cell lineage and pseudotime 66. H. Wickham, ggplot2, Elegant Graphics for Data Analysis
Y. Helariutta, APL regulates vascular tissue identity in inference for single-cell transcriptomics. BMC Genomics 19, (Use R!) (Springer, 2009), vol. 1, pp. VIII, 213.
Arabidopsis. Nature 426, 181–186 (2003). doi: 10.1038/ 477 (2018). doi: 10.1186/s12864-018-4772-0; pmid: 29914354
nature02100; pmid: 14614507 ACKNOWLEDGMENTS
53. L. Muñiz et al., ACAULIS5 controls Arabidopsis xylem
41. J. M. Alonso, T. Hirayama, G. Roman, S. Nourizadeh, J. R. Ecker, specification through the prevention of premature cell death. We thank V. Storme for help with statistical analyses and
EIN2, a bifunctional transducer of ethylene and stress Development 135, 2573–2582 (2008). doi: 10.1242/ D. Weijers for the use of unpublished materials. Funding: This work
responses in Arabidopsis. Science 284, 2148–2152 (1999). dev.019349; pmid: 18599510 was supported by funding from the European Research Council
doi: 10.1126/science.284.5423.2148; pmid: 10381874 (ERC Starting Grant TORPEDO; 714055); the Research
54. A. Carlsbecker et al., Cell signalling by microRNA165/6 directs gene Foundation – Flanders (FWO; Odysseus II G0D0515N); Ghent
42. T. Potuschak et al., EIN3-dependent regulation of plant dose-dependent root cell fate. Nature 465, 316–321 (2010). University (BOF20/GOA/012 and BOF18/PDO/151); and a Marie
ethylene hormone signaling by two Arabidopsis F box proteins: doi: 10.1038/nature08977; pmid: 20410882 Curie Fellowship (IEF-2009-252503). Author contributions: B.D.R.,
EBF1 and EBF2. Cell 115, 679–689 (2003). doi: 10.1016/ Y.S., and G.V.I. conceived the project; B.D.R., J.R.W., and B.Y.
S0092-8674(03)00968-1; pmid: 14675533 55. M. Kubo et al., Transcription switches for protoxylem and designed experiments; J.R.W., J.V.D., B.D.R., and G.V.I. generated
metaxylem vessel formation. Genes Dev. 19, 1855–1860 samples for scRNA-seq; N.V. performed scRNA-seq; N.V., K.V.,
43. M. K. Zhiponova et al., Helix-loop-helix/basic helix-loop-helix (2005). doi: 10.1101/gad.1331305; pmid: 16103214 and Y.S. analyzed scRNA-seq data and performed trajectory analyses;
transcription factor network represses cell elongation in K.V. generated the online browser tool; W.S., B.W., E.M., H.E.A., J.N.,
Arabidopsis through an apparent incoherent feed-forward loop. 56. M. Taylor-Teeples et al., An Arabidopsis gene regulatory B.Y., and J.R.W. generated reporter lines to validate the dataset;
Proc. Natl. Acad. Sci. U.S.A. 111, 2824–2829 (2014). network for secondary cell wall synthesis. Nature 517, 571–575 S.M. performed ploidy analysis; B.Y., C.V.d.V., M.M., and J.R.W.
doi: 10.1073/pnas.1400203111; pmid: 24505057 (2015). doi: 10.1038/nature14099; pmid: 25533953 analyzed effects on root hair growth; B.D.R. and Y.S. supervised the
project; and B.D.R. wrote the paper with input from all authors.
44. J. Goossens, G. Swinnen, R. Vanden Bossche, L. Pauwels, 57. S. Otero, Y. Helariutta, Companion cells: A diamond in the Competing interests: None declared. Data and materials availability:
A. Goossens, Change of a conserved amino acid in the MYC2 rough. J. Exp. Bot. 68, 71–78 (2017). doi: 10.1093/jxb/erw392; The single-cell data can be accessed through a freely accessible
and MYC3 transcription factors leads to release of JAZ pmid: 27811001 online browser tool (https://bioit3.irc.ugent.be/plant-sc-atlas/), and
repression and increased activity. New Phytol. 206, 1229–1237 raw data can be accessed at NCBI with GEO number GSE141730. All
(2015). doi: 10.1111/nph.13398; pmid: 25817565 58. V. López-Salmerón, H. Cho, N. Tonn, T. Greb, The phloem other data needed to evaluate the conclusions in the paper are
as a mediator of plant growth plasticity. Curr. Biol. 29, R173–R181 present in the paper or the supplementary materials. Material
45. A. Coego et al., The TRANSPLANTA collection of Arabidopsis (2019). doi: 10.1016/j.cub.2019.01.015; pmid: 30836090 requests should be directed to the corresponding authors.
lines: A resource for functional analysis of transcription factors
based on their conditional overexpression. Plant J. 77, 59. J. R. Wendrich, C. Y. Liao, W. A. van den Berg, B. De Rybel, SUPPLEMENTARY MATERIALS
944–953 (2014). doi: 10.1111/tpj.12443; pmid: 24456507 D. Weijers, Ligation-independent cloning for plant research.
Methods Mol. Biol. 1284, 421–431 (2015). doi: 10.1007/ science.sciencemag.org/content/370/6518/eaay4970/suppl/DC1
46. A. S. Iyer-Pascuzzi, P. N. Benfey, Fluorescence-activated cell 978-1-4939-2444-8_21; pmid: 25757785 Materials and Methods
sorting in plant developmental biology. Methods Mol. Biol. Figs. S1 to S23
655, 313–319 (2010). doi: 10.1007/978-1-60761-765-5_21; 60. R. Siligato et al., MultiSite gateway-compatible cell type-specific Tables S1 to S3
pmid: 20734270 gene-inducible system for plants. Plant Physiol. 170, 627–641 References
(2016). doi: 10.1104/pp.15.01246; pmid: 26644504 MDAR Reproducibility Checklist
47. A. T. Lun, K. Bach, J. C. Marioni, Pooling across cells to
normalize single-cell RNA sequencing data with many zero 61. S. J. Clough, A. F. Bent, Floral dip: A simplified method for View/request a protocol for this paper from Bio-protocol.
counts. Genome Biol. 17, 75 (2016). doi: 10.1186/ Agrobacterium-mediated transformation of Arabidopsis
s13059-016-0947-7; pmid: 27122128 thaliana. Plant J. 16, 735–743 (1998). doi: 10.1046/ 24 June 2019; resubmitted 17 July 2020
j.1365-313x.1998.00343.x; pmid: 10069079 Accepted 4 September 2020
48. A. T. Lun, D. J. McCarthy, J. C. Marioni, A step-by-step Published online 17 September 2020
workflow for low-level analysis of single-cell RNA-seq data 62. R. Ursache, T. G. Andersen, P. Marhavý, N. Geldner, A protocol 10.1126/science.aay4970
with Bioconductor. F1000Res. 5, 2122 (2016). doi: 10.12688/ for combining fluorescent proteins with histological stains
f1000research.9501.2; pmid: 27909575 for diverse cell wall components. Plant J. 93, 399–412 (2018).
doi: 10.1111/tpj.13784; pmid: 29171896
49. C. S. McGinnis, L. M. Murrow, Z. J. Gartner, DoubletFinder:
Doublet detection in single-cell RNA sequencing data using 63. E. Truernit et al., High-resolution whole-mount imaging of
artificial nearest neighbors. Cell Syst. 8, 329–337.e4 (2019). three-dimensional tissue organization and gene expression
doi: 10.1016/j.cels.2019.03.003; pmid: 30954475 enables the study of phloem development and structure in
Arabidopsis. Plant Cell 20, 1494–1503 (2008). doi: 10.1105/
tpc.107.056069; pmid: 18523061

Wendrich et al., Science 370, eaay4970 (2020) 13 November 2020 9 of 9

RESEARCH

◥ at least some herd protection (21, 25). This
genus also contains other viruses that cause
RESEARCH ARTICLE severe infections in humans, including Middle
East respiratory syndrome and SARS-CoV-1
CORONAVIRUS coronaviruses (21). Whereas humoral immu-
nity to SARS-CoV-1 is believed to last up to
Immune life history, vaccination, and the dynamics of 2 to 3 years (26, 27), antigen-specific T cells
SARS-CoV-2 over the next 5 years against this virus were found to be detectable
for at least 11 years after infection (28). Indeed,
Chadi M. Saad-Roy1*, Caroline E. Wagner2,3,4*, Rachel E. Baker2,3, Sinead E. Morris5, Jeremy Farrar6, T cell–mediated responses likely play a central
Andrea L. Graham2, Simon A. Levin2, Michael J. Mina7, C. Jessica E. Metcalf2,8, Bryan T. Grenfell2,8,9† role in controlling SARS-CoV-2 replication and
disease (14, 15). Recent evidence of preexisting
The future trajectory of the coronavirus disease 2019 (COVID-19) pandemic hinges on the dynamics of adaptive T cells (14, 15) and antibodies (29) capable of
immunity against severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2); however, salient features cross-reacting with SARS-CoV-2 suggests that
of the immune response elicited by natural infection or vaccination are still uncertain. We use simple immunological memory responses elicited dur-
epidemiological models to explore estimates for the magnitude and timing of future COVID-19 cases, given ing infection with seasonal coronaviruses may
different assumptions regarding the protective efficacy and duration of the adaptive immune response to also affect coronavirus disease 2019 (COVID-19)
SARS-CoV-2, as well as its interaction with vaccines and nonpharmaceutical interventions. We find susceptibility and disease risk. Finally, although
that variations in the immune response to primary SARS-CoV-2 infections and a potential vaccine can lead to it is currently unclear whether ADE influences
markedly different immune landscapes and burdens of critically severe cases, ranging from sustained the pathogenesis of SARS-CoV-2, it has been
epidemics to near elimination. Our findings illustrate likely complexities in future COVID-19 dynamics and hypothesized that severe COVID-19 cases may
highlight the importance of immunological characterization beyond the measurement of active infections for arise from the presence of nonneutralizing
adequately projecting the immune landscape generated by SARS-CoV-2 infections. antibodies from prior coronavirus infections
(30), in agreement with earlier proposals for
T he novel severe acute respiratory syn- (2) virus results in lifelong protection from the related coronaviruses (31–33).
drome coronavirus 2 (SARS-CoV-2) beta- reacquisition and retransmission of secondary
coronavirus (b-CoV) pandemic has resulted Various epidemiological models have been
in substantial morbidity and mortality, infections. Many other infections [e.g., influ- developed to capture how the diversity or
enza (3) and respiratory syncytial virus (RSV) variation in immune responses influences
with over 27 million confirmed cases (4)] confer imperfect or transient clinical and population-level infection dynamics. For in-
transmission-blocking immunity by either stance, the well-known Susceptible-Infected-
worldwide at the time of writing. To curb viral pathogen evolution or waning immunological Recovered (SIR) model is suitable for modeling
the dynamics of perfectly immunizing infections
transmission, nonpharmaceutical interventions memory. Finally, phenomena such as antibody- such as measles (34), whereas the Susceptible-
dependent enhancement (ADE) associated with Infected-Recovered-Susceptible (SIRS) model
(NPIs), including business and school closures, prior natural infection [e.g., dengue (5)] or a captures the epidemiology of imperfectly im-
vaccine [e.g., RSV (6)] could result in more clin- munizing infections such as influenza; here,
restrictions on movement, and total lockdowns, ically severe secondary infections. Furthermore, individuals eventually return to a fully or sub-
the immunity conferred by vaccines may not stantially susceptible class after a finite period
have been implemented to various degrees of immunity, because of either waning mem-
provide complete protection against reinfec- ory or pathogen evolution (35). More complex
around the world. Major efforts to develop tion and/or disease (7), and this protection compartmental models have also been devel-
may be inferior to that acquired after natural oped to study infections characterized by in-
effective vaccines and antivirals are ongoing. infection (8). Nevertheless, imperfect vaccines termediate immune responses lying between
that reduce both the clinical severity and trans- these two extremes, such as rotavirus (36)
Understanding the future trajectory of this missibility of subsequent infections (if they do and RSV (4).

disease requires knowledge of the population- occur) can still provide population-level disease Here, we adopt a generalization of these
protection (7, 9, 10). models, the SIR(S) model (35), outlined sche-
level landscape of immunity, generated by the matically in Fig. 1 and fig. S1, to explore how
The nature of the immune response after the pandemic trajectory might unfold for dif-
life histories of SARS-CoV-2 infection or vacci- natural SARS-CoV-2 infection remains an area ferent assumptions regarding the nature of the
of active investigation (11–18). Reports from adaptive immune response to SARS-CoV-2 in-
nation among individual hosts. We show that serological population- and individual-level fection. Because different adaptive immune re-
sponses may be associated with variations in
the nature of secondary infection, particularly studies demonstrate that detectable antibody the proportion of severe secondary cases, we
levels can wane over the first few months post- also consider a range of values for this fraction
the degree of acquisition, retransmission, and infection (19), yet recent findings demonstrate in order to explore the potential future clinical
robust antibody responses 4 months after in- burden of SARS-CoV-2 infections. The model
clinical severity of subsequent infections with fection (20). This is broadly consistent with assumes different infection and immune phe-
serum antibody levels against the seasonal notypes, depending on exposure history [see
the same pathogen, is particularly important. (37) for the full mathematical details]. Specif-
coronavirus human coronavirus OC43 (HCoV- ically, it interpolates between the fully immu-
The nature of acquired immune responses after OC43) [which belongs to the same b-CoV genus nizing SIR model, when immunity is lifelong,
as SARS-CoV-2 (21)], which wane on the time and the imperfectly immunizing SIRS mod-
natural infection varies substantially among scale of a few months (22) to 1 year (23). Such el via the degree of susceptibility to and
seasonal b-CoVs (which also include HCoV-
pathogens. At one end of this immune spectrum, HKU1) are thought to cause repeated infections
natural infection with measles (1) or smallpox throughout life (24), although a significant
biennial component in their dynamics implies
1Lewis-Sigler Institute for Integrative Genomics, Princeton
University, Princeton, NJ 08540, USA. 2Department of
Ecology and Evolutionary Biology, Princeton University,
Princeton, NJ 08544, USA. 3Princeton Environmental
Institute, Princeton University, Princeton, NJ 08544, USA.
4Department of Bioengineering, McGill University, Montreal,
Quebec H3A 0C3, Canada. 5Department of Pathology and
Cell Biology, Columbia University Medical Center, New York,
NY 10032, USA. 6Wellcome Trust, London, UK. 7Departments
of Epidemiology and Immunology and Infectious Diseases,
Harvard School of Public Health, Boston, MA 02115, USA.
8Princeton School of Public and International Affairs,
Princeton University, Princeton, NJ 08544, USA. 9Fogarty
International Center, National Institutes of Health, Bethesda,
MD 20892, USA.
*These authors contributed equally to this work.
†Corresponding author. Email: [email protected]

Saad-Roy et al., Science 370, 811–818 (2020) 13 November 2020 1 of 8

RESEARCH | RESEARCH ARTICLE

Fig. 1. Schematic of the SIR(S) model with a flowchart depicting flows IS denotes individuals with secondary infection that transmit at a reduced
rate ab; and m denotes the birth rate (37). Illustrations and flowcharts
between immune classes. Here, SP denotes fully susceptible individuals; of the limiting SIR and SIRS models are also shown (where individuals are
IP denotes individuals with primary infection that transmit at rate b; R denotes either fully susceptible (S), infected (I), or fully immune (R)), along with
fully immune individuals (a result of recovery from either primary or a representative time series for the number of infections in each scenario.
secondary infection); SS denotes individuals whose immunity has waned at The population schematics were made through use of (62).
rate d and are now again susceptible to infection, with relative susceptibility e;

transmissibility of secondary infections (quan- exploration of qualitative medium-term dynam- We find that decreases in the susceptibility
tified by the parameters e and a, respectively). ics under different immunological scenarios. to secondary infection, e, can delay secondary
As shown in the representative time series of peaks (compare individual time courses for
Fig. 1, the SIR model results in recurrent epi- Seasonal transmission rates and the different values of e in Fig. 2, A to C). However,
demics fueled by births following the pandemic deployment of NPIs delayed peaks may then be larger, because of
peak; by contrast, the SIRS model typically susceptible accumulation (through demogra-
generates shorter interepidemic periods owing Medium-term dynamics will be shaped by phy or immune waning) and dynamic reso-
to the possibility of reinfection and the buffer- changes in the magnitude of transmission. To nance. These nonmonotonicities in the timing
ing of the fully susceptible birth cohort by par- explore the effect of NPIs, we considered two and size of secondary peaks also occur with
tially immune individuals (35). different scenarios for timed reductions in the climate-driven seasonal transmission in the
force of infection to 60% of its original value absence of NPIs (37), and the trends are qual-
We begin by characterizing the effect of tem- [in agreement with intermediate levels of itatively similar if NPIs are assumed to be
poral changes in the transmission rate brought social distancing in (21)]. In Fig. 2, A to C, we relaxed more gradually (fig. S11). Notably,
about by climate and the deployment of NPIs show the time courses of primary and second- the delay that social distancing may cause
on the predictions of the SIR(S) model under ary infections, assuming single periods of NPI in the timing of the secondary peak can also
a range of immunity assumptions. Next, we lasting from weeks 16 to 67 (Fig. 2A) or 16 to allow for further accumulation of fully suscep-
examine the impact of a transmission-reducing 55 (Fig. 2B) and two shorter periods during tible individuals. This is illustrated in the top
vaccine of varying efficacy relative to natural weeks 16 to 55 and weeks 82 to 93 separated panels of Fig. 2D, where the average infection
immunity. Finally, we estimate the postpan- by normal interactions (Fig. 2C). We further rate per infected individual for fully (bSP; red
demic immunity landscape and clinical case assume a seasonal transmission rate derived curve) and partially (ebSS; green curve) sus-
burden for different possible futures (38) from the climate of New York City (37), al- ceptible individuals for the social distancing
shaped by the various aspects of SARS-CoV-2 though in principle this seasonality could also scenario outlined in Fig. 2C are shown. We
biology as well as the presence or absence of be derived from other nonclimate factors (25). contrast a reduction in susceptibility to sec-
these external drivers and interventions, as The weekly reproduction numbers correspond- ondary infection of 50% (e = 0.5, left panels)
well as vaccine refusal. To focus on the dynamic ing to these three scenarios are shown in fig. S2, with no reduction in susceptibility to sec-
impact of natural and vaccinal immunity, we D to F. Although these reproduction numbers ondary infection (e = 1, right panels). The
begin with a simple homogeneous model, are based on those obtained for the related corresponding fraction of primary (blue)
which averages across known heterogeneities b-CoV HCoV-HKU1 and are in general lower and secondary (purple) cases are presented
in COVID-19 transmission and severity [age than those estimated during the early stages in the bottom panels. As can be seen, when
(39), superspreading events (40), etc.]. We then of the SARS-CoV-2 pandemic (41), they may be the secondary peak does occur, the decrease
use heterogeneous model extensions to show more appropriate for considering the longer- in susceptibility to secondary infection (e < 1),
that such heterogeneities do not impact our term transmission dynamics.

Saad-Roy et al., Science 370, 811–818 (2020) 13 November 2020 2 of 8

RESEARCH | RESEARCH ARTICLE

Fig. 2. Seasonality in transmission rates and NPIs modulate disease dynamics. (left column) and e = 1 (right column) for the NPI scenario outlined in (C). (E) Time
(A to C) Effect of NPI adoption on the time series of primary (solid lines) and series of estimated numbers of severe infections for the NPI scenario defined in
secondary (dashed lines) infections with a seasonal transmission rate derived (C) for four different estimates of the fraction of severe cases during primary
from the climate of New York City with no lag between seasonality and epidemic infections (xsev,p) and secondary infections (xsev,s) with e = 0.5 (top row) and e =
onset. NPIs that reduce the transmission rate to 60% of the estimated climate 1 (bottom row). These are xsev,p = 0.14, xsev,s = 0 (solid red line); xsev,p = 0.14, xsev,s =
value are assumed to be adopted during weeks 16 to 67 (A), weeks 16 to 55 (B), 0.07 (dashed green line); xsev,p = 0.14, xsev,s = 0.14 (dashed and dotted blue line);
or weeks 16 to 55 as well as weeks 82 to 93 (C). Colors denote individual time and xsev,p = 0.14, xsev,s = 0.21 (purple line with long and short dashes). In all
courses for different values of e. (D) Time series of the average daily infection rate panels, the relative transmissibility of secondary infections and duration of natural
per infected individual of fully susceptible (red line) and partially susceptible (green immunity are taken to be a = 1 and 1/d = 1 year, respectively. The effects
line) individuals (top row) and the fraction of the population that is infected with of NPIs and other parameter variations can be explored interactively at https://
primary (blue line) and secondary (purple line) infections (bottom row), for e = 0.5 grenfelllab.princeton.edu/sarscov2dynamicsplots.

considered in the left panels, results in a solid red line); (ii) a reduced number of severe proportion of secondary infections increases
greater number of primary infections during cases with secondary infection relative to pri- during the later stages of the pandemic, these
the second peak relative to the panels on the mary infection (xsev,s = 0.07; dashed green line); findings stress that clinical epidemiological
right, where e is 1 [and the secondary infec- (iii) comparable proportions of severe cases studies of repeat infections will be critical for
tion rate per case (green curves) rises sharply]. (xsev,s = 0.14; dashed-dotted blue line); and proper planning of health care systems. We also
(iv) a hypothetical greater proportion of severe do not consider any long-term clinical impact
Next, an essential part of the planning and cases with secondary infection (xsev,s = 0.21; of infection here (43). The impact of increases
management of future SARS-CoV-2 infections purple line with short and long dashes), pos- in clinical severity with age is addressed below.
is the ability to characterize the magnitude sibly owing to phenomena such as ADE. When
and timing of severe cases requiring hospi- the assumed fraction of severe subsequent Vaccination
talization. In Fig. 2E we consider four possible infections is high, the fraction of the population
scenarios for the fraction of severe secondary with severe infections during subsequent The availability of an effective vaccine would
cases, xsev,s (37), on the basis of the scenario infection peaks is found to be comparable to be a key intervention against SARS-CoV-2, and
depicted in Fig. 2C and assuming 14% of pri- or even to slightly exceed that observed during numerous candidates are in development
mary cases are severe (42): (i) no severe cases the initial pandemic peak (Fig. 2E). As the (44, 45). Intuitively, if the effective vaccina-
associated with secondary infection (xsev,s = 0; tion rate is sufficiently high, then vaccinal

Saad-Roy et al., Science 370, 811–818 (2020) 13 November 2020 3 of 8

RESEARCH | RESEARCH ARTICLE

herd immunity generated by a transmission- of NPIs according to the scenario described in period of 8 months for persons aged at least
blocking vaccine could control or eliminate Fig. 2B. We assume that a constant proportion, 6 months in the United States, although rates
the infection. However, this becomes harder to n, ranging from 0% ≤ n ≤ 1% of the fully and between different nations varied (47). This
achieve when vaccinal and natural immunity partially susceptible populations (SP and SS), is crudely corresponds to a weekly vaccination
is imperfect and secondary infections occur, effectively vaccinated every week and acquires rate of 1% (48) (as with other parameter var-
or when logistical or other constraints limit transmission-blocking immunity for, on aver- iations, different scenarios for vaccination can
vaccine deployment. We extend the model (37) age, a period 1/dvax. For comparison, it was be explored with the accompanying Shiny
to include a vaccinated class, V, and make estimated that in response to the 2009 H1N1 application). Finally, we assume that the im-
the relatively optimistic assumption that a pandemic, one or more doses of the monoval- munity conferred from effective vaccination
transmission-reducing vaccine begins to be ent vaccine were administered to 80.8 million wanes at rate dvax, which in general may differ
introduced to general populations (the tvax) vaccinees during October 2009 to May 2010 from the waning rate of immunity from nat-
after 1.5 years. We also consider seasonal trans- in the United States (46), which implies a rate ural infection, d. The modified set of ordi-
mission rates, as in fig. S3, and the deployment of vaccination coverage of about 27% after a nary differential equations in this scenario

Fig. 3. Impact of vaccination and vaccinal immunity on disease dynamics. introduction of the vaccine) are plotted as a function of the weekly vaccination
(A) Modified model flowchart that incorporates a vaccinated class V (37). rate n and the duration of vaccinal immunity 1/dvax. (F to I) Time series
(B) Total infected fraction of the population at equilibrium as a function of the of the various immune classes plotted for different values of the vaccination
vaccination rate n for different values of the duration of vaccinal immunity rate n. The top row [(F) and (H)] contains the time series of primary (IP, solid
(1/dvax = 0.25 years, green lines; 1/dvax = 0.5 years, red lines; and 1/dvax = 1 year, lines) and secondary (IS, dashed lines) infections, whereas the bottom row
blue lines) and the susceptibility to secondary infection (e = 0.5, solid lines; e = [(G) and (I)] contains the time series of the fully susceptible (SP, solid lines), naturally
0.7, dashed lines; and e = 1, dotted lines). (C) Daily proportion of susceptibles immune (R, dashed lines), and partially immune (SS, dotted lines) subpopulations.
who must be vaccinated in order to achieve a disease-free state at equilibrium as a The duration of vaccinal immunity is taken to be 1/dvax = 0.5 years (shorter than
function of e for different values of the duration of vaccinal immunity (1/dvax = natural immunity) in (F) and (G), and 1/dvax = 1 year (equal to natural immunity) in
0.25 years, solid line; 1/dvax = 0.5 years, dashed line; and 1/dvax = 1 year, dotted (H) and (I). In (D) to (I), the relative susceptibility to secondary infection, relative
line). In (B) and (C), the relative transmissibility of secondary infections and transmissibility of secondary infections, and duration of natural immunity are taken to
duration of natural immunity are taken to be a = 1 and 1/d = 1 year, respectively, be e = 0.7, a = 1, and 1/d = 1 year, respectively. Vaccination is introduced 1.5 years after
and the transmission rate is derived from the mean value of seasonal New York the onset of the epidemic (i.e., during the 79th week) following a 40-week period of
City–based weekly reproduction numbers (R 0 = 1.75) (fig. S2C) (37). (D and E) The social distancing during which the force of infection was reduced to 60% of its original
ratio of the total number of primary (D) and secondary (E) infections with value during weeks 16 to 55 (i.e., the scenario described in Fig. 2B), and a seasonal
vaccination versus without vaccination, during years 1.5 to 5 (i.e., after the transmission rate derived from the climate of New York City with no lag is assumed.

Saad-Roy et al., Science 370, 811–818 (2020) 13 November 2020 4 of 8

RESEARCH | RESEARCH ARTICLE

corresponding to the flowchart in Fig. 3A is relatively low vaccination rates are sufficient and the nature of the adaptive immune re-
also presented in (37). to achieve zero infections at steady state. How- sponse (49).
ever, as immunity becomes more imperfect
In Fig. 3B, we begin by considering the long- (larger e), increasingly high vaccination rates We next explore the medium-term dynamic
term equilibrium infection burden (37) driven are required to eliminate infections, particu- effect of vaccination. Figure 3D shows the
by vaccination at a weekly rate n, for a variety larly when the duration of vaccinal immunity ratio of the total number of primary infections
of immunity assumptions. As expected, a re- is short. This is further emphasized in Fig. 3C, during years 1.5 to 5 (i.e., after the vaccine is
where the minimum vaccination rate n required introduced) relative to the zero vaccination
duction in the susceptibility to secondary in- to achieve a disease-free state at equilibrium case for different values of the vaccination rate
fections (e) results in a smaller number of (37) is shown as a function of e for different n and the duration of vaccinal immunity 1/dvax.
infections at steady state in the absence of values of the duration of vaccinal immunity. Figure 3E shows the equivalent for secondary
vaccination. Further, both e and the duration These results underline that reductions in in- infections. The burden of primary infection de-
of vaccinal immunity (1/dvax) affect the vacci- fection achievable through vaccination are in- creases with increasing vaccination rate for a
nation rate required to achieve a disease-free herently related to the efficacy of the vaccine given value of vaccinal immunity, 1/dvax. How-
ever, for the shortest durations of vaccinal im-
state at equilibrium. At the limit of fully immu- munity, achievable reductions in the number
nizing primary infections and vaccines (e = 0), of secondary cases begin to plateau even for
high vaccination rates. This saturation is due
Fig. 4. Time series of the fraction of the population with severe primary or secondary cases (top) to the rapid return of vaccinated individuals to
and area plots of the fraction of the population comprising each immune (SP, R, SS, V) or infection the partially susceptible class if vaccinal immu-
(IP, IS) class (bottom) over a 5-year time period under four different future scenarios. In all plots, the nity is short-lived. Further, if vaccinal immunity
relative transmissibility of secondary infections (a) is taken to be 1, the fraction of severe primary cases (xsev,p) wanes very rapidly, vaccination can transiently
is assumed to be 0.14, a seasonal transmission rate derived from the climate of New York City with no lag increase the total number of secondary cases.
is assumed, and a period of social distancing during which the force of infection is reduced to 60% of its To further emphasize the dependence of the
original value during weeks 16 to 55 (i.e., the scenario described in Fig. 2B) is enforced. (A and B) Two model results on the vaccination rate and du-
scenarios in which no vaccination occurs: a more pessimistic natural immunity scenario, with e = 0.7, 1/d = ration of vaccinal immunity, we present time
0.5 years, and 21% of secondary cases being severe (A) and a more optimistic natural immunity scenario, courses of infections and immunity for dif-
with e = 0.5, 1/d = 2 years, and 7% of secondary cases being severe (B). (C and D) Two scenarios in which ferent durations of vaccinal immunity and
vaccination is introduced at a weekly rate n of 1% at tvax of 1.5 years after the onset of the pandemic: vaccination rates in Fig. 3, F to I. In line with
with all the parameters in (A) along with vaccinal immunity lasting 1/dvax of 0.25 years (C) or with all the intuition, the model illustrates that both high
same parameters as in (B) along with vaccinal immunity lasting 1/dvax of 1 year (D). vaccination rates and relatively long durations
of vaccine-induced immunity are required to
achieve the largest reductions in secondary
infection burdens.

Infection, disease, and immunity landscape for
different possible futures

Figure 4 is a synoptic view of the medium-
term impact of vaccination and natural immu-
nity on the immune landscape and incidence
of severe disease. We consider four scenarios,
assuming seasonal transmission (as in fig. S3)
and social distancing according to the pat-
tern depicted in Fig. 2B. Figure 4, A and B,
corresponds to futures without vaccination,
with Fig. 4A illustrating a more pessimistic
scenario of greater susceptibility to secondary
infections (e = 0.7), a relatively short period
of natural immunity (1/d = 0.5 years), and a
greater proportion of severe cases with second-
ary infection. In contrast, the more optimistic
future of Fig. 4B assumes reduced susceptibil-
ity to secondary infections (e = 0.5), a longer
duration of natural immunity (1/d = 2 years),
and a smaller proportion of severe cases with
secondary infection. In both cases, the initial
pandemic wave is the same, but in the more
optimistic scenario (Fig. 4B), natural immunity
is longer lasting and, consequently, subsequent
infection peaks are delayed. Furthermore, the
reduction in susceptibility to secondary infec-
tion (smaller e) in Fig. 4B suppresses the later
peaks dominated by secondary infections (Fig.
4A), and substantially less depletion of fully
susceptible individuals occurs. In Fig. 4, C and D,

Saad-Roy et al., Science 370, 811–818 (2020) 13 November 2020 5 of 8

RESEARCH | RESEARCH ARTICLE

Fig. 5. Effect of vaccine refusal on disease dynamics. (A) Daily proportion immunity to still be achieved as a function of the contact rate among vaccine

of vaccine-adopting individuals from the partially and fully susceptible immune classes refusers c22 (37). In (A) and (B), the transmission rate is derived from the
mean value of seasonal New York City–based weekly reproduction numbers
who must be immunized in order to achieve R0 < 1 as a function of the fraction (R 0 = 1.75) (37) (fig. S2C). (C) Time series of the fraction of the population with
of the population that refuses the vaccine (37) for different values of the duration severe primary or secondary cases (top) and area plots of the fraction of the

of vaccinal immunity (1/dvax = 0.25 years, solid line; 1/dvax = 0.5 years, dashed population comprising each immune (SP, R, SS, V) or infection (IP, IS) class
line; 1/dvax = 1 year, dotted line) and different values of the susceptibility to (bottom) over a 5-year time period. The parameters in the left two series are
secondary infection e [e = 0.5 (left) e = 0.7 (middle) or e = 1 right)]. (Top row)
Homogeneous transmission between vaccine adopters and refusers (c11 = c12 = identical to those in Fig. 4C, and the parameters in the right two series are identical
c21 = c22 = 1). (Middle row) Increased transmission associated with vaccine refusers
(c11 = 1, c12 = 1.25, c21 = 1.25, and c22 = 1.5). (Bottom row) Decreased transmission to those in Fig. 4D. Additionally, the fraction of the population refusing vaccines
associated with vaccine refusers (c11 = 1, c12 = 0.825, c21 = 0.825, and c22 = 0.75).
(B) Maximum fraction of the population that can refuse vaccination for herd is taken to be N2 = 0.3. (Top row) Homogeneous mixing with c11 = c12 = c21 = c22 = 1.
(Bottom row) Increased contacts among vaccine refusers and c11 = 1, c12 = 1.25,
c21 = 1.25, and c22 = 1.5.

these pessimistic and optimistic scenarios are if a transmission-blocking vaccine confers a Impact of heterogeneity
translated into futures with vaccination, which relatively long period of protection, and if Transmission and clinical heterogeneity
is assumed to be introduced at a weekly rate we make optimistic assumptions regarding COVID-19 shows marked heterogeneity in
n of 1% after a tvax of 1.5 years. The future the nature of the adaptive immune response transmission and clinical severity with age and
described in Fig. 4C assumes all the same (Fig. 4D), a sufficient proportion of fully sus- other variables (40). There are also marked
outcomes as in Fig. 4A and incorporates vac- ceptible individuals can be immunized to individual heterogeneities, often associated
cination with short-lived vaccinal immunity suppress future outbreaks within the 5-year with superspreading events. It is useful to
(1/dvax = 0.25 years). The future presented in time period considered. These trends are distinguish “environmental” heterogeneity,
Fig. 4D assumes all the same outcomes as in qualitatively conserved for different vaccine where high transmission is associated with
Fig. 4B in addition to vaccinal immunity last- deployment strategies, such as a pulse of im- local environmental (or sociological) factors
ing for 1/dvax of 1 year. munization after a tvax of 1.5 years in which a such as low air exchange, and “intrinsic” het-
fixed percentage of the fully and partially sus- erogeneity, e.g., where certain individuals
Figure 4, C and D, emphasizes the impor- ceptible populations (SP and SS) are vacci- have consistently higher contact rates (40).
tant role that even an imperfect vaccine could nated (fig. S12). However, without sustained A number of studies have explored the pos-
have on SARS-CoV-2 dynamics and control immunization strategies, the waning of vac- sibility that intrinsically higher transmission
[compare with (7, 9, 10)]. Vaccination substan- cinal immunity results in a lower susceptible rates for some individuals could reduce the
tially reduces subsequent peaks in clinically depletion over time and larger future out- immune threshold for natural or vaccinal herd
severe cases, although in the pessimistic future breaks relative to the scenarios presented in immunity to COVID-19 (39, 50), echoing clas-
later infection peaks dominated by secondary Fig. 4, C and D. sical theory (51).
infections can still occur (Fig. 4C). Furthermore,

Saad-Roy et al., Science 370, 811–818 (2020) 13 November 2020 6 of 8

RESEARCH | RESEARCH ARTICLE

We approximate the impact of intrinsic het- for vaccine refusers (bottom). Notably, we find burden, and immunity landscape to COVID-19.
erogeneity using a two-subpopulation exten- that when vaccine refusers have increased con- In locations where we expect substantial cli-
sion of our homogeneous model (37) (figs. S13 tact rates relative to the rest of the population, matically driven seasonal variation in trans-
to S15). As well as varying transmission be- vaccination alone may not be able to prevent mission, such as New York City, the model
tween groups, the model assesses covariation an outbreak (Fig. 5B). Alternatively, a decrease predicts that a reduction in susceptibility to
between transmission rate and clinical sever- in contact rates for vaccine refusers decreases secondary infection or a longer duration of
ity. For example, this framing broadly reflects their impact. Finally, in Fig. 5C, we reproduce immunity may lead to a larger secondary in-
age-structured heterogeneities, in which more the area plots from Fig. 4, C and D, assuming fection peak, which may occur earlier if the
clinically threatened older groups might have that 30% of the population refuses the vac- duration of natural immunity is longer. With
a lower (because of fewer contacts or possible cine. This estimate is broadly consistent with smaller annual fluctuations in climate, we find
shielding) or higher [if long-term care facilities recent polls conducted in the United States that this nonmonotonic behavior is increasingly
are hit (40)] transmission rate. We show that and Canada (54, 55). We find that the overall suppressed; however, this effect is sensitive to
moderate heterogeneities do not affect our disease burden critically depends on the dura- the assumed form of climatic influences on
qualitative projections about the impact of tion and strength of immunity and is larger if SARS-CoV-2 transmission, which we have taken
partial natural or vaccinal immunity on epi- vaccine refusers have higher contact rates rela- here to be very similar to those of the related
demic dynamics (figs. S14 and S15). As ex- tive to the rest of the population (compare top b-CoV HCoV-HKU1. The subsequent pattern of
pected, intrinsic transmission heterogeneity and bottom rows). infection peaks is even more sensitive to the
does reduce future burden if there is strong relative fraction and transmissibility of primary
and durable immunity (39, 50) (compare Fig. Caveats and secondary cases, as well as the fraction of
4B with figs. S14B and S15B, particularly the severe cases for each category. Overall, whereas
subsequent epidemic peaks), because high- To focus on immune dynamics, we have made climatic effects or other seasonal modulators
transmission individuals would become im- several simplifying assumptions. First, we have of the transmission rate increase in importance
mune early, reducing average reproduction assumed that transmission of SARS-CoV-2 is as the pandemic progresses (25), our results
ratios [and modulating the herd immunity seasonal and similar to that of the related underline that understanding the immunol-
threshold (39, 50)]. However, this impact of in- b-CoV HCoV-HKU1, although we have also ogy of secondary infection (which modulates
trinsic heterogeneity is weakened [or “buffered” explored the effect of diminished seasonality susceptible supply) is even more dynamically
(35)] if immunity is imperfect (compare Fig. 4A (37). Second, we have simplified the important important, especially in the medium term.
with figs. S14A and S15A); this is because high- role for heterogeneities, such as age, clinical
ly transmissive individuals (e.g., those with a severity, transmissibility (40), and adaptive The pandemic trajectory can also be sub-
larger social network) contribute proportion- immune response (16) to primary and second- stantially altered by mass deployment of vac-
ately more to secondary transmission when ary (and beyond) infections. Notably, higher cines; however, the impact on burden is strongly
they enter the partially susceptible state. Again, viral loads or contact rates in some individuals dependent on the efficacy of the vaccine and
this subtlety illustrates the complexities of even can lead to superspreading events and heter- the nature of the adaptive immune response.
simple variations in immune life history. ogeneous transmission patterns (40). Addi- Recent vaccine trials in mice and rhesus ma-
tionally, the severity of an infection, especially caques indicate the generation of robust im-
Vaccine hesitancy if associated with higher viremia than in mild mune responses, clinical protection from severe
cases, could affect the nature of the subse- disease, and no evidence of ADE after viral
There is an extensive body of theory on vaccine quent adaptive immune response, via antigen- challenge, possibly indicating a more optimis-
hesitancy (52, 53). In the homogeneous case, driven expansion of the antibody response (17) tic immune scenario (44). Vaccine hesitancy
vaccine refusal essentially trades off against or exhaustion of the T cell response (18). We could also decrease vaccination rates (53), lead-
vaccine uptake; however, if refusers are spa- have explored the effect of these heteroge- ing to lower levels of population immunity.
tially or socially clustered, there may be more neities on disease dynamics via a simple model Nevertheless, even with imperfect vaccinal im-
impact of refusal, both epidemiologically and extension (figs. S13 to S15); we find that dy- munity and moderate vaccination rates, our
in terms of the social contagion which under- namic impacts of immune variation projected results indicate that vaccination may accel-
lies it (52). We use a simple adaptation of by our homogeneous model are qualitatively erate pandemic control. Ultimately, quanti-
our models (37) to explore how transmission robust to these inclusions. Finally, we have tatively projecting the impact of vaccination,
heterogeneity influences the epidemiological considered highly simplified scenarios for antivirals, and therapeutics will require more
impact of hesitancy (Fig. 5). As in the vaccina- NPI adoption and vaccination. granular immuno-epidemiological models;
tion rate titration presented in Fig. 3, B and however, parameterizing such models will con-
C, a larger fraction of vaccine refusers in a The dynamic impact of these and other pa- tinue to present huge challenges for this novel
homogeneous population increases the nec- rameter variations can be explored interac- virus. We argue that a family of simple and
essary vaccination rate for herd immunity tively at https://grenfelllab.princeton.edu/ more complex models, with a careful focus
(Fig. 5A, top). This effect is amplified when the sarscov2dynamicsplots. For example, strat- on model comparison and averaging, is the
susceptibility to secondary infection is high egies to suppress future outbreaks [e.g., (56)] way ahead (57).
(compare columns) or the duration of vaccinal could be simulated by increasing the duration
immunity is short (compare individual curves). and strength of NPIs, then exploring optimal Our work underlines that relying on the
In the heterogeneous case, where vaccine re- vaccine deployment as vaccines are developed status of infection of an individual as the main
fusers are assumed to have different transmis- and rolled out. See (37) for a full discussion of observable during an ongoing epidemic is in-
sion rates because of higher or lower adherence all caveats and future directions. sufficient to characterize the complex immune
to NPIs, the minimum vaccination rate to landscape generated by the pandemic. This is
achieve herd immunity is further altered. This Conclusion in line with ongoing calls for the development
can be seen by comparing the rows of Fig. 5A, of a Global Immunological Observatory for the
ordered on the basis of homogeneous contact We have examined how plausible variations surveillance of population-level susceptibility
rates (top), increased contact rates for vaccine in the natural immune response after SARS- and immunity to circulating pathogens, as well
refusers (middle), and decreased contact rates CoV-2 infection and vaccination could inter- as the emergence of new strains (58–60). Given
act with seasonal drivers and NPIs to shape the increasingly recognized importance of both
the medium-term epidemic dynamics, clinical

Saad-Roy et al., Science 370, 811–818 (2020) 13 November 2020 7 of 8

RESEARCH | RESEARCH ARTICLE

T cell–mediated (14, 15) and antibody-mediated 14. A. Grifoni et al., Cell 181, 1489–1501.e15 (2020). rate of 0.91% of the remaining unvaccinated population would
(11–13) adaptive immune responses in the rec- 15. J. Braun, L. Loyal, M. Frentsch, D. Wendisch, P. Georg, be required.
overy from SARS-CoV-2 infection, regular test- 49. M. E. Halloran et al., Design and Analysis of Vaccine Studies
ing of antibody presence, and correlates of F. Kurth, S. Hippenstiel, M. Dingledey, B. Kruse, F. Fauchere, (Springer, 2010).
protection such as neutralization, as well as E. Baysal, M. Mangold, L. Henze, R. Lauster, M. Mall, 50. R. Aguas, R. M. Corder, J. G. King, G. Goncalves, M. U. Ferreira,
T cell immunity, in parallel with viral testing, K. Beyer, J. Roehmel, J. Schmitz, S. Miltenyi, M. A. Mueller, M. G. M. Gomes, Herd immunity thresholds for SARS-CoV-2
will be required to adequately characterize M. Witzenrath, N. Suttorp, F. Kern, U. Reimer, H. Wenschuh, estimated from unfolding epidemics. medRxiv 20160762
population-level natural and vaccinal immu- C. Drosten, V. M. Corman, C. Gisecke-Theil, L.-E. Sander, [Preprint]. 31 August 2020. https://doi.org/10.1101/
nity to this pathogen. Specifically, our model A. Thiel, Presence of SARS-CoV-2 reactive T cells in COVID-19 2020.07.23.20160762.
indicates a key need to establish (i) the dura- patients and healthy donors. medRxiv 20061440 [Preprint]. 51. R. M. Anderson, R. M. May, Infectious Diseases of Humans:
tion and strength of transmission-blocking 22 April 2020. https://doi.org/10.1101/2020.04.17.20061440. Dynamics and Control (Oxford Univ. Press, 1991).
and clinical immunity after primary (and sub- 16. D. Mathew et al., Science 369, eabc8511 (2020). 52. M. Salathé, S. Bonhoeffer, J. R. Soc. Interface 5, 1505–1508
sequent) infection and vaccination; (ii) pop- 17. Y. Wang et al., J. Clin. Invest. 10.1172/JCI138759 (2020). (2008).
ulation and individual variations in these 18. S. De Biasi et al., Nat. Commun. 11, 3434 (2020). 53. E. Dubé et al., Hum. Vaccin. Immunother. 9, 1763–1773
parameters (age, sex, etc.); and (iii) the impact 19. Q.-X. Long et al., Nat. Med. 26, 845–848 (2020). (2013).
of viral evolution, coinfection, and other path- 20. D. F. Gudbjartsson et al., N. Engl. J. Med. 10.1056/ 54. S. M. O’Keefe, “One in Three Americans Would Not Get
ogen characteristics on COVID-19 infection NEJMoa2026116 (2020). COVID-19 Vaccine” (2020; https://news.gallup.com/poll/
and disease. Quantifying these parameters 21. S. M. Kissler, C. Tedijanto, E. Goldstein, Y. H. Grad, M. Lipsitch, 317018/one-three-americans-not-covid-vaccine.aspx).
will require long-term major investments in Science 368, 860–868 (2020). 55. K. Frank, R. Arim, “Canadians’ willingness to get a COVID-19
integrated viral and immune surveillance. 22. M. R. Macnaughton, Infect. Immun. 38, 419–423 (1982). vaccine when one becomes available: What role does trust
Moving beyond the current pandemic, these 23. K. A. Callow, H. F. Parry, M. Sergeant, D. A. J. Tyrrell, play?”(2020; https://www150.statcan.gc.ca/n1/pub/45-28-
structures (and associated developments in Epidemiol. Infect. 105, 435–446 (1990). 0001/2020001/article/00043-eng.htm).
biology, informatics, and translation) will be 24. E. R. Gaunt, A. Hardie, E. C. Claas, P. Simmonds, 56. M. G. Baker, A. Kvalsvig, A. J. Verrall, Med. J. Australia (2020);
powerful bases for understanding and combat- K. E. Templeton, J. Clin. Microbiol. 48, 2940–2947 (2010). https://www.mja.com.au/journal/2020/new-zealands-covid-
ing inevitable future microbial threats (58–60). 25. R. E. Baker, W. Yang, G. A. Vecchi, C. J. E. Metcalf, 19-elimination-strategy.
B. T. Grenfell, Science 369, 315–319 (2020). 57. C. Viboud et al., Epidemics 22, 13–21 (2018).
This work emphasizes the complex depen- 26. W. Liu et al., J. Infect. Dis. 193, 792–795 (2006). 58. C. J. E. Metcalf et al., Lancet 388, 728–730 (2016).
dence of the immune landscape generated by 27. W.-C. Cao, W. Liu, P.-H. Zhang, F. Zhang, J. H. Richardus, 59. C. J. E. Metcalf, M. J. Mina, A. K. Winter, B. T. Grenfell,
SARS-CoV-2 infection on the presently uncer- N. Engl. J. Med. 357, 1162–1163 (2007). Lancet 389, 252 (2017).
tain nature of the adaptive immune response 28. O.-W. Ng et al., Vaccine 34, 2008–2014 (2016). 60. M. J. Mina et al., eLife 9, e58989 (2020).
to this virus and the efficacy of potential future 29. K. W. Ng et al., Pre-existing and de novo humoral immunity to 61. C. M. Saad-Roy et al., Code for immune life-history, vaccination,
vaccines. Depending on how these unfold, the SARS-CoV-2 in humans. bioRxiv 095414 [Preprint]. and the dynamics of SARS-CoV-2 over the next five years.
model predictions for future clinical burdens 23 July 2020. https://doi.org/10.1101/2020.05.14.095414. Zenodo (2020; doi: 10.5281/zenodo.4025265).
range from sustained epidemics to near–case 30. J. A. Tetro, Microbes Infect. 22, 72–73 (2020). 62. Created with BioRender.com.
elimination. Consequently, accurately charac- 31. M. Jaume et al., Hong Kong Med. J. 18 (suppl. 2), 31–36
terizing the individual immune life histories (2012). ACKNOWLEDGMENTS
and the cumulative immune landscape of the 32. M. S. Yip et al., Virol. J. 11, 82 (2014).
population to SARS-CoV-2 primary and sec- 33. M. S. Yip et al., Hong Kong Med. J. 22 (suppl. 4), 25–31 We acknowledge useful discussions with B.Thompson and the
ondary infection and vaccination will be crit- (2016). Wellcome COVID-19 Futures Group and the members of the DELVE
ical for the management and control of the 34. A. D. Becker, B. T. Grenfell, PLOS ONE 12, e0185528 Immunology group. Funding: C.M.S.-R. acknowledges support
ongoing pandemic. (2017). from the Natural Sciences and Engineering Research Council of
35. S. E. Morris et al., J. R. Soc. Interface 12, 20141245 (2015). Canada through a Postgraduate-Doctoral Scholarship. C.E.W. is an
REFERENCES AND NOTES 36. V. E. Pitzer et al., Science 325, 290–294 (2009). Open Philanthropy Project fellow of the Life Sciences Research
37. Materials and methods are available as supplementary Foundation. R.E.B. is supported by the Cooperative Institute
1. B. M. Laksono, R. D. de Vries, S. McQuaid, W. P. Duprex, materials. for Modelling the Earth System (CIMES). S.A.L. and C.M.S.-R.
R. L. de Swart, Viruses 8, 210 (2016). acknowledge support from the James S. McDonnell Foundation
38. J. Bedford et al., Lancet 10.2139/ssrn.3678593 [Preprint] 21st Century Science Initiative Collaborative Award in Understanding
2. P. D. Ellner, Infection 26, 263–269 (1998). (2020). https://papers.ssrn.com/sol3/papers.cfm?abstract_ Dynamic and Multi-scale Systems. S.A.L. acknowledges support from
3. J. Dushoff, J. B. Plotkin, S. A. Levin, D. J. D. Earn, Proc. Natl. id=3678593 the C3.ai Digital Transformation Institute, the National Science
Foundation under grant CNS-2027908, and the National Science
Acad. Sci. U.S.A. 101, 16915–16916 (2004). 39. T. Britton, F. Ball, P. Trapman, Science 369, 846–849 Foundation Expeditions Grant CCF1917819. B.T.G. acknowledges
4. V. E. Pitzer et al., PLOS Pathog. 11, e1004591 (2015). (2020). support from the U.S. CDC and Flu Lab. Author contributions:
5. N. Ferguson, R. Anderson, S. Gupta, Proc. Natl. Acad. Sci. U.S.A. C.M.S.-R., C.E.W., C.J.E.M., and B.T.G. designed the study. C.M.S.-R.
40. B. M. Althouse, E. A. Wenger, J. C. Miller, S. V. Scarpino, and C.E.W. performed the simulations and wrote the manuscript.
96, 790–794 (1999). A. Allard, L. Hebert-Dufresne, H. Hu, Stochasticity and C.M.S.-R., C.E.W., R.E.B., C.J.E.M., and B.T.G. analyzed the results.
6. B. S. Graham, Science 368, 945–946 (2020). heterogeneity in the transmission dynamics of SARS-CoV-2. S.E.M. developed the Shiny application. C.M.S.-R., C.E.W., and
7. S. Gandon, M. J. Mackinnon, S. Nee, A. F. Read, Nature 414, arXiv:2005.13689 [q-bio.PE] (27 May 2020). S.A.L. developed the analytical equilibrium analysis. All authors
contributed to interpreting the results and editing the manuscript.
751–756 (2001). 41. S. W. Park et al., J. R. Soc. Interface 17, 20200144 Competing interests: The authors have no competing interests.
8. G. Anichini et al., Vaccines 8, 66 (2020). (2020). Data and materials availability: This manuscript contains no new
9. N. Arinaminpathy et al., Am. J. Epidemiol. 186, 92–100 (2017). data, and all referenced data sets are publicly available. Code
10. T. A. Perkins et al., PLOS Comput. Biol. 15, e1006710 (2019). 42. Z. Wu, J. M. McGoogan, JAMA 323, 1239–1242 (2020). for all analysis is available at Zenodo (61). This work is licensed
11. J. Huang, T. Mao, S. Li, L. Wu, X. Xu, H. Li, C. Xu, F. Su, J. Dai, 43. V. O. Puntmann et al., JAMA Cardiol. (2020). under a Creative Commons Attribution 4.0 International (CC BY 4.0)
license, which permits unrestricted use, distribution, and reproduction
J. Shi, J. Cai, C. Huang, X. Lin, D. Chen, X. Lin, B. Sun, 44. M. J. Mulligan, K. E. Lyke, N. Kitchin, J. Absalon, A. Gurtman, in any medium, provided the original work is properly cited. To
S. Tang, Long period dynamics of viral load and antibodies S. P. Lockhart, K. Neuzil, V. Raabe, R. Bailey, K. A. Swanson, view a copy of this license, visit https://creativecommons.org/
for SARS-CoV-2 infection: An observational cohort study. P. Li, K. Koury, W. Kalina, D. Cooper, C. Fonter-Garfias, licenses/by/4.0/. This license does not apply to figures/photos/
medRxiv 20071258 [Preprint]. 27 April 2020. https://doi.org/ P.-Y. Shi, O. Tuereci, K. R. Tompkins, E. E. Walsh, R. Frenck, artwork or other content included in the article that is credited to a
10.1101/2020.04.22.20071258. A. R. Falsey, P. R. Dormitzer, W. C. Gruber, U. Sahin, K. U. Jansen, third party; obtain authorization from the rights holder before
12. N. M. A. Okba et al., Emerg. Infect. Dis. 26, 1478–1488 Phase 1/2 Study to Describe the Safety and Immunogenicity using such material.
(2020). of a COVID-19 RNA Vaccine Candidate (BNT162b1) in
13. W. Tan, Y. Lu, J. Zhang, J. Wang, Y. Dan, Z. Tan, X, He, Adults 18 to 55 Years of Age: Interim Report. medRxiv SUPPLEMENTARY MATERIALS
C. Qian, Q. Sun, Q. Hu, H. Liu, S. Ye, X. Xiang, Y. Zhou, 20142570 [Preprint]. 1 July 2020. https://doi.org/10.1101/
W. Zhang, Y. Guo, X.-H. Wang, W. He, X. Wan, F. Sun, Q. Wei, 2020.06.30.20142570. science.sciencemag.org/content/370/6518/811/suppl/DC1
C. Chen, G. Pan, J. Xia, Q. Mao, Y. Chen, Viral Kinetics and Materials and Methods
Antibody Responses in Patients with COVID-19. medRxiv 45. N. van Doremalen et al., Nature 10.1038/s41586-020-2608-y Supplementary Text
20042382 [Preprint]. 26 March 2020. https://doi.org/10.1101/ (2020). Figs. S1 to S15
2020.03.24.20042382. References (63–70)
46. U.S. Centers for Disease Control and Prevention, “Final MDAR Reproducibility Checklist
estimates for 2009–10 Seasonal Influenza and Influenza A
(H1N1) 2009 Monovalent Vaccination Coverage—United States, View/request a protocol for this paper from Bio-protocol.
August 2009 through May, 2010” (2011); https://www.cdc.
gov/flu/fluvaxview/coverage_0910estimates.htm. 8 July 2020; accepted 16 September 2020
Published online 21 September 2020
47. H. Gilmour, N. Hofmann, H1N1 Vaccination (Statistics 10.1126/science.abd7343
Canada, 2015); https://www150.statcan.gc.ca/n1/pub/82-
003-x/2010004/article/11348-eng.htm.

48. Assuming a simplified scenario in which a constant proportion
of the total population is vaccinated weekly, the proportion of
the population that is not vaccinated, U, can be modeled to
decrease exponentially in time as U(t) = exp(–nt). Consequently,
in order to achieve 27% vaccination coverage after an 8-month
(or 34.7-week) period [i.e., U (34.7) = 0.73], a weekly vaccination

Saad-Roy et al., Science 370, 811–818 (2020) 13 November 2020 8 of 8

RESEARCH

PLANT SCIENCE rent with or soon after the formation of a
prebranch site.
Cell wall remodeling and vesicle trafficking mediate
the root clock in Arabidopsis RNA-sequencing screen identifies
oscillating genes
Guy Wachsman1,2, Jingyuan Zhang2, Miguel A. Moreno-Risueno3,
Charles T. Anderson4, Philip N. Benfey1,2* To identify genes with elevated expression in
the oscillation zone and/or in the most re-
In Arabidopsis thaliana, lateral roots initiate in a process preceded by periodic gene expression cently formed prebranch site, we used RNA-
known as the root clock. We identified the vesicle-trafficking regulator GNOM and its suppressor, sequencing (RNA-seq) analysis of sections from
ADENOSINE PHOSPHATE RIBOSYLATION FACTOR GTPase ACTIVATION PROTEIN DOMAIN3, as root individual plants using the modified Smart-
clock regulators. GNOM is required for the proper distribution of pectin, a mediator of intercellular seq2 method (11). In a spatial experiment, we
adhesion, whereas the pectin esterification state is essential for a functional root clock. In sites performed RNA-seq on transverse sections
of lateral root primordia emergence, both esterified and de-esterified pectin variants are from the oscillation zone, the first prebranch
differentially distributed. Using a reverse-genetics approach, we show that genes controlling pectin site, and their flanking regions (Fig. 1A). In a
esterification regulate the root clock and lateral root initiation. These results indicate that the complementary, temporal experiment, we per-
balance between esterified and de-esterified pectin states is essential for proper root clock function and the formed RNA-seq on DR5::LUC-expressing sec-
subsequent initiation of lateral root primordia. tions before, during, and after peak expression
(Fig. 1B and supplementary materials). To
B oth plants and animals use biological with LR phenotypes. These data suggest that validate the quality of the data, we compared
clocks to estimate time and space during the lack of LRs in the gnom mutant might RNA-seq reads of the luciferase transgene
development. In vertebrates, the forma- depend on other factors that are not directly (driven by the DR5 promoter) with luciferase
tion of presumptive vertebrae, known involved in auxin signaling. Vesicle trafficking imaging and found a high correlation (Fig. 1, C
as somitogenesis, is one of the best char- is necessary for the secretion of cell wall com- and D). Multidimensional scaling analysis and
acterized mechanisms involving periodic gene ponents including pectin, hemicelluloses, and other statistical analyses provided further evi-
expression. In plants, before the initiation of cell wall–modifying enzymes (10). Here, we show dence for the fidelity of the data (Fig. 1E and
lateral root primordia (LRP), the auxin re- that disrupting GNOM alters the distribution fig. S2). To identify putative root clock regula-
sponse marker DR5 and other genes exhibit of pectin and blocks the functioning of the root tors, we selected genes that were overrepresented
periodic expression in a region called the os- clock. Furthermore, higher levels of esterified in the oscillation zone-spatial, the oscillation
cillation zone (1, 2). This rhythmic gene ex- and de-esterified pectin are differentially lo- zone-spatial together with the oscillation zone-
pression, known as the root clock, produces calized around emerging LRP, whereas mod- temporal, or the prebranch site-spatial datasets
prebranch sites, DR5-expressing spots that ulation of either form of pectin leads to a loss (Fig. 1, F to I, and table S1, A to D) and used
develop into LRP (1). The root clock represents of prebranch site formation. On the basis of these three lists to perform gene ontology
a flexible oscillator, producing prebranch sites these data, we conclude that GNOM-dependent (GO) analyses (table S2). We identified CLE44
approximately every 6 hours. Most genes im- vesicle trafficking mediates the distribution as one of the genes expressed in incipient
plicated in the initiation of LRP are related to of pectin in the cell wall, which in turn is es- prebranch sites with similar frequency as
auxin signaling (3, 4) or metabolism (5). Re- sential for root clock function and subsequent DR5 (fig. S1, C to E, and movie S1) and used it
cently, it has been shown that auxin-dependent LR formation. as a prebranch site marker together with DR5
changes in endodermis cells are required be- in subsequent experiments. Some of the most
fore the initiation of LRP in underlying pericycle LRP cell divisions initiate soon common GO terms were related to cell wall
cells (6). One candidate protein that putatively after oscillation biogenesis and vesicle trafficking (Fig. 2A).
mediates signaling between cells and is also Of particular interest was a term for pectin-
involved in LR initiation is GNOM (7). The root clock is defined as a 6-hour periodic metabolic process, including genes that code
DR5 expression in a region known as the os- for pectin lyases, pectate lyases, pectin methyl-
GNOM is a trans-Golgi network-localized cillation zone near the border of the elongation esterases (PMEs), and their inhibitors (PMEIs)
(8) ARF-GEF [ADP RIBOSYLATION FACTOR and differentiation zones of the Arabidopsis (P = 0.005; Fig. 2, B and C, and table S2). PMEs
GUANINE NUCLEOTIDE EXCHANGE FACTOR thaliana root (1, 2). After peak DR5-driven de-methyl-esterify galacturonic acid residues
(9)] that regulates vesicle trafficking to the luciferase expression, the region shrinks to of homogalacturonan pectin. Pectin is synthe-
plasma membrane. It had been suggested that form a confined spot defined as a prebranch sized in the Golgi in a highly methyl-esterified
the reduced LR phenotype of the weak gnom site, where LRP emerge subsequently. All LRP form and delivered to the apoplast, where it
allele, fwr, is mediated by one of the PIN- emerge from prebranch sites, and most pre- can be de-methyl-esterified. Blockwise de-
FORMED (PIN) proteins. However, the fwr branch sites develop into LRP (1). To better esterified homogalacturonan chains can be
mutant has properly localized PIN1 (7). In understand the relationship between LR crossed-linked by calcium, leading to stiffen-
addition, there are no known PIN mutants initiation and prebranch site formation, we ing of the cell wall and/or promoting adhesion
examined roots expressing both the DR5:: between cells, whereas randomly de-esterified
1Howard Hughes Medical Institute, Duke University, Durham, LUC reporter, which marks prebranch sites, homogalacturonan can be degraded by pecti-
NC 27708, USA. 2Department of Biology, Duke University, and pCLE44::GFP, which marks early LRP nases, resulting in wall loosening and/or cell
Durham, NC 27708, USA. 3Centro de Biotecnología y (fig. S1, A and B; see below for details on the separation (12, 13). All nine oscillation zone-
Genómica de Plantas (Universidad Politécnica de Madrid – CLE44 marker). In most cases (21/26), the spatial PMEs belong to group 1 (14) within
Instituto Nacional de Investigación y Tecnología Agraria y most recent DR5::LUC-marked prebranch site the type I class of PMEs, which contain a pro
Alimentaria), 28223 Pozuelo de Alarcón (Madrid), Spain. was associated with an early-stage LRP. We domain at the N terminus that requires pre-
4Department of Biology, The Pennsylvania State University, conclude that LRP initiation starts concur- processing in the Golgi before secretion to
University Park, PA, 16802, USA. the apoplast (15). Another group of genes over-
*Corresponding author. Email: [email protected] represented in the oscillation zone-spatial
dataset is involved in auxin transport and

Wachsman et al., Science 370, 819–823 (2020) 13 November 2020 1 of 5

RESEARCH | RESEARCH ARTICLE

Fig. 1. Experimental design and analysis of A ss B RNA-seq
RNA-seq experiments. (A) Spatial experiment. Seven pbs
DR5::LUC roots were imaged using the Lumazone ss bp p ap
system (left). Once peak expression was detected in pbs s
the oscillation zone, the root was dissected into five s p RNA-seq
sections: r, a section proximal to the root meristem and
distal to the oscillation zone; p, the oscillation zone; p Space r Time
s, a section proximal to the oscillation zone and distal r
to the youngest prebranch site; pbs, the youngest
prebranch site; and ss, a section proximal to the C D E p5 p2 Spatial AND Temporal-385 genes
prebranch site. Each section was used to generate a
transcriptome profile (35 total). Arrowhead indicates 9 10.0 100 p4 p7 p1 s5 F G
the root tip. Scale bar, 0.3 mm. (B) Same as in (A) 6 7.5 p3 Spatial-1785 genes Temporal-3188 genes
but a root section was collected from the oscillation 3 5.0 0 r6 r2 p6
zone before peak expression (bp), during peak 2.5 r5 r1 6000 12000
expression (p), or after peak expression (ap). (C and CPM s4
D) CPM reads for luciferase RNA in the spatial CPM r7 4000CPM 9000
experiment (C) and the temporal experiment (D). dimension−2 r4 r3 s2 2000 CPM 6000
(E) Multidimensional scaling plot of the spatial experi- 3000
ment. Note that the regions (r, p, s, pbs, and ss) mostly −100 s3 s7
form distinctive clades. (F) Expression pattern of all 1785 −200 pbs5 s6
genes in the spatial experiment with a log2FC > 0.5
(FC > 1.414) in the p region compared with each of the s1
other zones. (G) Expression pattern of all 3188 genes in
the temporal experiment with a log2FC > 0.5 (FC > pbs2 pbs1
1.414) in the p region (peak luciferase expression)
compared with the bp region (before peak luciferase pbsssp4s5bsss63sp3pbsbsssssss77216 00
expression). (H) Expression pattern of all 244 genes in ss4
the spatial experiment with a log2FC > 0.5 (FC > 1.414) 0 0.0 r p s pbs ss bp p ap
in the pbs zone compared with each of the other zones. time point
(I) Expression profile of CLE44, which was later used r p s pbs ss bp p ap −100 0 100 region
as a prebranch site marker. CPM, counts per million.
region time point dimension−1

H I

Spatial 1st PBS-244 genes 5

2000CPM 4
1500
1000 CPM 3

500 p s pbs ss 2
0 region
1
r
0
r p s pbs ss

region

Fig. 2. Cell wall-, vesicle trafficking-, A B
and membrane-related genes are cell wall organization or biogenesis
overrepresented in the oscillation zone cell wall organization AT5G04960 RHS12−AT3G10710 AT1G02810
and prebranch site and required for its plant−type cell wall organization or biogenesis
formation. (A) GO term analysis of the BP CC MF 1000 400 80
1785 genes described in Fig. 1F (BP, biological polysaccharide metabolic process 750 300 60
process; CC, cellular component; MF, cell wall biogenesis 500 200 40
molecular function). The plot depicts
vesicle trafficking-, Golgi-, cell wall-, small plant−type cell wall organization 250 100 20
GTPase–mediated signal transduction-, and xyloglucan metabolic process 00
carbohydrate metabolism-related GO terms
with P ≤ 0.01 (hypergeometric test using hemicellulose metabolic process PME2−AT1G53830 ATPMEPCRD−AT2G43050 AT2G47550
the gProfiler R package; see table S2 for cell wall polysaccharide metabolic process
the complete list). (B and C) Expression cellular polysaccharide metabolic process CPM 500 150
pattern of nine PME (B) and four PMEI 400
(C) genes in the spatial experiment. glucan metabolic process 300 150
(D to F) pCLE44::LUC expression in WT cellular glucan metabolic process
[WS-4, (D)], PME5-overexpressing (E), cell wall macromolecule metabolic process 100
and PMEI3-overexpressing (F) seedlings polysaccharide biosynthetic process 100
after EtOH induction. Scale bar, 0.2 cm.
(G) Quantification of prebranch site number cellulose metabolic process 200 50 50
[(D) to (F)]. P < 10−14, Wilcoxon rank glycosaminoglycan biosynthetic process 100
sum test. (H) LR number of pme2;pme3
compared with Col-0 WT. P = 0.007, glycosaminoglycan metabolic process 000
Wilcoxon rank sum test. beta−glucan metabolic process
exocytosis AT3G43270 AT5G04970 ATPME3−AT3G14310
pectin metabolic process
galacturonan metabolic process 60 100
exocyst localization
secretion by cell 40 75 600

small GTPase mediated signal transduction 50 400
Golgi apparatus
cell wall 20 25
0 200
plant−type cell wall r p s pbs ss
Golgi subcompartment 0

Golgi membrane r p s pbs ss r p s pbs ss
exocyst
vesicle 10 20 30 40 50 region

Golgi apparatus part −log(p−value)
Golgi cisterna

Golgi cisterna membrane
secondary cell wall
Golgi stack
extracellular vesicle
intracellular vesicle

glucosyltransferase activity
xyloglucan:xyloglucosyl transferase activity
cellulose synthase (UDP−forming) activity

cellulose synthase activity
guanyl ribonucleotide binding

guanyl nucleotide binding
UDP−glucosyltransferase activity

0

C AT3G47670 AT1G23205 D E F G H

160 12.5

120 40

10.0 12

80 20
40
PBSs/cm
30
20 LRs
CPM 10 AT1G62770 0 AT4G25250 7.5
250
0 p s pbs ss 9
r 200
5.0
150
2.5 6

100

50 0.0
WS−4 PME5 PMEI3
0 Col−0 pme2;pme3

p s pbs ss r

region

Wachsman et al., Science 370, 819–823 (2020) 13 November 2020 2 of 5

RESEARCH | RESEARCH ARTICLE

signal transduction. Mutations in five of these A WT rhd3300 rhd3194 gnom184 seca1300 C
genes (LAX3, AUX1, ARF19, IAA14, and IAA28)
directly affect LR formation [fig. S3A and 15 8.1e−16
(3, 16–18)]. A third group overrepresented in p < 2.22e−16 2.9e−12
the oscillation zone-spatial dataset is involved p < 2.22e−16
in the transport, metabolism, and response to PBSs/cm p < 2.22e−16
nitrogen-related compounds (fig. S3B), all pro-
cesses that have been shown to affect LR den- 10
sity and outgrowth (19, 20). For example, the
tonoplast aquaporin TONOPLAST INTEGRAL dmso
PROTEIN2;1/DELTA-TIP1 (TIP2;1), which is bfa
involved in NH3 transport to the vacuole and
response to NH3 accumulation, is redundantly gr2633 5
required, together with two other TIP pro- shr474
teins, for LR initiation (21, 22). We also iden-
tified two additional genes that were not 0
included in the dataset because they did not
fulfill the log-fold change (logFC) > 0.5 criteria WgTnom184 agd3ga26ngodm3S1A8L4K 26 WTagd3a26gd3SALK
but still had high expression in the oscillation
zone: AMT2;1, an ammonium transporter re- agd3
quired for ammonium-dependent LR branch-
ing (20), and TGA4, a basic leucine zipper B DMSO BFA
transcription factor controlling nitrate-dependent
LR initiation (19). Another group of genes with gnom184;
high expression in the oscillation zone belong gnom184 agd326 agd3SALK agd326
to the cysteine-rich receptor-like protein kinase WT WT agd326 agd3SALK
(crRLK) subfamily (fig. S3C). FERONIA, one of
the members of this family, was recently shown Fig. 3. Perturbation of vesicle trafficking affects the root clock. (A) Root clock phenotype of six ethyl
to bind pectin and to regulate pavement cell
morphogenesis (23). Finally, our approach also methanesulfonate mutants. Line name is shown as a superscript and the gene name is in lowercase letters.
identified 10 genes in a GO term for LR
morphogenesis (fig. S3D), which involve auxin Arrowhead indicates DR5::LUC expression in the root tip. (B) Expression of pCLE44::LUC in different
transport (LAX3 and AUX1), auxin signaling
(ARF11/ARF19, IAA14 and IAA28), exocytosis genotypes under mock treatment [dimethyl sulfoxide (DMSO), left five panels] or BFA treatment (right three
(EXO70A1), and the promotion of protein de- panels). (C) Quantification of pCLE44::LUC prebranch site number in gnom184 and its suppressor, agd3 (SALK
gradation by ubiquitin ligase activity (XBAT32). and the EMS mutant agd326), with or without 3 mM BFA (t test). Scale bar, 0.2 cm.
Together, these results suggest that the oscillation
zone functions as a transcriptional hub for mul- for alterations in root clock function based [fig. S5J and (26)]. To better understand the
tiple pathways required for emergence of LRP. gnom184 phenotype, we grew wild-type (WT)
on imaging. We identified six mutations with seedlings on brefeldin A (BFA), a vesicle-
PMEs and PMEIs are required for prebranch
site and LR formation modified root clock function (table S3A) and trafficking inhibitor that targets a subset of
mapped the affected genes [table S3 and (25)]. plant ARF-GEFs, including GNOM (27). A single
Because PME5 and PMEI3 have been impli- Two independent mutations were found in amino acid substitution in the catalytic Sec7
cated in phyllotaxis (24), we decided to explore ROOT HAIR DEFECTIVE 3 (RHD3) and one
the role of pectin esterification in LR branch- each in GNOM, SHORTROOT, GLUTHATIONE domain of GNOM can convert WT seedlings
ing. To determine whether changing the pec- REDUCTASE, and SECA1 (Fig. 3A and table to BFA-resistant seedlings (27), which produce
tin esterification state also affects rhyzotaxis, S3). Because half of the mutations affect genes LRs despite exposure to BFA, suggesting that
we inducibly overexpressed PME5 and PMEI3
and found a severe reduction in prebranch site coding for vesicle-trafficking-related proteins BFA specifically targets GNOM. Seedlings grown
number in both cases (Fig. 2, D to G, and fig. (RHD3 and GNOM) and the gnom mutant on low BFA concentrations (3 to 5 mM) have
S4, A to D), as well as a decrease in LR number allele (gnom184) did not appear to be highly a similar phenotype to gnom184 with no LRs
(fig. S4, E to G). On the basis of our RNA-seq pleiotropic, we focused our analysis on this [fig. S5I and (27)]. We next tested the effect
results, we identified two PME genes, the mu- of BFA on the root clock using the DR5::LUC
tants of which, pme2 and pme3, as well as the mutant (Fig. 3 and fig. S5, A to I). This allele
double mutant combination, showed reduced and the pCLE44::LUC lines. The expression
numbers of LRs (Fig. 2H and fig. S4, H to K). could be partially complemented by its native
These results suggest that the balance between promoter and 3′ untranslated region or with pattern of these markers was strongly altered,
pectin esterification and de-esterification under- the G1090-inducible overexpression promoter phenocopying the gnom184 allele (Fig. 3, B and
pins the functionality of the root clock.

Vesicle-trafficking genes mediate root clock
and LR formation

To complement our reverse-genetics approach,
we mutagenized DR5::LUC seeds and screened

Wachsman et al., Science 370, 819–823 (2020) 13 November 2020 3 of 5

RESEARCH | RESEARCH ARTICLE

Fig. 4. Low-esterified homogalacturonan distribution is affected in gnom184
and negatively correlates with sites of LR inception. (A) Immunohistochemical
quantification of cell wall area per root cross-section area with the antibody against
de-esterified and Ca2+ cross-linked homogalacturonan 2F4 (P2F4 = 0.31, Wilcoxon
rank sum test), de-esterified homogalacturonan LM19 (PLM19 = 0.0028, t test), and
esterified homogalacturonan LM20 (PLM20 = 0.042, t test). False discovery rate
method adjustment was applied. (B and C) Transmission electron microscopy
(TEM) images of LM19 immunogold localization. De-esterified homogalacturonan is
marked by round black gold particles. Cell types (end, endodermis; per, pericycle)
are indicated in the cytoplasmic region of each cell. Insets show the three-way
dense (B) and particle-free (C) junctions. Note the collapsed and antibody-free
endodermis-endodermis cell wall marked by the arrowhead. (D) Quantification of
gold particles in pericycle-endodermis junctions; genotypes show different particle
densities (P = 0.001, Wilcoxon test). (E and F) Immunohistochemistry with LM19 (E)
and LM20 (F) antibodies of root cross-sections in regions of emerging LRP. White
arrowheads indicate the side facing LRP; blue arrowheads indicate the opposite
(control) side. Scale bar, 20 mm. (G and H) Quantification and statistics of (E)
and (F), respectively. Wilcoxon paired test statistics are given above boxes.

C, and fig. S5, G and H). To further identify ARF-GTP ACTIVATING PROTEIN DOMAIN vesicle-trafficking GNOM-AGD3 circuit is
root-clock-regulated genes that function in (AGD3), is required for scission during vesicle required for LRP initiation and root clock
the GNOM pathway, we used the mutagenized formation (29). GNOM is an ARF-GEF that function.
DR5::LUC population and screened for seed- facilitates the exchange of GDP to GTP, lead-
lings that could develop LRs under BFA treat- GNOM regulates homogalacturonan distribution
ment. We identified two BFA-insensitive mutants ing to membrane recruitment of the ARF
(table S3) displaying both LRs and prebranch Because pectic homogalacturonan is synthe-
sites upon BFA treatment. The first affected protein. AGD3 acts in the opposite direction sized in its esterified form in the Golgi (covalently
gene, LEUCINE CARBOXYLMETHYLTRANS- linked to a methyl group through an ester
FERASE1/SUPPRESSOR OF BRI1 (LCMT/SBI1), by activating ARF-dependent GTP hydrolysis, bond) and then secreted to the cell wall,
is involved in terminating the brassinosteroid where it can be de-esterified by PMEs, we
signaling cascade during endocytosis by pro- allowing vesicle budding and eventually sought to determine whether its involvement
moting degradation of the hormone receptor membrane dissociation of ARF-GDP (30). in the root clock and LRP initiation requires
BRI1 [fig. S5, K to M, and (28)]. The second gene, When introduced into the gnom184 back- GNOM-dependent vesicle trafficking. We used
ground, our mutant agd3 allele (agd326) partially immunohistochemistry to determine whether
rescued the loss of root clock function and

the LR phenotype (Fig. 3, B and C, and fig.

S5, G to I). These results indicate that the

Wachsman et al., Science 370, 819–823 (2020) 13 November 2020 4 of 5

RESEARCH | RESEARCH ARTICLE

gnom184 has altered distribution of three at prebranch sites, providing a competence to 17. B. De Rybel et al., Curr. Biol. 20, 1697–1706 (2010).
homogalacturonan forms (31): de-esterified form LRP. 18. K. Swarup et al., Nat. Cell Biol. 10, 946–954 (2008).
and Ca2+ cross-linked [2F4 antibody; (32)], 19. J. M. Alvarez et al., Plant J. 80, 1–13 (2014).
de-esterified (LM19 antibody), and esterified Discussion 20. J. E. Lima, S. Kojima, H. Takahashi, N. von Wirén, Plant Cell 22,
[LM20 antibody; (33)]. The mean intensity
signal of homogalacturonan in all three ex- A minimum requirement for an oscillating 3621–3633 (2010).
periments did not differ between WT and system is delayed negative feedback, which 21. D. Loqué, U. Ludewig, L. Yuan, N. von Wirén, Plant Physiol. 137,
gnom184 (fig. S6). Using xyloglucanase to digest can produce a periodic pattern with as few as
hemicellulose did not unmask additional pectin two components (39). However, the somito- 671–680 (2005).
sites or change the intensity of the antipectin genesis clock requires the function of multiple 22. H. Reinhardt et al., Plant Physiol. 170, 1640–1654
antibody staining (fig. S6), as previously shown components arranged in three complex net-
for mannan polysaccharide (34). However, the works (40). In Arabidopsis, more than 30 genes (2016).
cell wall area occupied by the low-esterified have been implicated in LR development, 23. W. Lin, W. Tang, C. T. Anderson, Z. Yang, FERONIA’s sensing
form was significantly lower in gnom184 than most of which are related to auxin signaling,
in WT (P = 0.0028; Fig. 4A). A severe gnom transport, and homeostasis (41). A recent of cell wall pectin activates ROP GTPase signaling in
allele has been shown to alter the distribution study showed that endodermis cells overlying Arabidopsis. bioRxiv 269647 [Preprint]. 22 February 2018.
of de-esterified homogalacturonan in the the pericycle undergo volume reduction, which https://doi.org.10.1101/269647.
cell wall (35). Because the localization of de- is required for the first cell division during LRP 24. A. Peaucelle et al., Curr. Biol. 18, 1943–1948 (2008).
esterified pectin is altered in gnom184 and initiation (6). Although the precise mechanism 25. G. Wachsman, J. L. Modliszewski, M. Valdes, P. N. Benfey,
the boundary between the pericycle and the by which pectin modification modulates the Plant Physiol. 174, 1307–1313 (2017).
endodermis is relevant for the emergence root clock and LRP initiation is still not known, 26. R. Siligato et al., Plant Physiol. 170, 627–641 (2016).
of LRP, we focused our comparison of de- we show here that LRP initiation involves 27. N. Geldner et al., Cell 112, 219–230 (2003).
esterified homogalacturonan distribution on differential distribution of pectin species, with 28. G. Wu et al., Sci. Signal. 4, ra29 (2011).
pericycle-endodermis junctions. Using an anti- reduced de-esterified pectin levels in cells 29. K. Koizumi et al., Development 132, 1699–1711 (2005).
body that preferentially binds low-esterified overlying primordia. This distribution could 30. J. G. Donaldson, C. L. Jackson, Nat. Rev. Mol. Cell Biol. 12,
homogalacturonan (36), immunogold labeling allow the physical emergence of LRP while 362–375 (2011).
(37, 38) followed by transmission electron preventing emergence on the opposite side of 31. Y. Rui et al., Plant Cell 29, 2413–2432 (2017).
microscopy identified two types of gold particle the root. Additional support for the hypothesis 32. F. Liners, J.-J. Letesson, C. Didembourg, P. Van Cutsem, Plant
distribution in three-way pericycle-endodermis that a balanced esterification level is essential Physiol. 91, 1419–1424 (1989).
junctions (junctions of two pericycle cells and for the root clock and LRP initiation is demon- 33. Y. Verhertbruggen, S. E. Marcus, A. Haeger, J. J. Ordaz-Ortiz,
one endodermis cell or one pericycle cell and strated by overexpression of either PME5 or J. P. Knox, Carbohydr. Res. 344, 1858–1862 (2009).
two endodermis cells). gnom184 junctions were PMEI3, which confers similar phenotypes in 34. Y. Verhertbruggen, J. L. Walker, F. Guillon, H. V. Scheller, Front.
more particle dense compared with WT junc- the root (this study) and in pavement cells Plant Sci. 8, 1505 (2017).
tions (Fig. 4, B to D, and fig. S7, B to E), which (42). The fact that two proteins (GNOM and 35. D. E. Shevell, T. Kunkel, N. H. Chua, Plant Cell 12, 2047–2060
may prevent LR emergence. A more qualita- AGD3) with opposing functionalities also have (2000).
tive analysis showed a similar trend (fig. S7A). an opposite phenotype provides additional evi- 36. S. E. Marcus et al., Plant J. 64, 191–203 (2010).
This indicates that GNOM mediates homoga- dence for the importance of balanced pectin 37. K. Ruel, Y. Nishiyama, J.-P. Joseleau, Plant Sci. 193-194, 48–61
lacturonan distribution in pericycle-endodermis components. One of the remaining questions (2012).
junctions. is how the status of extracellular pectin is con- 38. J. K. Polko et al., Curr. Biol. 28, 3174–3182.e6 (2018).
veyed to the cell to initiate LR formation. One 39. A. Gierer, H. Meinhardt, Kybernetik 12, 30–39 (1972).
Homogalacturonan subtypes are differentially possibility is that biomechanical changes in 40. A. Goldbeter, O. Pourquié, J. Theor. Biol. 252, 574–585 (2008).
distributed at sites of LRP initiation the cell wall are translated into signal transduc- 41. B. Péret et al., Trends Plant Sci. 14, 399–408 (2009).
tion that leads to LRP initiation. Alternatively, 42. K. T. Haas, R. Wightman, E. M. Meyerowitz, A. Peaucelle,
We next investigated whether we could iden- pectin might bind and activate a membranous Science 367, 1003–1007 (2020).
tify a direct link between homogalacturonan receptor such as the FERONIA crRLK to ini-
distribution and LRP inception. Using immuno- tiate a signaling cascade that eventually re- ACKNOWLEDGMENTS
histochemistry on transverse sections at the sults in the formation of LRP.
site of LRP inception, we found that pericycle- We thank C. Wilson and Y. Rui for technical support and advice;
endodermis and endodermis-cortex circum- REFERENCES AND NOTES A. Peaucelle for seeds; M. Plue (Duke, SMif), R. Vancini and H. Mekeel
ference boundaries adjacent to the LRP have (Duke Pathology), and K. Pryer for assistance with immunotissue
reduced levels of de-esterified homogalactur- 1. M. A. Moreno-Risueno et al., Science 329, 1306–1311 processing and TEM microscopy; the Duke Center for Genomic and
onan compared with the opposite (control) side (2010). Computational Biology; and I. W. Taylor, R. Shahan, J. Vermeer,
(Fig. 4, E and G). By contrast, we identified and J. K. Polko for critical review of the manuscript. Funding: This
higher levels of high-esterified homogalactur- 2. I. De Smet et al., Development 134, 681–690 (2007). work was supported by grants to P.N.B. from the National Institutes
onan in the endodermis-cortex boundary (but 3. H. Fukaki, S. Tameda, H. Masuda, M. Tasaka, Plant J. 29, of Health (grant nos. R01-GM043778 and R35-GM131725), the
not in the pericycle-endodermis) adjacent to Howard Hughes Medical Institute, and the Gordon and Betty Moore
the LRP compared with the opposite (control) 153–168 (2002). Foundation (grant no. GBMF3405). M.A.M.-R. was funded by
side (Fig. 4, F and H). These results show that 4. J. C. Wilmoth et al., Plant J. 43, 118–130 (2005). Ministerio de Economía y Competitividad of Spain (MINECO) and
de-esterified homogalacturonan, which is usu- 5. W. Xuan et al., Science 351, 384–387 (2016). ERDF (grant no. BFU2016-80315-P). Author contributions: G.W. and
ally associated with increased cell wall stiffness, 6. J. E. M. Vermeer et al., Science 343, 178–183 (2014). P.N.B. conceived the project. G.W. and J.Z. performed the
is reduced near emerging LRP, and that pectin 7. K. Okumura et al., Plant Cell Physiol. 54, 406–417 (2013). experiments. M.A.M.-R. assisted with mapping. C.T.A. provided
subtypes are differentially distributed at sites 8. S. Naramoto et al., Plant Cell 26, 3062–3076 (2014). supervision and resources for immunohistochemistry. G.W. wrote
of initiating LRP. Given that LRP emerge by 9. D. E. Shevell et al., Cell 77, 1051–1062 (1994). an initial draft of the manuscript. P.N.B., J.Z., M.A.M.-R., and C.T.A.
pushing through overlying cell layers, loss of 10. S.-J. Kim, F. Brandizzi, Glycobiology 26, 940–949 (2016). reviewed and edited the manuscript. Competing interests:
de-esterified pectin could reduce cell adhesion 11. S. Picelli et al., Nat. Protoc. 9, 171–181 (2014). The authors declare no competing interests. P.N.B. is the
12. F. Micheli, Trends Plant Sci. 6, 414–419 (2001). cofounder and chair of the scientific advisory board of Hi Fidelity
13. W. G. Willats et al., J. Biol. Chem. 276, 19404–19413 Genetics, Inc., a company that works on crop root growth. Data
and materials availability: All materials are available upon
(2001). request. All data are available in the main paper or the supplementary
14. R. Louvet et al., Planta 224, 782–791 (2006). materials. The RNA-seq and DNA-seq datasets are available in the
15. S. Wolf, T. Rausch, S. Greiner, Plant J. 58, 361–375 Sequence Read Archive repository with the accession numbers
PRJNA478564 and PRJNA490692, respectively.
(2009).
16. K. Tatematsu et al., Plant Cell 16, 379–393 (2004). SUPPLEMENTARY MATERIALS

science.sciencemag.org/content/370/6518/819/suppl/DC1
Materials and Methods
Figs. S1 to S7
Table S1 to S5
References
Movie S1
Datasets S1 to S3
Code S1 to S9
MDAR Reproducibility Checklist

View/request a protocol for this paper from Bio-protocol.

15 March 2020; accepted 1 October 2020
10.1126/science.abb7250

Wachsman et al., Science 370, 819–823 (2020) 13 November 2020 5 of 5

RESEARCH

MARTIAN ATMOSPHERE samples neutral and ionic species in situ as
MAVEN flies through the martian upper atmo-
Hydrogen escape from Mars is driven by seasonal sphere. We search for the chemical intermedi-
and dust storm transport of water ates that lie between H2O delivered from the
lower atmosphere and H escape from the top of
Shane W. Stone1*, Roger V. Yelle1, Mehdi Benna2,3, Daniel Y. Lo1, Meredith K. Elrod2,4, Paul R. Mahaffy5 the atmosphere; for example, H2O+ and H3O+.

Mars has lost most of its once-abundant water to space, leaving the planet cold and dry. In standard In situ MAVEN measurements
models, molecular hydrogen produced from water in the lower atmosphere diffuses into the upper
atmosphere where it is dissociated, producing atomic hydrogen, which is lost. Using observations The position of MAVEN’s periapsis (i.e., closest
from the Neutral Gas and Ion Mass Spectrometer on the Mars Atmosphere and Volatile Evolution approach to Mars) varied over the course of
spacecraft, we demonstrate that water is instead transported directly to the upper atmosphere, then the mission, as shown in Fig. 1. Regional dust
dissociated by ions to produce atomic hydrogen. The water abundance in the upper atmosphere storms occurred during Mars years (MYs) 32
varied seasonally, peaking in southern summer, and surged during dust storms, including the 2018 and 33, and a global dust storm occurred in
global dust storm. We calculate that this transport of water dominates the present-day loss of atomic MY 34, also shown in Fig. 1. MAVEN instru-
hydrogen to space and influenced the evolution of Mars’ climate. ments directly sample the atmosphere during
the periapse segment of its 4.5-hour orbit, in
T he enrichment of D/H in the martian at- of H in the exosphere, the outermost region which periapsis occurs at roughly 150-km
mosphere and surface, relative to Earth, of the atmosphere, indicating that H escape altitude (19, 20). During the observations of
indicates that most of Mars’ initial res- varies seasonally and during dust storms (8–10). the last regional dust storm of MY 32, MAVEN’s
ervoir of H2O has been lost to space (1–4). A possible cause for rapid variation in the exo- periapsis precessed from 18°N to 8°N latitude,
The standard scenario to explain this with measurements at the onset of the dust
escape is that H2O is confined to low altitudes spheric H abundance and escape rate is direct storm obtained near 13°N. This regional storm
below a hygropause (a cold layer at 40- to 50-km transport of H2O—the most abundant hydrogen- arose in the southern hemisphere near martian
altitude at which H2O condenses), where photo- bearing species in the martian atmosphere (2)— solar longitude [Mars-Sun angle (LS)] 313°,
dissociation and HOx chemistry produce H2. into the middle and upper atmosphere, where well after perihelion (LS=251°) and southern
The H2 then diffuses through the hygropause summer solstice (LS = 270°). Amid the global
into the upper atmosphere where it is destroyed, it could be destroyed by reactions with ions to dust storm of MY 34, MAVEN’s periapsis pre-
producing H, which escapes to space (2, 5–7). produce H (11). This transport would require a cessed from 21°S to 17°S latitude, with measure-
This model would produce steady H2 diffusion weakening of the hygropause, which could be ments at the onset of the dust storm obtained at
and H escape. Observations have shown rapid warmed and raised in altitude as a result of ~18°S. This global dust storm occurred around
LS = 189°, before perihelion and just after
order-of-magnitude variations in the abundance increased solar irradiation around perihelion southern spring equinox (LS = 180°). Figure 1A
and/or heating caused by dust in the atmo- shows the total column dust optical depth from
1Lunar and Planetary Laboratory, University of Arizona, a combination of instruments: the Thermal
Tucson, AZ 85711, USA. 2Planetary Environments Laboratory, sphere. The middle-atmospheric (~15- to 90-km Emission Spectrometer on Mars Global Sur-
NASA Goddard Space Flight Center, Greenbelt, MD 20771, altitude) H2O abundance is often higher than veyor, the Thermal Emission Imaging System
USA. 3Center for Research and Exploration in Space Science that implied by the equilibrium vapor pressure on Mars Odyssey, and the Mars Climate Sounder
and Technology, University of Maryland Baltimore County, at the hygropause temperature (12, 13) and (MCS) on Mars Reconnaissance Orbiter (21, 22).
Baltimore, MD 21250, USA. 4Center for Research and varies seasonally (14, 15), and transient events Figure 1B shows the dayside temperature in
Exploration in Space Science and Technology, University of substantially alter the vertical distribution of the middle atmosphere at the 50-Pa pres-
Maryland College Park, College Park, MD 20742, USA. H2O (14, 16, 17). sure level measured by MCS (23, 24). These
5Solar System Exploration Division, NASA Goddard Space quantities indicate the onset, evolution,
Flight Center, Greenbelt, MD 20771, USA. We test this hypothesis using data from the and distribution in latitude of the dust events
*Corresponding author. Email: [email protected] and the atmospheric response.
Neutral Gas and Ion Mass Spectrometer (NGIMS)
(18) on the Mars Atmosphere and Volatile
EvolutioN (MAVEN) spacecraft (19). NGIMS

Table 1. H2O budget and H production. Two models with low and high H2O are listed, representing periods between and during dust storms, respectively

(32). The H2O mixing ratios at 80 km for the low- and high-H2O cases are 2 and 430 ppm. Total H production from H2O is twice the difference between
all H2O destruction and all H2O production. Total H production from H2 is twice the rate of destruction of H2 by CO2+.

Reaction Column rate (cm−2 s−1)

Low H2O High H2O

H O production..........................................................................................................................................................2..................................................................................................................................................................................
+−
78
H O + e → H O + H 0.1 × 10 1.99 × 10....3.............................2...........................................................................................................................................................................................................................................................................................................

H O destruction.........................................................................................................................................................2...................................................................................................................................................................................
++
78
CO + H O → H O + CO 1.6 × 10 6.35 × 10.......2.............2.................2....................2...................................................................................................................................................................................................................................................................................
++
78
HCO + H O → H O + CO 0.5 × 10 7.56 × 10.....................2.................3......................................................................................................................................................................................................................................................................................................
++
78
CO + H O → HCO + OH 0.5 × 10 2.18 × 10.......2.............2.......................2.................................................................................................................................................................................................................................................................................................

H production

............................................................................................................................................................................................................................................................................................................................................

79

Total H production from H O 5.0 × 10 2.85 × 10........................................................2....................................................................................................................................................................................................................................................................................

88

Total H production from H 1.9 × 10 1.9 × 10........................................................2....................................................................................................................................................................................................................................................................................

Stone et al., Science 370, 824–831 (2020) 13 November 2020 1 of 8

RESEARCH | RESEARCH ARTICLE

Figure 2 shows the variation of the abun- First, there is diurnal variation across two or- in the dayside ionosphere varies by more than

dance in the ionosphere (the charged portion ders of magnitude. The solar zenith angle an order of magnitude over a martian year, from
of the atmosphere) of H2O+ and H3O+, nor-
(the angle between the vertical and the Sun) tens of parts per million (ppm) to hundreds of
malized by the abundance of electrons, near
in Fig. 2, A to C, corresponds to the time of ppm, peaking in southern summer near peri-
MAVEN’s periapsis over the course of the mis-
day at which the measurement was made (0° helion and declining during northern summer.
sion. These two species are produced mainly The peak in ratio of H2O+ abundance to electron
by charge transfer between CO2+ and H2O, is noon at the equator, 90° dawn or dusk, and abundance ([H2O+]/[e−]) in southern summer
producing H2O+, and proton transfer from 180° midnight). Second, H2O+ varies with sea- coincides with the peak in observed seasonal
HCO+ to H2O, producing H3O+ (see below). son (LS), with repeatable maxima near LS =
270°, the peak of southern summer, and minima trends in middle-atmospheric H2O abundance
Three forms of variations are visible in Fig. 2A. near LS = 40°. The relative abundance of H2O+ (14, 15), the exospheric H and D abundances

Fig. 1. MAVEN location and atmospheric conditions during martian dust and November 2018. MAVEN’s periapsis positions are indicated with blue
storms. (A) Gridded total column dust optical depth (red to yellow) (21, 22) and crosses. Gaps in MAVEN’s periapsis location correspond to gaps in the NGIMS
(B) mean dayside temperature (red to yellow with black contour at 200 K) at dataset (32). Ticks in the MY axis occur at the beginning of the year (LS = 0°).
the 50-Pa pressure level from MCS measurements (23, 24) between October 2014 Dust storms correspond with heating in the middle atmosphere.

Stone et al., Science 370, 824–831 (2020) 13 November 2020 2 of 8

RESEARCH | RESEARCH ARTICLE

Fig. 2. Variation of H2O ions measured using NGIMS. Mean [H2O+]/[e−] for [O2+]/[e−] [green squares in (B) and (C)] between CO2 densities of 5 × 108
each MAVEN orbit (A) between October 2014 and November 2018, (B) during a to 109 cm−3 (MAVEN’s nominal periapsis) for a single orbit. The value of the

regional dust storm in MY 32, and (C) during the MY 34 global dust storm. solar zenith angle is indicated by the color of the circle, as indicated in the
colorbar. The transition from red to blue for [H2O+]/[e−] corresponds to the shift
Dotted boxes in (A) indicate the regions shown in (B) and (C). Each data point is
the mean [H2O+]/[e−] (circles), [H3O+]/[e−] [blue triangles in (B) and (C)], or from day to night.

(9, 10, 25), and H escape rate (26–28). Third, Fig. night side of Mars shortly after the onset of Large diurnal variations are also observed
2A shows large and rapid increases in the H2O+ the global dust storm, and measurements in in the relative abundance of H2O+ as MAVEN’s
abundance coincident with dust storms near this period are affected by the diurnal variation periapsis repeatedly precessed from the day
LS = 315° in MY 32 and LS = 190° in MY 34. of the ionosphere (see below). Nevertheless, side to the night side, a process which occurs
More detailed views of these events are shown there is an abrupt increase in the relative abun-
dance of these ions, which coincides with the over several months, as seen in Fig. 2A. The
in Fig. 2, B and C. first effects of the dust storm on the upper at- relative abundance of H2O+ increases on the
mosphere on 8 June 2018, as measured by the night side because, relative to other species
The dust storms shown in Fig. 2, B and C, Imaging Ultraviolet Spectrograph on MAVEN in the martian atmosphere, H2O has a low
(29) and neutral atmosphere density measure- ionization potential (IP), 12.6 eV, and OH has
are substantial perturbations to the observed ments from NGIMS (30). The regional dust a high proton affinity (PA), 6.14 eV. In gen-
storm observed in March 2015 (MY 32, Fig. 2B)
seasonal trend. At the onset in the upper at- occurs during the declining seasonal trend, with eral, atmospheric ionization flows by electron
an upward perturbation caused by the storm: transfer from neutral species with lower IPs
mosphere of the global dust storm of June 2018 Over 2 days, [H2O+]/[e−] increases from 32 to
73 ppm, a factor of 2.3, and [H3O+]/[e−] in- to positive ions whose parent neutrals have
(MY 34, Fig. 2C), the relative abundance of creases from 2.4 to 3.9 ppm, a factor of 1.6. higher IPs and, by proton transfer, from spe-
H2O+ in the ionosphere rises sharply in a short
period of time: [H2O+]/[e−] increases from 106 cies whose nonprotonated forms have lower
to 327 ppm, a factor of 3.1, over ~2 days, from PAs to species that have higher PAs (31). On
the night side, away from a constant source
8 June 2018 to 10 June 2018. Simultaneously,
[H3O+]/[e−] increases from 11 to 28 ppm, a factor
of 2.5. MAVEN’s periapsis precessed onto the

Stone et al., Science 370, 824–831 (2020) 13 November 2020 3 of 8

RESEARCH | RESEARCH ARTICLE

Fig. 3. Variation of H2O and H2. (A to C) Same as in Fig. 2 but for the H2O mixing ratio (circles) and H2 mixing ratio [blue triangles in (B) and (C)].

of photoionization driving the production of The results of these calculations are shown in 4.3 ppm at LS = 259° and a minimum of 2.0 ppm
short-lived ions (such as O2+), H2O+ becomes at LS = 86°, for a variation of more than a
more abundant. Thus, we can only observe Fig. 3. The diurnal variation of H2O is much factor of two. The H2O mixing ratio is never
the effects of dust storms on the day side; smaller than that of H2O+, as expected in the negligible, contrary to the assumption of 0
otherwise, the diurnal variation overwhelms neutral atmosphere, which is more chemically to 4 parts per billion made in numerous mod-
the signal because of the storm. eling studies (5, 33, 35).
stable than the ionosphere.
Water in the upper atmosphere During the global dust storm of June 2018,
A repeatable seasonal trend occurs in the
Assuming that the increases in H2O+ and H3O+ the mean H2O mixing ratio increased by a
are due to an influx of H2O into the upper at- H2O abundance in the upper atmosphere, which factor of 2.4, from a mean value of 3.0 to 7.1 ppm
mosphere, we calculate H2O abundances from peaks in southern summer between LS = 250° over 2 days. The water abundance then con-
NGIMS density measurements of H2O+, H3O+, and 270°, shown in Figs. 3A and 4. This trend
HCO+, and e− obtained over 4254 MAVEN or- tinued to increase following the seasonal and
bits. We assume photochemical equilibrium in upper-atmospheric H2O coincides with the dust storm trends, reaching values >60 ppm
for H3O+; i.e., its production rate is equal to seasonal trends seen in previous observations at LS = 204°. This event included the highest
its loss rate. This assumption is known to be persistent H2O abundances we observed, hold-
valid in the thermosphere on the day side of the H2O abundance in the middle atmo- ing at tens of ppm for more than 5 months
but has not been substantiated on the night sphere (14, 15), the exospheric H and D abun- at the end of 2018. During the regional storm
side (32). H3O+ is mainly produced by pro- dances (9, 10, 25), and H escape flux (26–28).
ton transfer to H2O and is only destroyed by The mean upper-atmospheric H2O mixing observed in MY 32, the H2O mixing ratio
dissociative recombination with e− (31–34). ratio is >1 ppm at all times and can exceed increases by 7%, from 4.6 to 4.9 ppm in 2 days.

10 ppm in southern summer. A sinusoidal Thus, dust storms drive perturbations in upper-
atmospheric H2O that are far more rapid than
model fitted to the dayside H2O abundances the diurnal variations, and even small dust
from all MYs, but excluding regional and

global dust events, is shown in Fig. 4. This

model has a maximum H2O mixing ratio of

Stone et al., Science 370, 824–831 (2020) 13 November 2020 4 of 8

RESEARCH | RESEARCH ARTICLE

storms have a measurable impact. The MY 32 59°N), indicating an injection of water into of H in the upper atmosphere. Measured H2
dust storm occurred when MAVEN was near the middle atmosphere coincident with our variations obtained by NGIMS are shown in
the subsolar point and is superimposed on a observations (16). Both of these measurements Fig. 3, B and C. H2 reacts with numerous ions,
declining seasonal trend. We interpret these were obtained at LS = 196.64°. Further TGO mainly CO2+, to produce H atoms that may
variations as being due to enhanced transport measurements showed that H2O abundances escape the planet’s atmosphere. H2 itself can
of H2O into the upper atmosphere during the of ~5 to 10 ppm extended to altitudes up to 90 escape, although much more slowly than H
dust storms. to 100 km (13, 17). The time scales of dust storm– because H2 is twice the mass of H and lighter
induced variations in the upper-atmospheric species have lower escape energies. The abso-
Previous and coincident measurements of water H2O mixing ratio shown in Fig. 3 are con- lute sensitivity of NGIMS to H2 is unknown, so
sistent with these simultaneous measurements we cannot derive an absolute H2 abundance.
The effects of both regional and global dust of the middle-atmospheric H2O mixing ratio However, relative variations of H2 are mea-
storms on the H2O in the lower and middle during the 2018 global storm. MCS observed surable. No seasonal trend or impact of dust
atmosphere have been observed before. In an H2O mixing ratio close to 300 ppm near storms are apparent in the upper-atmospheric
MY 32, Mars Express observed large seasonal 50-km altitude in the middle atmosphere (36). H2 abundance (Fig. 3, B and C). H2 is there-
increases in the H2O mixing ratio, including At the same time, we measure a mean H2O fore not the source of H responsible for sea-
an increase around LS = 315° (14), coinciding mixing ratio of 7.5 ppm near MAVEN’s peri- sonal and dust storm–induced variations in
with the increase we observe in Figs. 2 and apsis (~150-km altitude) over an LS range of the upper atmosphere.
3. These measurements indicate an increase 191.05° to 193.21° and a latitude range of 13°S
in the hygropause altitude and an injection to 18°S. These measurements in the middle Water as a source of escaping hydrogen
of water into the middle atmosphere, coincid- atmosphere, below the NGIMS sampling range,
ing with our observed increases in the abun- provide context for the NGIMS measurements. We constructed a photochemical model to in-
dances of H2O, H2O+, and H3O+ in the upper Seasonally and during dust storms, the H2O vestigate the effects of the H2O abundances in
atmosphere. Similarly, MCS detected an in- abundance in the middle atmosphere increased the Mars ionosphere on the H escape rate. We
crease in the hygropause altitude, rising to substantially in the span of a few martian days, constrain the model with data from a MAVEN
65 to 70 km, in this period (15). During the coincident with NGIMS observations in the Deep Dip (DD) campaign. During weeklong
global dust storm of MY 34, Trace Gas Orbiter upper atmosphere. DD campaigns, the MAVEN spacecraft pene-
(TGO) measured a peak H2O abundance of trates to lower altitudes, ~125 km. The mea-
~100 ppm at 50-km altitude in the southern Molecular hydrogen in the upper atmosphere sured altitude profiles provide constraints on
hemisphere (80°S to 83°S) and a peak H2O ionospheric composition. We use the second
abundance of ~250 ppm just below 40-km Molecular hydrogen, H2, may contribute to DD campaign (DD2) executed around LS = 329°
altitude in the northern hemisphere (51°N to variations in protonated ions and is a source in MY 32, because MAVEN’s periapsis was
then close to the subsolar point. We average
Fig. 4. Seasonal variation of upper-atmospheric H2O. Each data point is the mean dayside (solar zenith the measurements over the DD using a pre-
angle <90°) H2O mixing ratio between CO2 densities of 5 × 108 and 109 cm−3 for a single MAVEN orbit viously described technique to remove wave
over MYs 32 (red circles), 33 (blue triangles), and 34 (green squares). The solid line is a sinusoidal model structure in the altitude profiles (20). The solar
zenith angles and latitudes do not vary appre-
fitted to the data from all three MYs to illustrate the seasonal variation. Dust storm data (dotted boxes) ciably during a DD, so the resulting profiles
are effectively averages over longitude for a
are included in the plot but excluded from the model fitting. latitude of 4°S and local time of 12 p.m.

The photochemical calculations are one di-
mensional (1D) and use the DD2 geometry.
Neutral and ionic composition is calculated
from 80 to 250 km, including photolysis, chem-
ical reactions, and diffusion of neutrals (32).
We include both molecular and eddy diffu-
sion, adopting an eddy diffusion coefficient
of 108 cm2 s−1. Ions are assumed to be in local
photochemical equilibrium. Rate coefficients
are discussed below.

The model is constructed to match the pri-
mary neutral constituents measured by NGIMS,
as shown in Fig. 5A. We calculate H2O by
assuming a mixing ratio of 2 ppm at 80 km,
consistent with our measurements obtained
near aphelion (LS = 71°) in the absence of dust
storms (Fig. 4). Photolysis causes the H2O
mixing ratio to decrease slightly with altitude
until ~160 km, where molecular diffusion be-
gins to dominate and the mixing ratio starts to
increase with height. For O, we assume a mix-
ing ratio of 2% at 80 km, chosen to produce a
vertical profile consistent with NGIMS mea-
surements at higher altitudes.

Figure 5B shows that the altitude profiles of
H2O+ and H3O+ in the model are consistent

Stone et al., Science 370, 824–831 (2020) 13 November 2020 5 of 8

RESEARCH | RESEARCH ARTICLE

Fig. 5. Calculated 1D model abundances and reaction rates. (A) Neutral species abundances from our low-H2O photochemical model (lines)
compared with the NGIMS DD2 mean data (crosses). (B) Same as in (A), but for ionic species. Chemical reaction rates for selected reactions in the
(C) low- and (D) high-H2O models. The legend in (D) also applies to (C).

with the observations, validating our earlier pancies at higher altitudes and near MAVEN’s for the low-H2O model is 5 × 107 cm−2 s−1, at
periapsis, but the observed densities near pe- the low end of the range of H escape rates
inferences. The model also matches measure- riapsis are uncertain because the spacecraft is inferred by previous studies, 107 to 5 × 109 cm−2
ments of CO2+, O2+, and HCO+, species involved moving primarily horizontally (20) and the s−1 (26–28). The column-integrated H2O de-
model densities are uncertain at higher alti- struction in the ionosphere for the high-H2O
in the chemical destruction of H2O. Our calcu- tudes because ion diffusion has been neglected model is 1.4 × 109 cm−2 s−1, corresponding to
lations also match the O+ observations. Models in the model. an H production rate of 2.8 × 109 cm−2 s−1, at

of the Mars ionosphere have often assumed We use the model to calculate the net de- the high end of the range determined in pre-
that O+ was produced by CO2+ + O → O+ + CO2 struction rate of H2O and the production rate vious studies (26–28). H production from de-
with a rate coefficient of 10−10 cm3 s−1 (37); of H in the ionosphere (Table 1). The NGIMS struction of H2 by CO2+ is 9.6 × 107 cm−2 s−1,
measurements are only available above the less than twice as large as H production from
however, laboratory work has placed an upper ionospheric peak, the location of the max-
limit of 6 × 10−13 cm3 s−1 on this rate coefficient imum total ion density (Fig. 5B, 135 km), so we ionospheric destruction of H2O in the low-
(38, 39). Adopting this value, O+ is produced rely on model predictions to investigate chem- H2O models, and substantially less than that
primarily by CO2 + hn → CO + O+ (hn, photon ical processes down to 80 km. In addition to in the high-H2O models.
energy) and lost by O+ + CO2 → CO2+ + O and the results for the DD2 model, we present
O+ + CO2 → O2+ + CO. Previous models as- rates for a similar model calculated for the It is possible that ionospheric destruction of
same conditions and input data, but in which
sumed a combined rate coefficient for these we adopt an H2O mixing ratio of 430 ppm H2O is equally or more important than neutral
reactions of 10−9 cm3 s−1 (40). However, state- at 80 km, consistent with our measurements photolysis of H2O for H escape. We cannot
obtained during the global dust storm of MY compare the upper- and lower-atmosphere
specific measurements indicate a value of 4 × 34 (Fig. 4). H production in the ionosphere
10−10 cm3 s−1 (41). Our model uses a value of 3 × contributions to escape in a more quantita-
10−10 cm3 s−1, chosen to match DD2 observa-
tive way because escape rates in these models
tions. Between 140 and 150 km, the predicted
are closely dependent on the boundary condi-
and observed protonated ion densities agree to
tions on the H and H2 mixing ratios at 80 km,
within 25% (Fig. 5B). There are larger discre-

Stone et al., Science 370, 824–831 (2020) 13 November 2020 6 of 8

RESEARCH | RESEARCH ARTICLE

which are assumed values chosen to replicate through the middle atmosphere. In our in- cluded in our 1D model. The largest peak in the
observations at higher altitudes. Boundary terpretation, H escape depends on how H2O lower-atmospheric H2O mixing ratio occurs in
conditions on mixing ratios can introduce spu- moves past the hygropause and the rate at northern summer (2), the period in which we
rious fluxes into photochemical models. Phys- which it is transported from the middle at- observe a minimum in the upper-atmospheric
ical boundary conditions, such as deposition mosphere to the ionosphere. H2O mixing ratio (Fig. 4). The cause of this
velocities at the surface, are required to accu- discrepancy between the seasonal variations
rately calculate fluxes. Our calculations and The relative importance of neutral and ion in lower- and upper-atmosphere H2O abun-
the high H2O densities in the thermosphere chemistry in the production and escape of dances is unknown and cannot be explained
found in the MAVEN observations show that thermal and nonthermal H remains unclear. by a 1D model.
ionospheric chemistry is a source of H, which Because we observe that the H2 abundance is
can escape from the Mars atmosphere. constant during the dust storms whereas the The seasonal and dust storm–mediated de-
H2O abundance increases strongly, ionospheric livery of water to the upper atmosphere could
Implications for climate evolution destruction drives the elevated H escape rate have played a substantial role in the evolution
during dust events. The importance of ion of the martian climate from its warm and wet
Early studies of H production and escape ne- chemistry to H production has been dis- state billions of years ago to the cold and dry
glected ionospheric destruction of H2O be- cussed previously (34), although without di- planet we observe today. Mars has likely lost
cause H2O was assumed to be confined to rect information on the H2O abundance in enough H2O to cover the planet’s surface with
low altitudes by a hygropause (5, 6, 42). Later, the ionosphere. The H escape flux can be ex- an ocean tens to hundreds of meters deep, and
it was found that the hygropause varies in pressed as loss rates must have been higher in the past
altitude with season, leading to speculation (4). In a martian year with no dust storms, as-
that H2O saturation may not occur at all dur- F = 1.6 × 108 + 1.4 × 1013 f cm−2 s−1 (1) suming the H2O mixing ratio at 80 km is con-
ing dust storms, owing to elevated tempera- stant at 2 ppm, we calculate that 3.0 × 1015 cm−2
tures (43). More recent 1D photochemical where f is the H2O mixing ratio at 80 km H would be produced from H2O and 1.1 × 1016
models (35, 44, 45) and 3D global circulation (34). The constant term of 1.6 × 108 cm−2 s−1 cm−2 H by destruction of H2 by CO2+, assuming
models (46–48) have likewise not included is the limiting flux for an H2 mixing ratio of all H, including that in HCO2+, is eventually
chemical destruction of H2O in the ionosphere. 15 ppm, the value imposed at the lower boun- liberated. During a global dust storm such as
Occultation and limb profile measurements dary in those models (34). Adopting an H2O the one observed in MY 34, in which the H2O
in the middle atmosphere demonstrated that mixing ratio of 2 ppm in Eq. 1 produces a mixing ratio at 80 km reaches 430 ppm, 1.1 ×
the hygropause does not confine H2O to low contribution to escape from ion chemistry of 1016 cm−2 H would be produced from H2O in
altitudes (13, 15–17), and it has been postulated 2.8 × 107 cm−2 s−1, roughly half the column- just 45 days. Thus, a single global dust storm
the elevated exospheric H densities during integrated production rate in Table 1, sug- can produce more than an entire martian
dust storms could be explained by these en- gesting an escape probability of 50% for the year’s worth of H production and escape. If
hancements in the middle-atmospheric H2O H produced in the ionosphere. This result we extrapolate these numbers into the past,
abundance (11). However, this model neglected is strongly dependent on the H mixing ratio we find that over a billion years, a 44-cm-deep
ionospheric chemistry, assuming the H was at 80 km, which is an assumed value. Thus, global layer of H2O would be lost owing to
produced solely by neutral photolysis. We have in the previous model (34) and ours, the rela- destruction of H2O in the ionosphere, and, as-
shown that H2O is present in the ionosphere tive contributions of neutral and ion chemistry suming a rate of about one global dust storm
with substantial abundance throughout the to escape depend on the boundary conditions every 10 years (50), an additional 17 cm of H2O
martian year continuously since the arrival of for H and H2 at 80 km. The abundance of, would be lost owing to global dust storms.
MAVEN and that H is produced from this exact mechanism of production for, and con- Larger loss estimates would be obtained if this
H2O by ion chemistry. tribution to escape of a nonthermal compo- extrapolation included the seasonal fluctua-
nent of the H corona at Mars have not been tion of H2O in the upper atmosphere (Fig. 4)
The time scales for destruction of H2O and determined because of degeneracies in the and regional dust storms, which, relative to
production of H by ionospheric chemistry are models used (27, 28) and cannot be deter- global dust storms, likely have a smaller but
short. Once in the ionosphere, H2O is destroyed mined from our results. non-negligible impact on H escape. These ef-
primarily by reactions with CO2+ with a rate fects, though, are smaller than uncertainties
coefficient of 2.4 × 10−9 cm−3 s−1 (40). Adopting NGIMS data are confined to altitudes above related to our lack of knowledge of the martian
a CO2+ density of 3 × 104 cm−3 (Fig. 5B, 135 km) 125 km, whereas occultation and limb-sounding climate billions of years ago (4). H escape driven
implies a lifetime for H2O of 4 hours. This is experiments have measured H2O up to ~100 km, by direct transport of H2O to the upper atmo-
more than an order of magnitude shorter than leaving a large gap in coverage between 100 sphere would have accelerated the loss of water
the H2O photolysis time constant of 2 × 105 s and 125 km. Figure 5A shows our predicted from Mars as the martian hygropause weakened
(48). Assuming an electron density of 105 cm−3 H2O mole fraction from the middle atmo- owing to a decrease in atmospheric pressure,
(Fig. 5B, 135 km) and electron recombination sphere through the ionosphere, but this is itself driven by atmospheric escape and the
rates of 4.4 × 10−7 cm−3 s−1 (49) for H2O+ and based on a 1D model, and the real atmosphere development of a dry, dusty surface.
H3O+ yields a time constant of ~20 s for pro- is likely to be more complex. The mixing ratios
duction of H from electron recombination. inferred for the middle atmosphere are often REFERENCES AND NOTES
Thus, H2O transported to the ionosphere is far greater than the values we calculate for the
quickly broken apart and converted to H. We ionosphere (13, 16, 17). This contrast between 1. R. D. Wordsworth, Annu. Rev. Earth Planet. Sci. 44, 381–408
expect the H produced at these high altitudes the inhomogeneous H2O distribution derived (2016).
to have a high probability for escape. This im- from the middle-atmosphere measurements
plies that the limiting factor is the supply rate and the roughly constant altitude profile pre- 2. R. M. Haberle, R. T. Clancy, F. Forget, M. D. Smith, R. W. Zurek,
of H2O to the ionosphere from lower alti- dicted by the 1D model (Fig. 5) suggests that the Eds., The Atmosphere and Climate of Mars (Cambridge Univ.
tudes. This has implications for our view of middle-atmosphere measurements are sensing Press, 2017).
H escape from Mars. In previous models, H a local enhancement of H2O that is transported
escape was controlled by production of H2 in horizontally around the planet as it diffuses 3. R. M. Ramirez, R. A. Craddock, Nat. Geosci. 11, 230–237 (2018).
the lower atmosphere and its slow diffusion upward. Such horizontal transport is not in- 4. B. M. Jakosky et al., Icarus 315, 146–157 (2018).
5. D. M. Hunten, M. B. McElroy, J. Geophys. Res. 75, 5989–6001

(1970).
6. T. D. Parkinson, D. M. Hunten, J. Atmos. Sci. 29, 1380–1390

(1972).

Stone et al., Science 370, 824–831 (2020) 13 November 2020 7 of 8

RESEARCH | RESEARCH ARTICLE

7. M. B. McElroy, T. M. Donahue, Science 177, 986–988 (1972). 30. M. K. Elrod, S. W. Bougher, K. Roeten, R. Sharrar, J. Murphy, ACKNOWLEDGMENTS
8. M. S. Chaffin et al., Geophys. Res. Lett. 41, 314–320 (2014). Geophys. Res. Lett. 47, e2019GL084378 (2020).
9. D. Bhattacharyya, J. T. Clarke, J.-L. Bertaux, J.-Y. Chaufray, We acknowledge the contributions of the NASA MAVEN teams
31. J. L. Fox, Icarus 252, 366–392 (2015). responsible for designing, constructing, and operating the
M. Mayyasi, Geophys. Res. Lett. 42, 8678–8685 (2015). 32. Materials and methods are available as supplementary materials. spacecraft and its instruments, without which this work would be
10. J. T. Clarke et al., J. Geophys. Res. Space Phys. 122, 33. J. L. Fox, M. Benna, P. R. Mahaffy, B. M. Jakosky, Geophys. Res. Lett. impossible. Funding: All authors were supported by funding
from the MAVEN mission, part of the NASA Mars Exploration
2336–2344 (2017). 42, 8977–8985 (2015). Program. Author contributions: S.W.S. and R.V.Y. conceived the
11. M. S. Chaffin, J. Deighan, N. M. Schneider, A. I. F. Stewart, 34. V. A. Krasnopolsky, Icarus 321, 62–70 (2019). study and wrote the draft manuscript. S.W.S., R.V.Y., M.B.,
35. V. A. Krasnopolsky, J. Geophys. Res. Planets 107, 5128 (2002). M.K.E., and P.R.M. collected, calibrated, and analyzed the NGIMS
Nat. Geosci. 10, 174–178 (2017). 36. N. G. Heavens, D. M. Kass, J. H. Shirley, J. Geophys. Res. data. R.V.Y. and D.Y.L. constructed the 1D photochemical
12. L. Maltagliati et al., Science 333, 1868–1871 (2011). model and analyzed the output, with the assistance of S.W.S.
13. A. A. Fedorova et al., Science 367, 297–300 (2020). Planets 124, 2863–2892 (2019). All authors contributed to the interpretation of results and
14. A. A. Fedorova et al., Icarus 300, 440–457 (2018). 37. F. C. Fehsenfeld, D. B. Dunkin, E. E. Ferguson, Planet. Space Sci. preparation of the manuscript. Competing interests: We declare
15. N. G. Heavens et al., New Astron. 2, 126–132 (2018). no competing interests. Data and materials availability:
16. A. C. Vandaele et al., Nature 568, 521–525 (2019). 18, 1267–1269 (1970). MAVEN data are available on the NASA Planetary Data System at
17. S. Aoki et al., J. Geophys. Res. Planets 124, 3482–3497 (2019). 38. J. L. Fox, A. S. Johnson, S. G. Ard, N. S. Shuman, A. A. Viggiano, https://atmos.nmsu.edu/data_and_services/atmospheres_data/
18. P. R. Mahaffy et al., Space Sci. Rev. 195, 49–73 (2015). MAVEN/maven_main.html and the MAVEN Science Data Center at
19. B. M. Jakosky et al., Space Sci. Rev. 195, 3–48 (2015). Geophys. Res. Lett. 44, 8099–8106 (2017). https://lasp.colorado.edu/maven/sdc/public/. We used NGIMS
20. S. W. Stone, R. V. Yelle, M. Benna, M. K. Elrod, P. R. Mahaffy, 39. J. E. Tenewitz et al., J. Chem. Phys. 148, 084305 (2018). Level 2, version 8, revision 1 data (32), which can be searched at
40. V. G. Anicich, J. Phys. Chem. Ref. Data 22, 1469–1569 (1993). https://lasp.colorado.edu/maven/sdc/public/pages/search/
J. Geophys. Res. Planets 123, 2842–2867 (2018). 41. C. Alcaraz, C. Nicolas, R. Thissen, J. Zabka, O. Dutuit, search.html. The 1D photochemical model code and output files
21. L. Montabone et al., Icarus 251, 65–95 (2015). are available at Zenodo (51).
22. L. Montabone et al., J. Geophys. Res. Planets 125, J. Phys. Chem. A 108, 9998–10009 (2004).
42. M. B. McElroy, Science 175, 443–445 (1972). SUPPLEMENTARY MATERIALS
e2019JE006111 (2020). 43. B. M. Jakosky, Space Sci. Rev. 41, 131–200 (1985).
44. H. Nair, M. Allen, A. D. Anbar, Y. L. Yung, R. T. Clancy, Icarus science.sciencemag.org/content/370/6518/824/suppl/DC1
23. A. Kleinböhl et al., J. Geophys. Res. Planets 114, E10006 (2009). Materials and Methods
111, 124–150 (1994). Fig. S1
24. D. M. Kass et al., Geophys. Res. Lett. 46, 2019GL083931 45. J. L. Fox, A. B. Hać, Icarus 204, 527–544 (2009). Table S1
(2019). 46. F. González-Galindo et al., J. Geophys. Res. Planets 120, References (52–63)

25. M. Mayyasi et al., J. Geophys. Res. Space Phys. 124, 2152–2164 2020–2035 (2015). 11 December 2019; accepted 11 September 2020
(2019). 47. J.-Y. Chaufray et al., Icarus 245, 282–294 (2015). 10.1126/science.aba5229
48. F. Daerden et al., Icarus 326, 197–224 (2019).
26. J. S. Halekas, J. Geophys. Res. Planets 122, 901–911 (2017). 49. M. J. Jensen et al., Astrophys. J. 543, 764–774 (2000).
27. D. Bhattacharyya et al., J. Geophys. Res. Space Phys. 122, 50. J. H. Shirley, C. E. Newman, M. A. Mischna, M. I. Richardson,

11756–11764 (2017). Icarus 317, 197–208 (2019).
28. M. S. Chaffin et al., J. Geophys. Res. Planets 123, 2192–2210 51. S. W. Stone et al., Hydrogen Escape from Mars is Driven by

(2018). Seasonal and Dust Storm Transport of Water, Version 2,
29. J.-Y. Chaufray et al., Geophys. Res. Lett. 47, e2019GL082889 Zenodo (2020); doi: 10.5281/zenodo.4010623.

(2020).

Stone et al., Science 370, 824–831 (2020) 13 November 2020 8 of 8

RESEARCH

◥ (8). The specimen consisted of irregularly shaped
aggregates or chains connected to each other
REPORT that form continuous networks. By increasing
the magnification, we observed that the aggre-
M E TA L L U R GY gates are made up of several individual grains
of a few nanometers in size. Among grains of
Constrained minimal-interface structures in similar contrast in an aggregate, we detected
polycrystalline copper with extremely fine grains either twin relationships or coincidence site
lattice (CSL) boundaries. For example, as indi-
X. Y. Li1*, Z. H. Jin1,2, X. Zhou1, K. Lu1* cated by the fast Fourier transform (FFT) pat-
terns of an aggregate consisting of five grains
Metals usually exist in the form of polycrystalline solids, which are thermodynamically unstable because (Fig. 1B), we observed two sets of S3 bound-
of the presence of disordered grain boundaries. Grain boundaries tend to be eliminated through aries on ð1 11Þ and ð 1 11Þ planes, respectively,
coarsening when heated or by transforming into metastable amorphous states when the grains are small and a CSL boundary with a misorientation
enough. Through experiments and molecular dynamics simulations, we discovered a different type of of 21° between G4 and G5.
metastable state for extremely fine-grained polycrystalline pure copper. After we reduced grain sizes to a
few nanometers with straining, the grain boundaries in the polycrystals evolved into three-dimensional TEM images obtained in high-resolution
minimal-interface structures constrained by twin boundary networks. This polycrystalline structure mode of the ultrathin foil specimens showed
that underlies what we call a Schwarz crystal is stable against grain coarsening, even when close nanosized crystallites with various orienta-
to the equilibrium melting point. The polycrystalline samples also exhibit a strength in the vicinity of tions in the as-prepared Cu (Fig. 1C). Grains
the theoretical value. were roughly equiaxed with a uniform size
distribution, with most grains <10 nm in size.
T he end-member structures for solids are states are compositionally selective and rarely Selected area electron diffraction covering a
single crystals, where atoms are arranged form for most metallic alloys and pure metals large number of grains showed nearly con-
in ordered crystal lattices throughout the under conventional conditions. A fundamen- tinuous diffraction rings of face-centered cubic
sample, and amorphous solids or glasses, tal question to be answered is whether other (fcc) Cu, an indication of seemingly random
where atomic arrangements are disordered metastable structures may be adopted when orientations of the polycrystallites. The tiny
with short- or medium-range ordering only (1). polycrystalline grains are steadily refined to crystallites are connected with each other by
Metals usually exist in the form of polycrystalline extremely small scales. atomically distinct boundaries. We did not
solids, which are between these two extremes. detect amorphous phases or pores. The full
They are made up of smaller crystallites (called We have recently found that when grains crystallinity of the as-prepared Cu sample
grains) that are separated by a variety of of pure Cu and Ni are refined to a few tens of is in concert with x-ray diffraction patterns,
boundaries where atomic arrangements are nanometers in size through plastic deforma- where sharp diffraction peaks show up with-
normally disordered. The disordered grain tion, the autonomous GB relaxation into low- out a diffused amorphous peak.
boundary (GB) is what makes polycrystalline energy states is triggered with GB dissociations
metals thermodynamically unstable. (6). The corresponding GB energy decreases Our procession electron diffraction mea-
accordingly, which leads to a substantial ele- surements under TEM (Fig. 1D) revealed that
Polycrystalline materials tend to become vation in the thermal and mechanical stabil- a considerable fraction of GBs (~25%) are S3
more stable when GBs are eliminated, until ities of nanograins against coarsening at twin boundaries with a misorientation of 60°
they ultimately become single crystals. An smaller sizes (6, 7). This observation suggests in the sample. Other interfaces separating
example of this process is how the grains in the possibility that nanograined structures may grains have randomly distributed misorien-
polycrystals may coarsen through GB migra- evolve into more-stable states by approaching tations, among which different types of low-S
tion that usually occurs at temperatures below the grain-size extreme. We discovered, through CSL boundaries were detected as well, con-
half of the melting point (2). The coarsening experimental and molecular dynamics (MD) stituting ~20%. This is in agreement with our
temperature drops with decreasing grain size, simulations, a metastable state in polycrystal- other observations (Fig. 1, A and B). As in the
going as far down as room temperature in line pure Cu with grain sizes of a few nano- magnified images (Fig. 1E), some boundaries are
some nanograined metals (3). For fine-grained meters. The state is formed by the evolution of associated with GB structural units as pre-
polycrystals with a high enough GB density, GBs into three-dimensional (3D) minimal- viously observed, underlying the typical struc-
transforming into a metastable amorphous interface structures constrained by twin bound- tures of relaxed GBs. All of these structural
state is an alternative option to stabilization. ary networks. In this type of polycrystal, grain characteristics support the assumption that
This state is anticipated from a thermodynamic coarsening is effectively suppressed, even close the GBs are in relaxed states, which is consistent
point of view, and amorphization has been to the melting point. The strength is also close with the decreased GB energy in the relaxed
observed in a number of metallic alloys [e.g., to the theoretical value of Cu. nanograined Cu below the critical grain size (6).
Ni-P (4) and Ni-W (5)] as grain sizes were
reduced below certain limits (typically a few With a two-step plastic deformation process We characterized individual grains by prop-
nanometers). However, under the thermody- of surface mechanical grinding treatment fol- erly tilting the specimens under high-resolution
namic and kinetic constraints, amorphous lowed by high-pressure torsion in liquid nitro- TEM so as to resolve their lattice images. We
gen, grains of polycrystalline Cu with a purity of identified diverse geometries for plenty of in-
1Shenyang National Laboratory for Materials Science, 99.97 wt % were refined into the nanometer dividual grains. Rather than curved bounda-
Institute of Metal Research, Chinese Academy of Sciences, scale (see the supplementary materials). ries, many grains distinguish themselves by
Shenyang 110016, China. 2School of Materials Science and Bright-field transmission electron microscopy faceted boundary planes, such as (111) and
Engineering, Shanghai Jiaotong University, Shanghai (TEM) images of the extremely fine grains in (100) atomic planes. We show a typical grain
200240, China. the as-prepared specimens showed a typical of ~2 nm in diameter imaged along the [110]
*Corresponding author. Email: [email protected] (X.Y.L.); lu@imr. manifold structure (Fig. 1A), as is normally zone axis that is shaped by four faceted {111}
ac.cn (K.L.) observed in the bicontinuous lipid-water phases and two {001} boundaries (Fig. 2A). Not all
faceted boundaries are twin boundaries. Some
grains are defect free, whereas others contain

Li et al., Science 370, 831–836 (2020) 13 November 2020 1 of 6

RESEARCH | REPORT B

A

C D 60 49 51 44
60 54
45
60 25 60 60

45

50 60 40 60 60 E
39

30
30 47 36 53

40 60 60 36 27
48 55 57
48 45

34 37 30 50 21

33 25
54

46 45 55
52 27
60

60 44 50 38

60 41 32 43
40 47 57

52 36 58
60
42 60

60 60

Fig. 1. Microstructures of the as-prepared Cu sample with extremely schematic on the right. (C) A typical high-resolution TEM image. (D) A
fine grains. (A) A typical bright-field TEM image. (B) (Left) A magnified typical inverse pole figure (IPF) image acquired from a region in (C)
image of a selected area in (A). Dashed lines represent {111} planes and solid from the precession electron diffraction analysis. Numbers indicate
lines show CTBs. (Right) Corresponding FFT images of grains (G1, G2, G3, misorientation angles of GBs. (E) A typical structural-unit–type boundary
G4, and G5) labeled in the left panel. G-All indicates all grains, with a as outlined between two tiny grains.

twins or stacking faults that cut through the match the twinned or faulted polyhedra very (Fig. 3, B and C). We still observed polyhedron-
grains. The twinned or faulted grains are also well. Lattice spots on the two corners of the shaped grains with an average size of ~11 ±
bounded by {111} and {001} atomic planes (Fig. grain in Fig. 2C (arrows) are blurred or absent 2.3 nm with distinct boundaries—akin to what
2, C and E). The {111} twin boundaries in the in the TEM image. We identified a stacking we observed before annealing. We detected
grains are fully coherent, with spacing of only a fault zone between two truncated-octahedron– more twins in the annealed grains (Fig. 3C),
few atomic layers. Statistically, for grains <3 nm shaped grains (Fig. 2E), which is consistent presumably because of the further dissoci-
in size, 62% of faceted boundaries are found to with other observations (Fig. 1). Along the ation of GBs during annealing at elevated
be {111} and 33% are found to be {100}—a [110] projection of a truncated octahedron, temperatures. Elevating the temperature
frequency ratio of about 2:1. For grains between four {111} and two {100} planes exist, with a above 1357 K induced melting, at which point
3 and 10 nm, the numbers are 52% {111} and frequency ratio of 2:1. This ratio is the same all nanograins disappeared. Our observations
28% {100}, respectively. This feature disappears as what we observed in the experiments. The lead us to the conclusion that no substantial
in grains >10 nm, which supports the idea truncated-octahedral shape is apparently a grain coarsening seems to occur before melting.
that grain refining facilitates the formation of favorable option for grains <10 nm.
faceted {111} and {100} boundaries at the We prepared a separate sample for compar-
expense of other boundaries. We determined the thermal stability of the ison with larger grains (average size of 25 nm)
as-prepared Cu samples with an average grain using the same process (see the supplemen-
We also found that the shapes of those grains size of 10 nm by isothermal annealing at var- tary materials) but with a smaller strain. We
resemble that of the truncated octahedron (9). ious temperatures. We detected no apparent observed the onset of grain coarsening in this
The atomic images of the defect-free grain grain coarsening with increasing the anneal- sample at ~985 K (Fig. 3A). The grain coarsen-
(Fig. 2A) can be perfectly fitted with a part of ing temperature up to 1348 K, which is only ing temperature is higher than that of 50-nm
a truncated octahedron projected along the 9 K below the equilibrium melting point of grains, which is a trend that runs contrary to the
[110] zone axis, except for several corners Cu (1357 K) (Fig. 3A and fig. S3). Holding conventional belief that smaller grains coarsen
where atomic images are missing (circled in for 15 min at this temperature and cooling at lower temperatures. Our observations support
Fig. 2A). Other grains (Fig. 2, C and E) also down, the extremely fine grains remained the idea that GB relaxations in polycrystals of

Li et al., Science 370, 831–836 (2020) 13 November 2020 2 of 6

RESEARCH | REPORT

smaller grain size may operate more effectively Petch relationship for grains >30 nm. For the basis of the well-known Kelvin conjecture—
to enhance the thermal stability (6). The Cu finer grains, a strength plateau or even soften- that is, a topological form of grains represent-
samples in this study are the most stable ing occurs (20–23). By contrast, we did not ing polycrystalline metals constructed with
compared with the literature data for Cu (10–19), find this trend in our hardness measurements. equal-sized grains of truncated-octahedral
including nanograined, deformed coarse-grained, The strengthening we observed suggests that shape. This model provides a simple geomet-
or the metastable amorphous Cu-based alloys plastic deformation events, such as GB migra- rical solution of idealized 3D polycrystals with
with crystallization temperatures in the range tion or sliding, must be effectively prohibited nearly the minimum total area of GBs (28).
of 0.33 to 0.36 Tm. in the present Cu sample. We also found no Because a considerable fraction of GBs are
trace of dynamic recrystallization. S3-type twin boundaries in our as-prepared
For the 10-nm grain–sized sample, we mea- sample, the structure of a 3D space–filling twin
sured nanoindentation hardness to be as high The unusual stability we observed supports boundary network must be considered as a
as 4.7 ± 0.2 GPa (fig. S2). This value is equivalent the belief that when grains are refined to the basic precept as well (29).
to a yield strength of 1.57 GPa. This value is also extreme scale, the polycrystalline structure
higher than reported data on nanograined Cu seems to readily adopt one or more specific We set up an atomistic model aimed at
specimens as well as amorphous Cu-based states. Our TEM observations, combined with addressing the outstanding stability of the
alloys (20–26). The ideal shear strength of the reconstructed atomic models, indicate that extremely fine Cu grains. Referring to the
Cu at 0 K is 2.16 GPa. The shear strength we the truncated octahedron—either perfect or Kelvin model, which consists of two ideal
estimated from the nanoindentation is nearly containing planar faults—is a prototype struc- truncated-octahedra of equal size with their
comparable to the ideal shear strength when ture that represents a large portion of individ- centers (K1 and K2; Fig. 4A) in a body-centered
temperature effects are considered (27). From ual grains during the deformation-induced cubic array, we constructed an extended
previous measurements, the strength of dif- refinement process. Accordingly, potential Kelvin supercell with 16 truncated-octahedral–
ferent Cu samples follows the classical Hall- structures and GB networks can be perceived on shaped grains of equal size. We can recognize

AAA B E

CD F

Fig. 2. High-resolution TEM images of individual grains with truncated- rotated by 25.5° about the ½01 1Š axis after introducing twins (lower right).
octahedral geometries. (A) A tiny grain of ~2 nm in size. (B) A part of an Projected atomic positions (bottom left) agree with the TEM image in (C) (where
ideal truncated-octahedron with 1154 atoms (top), rotated by 49° along the only border atoms in orange and twin boundary atoms in red are shown).
[110] axis (lower right). The projected atomic positions on the (001) plane (lower Missing corners are indicated by orange arrows in (C). (E) Two grains
left) are coincident with the TEM image in (A) (where only border atoms are containing stacking faults (SFs) and twins. (F) Two attached truncated-
shown in orange). Corner atoms in blurred contrast are circled in (A). (C) A grain octahedral grains of different sizes with projected atomic positions agreeing
containing twins. (D) An ideal truncated-octahedron of 11,817 atoms (top), with the TEM image in (E).

Li et al., Science 370, 831–836 (2020) 13 November 2020 3 of 6

RESEARCH | REPORT

0.2 0.4 T/Tm 0.8 1.0 D

A 104 50 nm 0.6 1200
25 nm 1000
103 10 nm 1.0
800
600 0.8
400
Grain size (nm) T (K) 0.6 TGC/ Tm
GC 2.0
102
1.5
0.4
1.0
101 400 600 800 1000 1200 Present work (Exp.) 0.2
0.5 1.0
B Annealing temperature (K) Present work (MD)
0.0 0.8
1 SMGT
0.6
SMGT(Annealed)
0.4
C IGC ECAP

Strength (GPa) HPT DPD

ED(NT) CR

Strain machining τ/τmax

Amorphous

Hall-Petch

0.2

0.0

10 100 1000 10000

Grain size (nm)

Fig. 3. Extremely high thermal stability and strength. (A) Grain size variations as a through various processes in (6, 10–17, 20–24) are included. Data for amorphous
function of annealing temperature for three samples with initial average grain sizes of Cu alloys are from (18, 19, 25, 26). Tm, melting point of Cu; tmax, ideal shear strength
50 nm, 25 nm, and 10 nm, respectively. Each point of grain size was averaged of Cu (27). Each grain coarsening temperature was obtained from three independent
from >300 grains. (B) A TEM image of the sample with an initial grain size of 10 nm experiments, and each strength datum was obtained from 10 independent experiments.
after annealing at 1348 K for 15 min. (C) A high-resolution TEM image of a grain in (B). Exp., experimental; SMGT, surface mechanical grinding treatment; IGC, inert gas
Red lines indicate twin boundaries. (D) Grain coarsening temperatures (TGC) and condensation; ECAP, equal channel angular pressing; HPT, high-pressure torsion; DPD,
strength as a function of grain size in pure Cu. Literature data for Cu samples prepared dynamic plastic deformation; ED(NT), electrodeposition (nanotwin); CR, cold rolling.

the fundamental characteristics of GBs, and at a rate of 8 K/ns. We based the interatomic migration, the entire GB network failed to col-
twin boundary networks can be recognized. potential on the semiempirical model devel- lapse. The GBs merged and developed into a
By taking polyhedron O1 as the center cell, we oped with the embedded atom methods (EAM) different form that topologically resembles the
can specify the lattice orientation of its eight (30). For comparison, we also conducted MD tests Schwarz D-surface (32, 33). As a triply periodic
nearest neighboring grains such that each on a number of samples representing cases for flat minimal surface, we can identify the shape of
{111} face of O1 can form a S3 coherent twin and curved GBs, in which both grain size and the Schwarz D-interface in our calculations
boundary (CTB) with them—i.e., A1, B1, C1, orientations are treated as free variables (fig. S6). (Fig. 4B and fig. S5).
D1 (upper) and A2, B2, C2, D2 (lower). The
lattice orientations of the six second-nearest During MD relaxation and subsequent heat- This transformation is thermodynamically
neighboring grains to O1 (I1, E1, F1 and I2, ing, GBs in the extended Kelvin polycrystal driven by minimizing the interface area, as
E2, F2) are not strongly limited. We chose transform into different structures through indicated by the MD results. Although the
them such that in each <001> orientation, a extensively GB reaction and migration events. Kelvin polycrystal provides a reasonable solu-
S9-type tilt GB is formed with O1 (or O2) (fig. With increasing temperature, GB reactions— tion for an efficient, space-filling arrangement
S4 and table S2). Our result confirms the the- such as dissociation and coalescence of GBs as- of grains, the total boundary area was further
oretical prediction (29) that a 3D space–filling sociated with some S9-type GBs—are triggered minimized by forming the Schwarz D-interface
CTB network can be constructed by a spe- at triple lines or junctions of GBs, as expected structure (34). Our observation indicates that
cific arrangement of truncated octahedra in experimentally (31) and theoretically (29). Mi- the polycrystalline structure with Schwarz D-
fcc solids. gration of twist GBs becomes substantial at interfaces is more stable than the Kelvin poly-
mediate temperatures (640 K), and it is close crystals, but this does not necessarily mean that
By choosing an extended Kelvin polycrystal to that for curved GBs observed in separate such a transformation occurs at a distinct point
with an initial grain size of 3.27 nm as a simulations (fig. S6). The preexisting CTB net- during the refinement of coarse grains into the
starting structure for simplicity, we carried work is stable against thermal fluctuations. nanograins. Our MD simulations also confirmed
out MD simulations to relax the sample by After heating to 760 K, although some grains that the Schwarz D-structure can be obtained
heating it up to different target temperatures shrank and finally disappeared because of GB by imposing strain at low temperatures. The

Li et al., Science 370, 831–836 (2020) 13 November 2020 4 of 6

RESEARCH | REPORT C

A

C1 D1

I1
E1 O2

C2 A1 B1
D2
F1
E2 O1 I2

A2 B2

O2 F2

B T/Tm

S1 S2 D 0.15 0.29 0.44 0.59 0.74 0.88

S3 2.5

Yield strength (GPa) 2.0

1.5

1.0

0.5

0.0 200 400 600 800 1000 1200
0
T (K)

Fig. 4. Atomistic model and MD simulations of Schwarz crystals. (A) The MD-obtained twin-bounded polycrystalline structure at 0 K, demonstrated by
original Kelvin model of two ideal truncated-octahedra of equal volume 2 by 2 by 2 supercells where atoms in fcc lattice sites are removed. (Right)
(K1 and K2) in 1 by 1 packing (top left). A polycrystal of 16 grains (right) was GBs resembling the Schwarz D-interface in a 1 by 1 by 1 supercell. (C) A
constructed using a 4 by 4 packing Kelvin model (initial grain size, 6.6 nm). section-view of the Schwarz crystal showing Schwarz D-GBs constrained by
A space-filling 3D CTB network was constructed with a specified lattice CTB networks. (D) The MD-obtained yield stress as a function of temperature.
orientation for individual grains (see Supplementary Materials). (B) (Left) Error bars quantify uncertainty caused by rate effects and thermal fluctuations.

adoption of the minimal-interface structure point the liquid phase is nucleated hetero- incipient twinning is temperature-dependent
constitutes an energetically more-favorable geneously at 1321 K (fig. S6). This suggests that (Fig. 4D). The yield strength is 1.8 to 2 GPa at
GB relaxation process under thermal and the upper thermal stability is limited kineti- 300 K, which is higher than our experimen-
mechanical stimuli. cally by GB melting. This notable stability agrees tally measured data, possibly owing to the grain
fully with our experimental results (Fig. 3). size distribution in real samples. The strength
The Schwarz D-interface is constrained by Clearly, preventing the migration of GBs at 10 K is comparable to the ideal pure shear
high-density CTBs on both sides (Fig. 4C)—a through a minimal interface constrained by a strength obtained through first-principle cal-
structure meant to suppress GB migration. 3D CTB network constitutes the key mecha- culations (27).
With the minimal-interface area and the zero- nism to achieve the pronounced stability for
mean curvature (another characteristic of mini- polycrystals at such extreme sizes. Evidence based on our experiments and
mal interface structures), the driving force to MD simulations confirms that a pronounced
migrate the GBs tends to vanish, which is pro- We conducted uniaxial tensile-loading tests stability is achievable if the minimal-interface
portional to the GB curvature. CTBs are intrin- on the CTB-constrained Schwarz D-structure structure constrained by a CTB network is
sically stable at elevated temperatures. The stable at various temperatures and strain rates (105 adopted in polycrystalline Cu with nanosized
CTB networks and Schwarz D-interface inter- to 107 s−1) (fig. S7). GB activities such as sliding grains. The structure, which we refer to as a
lock with each other, preventing GBs from de- or migration can be suppressed entirely. The Schwarz crystal, is a different kind of meta-
veloping substantial curvatures to escape and primary mode of deformation is twinning. stable state for polycrystalline solids. This state
CTBs from migrating. Such a pronounced stability For example, a twinning partial dislocation differs fundamentally from the amorphous solid
cannot be achieved with the Schwarz D-interface is generated from one GB site and absorbed states. Because the metastable structure is formed
only, nor with only the nanotwinned structures. at the opposite GB, which results in migra- when twinning—either thermally induced or
tion of the twin boundary plane. Generating deformation induced—and is effective in relax-
The obtained Schwarz D-structure stays a stacking fault to a glide plane parallel to the ing GB structures, the appearance of the Schwarz
stable at elevated temperatures. Rather than CTB is also possible during the plastic defor- crystal in different metals and alloys is expected
grain coarsening, roughening of GBs occurs mation. The critical stress corresponding to with the activation of twinning mechanisms at
as the melting point is approached, at which

Li et al., Science 370, 831–836 (2020) 13 November 2020 5 of 6

RESEARCH | REPORT

the nanoscale. The pure Cu Schwarz crystal, 4. P. Färber, E. Cadel, A. Menand, G. Schmitz, R. Kirchheim, Acta 29. V. Y. Gertsman, B. W. Reed, Z. Metallk. 96, 1106–1111
which contains a very high density of inter- Mater. 48, 789–796 (2000). (2005).
faces, exhibits a thermal stability as high as
that of a single crystal and much higher than 5. C. A. Schuh, T. G. Nieh, H. Iwasaki, Acta Mater. 51, 431–443 (2003). 30. V. Borovikov, M. I. Mendelev, A. H. King, R. LeSar, Model. Simul.
that of amorphous solids. Therefore, the struc- 6. X. Zhou, X. Y. Li, K. Lu, Science 360, 526–530 (2018). Mater. Sci. Eng. 23, 055003 (2015).
ture provides emerging opportunities for ex- 7. X. Zhou, X. Y. Li, K. Lu, Phys. Rev. Lett. 122, 126101 (2019).
ploring physical and chemical phenomena of 8. W. Longley, T. J. McIntosh, Nature 303, 612–614 (1983). 31. T. Yu, N. Hansen, X. Huang, Proc. R. Soc. A. 467, 3039–3065
metals, particularly with respect to transport 9. C. Chieh, Acta Cryst. A35, 946–952 (1979). (2011).
dynamics of atoms and electrons at interfaces 10. C. Saldana, A. H. King, S. Chandrasekar, Acta Mater. 60,
and interactions of various defects at high tem- 32. H. A. Schwarz, Gesammelte Mathematische Abhandlungen,
peratures. As a benchmark structure capable 4107–4116 (2012). vols. 1 and 2 (Springer, 1890).
of suppressing grain coarsening under both 11. X. Y. Li, X. Zhou, K. Lu, Sci. Adv. 6, eaaz8003 (2020).
thermal fluctuations and external forces, the 12. A. Kumpmann, B. Günther, H. D. Kunze, Mater. Sci. Eng. A 168, 33. A. L. Mackay, Nature 314, 604–606 (1985).
Schwarz crystal allows for elevated stability 34. L. E. Scriven, Nature 263, 123–125 (1976).
and strength that goes with refining grains 165–169 (1993).
into the extremely fine scale. This overcomes 13. N. Lugo, N. Llorca, J. J. Suñol, J. M. Cabrera, J. Mater. Sci. 45, ACKNOWLEDGMENTS
some of the difficulties present in traditional
strategies for materials development. In prin- 2264–2273 (2010). We thank C. Y. Wu for help with sample preparation.
ciple, the Schwarz crystal should be accessible 14. C. F. Gu, C. H. J. Davies, Mater. Sci. Eng. A 527, 1791–1799 (2010). Funding: This work is funded by the Ministry of Science and
in other materials and may provide a different 15. Y. Zhang, N. R. Tao, K. Lu, Acta Mater. 56, 2429–2440 (2008). Technology of China (grant nos. 2017YFA0204401 and
direction for developing strong and stable ma- 16. P. Jenei, J. Gubicza, E. Y. Yoon, H. S. Kim, J. L. Lábár, 2017YFA0700700) and the Chinese Academy of Sciences.
terials for high-temperature applications. Author contributions: X.Y.L. and K.L. developed the concept,
Compos. A: Appl. Sci. Manuf. 51, 71–79 (2013). designed the experiments, and supervised the projects. X.Z.
REFERENCES AND NOTES 17. Z. N. Mao et al., Mater. Sci. Eng. A 674, 186–192 (2016). and X.Y.L. prepared the samples and characterized structures.
1. R. Zallen, The Physics of Amorphous Solids (Wiley, 1983). 18. B. W. Zhang, X. L. Shu, S. Y. Zhu, S. Z. Liao, J. Mater. X.Z. measured properties. Z.H.J. performed modeling and
2. P. Haasen, B. L. Mordike, Physical Metallurgy (Cambridge Univ. simulations. All authors discussed and analyzed the results.
Proc. Tech. 91, 90–94 (1999). X.Y.L., Z.H.J., and K.L. wrote the paper. Competing interests:
Press, 1996.) 19. H. Zhang, Z. Tan, B. Zhang, X. Shu, J. Yu, J. Mater. Sci. Lett. 10, The authors declare no competing interests. Data and
3. K. Lu, Nat. Rev. Mater. 1, 16019 (2016). materials availability: All data are available in the manuscript
45–46 (1991). or the supplementary materials.
20. L. Lu, X. Chen, X. Huang, K. Lu, Science 323, 607–610 (2009).
21. A. H. Chokshi, A. Rosen, J. Karch, H. Gleiter, Scr. Metall. 23, SUPPLEMENTARY MATERIALS

1679–1683 (1989). science.sciencemag.org/content/370/6518/831/suppl/DC1
22. G. E. Fougere, J. R. Weertman, R. W. Siegel, Nanostruct. Mater. Materials and Methods
Supplementary Text
3, 379–384 (1993). Figs. S1 to S7
23. P. G. Sanders, J. A. Eastman, J. R. Weertman, Acta Mater. 45, Tables S1 to S3
References (35–44)
4019–4025 (1997).
24. H. G. Jiang, Y. T. Zhu, D. P. Butt, I. V. Alexandrov, T. C. Lowe, 1 August 2020; accepted 8 October 2020
10.1126/science.abe1267
Mater. Sci. Eng. A 290, 128–138 (2000).
25. W. H. Wang, J. Appl. Phys. 99, 093506 (2006).
26. Y. P. Deng et al., Intermetallics 15, 1208–1216 (2007).

27. S. Ogata, J. Li, S. Yip, Science 298, 807–811 (2002).

28. F. N. Rhines, K. R. Craig, R. T. DeHoff, Metall. Trans. 5, 413–425
(1974).

Li et al., Science 370, 831–836 (2020) 13 November 2020 6 of 6

RESEARCH

COSMOCHEMISTRY phase) (1, 6) or with the pre–main sequence,
T Tauri phase (class II phase) of the proto-
Astronomical context of Solar System formation Sun, which would have occurred up to ~1 Myr
from molybdenum isotopes in meteorite inclusions later (10, 11). Distinguishing between these
interpretations would provide a link between
Gregory A. Brennecka1,2*, Christoph Burkhardt2, Gerrit Budde2,3, Thomas S. Kruijer1,4, astronomical observations of YSOs and the
Francis Nimmo5, Thorsten Kleine2 physical samples defining the age of the Solar
System.
Calcium-aluminum–rich inclusions (CAIs) in meteorites are the first solids to have formed in the Solar
System, defining the epoch of its birth on an absolute time scale. This provides a link between We investigate this link using the isotopic
astronomical observations of star formation and cosmochemical studies of Solar System formation. compositions of CAIs. Previous work has shown
We show that the distinct molybdenum isotopic compositions of CAIs cover almost the entire that CAIs are isotopically distinct from subse-
compositional range of material that formed in the protoplanetary disk. We propose that CAIs formed quently formed solids in the solar protoplan-
while the Sun was in transition from the protostellar to pre–main sequence (T Tauri) phase of star etary disk (12, 13), but any “genetic” connections
formation, placing Solar System formation within an astronomical context. Our results imply that between CAIs and later-formed solids are
the bulk of the material that formed the Sun and Solar System accreted within the CAI-forming epoch, unclear. Compared with bulk meteoritic mate-
which lasted less than 200,000 years. rials, CAIs are 16O-rich and more similar to
the Sun than later-formed objects (14). How-
C alcium-aluminum–rich inclusions (CAIs) percent-level isotope differences from other ever, because O-isotope variations reflect the
are the oldest dated samples of the Solar Solar System materials, may have formed at a complex interplay of mixing and fractiona-
System and provide a direct record of slightly earlier stage [e.g., (5)], the majority of tion processes within the protoplanetary disk,
its formation. These micrometer- to CAIs formed 4567.2 ± 0.2 million years (Myr) and perhaps the parent molecular cloud (14),
centimeter-sized inclusions in meteorites ago, over a period of ~40,000 to 200,000 years O isotopes alone are of limited use for linking
formed in a high-temperature environment (6–8). Astronomical observations of young stel- CAI formation to stellar accretion and disk
(>1300 K), probably near the young Sun (1), lar objects (YSOs) show that solar-mass stars formation. Another type of isotope variation,
and were subsequently transported outward accrete by infall from their parental molecular known as nucleosynthetic variation, is due
to the region where carbonaceous chondrite cloud cores before transitioning into pre–main solely to inherent isotopic differences in the
meteorites (and their parent bodies) formed, sequence stars surrounded by a disk of gas and starting material and can be used to identify
in which they are found today (2–4). Although dust (referred to as classes 0 to II in the star temporal and spatial differences in the prov-
some peculiar CAI-like objects, with intrinsic formation literature) within ~1 Myr (9). This is enance of Solar System materials. These iso-
longer than CAIs took to form, raising the ques- tope variations arise from the heterogeneous
1Lawrence Livermore National Laboratory, Livermore, CA, tion of which astronomical phase in the Solar distribution of presolar material within the
USA. 2Institut für Planetologie, University of Münster, System’s formation is recorded by the forma- disk and show that the protoplanetary disk
Münster, Germany. 3Division of Geological and Planetary tion of CAIs. Previous work modeling the for- can be divided into noncarbonaceous (NC)
Sciences, California Institute of Technology, Pasadena, CA, mation of the Solar System and observations and carbonaceous (CC) reservoirs, which have
USA. 4Museum für Naturkunde, Leibniz Institute for Evolution of other stellar systems have associated the been suggested to represent the inner and outer
and Biodiversity Science, Berlin, Germany. 5Department of formation of CAIs either with the period of Solar System, respectively (15–17). This NC-CC
Earth & Planetary Sciences, University of California Santa collapse and accretion of a protostar (class 0 isotopic dichotomy has been demonstrated
Cruz, Santa Cruz, CA, USA. in several elements, including Mo, which
*Corresponding author. Email: [email protected] can distinguish between isotope variations
arising from the heterogeneous distribution

Fig. 1. Molybdenum isotopic data for all CAIs in B.
our sample set. (A) e94Mo versus e95Mo [where e
notation represents parts per 10,000 deviation from
a terrestrial standard, (20)]. Data plotted are from
this work and other studies (12, 18, 25). Samples
with half-gray symbols have been excluded from
calculations of D95Mo (20). The CC (blue) and NC
(red) correlations (21) are shown for reference.
(B) Magnified view of the region within the dashed
box in (A). (C) CAI isotopic variability due to just
the r-process, presented using D95Mo notation (see
text for details). The left axis lists the CAIs from
literature sources (12, 18, 25) in plain text and those
from this study in bold. We have excluded CAIs from
the literature that do not have comparable method-
ologies and CAIs with terrestrial contamination of
their Mo isotopic composition. Error bars show 95%
confidence intervals. The CC (blue) and NC (red)
correlations (21) are shown for reference. In (A) to
(C), unfractionated CAIs are plotted with orange
squares and group II CAIs with yellow circles.

Brennecka et al., Science 370, 837–840 (2020) 13 November 2020 1 of 4

RESEARCH | REPORT

Fig. 2. Isotopic and concentration correlations in CAIs. (A) D95Mo plotted D95Mo and e183W values do not correlate with one another or show a mixing
relationship with the meteorite matrix. This indicates that parent body
against concentration of Mo in the CAIs. In general, samples with a lower alteration does not control these signatures. Bulk NC and CC meteorite
concentration of Mo also have a lower D95Mo, suggesting at least some data for Mo are from (21), Ti data are from (31, 32), and the matrix of
the CC meteorite Allende (blue diamonds) is from (16, 28). Other plotting
disk processing or CAI formation effect on the r-process signature of symbols are as in Fig. 1.
the CAI. (B) D95Mo plotted against e50Ti in CAIs. See text for discussion
of these isotopic signatures in CAIs. (C) D95Mo plotted against e183W. The

of matter produced in the slow (s-process) data from three previous studies (12, 18, 25), 1C). Unlike the absolute Mo isotope anoma-
which are summarized in table S1. CAIs are lies (i.e., eiMo; Fig. 1, A and B) and the iso-
and rapid (r-process) neutron capture pro- tope anomalies for other elements (Fig. 2, B
cesses of stellar nucleosynthesis (15, 16, 18, 19). classified by their condensation histories, as and C, and tables S2 and S4), the D95Mo
Whereas there are large s-process variations inferred from their rare earth element (REE) values provide a link between CAIs and the

among materials within each group, all CC systematics. Hereafter, we refer to samples with solid material of the disk as sampled by
unfractionated REE patterns as “unfraction-
materials are characterized by an approxi- ated CAIs,” as distinct from “group II CAIs,” bulk meteorites. That is, the composition of
which have fractionated REE patterns, indicat-
mately constant r-process excess over NC CAIs varies from larger r-process excesses
ing the loss of an ultrarefractory component (more positive D95Mo) toward more NC-like
materials. Deviations from pure s-process varia- (26). In addition to Mo, isotopic compositions compositions.
tion are quantified using the D95Mo notation of the CAIs were determined for several other
[D95Mo ≡ (e95Mo – 0.596 × e94Mo) × 100, where elements [Ti, Fe, Ni, Sr, Ba, Nd, Sm, Er, Yb, We interpret the distinct D95Mo values for
e represents parts per 10,000 deviation from a Hf, and W; (20)], all of which are isotopically these CAIs by relating them to models of how
terrestrial standard (20)], which isolates vari- anomalous in CAIs. The 15 samples newly mea- the D95Mo values of the NC and CC reservoirs
ation in the amount of r-process Mo from sured in this study are isotopically consist- may have been established. These models demon-
variation in the amount of s-process Mo. strate that during infall of the Solar System’s
Previous studies have shown that the D95Mo ent with previously published studies of parental molecular cloud, the isotopic com-
values for the NC reservoir, thought to re- CAIs for these elements and exhibit larger
present the inner Solar System (D95Mo = −9 ± 2), position of the infalling material changed
and CC reservoir, thought to represent the anomalies than bulk meteorites; however, from high D95Mo values in the early stages to
outer Solar System (D95Mo = 26 ± 2), are we concentrate on their Mo isotopes (Fig. 1). lower, NC-like D95Mo values toward the final
distinct from one another (21). The NC-CC stage; and that as a result of viscous spreading
isotopic difference was established within the In general, unfractionated CAIs exhibit less
first ~1 Myr of Solar System history (15), so it variation in s-process Mo signatures (i.e., of the initial disk, the bulk of the mass of the
likely reflects the incomplete mixing of iso- eiMo values; Fig. 1, A and B) and mostly
have higher and less variable D95Mo values inner protoplanetary disk comes from the
topically variable material infalling from the (Fig. 1C). By contrast, group II CAIs tend to
Sun’s protostellar envelope onto the disk at have more variable s-process Mo signatures later infall, whereas the outer disk contains a
different times (22–24). Molybdenum isotopes, (Fig. 1, A and B) and lower D95Mo values
therefore, provide a tool for linking the for- (Fig. 1C). However, a similar overall range in higher proportion of early infalling material
r-process material (>100 parts per million; (4, 22–24). The rapid outward transport of
mation of CAIs to the early processes of infall Fig. 1C) is present in both types regardless material during the initial disk-building pro-

and disk building and, hence, to a specific of how much s-process variability exists, cess provides a mechanism for the transport
showing that variations in s-process and
stage of star formation. r-process Mo nuclides in CAIs are not directly of CAIs from near the Sun to the outer Solar
coupled. System (2–4). Within this framework, the
We measured the Mo isotopic and trace range of D95Mo in CAIs extending from values
The CAIs exhibit different D95Mo values, characteristic of early infall (higher than the
element compositions for a variety of CAIs only some of which overlap with the values
CC reservoir) to that of bulk meteorites may
taken from the carbonaceous chondrite (CV3 observed for bulk CC and NC meteorites (Fig.
have three potential origins: (i) exchange of
group) meteorites Allende (13 CAIs), NWA
Mo with the host meteorite during altera-
6870 (CAI AI01), and NWA 6717 (CAI AI02).
tion on the parent body; (ii) changes in the
We combine these measurements with similar
isotopic composition of infall, and, hence,

Brennecka et al., Science 370, 837–840 (2020) 13 November 2020 2 of 4

RESEARCH | REPORT CAI formation Dynamic +150 Δ95Mo
infall +100
Fig. 3. Schematic diagram of (<200,000 years) Class Evolution of infall
the early evolution of the Solar I CAIs +50 over time
System. The diagram is modeled Infall period ? 0
after (22) but modified to reflect Class
our interpretation of the CAI data I CAIs -25
and associated stages of stellar
evolution. Heterogeneous infall Time Class Mixing
from the Sun’s parent molecular I CAIs
cloud evolves from a higher to NC CC
lower amount of r-process Class
material (i.e., higher to lower II Jupiter
D95Mo, indicated by the color bar).
The isotopic compositions of
CAIs record this change in
composition and subsequent
processing within the disk. The
formation of CAIs occurs during
the transition of the Sun from a
class I to class II YSO in less
than 200,000 years.

the gas from which CAIs condensed; or (iii) CAIs condensed, CAIs should also directly served among bulk meteorites, suggesting that
CAIs acquired their isotopic signatures when
processing of CAIs within the disk before record changes in other isotope systems that the Sun’s protoplanetary disk had largely
formed. This links CAI formation to the
incorporation into their host meteorite. distinguish between NC and CC compositions, transition of the Sun from a class I to a class II
protostar, which is consistent with evidence of
Group II CAIs have substantially lower con- such as Ti. Bulk NC meteorites are charac- the short-lived isotope 10Be in CAIs (11) and
terized by negative e50Ti anomalies down to noble gas signatures from hibonite inclusions
centrations of Mo than unfractionated CAIs about –2, whereas the majority of CAIs have (30)—refractory inclusions thought to have
e50Ti excesses around +9 [e.g., (29)]. Thus, if formed before the more common CAIs—
and so are more susceptible to alteration in evolving infall from the molecular cloud was indicating that these refractory objects formed
when the Sun was a T Tauri star.
their Mo isotopic composition than unfrac- the sole cause of isotopic anomalies in CAIs,
then D95Mo should coevolve with e50Ti in Because the vast majority of CAIs are
tionated CAIs. This may explain why group CAIs, assuming that Mo and Ti had similar thought to have formed within ~40,000 to
carriers. However, with one exception—CAI 200,000 years (6–8), the recorded isotopic
II samples tend to have lower (more bulk R2, which exhibits both low D95Mo and low character of CAIs also represents an amount
meteorite-like) D95Mo values, although two e50Ti—no such clear coevolution is observed of time in which that isotopic character was
unfractionated CAIs (AI05 and R2) also show (Fig. 2B). A change of infalling matter toward established. Interpreting the isotopic com-
positions of CAIs in the framework of disk
such isotopic compositions. Nevertheless, be- NC-like compositions, which is characterized evolution models, and with the knowledge
cause the D95Mo values of many group II CAIs by e183W ~ 0, also does not provide a viable that CAIs formed close to the young Sun in a
are similar to those of the matrix in Allende explanation for the positive e183W values of known time window, we constrain the primary
(16), their D95Mo values may potentially reflect group II CAIs. Together, these observations accretion phase (i.e., classes I and II) of stars
exchange with the host meteorite on the parent like the Sun to less than 200,000 years, consist-
suggest that dynamic infall from the molec- ent with astronomical observations of YSOs (9).
body. However, W, which is thought to behave
ular cloud (Fig. 3) cannot be the sole cause of REFERENCES AND NOTES
similarly to Mo during parent-body alteration
(27), is also strongly depleted in group II CAIs, the observed isotopic signatures in CAIs. We 1. J. A. Wood, Geochim. Cosmochim. Acta 68, 4007–4021 (2004).
yet all group II CAIs with low D95Mo exhibit 2. E. Jacquet, S. Fromang, M. Gounelle, Astron. Astrophys. 526,
positive e183W signatures, unlike the negative propose that even though dynamic infall ini-
e183W of the matrix of CV3 chondrites (28). tially generated the range of D95Mo varia- L8 (2011).
This indicates that parent-body alteration is tions within the disk, additional processing 3. F. C. Pignatale, S. Charnoz, M. Chaussidon, E. Jacquet,

not the source of the signatures. As such, the within the protoplanetary disk established Astrophys. J. Lett. 867, 23 (2018).
4. L. Yang, F. J. Ciesla, Met. Planet. Sci 47, 99–119 (2012).
combined Mo-W isotopic signatures of group the measured isotopic signatures of CAIs. This 5. L. Kööp et al., Geochim. Cosmochim. Acta 189, 70–95 (2016).
processing evidently did not affect e50Ti, 6. J. N. Connelly et al., Science 338, 651–655 (2012).
II CAIs are difficult to account for solely by which shows little variation among the CAIs 7. N. Kawasaki et al., Earth Planet. Sci. Lett. 511, 25–35 (2019).
8. G. J. MacPherson, N. T. Kita, T. Ushikubo, E. S. Bullock,
parent-body processes. This is also true for the we studied but resulted in substantial and
large range of eiMo values in group II CAIs uncorrelated anomalies of eiMo and e183W. A. M. Davis, Earth Planet. Sci. Lett. 331-332, 43–54 (2012).
(Fig. 1, A and B) and for the isotopic variation The reasons for this disparate behavior during 9. L. Cieza et al., Astrophys. J. 667, 308–328 (2007).
10. F. J. Ciesla, Icarus 208, 455–467 (2010).
in unfractionated CAIs. Thus, although parent- processing are unclear, but they may include 11. N. Dauphas, M. Chaussidon, Annu. Rev. Earth Planet. Sci. 39,

body processes may have modified Mo isotope different thermal properties of the carriers in- 351–386 (2011).
12. G. A. Brennecka, L. E. Borg, M. Wadhwa, Proc. Natl. Acad.
signatures in some group II CAIs, they cannot volved, size or density sorting of these carriers,
Sci. U.S.A. 110, 17241–17246 (2013).
account for most of the isotopic variations that and chemical fractionations during CAI for-

we see (see supplementary text). Consequently, mation, which made Mo and W more suscep-
the range of D95Mo in CAIs was likely acquired
before chondrite parent body accretion, either tible to isotopic modifications by processes

during CAI condensation or subsequent pro- within the disk.
The range of D95Mo values in CAIs covers
cessing of CAIs within the disk.
If the range of D95Mo in CAIs directly almost the entire compositional range ob-

recorded the change of the isotopic composition

of infall and, hence, of the gas from which

Brennecka et al., Science 370, 837–840 (2020) 13 November 2020 3 of 4

RESEARCH | REPORT

13. N. Dauphas, E. A. Schauble, Annu. Rev. Earth Planet. Sci. 44, 29. A. M. Davis et al., Geochim. Cosmochim. Acta 221, 275–295 (2018). was performed under the auspices of the U.S. Department of Energy by
709–783 (2016). 30. L. Kööp et al., Nat. Astron. 2, 709–713 (2018). Lawrence Livermore National Laboratory under contract DE-AC52-
31. A. Trinquier et al., Science 324, 374–376 (2009). 07NA27344 with release number LLNL-JRNL-802017. Author
14. K. D. McKeegan et al., Science 332, 1528–1532 (2011). 32. J. Zhang, N. Dauphas, A. M. Davis, I. Leya, A. Fedkin, contributions: G.A.B., C.B., and G.B. analyzed samples; G.A.B., C.B.,
15. T. S. Kruijer, C. Burkhardt, G. Budde, T. Kleine, Proc. Natl. T.S.K., and T.K. devised the project; and all authors contributed to
Nat. Geosci. 5, 251–255 (2012). interpretation of the data and writing the manuscript. Competing
Acad. Sci. U.S.A. 114, 6712–6716 (2017). interests: The authors declare no competing interests. Data and
16. G. Budde et al., Earth Planet. Sci. Lett. 454, 293–303 (2016). ACKNOWLEDGMENTS materials availability: All data are available in tables S1 to S6. Newly
17. P. H. Warren, Earth Planet. Sci. Lett. 311, 93–100 (2011). measured CAIs were taken from fragments of the meteorites Allende,
We thank A. Bischoff for helpful discussions and J. Render and NWA 6870, and NWA 6717, all of which are curated at the Institut
18. C. Burkhardt et al., Earth Planet. Sci. Lett. 312, 390–400 (2011). Q. Shollenberger for support and helpful comments on an early für Planetologie, University of Münster, Germany, except samples
19. G. M. Poole, M. Rehkämper, B. J. Coles, T. Goldberg, draft of this manuscript. We thank N. Marks and B. Jacobsen for CAI 43 and BB8, which were obtained from a fragment of Allende
assistance with the petrography. We acknowledge the efforts of curated at the AMNH, New York, USA. The CAI samples were consumed
C. L. Smith, Earth Planet. Sci. Lett. 473, 215–226 (2017). U. Heitmann for the initial preparation of many of these samples. during the experiments.
20. Materials and methods are available as supplementary materials. Allende CAI samples CAI 43 and BB8 were obtained from the
21. G. Budde, C. Burkhardt, T. Kleine, Nat. Astron. 3, 736–741 (2019). collection of the American Museum of Natural History (AMNH) SUPPLEMENTARY MATERIALS
22. J. A. M. Nanne, F. Nimmo, J. N. Cuzzi, T. Kleine, Earth Planet. (New York, USA). D. S. Ebel and S. Wallace (Earth and Planetary
Sciences, AMNH); R. Matthes, L. Schultz, and L. L. Pryer (Siemens science.sciencemag.org/content/370/6518/837/suppl/DC1
Sci. Lett. 511, 44–54 (2019). Medical Solutions, Cary, NC); R. Smith (NC Museum of Natural Materials and Methods
23. C. Burkhardt, N. Dauphas, U. Hans, B. Bourdon, T. Kleine, Sciences); and J. Thorstenson (Duke University) are thanked for Supplementary Text
their support and for providing access to computerized Figs. S1 to S7
Geochim. Cosmochim. Acta 261, 145–170 (2019). tomography scanners. Funding: G.A.B. was supported by a Tables S1 to S6
24. E. Jacquet, F. C. Pignatale, M. Chaussidon, S. Charnoz, Sofja Kovalevskaja award from the Alexander von Humboldt References (33–63)
Foundation. The Deutsche Forschungsgemeinschaft as part of the
Astrophys. J. 884, 32 (2019). Research Priority Program SPP 1385 supported T.S.K. (grant KR 14 October 2019; accepted 16 September 2020
25. Q. R. Shollenberger et al., Geochim. Cosmochim. Acta 228, 4463/1) and T.K. (grant KL 1857/4). An Annette Kade Fellowship 10.1126/science.aaz8482
(to G.B.) supported characterization of two CAI samples. This study
62–80 (2018).
26. A. M. Davis, L. Grossman, Geochim. Cosmochim. Acta 43,

1611–1632 (1979).

27. P. J. Sylvester, B. J. Ward, L. Grossman, I. D. Hutcheon,
Geochim. Cosmochim. Acta 54, 3491–3508 (1990).

28. G. Budde, T. Kleine, T. S. Kruijer, C. Burkhardt, K. Metzler,
Proc. Natl. Acad. Sci. U.S.A. 113, 2886–2891 (2016).

Brennecka et al., Science 370, 837–840 (2020) 13 November 2020 4 of 4

RESEARCH

OPTOMECHANICS phononic shield, with breathing-mode quality
factors reaching Q ¼ 4:9 Â 1010, corresponding
Nano-acoustic resonator with ultralong to a frequency-Q product of f À Q ¼ 2:6 Â 1020.
phonon lifetime
The devices studied in this work are fab-
Gregory S. MacCabe1,2,3*, Hengjiang Ren1,2*, Jie Luo1,2†, Justin D. Cohen1,2‡, Hengyun Zhou1,2§, ricated from the 220-nm device layer of a
Alp Sipahigil1,2, Mohammad Mirhosseini1,2, Oskar Painter1,2,3# silicon-on-insulator (SOI) microchip. [See (17)
for materials and methods.] In Fig. 1, A and B,
The energy damping time in a mechanical resonator is critical to many precision metrology applications, we show scanning electron microscope (SEM)
such as timekeeping and force measurements. We present measurements of the phonon lifetime images of a single fabricated device, which
of a microwave-frequency, nanoscale silicon acoustic cavity incorporating a phononic bandgap acoustic consists of a coupling optical waveguide, a
shield. Using pulsed laser light to excite a colocalized optical mode of the cavity, we measured the pair of nanobeam OMC cavities that support
internal acoustic modes with single-phonon sensitivity down to millikelvin temperatures, yielding both a microwave acoustic and an optical reso-
a phonon lifetime of up to tph;0 ≈ 1:5 seconds (quality factor Q ¼ 5 Â 1010) and a coherence time of nant mode, and the acoustic shield that con-
tcoh;0 ≈ 130 microseconds for bandgap-shielded cavities. These acoustically engineered nanoscale nects the cavities to the surrounding chip
structures provide a window into the material origins of quantum noise and have potential applications substrate. Figure 1C shows finite-element
ranging from tests of various collapse models of quantum mechanics to miniature quantum memory method (FEM) simulations of the microwave
elements in hybrid superconducting quantum circuits. acoustic breathing mode and fundamental op-
tical mode of the nanobeam cavity. The design
I n optics, geometric structuring at the smaller than the acoustic frequency ( w ), a of the OMC cavities, detailed in (5), provide
nanoscale has become a powerful method quantum model of three-phonon scattering strong localization and overlap of the breath-
can be used to describe phonon-phonon mix- ing mode and the fundamental optical mode,
for modifying the electromagnetic proper- ing that results in damping and thermaliza- resulting in a vacuum optomechanical coupling
tion of acoustic modes (12, 13). Landau-Rumer rate (5) between photons and phonons of
ties of a bulk material, leading to meta- damping scales approximately as T a, where g0=2p ≈ 1 MHz.
a ≈ 4 depends upon the phonon dispersion
materials capable of manipulating light in and density of states (DOS) (13). At the very To minimize mechanical clamping losses,
unprecedented ways (1). In the most extreme lowest lattice temperatures (≲ 10 K), where the nanobeam is anchored to the Si bulk with
case, photonic bandgaps can emerge in which Landau-Rumer damping has dropped off, a a periodic cross structure (5). FEM numerical
residual damping emerges due to material de- modeling indicates that the cross shield must
light is forbidden from propagating, markedly fects. These two-level system (TLS) defects (14), be carefully designed to provide protection
typically found in amorphous materials, cor- against nanometer-scale disorder, which is in-
altering the emission of light from within such respond to a pair of nearly degenerate local herently introduced during device fabrication
materials (2). More recently, a similar phono- arrangements of atoms in the solid, which (see materials and methods, section II). As
nics revolution (3) in the engineering of acoustic can have both an electric and an acoustic tran- shown in Fig. 1, D to F, by including the finite
waves has led to a variety of new devices, from sition dipole, and couple to both electric and radius of curvature (35 nm) of the corners of
strain fields. Recent theoretical analysis shows the cross shape present in fabricated struc-
thermal crystals for controlling the flow of that TLS interactions with acoustic waves can tures, we could increase the acoustic bandgap
heat (3) to phononic topological insulators for be substantially altered in a structured mate- to a 3-GHz bandwidth and accurately align
scattering-free transport of acoustic waves (4). rial (10). the breathing acoustic mode of the nanobeam
to the center of the shield gap. Reevaluation
Phononic bandgap structures, similar to Here we explore the limits of acoustic damp- of devices measured in (16), which did not
ing and coherence of a microwave-frequency factor in the finite radius curvature of the
their electromagnetic counterparts, can be acoustic nanocavity with a phononic crystal fabricated cross shield, indicate a misalign-
shield that possesses a wide bandgap for all ment of the acoustic bandgap with the nano-
used to modify the emission or scattering of polarizations of acoustic waves. Our nano- beam mode.
cavity, formed from an optomechanical crystal
phonons. These ideas have recently been ex- (OMC) nanobeam resonator (5), supports an To investigate the efficacy of the newly de-
plored in quantum optomechanics (5–8) and acoustic breathing mode at wm=2p ≈ 5 GHz signed acoustic shielding in practice, we fab-
electromechanics (9) experiments to greatly and a colocalized optical resonant mode at ricated and characterized arrays of devices
reduce the mechanical coupling to the ther- wc=2p ≈ 195 THz (lc ≈ 1550 nm), which allows with a scaling of the cross period number from
us to excite and read out mechanical motion Nshield ¼ 0 to 10, with all other design param-
mal environment through acoustic radiation. using radiation pressure from a pulsed laser eters held constant. Optical measurements of
source. This minimally invasive pulsed mea- the acoustic properties of the OMC cavities are
To date, far less attention has been paid to the surement technique avoids a slew of parasitic performed at millikelvin temperatures in a di-
damping effects—typically associated with elec- lution refrigerator, with a lensed optical fiber
impact of geometry and phononic bandgaps trode materials and mechanical contact (15), used to couple light into and out of each device
or probe fields for continuous readout—and (16). In a first set of measurements of acoustic
on acoustic material absorption and related allows for the sensitive measurement of mo- energy damping, we use a single pulsed laser
noise effects (10, 11). Fundamental limits to tion at the single-phonon level (16). The results scheme to perform both excitation and read-
sound absorption in solids are known to result of acoustic ringdown measurements at milli- out of the breathing mode. In this scenario,
kelvin temperatures show that damping due depicted in Fig. 2, A and B, the laser fre-
from the anharmonicity of the host crystal to radiation is effectively suppressed by the quency ( wl ) is tuned to the red motional
lattice (12, 13). At low temperatures T, in the sideband of the OMC cavity optical resonance,
Landau-Rumer regime (wtth ≫ 1), where the D ≡ wc À wl ≈ þ wm, and is pulsed on for a du-
thermal phonon relaxation rate (tÀth1) is much ration Tpulse and then off for a variable time t.
This produces a periodic train of photon pulses
1Kavli Nanoscience Institute, California Institute of
Technology, Pasadena, CA 91125, USA. 2Institute for Quantum
Information and Matter and Thomas J. Watson, Sr., Laboratory
of Applied Physics, California Institute of Technology, Pasadena,
CA 91125, USA. 3AWS Center for Quantum Computing,
Pasadena, CA 91125, USA.
*These authors contributed equally to this work.
†Present address: Advanced Quantum Testbed, National Energy
Research Scientific Computing Center, Lawrence Berkeley National
Laboratory, Berkeley, CA 94720, USA. ‡Present Address: Booz Allen
Hamilton Inc, McLean, VA 22102, USA. §Present Address: Department
of Physics, Harvard University, Cambridge, MA 02138, USA.
#Corresponding author. Email: [email protected]

MacCabe et al., Science 370, 840–843 (2020) 13 November 2020 1 of 4

RESEARCH | REPORT D F z-even
z-odd
A coupling waveguide 10
9 XMΓ
nanobeam OMC 2 μm acoustic frequency (GHz) 8
7
acoustic shielding E 6
5
B coupling waveguide r1 4
hc 3
C r 2
2 1
0
hh wc
wh Γ

Fig. 1. Nanobeam optomechanical crystal and phononic shield design. displacement profile is exaggerated for clarity. (D) SEM image showing the
(A) Scanning electron microscope (SEM) image of a full nanobeam optomechanical nanobeam clamping geometry. (E) SEM image of an individual unit cell of the
crystal (OMC) device fabricated on SOI with N ¼ 7 periods of acoustic shielding. cross-crystal acoustic shield. The dashed lines show fitted geometric parameters
A central coupling waveguide allows for fiber-to-chip optical coupling as well as used in simulation, including cross height (hc ¼ 474 nm), cross width (wc ¼ 164 nm),
side-coupling to individual nanobeam OMC cavities. (B) SEM image of an individual inner fillet radius (r1), and outer fillet radius (r2). (F) Simulated acoustic band
nanobeam OMC and the coupling waveguide, with enlarged illustration of an structure of the realized cross-crystal shield unit cell, with the full acoustic
individual unit cell in the end-mirror portion of the nanobeam. (C) FEM simulations bandgap highlighted in pink. Solid (dotted) lines correspond to modes of even
of the mechanical (top; total displacement) and optical (bottom; transverse (odd) symmetry in the direction normal to the plane of the unit cell. The dashed red
electric field) modes of interest in the nanobeam. Distortion of the mechanical line indicates the mechanical breathing-mode frequency at wm=2p ¼ 5:0 GHz.

AB C

D EF

Fig. 2. Ringdown measurements of the acoustic breathing mode. corresponds to the range of simulated Q values (ensemble size 10) within
(A) Illustration showing the red-detuned (D ¼ þwm) optical drive and sideband 1 SD of the mean. The square purple data point represents the measured Q
optical filtering used to realize effective phonon counting measurement. in (F). (E) Acoustic excitation is performed coherently by using either a
(B) Illustration of the ringdown measurement performed using a single pulsed blue-detuned pump (upper diagram) to drive the breathing mode into
laser for both excitation and readout. Top (bottom): Intracavity photons (phonons) self-oscillation, or using a radiofrequency-modulated red-detuned pump
versus time. (C) Ringdown measurements of a seven-shield device (device C) (18) (lower diagram). See materials and methods, section V, for details
for readout pulse amplitude of nc ¼ 320. The series of inset panels show the of the coherent excitation and readout parameters. (F) Ringdown measure-
measured (and fit; solid blue curve) anti-Stokes signal during the optical pulse at ments performed on an eight-shield device (device D) at large phonon
a series of pulse delays. (D) Plot of the measured breathing mode Q-factor versus amplitude. For blue-detuned driving (red squares), the fit decay rate is
number of acoustic shield periods Nshield. The solid green line is a fit to the g0=2p ¼ ð0:122 T 0:020Þ Hz. For modulated-pump driving (purple circles),
corresponding simulated radiation-limited Q-factor for devices with standard the fit decay rate is g0=2p ¼ ð0:108 T 0:006Þ Hz. The error bars are 90%
deviation (SD) s ¼ 4 nm disorder in hole position and size, similar to the value confidence intervals of the measured values of nmi . The shaded regions are
measured from device SEM image analysis. The shaded green region the 90% confidence intervals for the exponential fit curves.

MacCabe et al., Science 370, 840–843 (2020) 13 November 2020 2 of 4

RESEARCH | REPORT

A B 2.5 measured temperature scaling of the damp-
ing. Analysis of the interactions of TLS in
102 108 2.0 the surface of the silicon with the localized
γ0/2π (Hz) acoustic modes in the confined geometry of
Qm the OMC cavity sructure, however, show that
TLS interactions can explain all of the ob-
Count rate (×105 c.p.s.) served breathing-mode behavior when tak-
101 109 1.5 ing into account the highly modified density
of phonon states in the shielded OMC cavity
100 10-1 1010 1.0 (see supplementary text, sections VII and
Tf (K) 100 VIII). In particular, we find that damping due
10-1 1.4 to nearly resonant TLS is suppressed owing
10-2 1.2 Δ1/2 to the bandgap of the phononic shield and
1.0 -8 -4 0 4 8 that relaxation damping from nonresonant
TLS can explain the magnitude and low-
f - f0 (kHz) temperature dependence of the breathing-
mode damping and the lack of saturation
Fig. 3. Temperature dependence of acoustic damping and frequency jitter. (A) Plot of the measured of the damping with both temperature and
acoustic amplitude.
breathing-mode energy damping rate, g0=2p, as a function of fridge temperature (Tf). Dashed green
(magenta) curve is a fit with temperature dependence g0 $ Tf1:01 (g0 $ Tf2:39). Error bars are 90% confidence In addition to energy damping, the co-
intervals of the exponential fit to measured ringdown curves. (B) Two-tone coherent spectroscopy signal. herence of the high- Q breathing modes is
also of interest. Using the same pump-probe
Upper plot: Three individual spectra of rapid frequency sweeps with a frequency step size of 500 Hz and dwell excitation method (18) as in the ringdown
measurements of Fig. 2, E and F, we can mea-
time of 1 ms (resolution bandwidth RBW ≈0:5 kHz). Lower plot: Average spectrum of rapid-scan spectra taken sure the breathing-mode acoustic spectrum by
over minutes, showing broadened acoustic response with FWHM linewidth of D1=2=2p ¼ 4:05 kHz. The large sweeping the probe tone across the acoustic
on-resonance response corresponds to an estimated optomechanical cooperativity of C gOM=ðg0 þ gpÞ ≳ 1:1, resonance. Figure 3B shows the measured
consistent with the predicted magnitude of back-action damping gOM=2p ≈ 817 Hz and bath-induced damping acoustic spectrum of device D at the fridge base
gp=2p ≈ 120 Hz at the measurement pump power level nc ¼ 0:1. Data presented in (A) and (B) are for device D. temperature Tf ¼ 7 mK and a low pump power
of nc ¼ 0:1. The top plot shows a series of rapid
due to anti-Stokes scattering of the probe laser mode using a Â1000 weaker excitation and spectral scans (40 ms per scan), which captures
that are on-resonance with the optical cavity. readout optical pulse amplitude (nc ≲ 0:3). The a jittering narrow acoustic line. Scanning more
The anti-Stokes scattered photons are filtered measured ringdown curves, displayed in Fig. slowly than the 40-ms per scan begins to in-
from the probe laser and sent to a single pho- troduce broadening into the single-scan spec-
ton detector, producing a photon count rate 2F, show the decay from initial phonon occu- tral line, providing a rough estimate for the
proportional to the number of phonons in the pancies of 103 to 104 of an eight-shield device (inverse) bandwidth of the frequency jitter
acoustic resonator (see materials and meth- noise of the breathing mode. An ensemble av-
ods, section III). The resulting intracavity pho- (device D; square purple data point in Fig. 2D). erage of these scans, taken over several min-
non pulse shape, illustrated in Fig. 2B, has a utes, yields a broadened acoustic full width at
shape determined by a competition between The two methods yield similar breathing-mode half-maximum (FWHM) linewidth of D1=2=2p ¼
optical absorption–induced heating and opto- energy decay rates of g0=2p ¼ 0:108 Hz and 4:05 kHz (bottom plot of Fig. 3B). This be-
mechanical back-action cooling (16). The net 0:122 Hz, the smaller of which corresponds to a havior is characteristic of coherent-like fre-
result is that the acoustic mode is excited to a Q-value of Qm ¼ 4:92Àþ00::3269 Â 1010 and a pho- quency noise, such as flicker or low-frequency
small thermal population at the end of the non lifetime of tph;0 ¼ 1:47Àþ00::0089 s. Comparing telegraph noise. In this case, the coherence time
readout pulse, with free decay of the phonon all three excitation methods with widely vary- is given by tcoh;0 ≈ 4ðlnð2ÞÞ1=2=D1=2 ≈ 130 ms
number amplitude in between pulses (see ing optical absorption–heating and phonon (19, 20). The likely source of this noise is again
materials and methods, section IV). Plotting amplitude, we consistently measure Qm ≳ 1010 fluctuating TLS (see supplementary text, sec-
the initial mode occupancy at the beginning for devices with Nshield ≥ 5. This represents an tion VIII).
of the readout pulse (nmi ) versus delay time t improvement by more than three orders of
between pulses results in a ringdown curve Several subtle points regarding the energy
of the stored phonon number in the breath- magnitude compared to previous measure- decay and coherence measurements should be
ing mode, an example of which is displayed noted. First, the breathing-mode spectral line-
in Fig. 2C for a device with Nshield ¼ 7. ments of devices with nonoptimized acoustic width (and thus coherence time) is measured
shielding (16). in the presence of a weak laser pump field, in
Performing a series of ringdown measure- contrast to energy decay measurements, in
ments over a range of devices with varying To understand the origin of the residual which the laser is pulsed off. Although we did
Nshield, and fitting an exponential decay curve damping for large Nshieldwe also measured the not observe a power dependence of the mea-
to each ringdown, we produce the Q -factor temperature dependence of the energy damp- sured time-average acoustic linewidth, the in-
plot in Fig. 2D. We observe an initial trend in ing rate for a number of the high- Q , large fluence of the laser field on the acoustic mode
Q -factor versus shield number that rises on Nshield devices. In Fig. 3A, we plot the energy coherence time cannot be ruled out. Second,
average exponentially with each additional damping rate versus fridge temperature (Tf ) the pump-probe technique used to measure
shield period (Â5:5 per period, in good corre- for the highest-Q eight-shield device (device the acoustic spectrum excites the breathing
spondence with the FEM modeling) and then D), typical of all of these types of devices that mode to an estimated mode occupancy of order
saturates for Nshield ≥ 5 to Qm ≳ 1010. As indi- 1000; of relevance to many quantum applica-
cated in Fig. 2C, these Q values correspond to we measured, which shows an approximately tions is the single-phonon coherence time.
ringdown of small, near-single-phonon–level linear rise in temperature up to Tf ≈ 100 mK
amplitudes. We also perform ringdown mea- and then a much faster ∼ ðTf Þ2:4 rise in the
surements at high phonon amplitude using damping above this temperature. This rather
two additional methods displayed schemat-
ically in Fig. 2E (see materials and methods, anomalous temperature scaling of the damp-
section V). These methods use two laser tones
to selectively excite the acoustic breathing ing does not fit the standard damping models

relating to three-phonon scattering, and esti-

mates of the magnitude of Landau-Rumer

damping of the breathing mode indicate that

three-phonon scattering in Si is far too weak
at Tf ≲ 1 K to explain the absolute level of
measured damping (see materials and meth-

ods, section VI). Canonical models of TLS-

induced damping also fail to predict the

MacCabe et al., Science 370, 840–843 (2020) 13 November 2020 3 of 4

RESEARCH | REPORT

With the recent demonstration of strong dis- small scale, reduced cross-talk, and ultralong 28. M. S. Hanay et al., Nat. Nanotechnol. 10, 339–344
persive coupling of a superconducting qubit to (2015).
similar nanomechanical acoustic cavities (21), coherence time of quantum acoustic devices
measurement of single-phonon coherence time 29. S. Nimmrichter, K. Hornberger, K. Hammerer, Phys. Rev. Lett.
in the absence of light fields should be possible. may provide substantial improvements in con- 113, 020405 (2014).
Additionally, the low-frequency character of
the frequency noise measured in this work nectivity and performance of current quan- 30. M. H. Devoret, R. J. Schoelkopf, Science 339, 1169–1174
indicates that the acoustic coherence time (2013).
may be substantially increased toward the en- tum hardware.
ergy decay time through techniques such as 31. G. S. MacCabe, H. Ren, J. Luo, J. D. Cohen, H. Zhou,
dynamic decoupling (22). Finally, the attribu- REFERENCES AND NOTES A. Sipahigil, M. Mirhosseini, O. Painter, Data for “Nano-acoustic
tion of the residual damping and the frequen- resonator with ultralong phonon lifetime,” Version 1, Zenodo
cy jitter noise to surface TLS indicates that 1. J. D. Joannopoulos, S. G. Johnson, J. N. Winn, R. D. Meade, (2020); http://doi.org/.
further study of the silicon surface compo- Photonic Crystals: Molding the Flow of Light (Princeton Univ.
sition may lead to a better understanding of Press, ed. 2nd, 2008). ACKNOWLEDGMENTS
the microscopic nature of the TLS and help
identify further surface preparation methods 2. S. John, J. Wang, Phys. Rev. B Condens. Matter 43, Funding: This work was supported by the ARO Quantum
for their removal (see materials and methods, 12772–12789 (1991). Opto-Mechanics with Atoms and Nanostructured Diamond
section I, and supplementary text, section VIII). MURI program (grant N00014-15-1-2761), the ARO-LPS Cross-
3. M. Maldovan, Nature 503, 209–217 (2013). Quantum Systems Science and Technology program (grant
More generally, the results presented here 4. J. Cha, K. W. Kim, C. Daraio, Nature 564, 229–233 W911NF-18-1-0103), the Institute for Quantum Information and
show that the advanced methods of nano- Matter, an NSF Physics Frontiers Center (grant PHY-1733907)
fabrication and cavity optomechanics provide (2018). with support of the Gordon and Betty Moore Foundation, and
a new toolkit to explore quantum acoustody- 5. J. Chan, A. H. Safavi-Naeini, J. T. Hill, S. Meenehan, O. Painter, the Kavli Nanoscience Institute at Caltech. H.R. gratefully
namics in solid-state materials (23). For exam- acknowledges support from the National Science Scholarship
ple, continued studies of phononic engineered Appl. Phys. Lett. 101, 081115 (2012). from A*STAR, Singapore. Author contributions: G.S.M., H.R.,
structures in the quantum regime may lead 6. P.-L. Yu et al., Appl. Phys. Lett. 104, 023510 (2014). J.D.C., and O.P. came up with the concept and planned the
to techniques for modifying the behavior of 7. Y. Tsaturyan, A. Barg, E. S. Polzik, A. Schliesser, Nat. Nanotechnol. experiment. G.S.M., H.R., J.L., H.Z., and J.D.C. performed
quasiparticles in superconductors (24), new the device design and fabrication. G.S.M. and H.R. performed
approaches for mitigating decoherence (noise) 12, 776–783 (2017). the measurements. G.S.M., H.R., J.L., A.S., M.M., and O.P.
in superconducting (25) and color center (26) 8. A. H. Ghadimi et al., Science 360, 764–768 (2018). analyzed the data. All authors contributed to the writing of
qubits, and even new TLS–acoustic cavity qubit 9. M. Kalaee et al., Nat. Nanotechnol. 14, 334–339 (2019). the manuscript. Competing interests: G.S.M, H.R., J.L., A.S.,
architectures (27). The extremely small mo- 10. R. O. Behunin, F. Intravaia, P. T. Rakich, Phys. Rev. B 93, M.M., and O.P. acknowledge two related patent applications that
tional mass [meff ¼ 136 fg (5)] and narrow draw on the work presented in this article: U.S. Patent
linewidth of the OMC cavity also make it ideal 224110 (2016). Application no. 16/293,457, “Techniques for Bidirectional
for precision mass sensing (28) and in explor- 11. B. D. Hauer, P. H. Kim, C. Doolin, F. Souris, J. P. Davis, Phys. Rev. B Transduction of Quantum Level Signals Between Optical and
ing limits to alternative quantum collapse Microwave Frequencies Using a Common Acoustic
models (29). Perhaps most intriguing is the 98, 214303 (2018). Intermediary”; U.S. Patent Application no. 16/293,455,
possibility of creating a hybrid quantum archi- 12. L. Landau, G. Rumer, Phys. Z. Sowjet. 11, 18 (1937). “Techniques for Transduction and Storage of Quantum Level
tecture consisting of acoustic and supercon- 13. G. P. Srivastava, The Physics of Phonons (Taylor & Francis, Signals”. J.L. is also affiliated with Anyon Computing Inc.,
ducting quantum circuits (21, 30), where the Pasadena, CA 91101, USA. Data and materials availability:
1990). Data and data analysis code are available through Zenodo (31).
14. W. A. Phillips, Rep. Prog. Phys. 50, 1657–1708 (1987). All other data that support the findings of this study are in
15. S. Galliou et al., Sci. Rep. 3, 2132 (2013). the main text and supplementary materials.
16. S. M. Meenehan et al., Phys. Rev. X 5, 041002 (2015).
17. See supplementary materials. SUPPLEMENTARY MATERIALS
18. A. H. Safavi-Naeini et al., Nature 472, 69–73 (2011).
19. G. Di Domenico, S. Schilt, P. Thomann, Appl. Opt. 49, science.sciencemag.org/content/370/6518/840/suppl/DC1
Materials and Methods
4801–4807 (2010). Supplementary Text
20. P. J. J. O’Malley, thesis, University of California, Santa Barbara Figs. S1 to S19
Tables S1 to S4
(2016). References (32–56)
21. P. Arrangoiz-Arriola et al., Phys. Rev. X 8, 031007 (2018).
22. J. Bylander et al., Nat. Phys. 7, 565–570 (2011). 10 May 2020; accepted 5 October 2020
23. W. H. Renninger, P. Kharel, R. O. Behunin, P. T. Rakich, 10.1126/science.abc7312

Nat. Phys. 14, 601–607 (2018).
24. K. Rostem, P. J. de Visser, E. J. Wollack, Phys. Rev. B 98,

014522 (2018).
25. J. M. Martinis, A. Megrant, UCSB final report for the CSQ

program: Review of decoherence and materials physics
for superconducting qubits. arXiv:1410.5793 (2014).
26. Y.-I. Sohn et al., Nat. Commun. 9, 2012 (2018).
27. T. Ramos, V. Sudhir, K. Stannigel, P. Zoller, T. J. Kippenberg,
Phys. Rev. Lett. 110, 193602 (2013).

MacCabe et al., Science 370, 840–843 (2020) 13 November 2020 4 of 4

RESEARCH

NEUROSCIENCE surround the first-order auditory thalamus
[ventral division of the MG (MGv)] like a shell.
A thalamocortical top-down circuit To selectively target protein expression to the
for associative memory HO-MG while avoiding the MGv, we used con-
ditional adeno-associated viral vector (AAV)–
M. Belén Pardi1, Johanna Vogenstahl1, Tamas Dalmay1,2, Teresa Spanò1,3, De-Lin Pu1, mediated expression in CR-IRES-Cre mice (9)
Laura B. Naumann4,5, Friedrich Kretschmer1, Henning Sprekeler4,5, Johannes J. Letzkus1,6* (Fig. 1, A and B, and fig. S1; for full results of all
statistical tests in this study, see supplementary
The sensory neocortex is a critical substrate for memory. Despite its strong connection with the text section of the supplementary materials).
thalamus, the role of direct thalamocortical communication in memory remains elusive. We performed Anterograde tracing revealed that HO-MG
chronic in vivo two-photon calcium imaging of thalamic synapses in mouse auditory cortex layer 1, a axons display a mediotemporal density gradi-
major locus of cortical associations. Combined with optogenetics, viral tracing, whole-cell recording, ent that closely matches the location of cor-
and computational modeling, we find that the higher-order thalamus is required for associative learning tical areas critical for associative memory (8),
and transmits memory-related information that closely correlates with acquired behavioral relevance. In turn, with the highest density in temporal cortical
these signals are tightly and dynamically controlled by local presynaptic inhibition. Our results not only areas, including the secondary auditory cortex
identify the higher-order thalamus as a highly plastic source of cortical top-down information but also reveal (AuV) and the temporal association cortex (Fig.
a level of computational flexibility in layer 1 that goes far beyond hard-wired connectivity. 1C). We therefore focused on the most temporal
area of the auditory cortex (i.e., AuV), where axons
A ccurate perception of the world requires inent yet little understood projection to L1 were strongly enriched in upper L1 (Fig. 1C).
integration of external sensory signals arises from the higher-order thalamus (4–6).
with internally generated information We hypothesized that these afferents consti- To identify the postsynaptic targets of HO-MG
representing the previously acquired tute a source of experience-dependent top-down afferents in L1, we used channelrhodopsin-2
relevance of stimuli (1, 2). Whereas de- information, distinct from the classical bottom- (ChR2)–assisted circuit mapping in acute brain
cades of work have elucidated how the sensory up encoding that has been widely observed in slices (10) while recording from L1 interneu-
neocortex processes physical stimulus fea- first-order thalamic nuclei (7). rons (L1 INs) and pyramidal neurons (Pyrs) in
tures, encoding of memory-related top-down L2/3 and L5 (Fig. 1, D to F). Local light stim-
information by brain-wide afferents remains To investigate the function of thalamocor- ulation in L1 under action potential block (fig.
poorly understood. Neocortical layer 1 (L1) re- tical communication in memory, we focused S2, A and B) excited the three neuron types
ceives a range of top-down inputs from diverse on afferents to the auditory cortex, a critical with similar amplitudes (Fig. 1, E and F), albeit
sources (1, 3) and could therefore be a key area for associative memory (8). The higher- with different probability (fig. S2C). In line
node for integration of these signals. One prom- order auditory thalamus [HO–medial genicu- with synaptic contacts residing on distal den-
late (HO-MG)] comprises several nuclei that drites of Pyrs, excitatory currents recorded in
L5 displayed longer latencies (Fig. 1F). These
results indicate that HO-MG L1 input provides

Fig. 1. The higher-order auditory thalamus pro- Genetic targeting GCaMP6s HO-MG 200 µm GCaMP6s Intensity (a.u.)
vides major top-down input to cortical layer 1. AAV CAG MGd
(A) Experimental strategy for targeting of HO-MG in SG 0
calretinin (CR)–IRES-Cre mice. (B and C) Immuno- flex-GCaMP6s L1
staining against GCaMP6s and neuronal marker
NeuN. (B) Auditory thalamus, posterior intralaminar Higher-order auditory 100
nucleus (PIN) detail, and quantification of NeuN+ thalamus (HO-MG)
cells expressing GCaMP6s (Wilcoxon test, n = 10 slices, CR-IRES-Cre mice AuD 200
5 mice). (C) HO-MG axons in cortex and mean A1
fluorescence intensity (arbitrary units, a.u.) versus GCaMP6s+ (%) 80 ** 300
distance from pia (in micrometers, n = 3 slices, AuV
3 mice). (D) ChR2-assisted mapping of connections MGm MGv 60 TeA AuV 400
between HO-MG axons in AuV and L1 INs and Ect 500
L2/3 and L5 Pyrs in acute slices. Light stimulation 40
(5 ms) was restricted to L1. (E) Recorded cells NeuN NeuN 600
and ChR2+ HO-MG axons in L1 (pseudocolored; scale 20
bars, 20 mm), and light-evoked excitatory post- PIN 700
PP 0 100 µm 800
HO MGv
MG

Output mapping HO-MG L1 IN L1 IN
terminals L2/3 Pyr
AAV CAG-flex Blue light on layer 1 L5 Pyr **
ChR2-mCherry n.s.

n.s.
L1 IN L2/3
AuV HO-MG Pyr
ChR2 L2/3
Pyr L5
Pyr
CR-IRES-Cre mice L5 Pyr 40 pA
20 ms 0 200 400 0 20 40 60 80
Amplitude (pA) Latency (ms)

synaptic currents (gray, single trials; black, Input mapping DAPI mCherry eGFP Midbrain Ipsi. Cortex Thalamus Other
mean; blue bar, light). (F) Current amplitude and Presynaptic cells (log %) Contra.
latency to peak (quartiles, median and range, AAV synP-DIO RV 00.1 0.1 1 10 100
Kruskal-Wallis test with Dunn’s multiple comparisons sTpEpB-eGFP mCherry 200 µm
test for latency, n = 9 L1 INs, n = 12 L2/3 Pyrs, SC
n = 11 L5 Pyrs). (G) Transsynaptic rabies (RV) Presynaptic IC
tracing of inputs to HO-MG neurons that project to cells MRN
PAG
AuV M
6 w. NB
SNc
Starter APN
cells A1
TeA
CR-IRES-Cre mice AuV
AuD
AuV. Glycoprotein, EnvA receptor (TVA), and PTLp
Ect
VIS
AUDpo
SSp-bfd
RSP
MG
PIN
SG
RT
POL
PP
VPM
LGv
LP
SOC
GPe
ZI

enhanced green fluorescent protein (eGFP) were conditionally expressed in the HO-MG. RV-mCherry was injected into the auditory cortex 6 weeks later, leading to retrograde

transduction of starter cells. Right: Representative expression in the auditory thalamus [4′,6-diamidino-2-phenylindole (DAPI) counterstain]. Inset from PIN: eGFP,

mCherry, and overlay; arrowheads indicate starter cells. (H) Distribution of presynaptic cells ipsilateral and contralateral to injection, normalized per animal (logarithmic

scale, n = 3 mice). Abbreviations used in (B), (C), and (H) are defined in table S1. Data shown as mean ± SEM, except in (F). n.s. P > 0.05, **P < 0.01.

Pardi et al., Science 370, 844–848 (2020) 13 November 2020 1 of 5

RESEARCH | REPORT

widespread control over cortical activity that analyzed (Fig. 2C and fig. S4), enabling an un- with most of the response occurring after each
can affect circuit processing in diverse ways: biased, longitudinal characterization. FM sweep (Fig. 2, D and G)—very different from
Recruitment of L1 INs can generate disinhibi- the short-latency responses typically found in
tion of lower layers that is particularly relevant CS+ stimulation during habituation in awake first-order MGv (7).
for associative learning (11), and it can also head-fixed mice elicited clear responses in a
directly inhibit distal dendrites (3). Further- subset of boutons (Fig. 2, D and E). However, In contrast, CS+ responses in memory recall
more, excitation of Pyr distal dendrites can HO-MG synaptic responses decayed rapidly to
trigger dendritic spikes and facilitate long- baseline after a few trials (Fig. 2F), indicating displayed diverse learning-related plastic changes
term potentiation (4, 12). that stimulus encoding by HO-MG afferents is
highly sensitive to novelty. Another distinctive leading to a much stronger activation of HO-MG
These observations are consistent with the feature was their markedly delayed kinetics, boutons. First, while CS+ responses of individ-
hypothesized role of the HO-MG as a major top-
down projection. Top-down inputs are pre- ual boutons either increased or decreased, the
dicted to convey feedback information from
diverse sources. We therefore identified the total number of responsive boutons grew by
brain-wide inputs specifically to auditory cortex–
projecting HO-MG neurons, using dual-site- AAV CAG-flex Habituation Conditioning Recall 1 Recall 2 Freezing (%) 100 ****
injection Cre-dependent transsynaptic retrograde GCaMP6s Chronic Day 1 Day 2 Day 3 Day 4 80 **
rabies tracing (13, 14) (Fig. 1G). This experiment window CS 60
revealed diverse input sources (Fig. 1H and CS CS CS 40 *
fig. S3), including cortical areas, intermediate AuV US
and deep layers of the superior colliculus, ex- Pupil
ternal capsule of the inferior colliculus, and the track
periaqueductal gray. The HO-MG thus trans-
mits integrated information to cortical L1, in CR-IRES-Cre mice CS+ 20
line with previous work on input to the entire CS-
HO-MG (7, 15). 0
Baseline CS- CS+
To directly test whether HO-MG afferents
convey memory-related information to the cor- Habituation Recall 1 CS+ in habituation ID CS+ in recall ID
tex, we used discriminative threat conditioning 0.7
(DTC) to endow auditory stimuli with experience- Boutons 50 0.6 DeltaF/F
dependent relevance in a controlled manner 100 0.5
(16). Conditioning was performed by repeatedly 150 0.4
pairing one of two initially neutral conditioned 200 0.3
stimuli (CSs) with mild foot-shocks. We used 250 0.2
trains of frequency-modulated sweeps as CSs 0.1
[trains are termed CSs, individual sweeps are 300 5 10 0 5 10 0
termed frequency-modulated (FM) sweeps], nat- 0 Time (s) Time (s)
uralistic stimuli that critically require cortical 20 z-score=1.96 Habituation n.s. 5% ∆F/F
processing (8, 17). We combined this approach CS+ 0.6 *** Habituation
with two-photon calcium imaging of HO-MG 0.5 Recall 2s
boutons expressing GCaMP6s in L1 of awake, Recall (z-score) Responsive Response (∆F/F*s)
head-fixed mice after identification of AuV by 15 Non-responsive 0.4 CS+
intrinsic imaging (Fig. 2A and fig. S4) (3, 8).
We imaged boutons in response to the rein- 10 CS+ 0.3
forced CS+ and the unreinforced CS− 1 day be-
fore (habituation) and 1 day after conditioning 5 Recall 0.2 ****
(recall 1, Fig. 2A). CSs elicited different levels of
threat memory, measured as the change in eye 0 100% 0.1
pupil dilation during head fixation (from habi- 0 5 10 15 20 ∆F/F
tuation to recall), and as freezing in freely be- Habituation (z-score) 5s 0.0
having conditions (recall 2, Fig. 2, A and B, and
fig. S5, A to C), readouts that displayed a robust -0.1
positive correlation (fig. S5D) (3, 8). Only bou-
tons identified in habituation and recall were 1234
Trials (binned/2)
1Max Planck Institute for Brain Research, 60438 Frankfurt,
Germany. 2Donders Centre for Neuroscience, Faculty of 0.5 n.s. **** Pupil dilation change (∆D/D*s) R2= 0.56 Habituation 4 r = 0.97 0.4
Science, Radboud University, 6525 AJ Nijmegen, Netherlands. 4 p< 0.0001 Bouton 1 ****
3Faculty of Biological Sciences, Goethe Universität Frankfurt, Response (∆F/F*s) 2 Noise correlation (r) ****
60438 Frankfurt, Germany. 4Bernstein Center for Computational 2
Neuroscience Berlin, 10115 Berlin, Germany. 5Department of 0.4 0 0.3 ****
Electrical Engineering and Computer Science, Technische 0 n.s.
Universität Berlin, 10587 Berlin, Germany. 6Institute for -2
Physiology I, Faculty of Medicine, University of Freiburg, 0.3 -2 -1 0 1
79104 Freiburg, Germany. -1 0 1 2
*Corresponding author. Email: [email protected] 2 r = -0.09 0.2
Bouton response change (∆F/F*s) 1
0.2 0
-1
0.1 Recall -2 0.1
Bouton 1 -1 0 1
0.0 0.0
Hab. Rec. Hab. Rec. Bouton 2 Hab. Rec. Hab. Rec.
CS- CS+
CS- CS+

Fig. 2. Experience-dependent signaling by HO-MG afferents to cortical L1. (A) Experimental strategy.

US, unconditioned stimulus. (B) Freezing behavior of learner mice in recall 2 [imaging of these animals

is shown in (D) to (H) and (J); one-way repeated measures analysis of variance (RM ANOVA) with Tukey’s
multiple comparisons test, n = 10 mice]. (C) Representative boutons in habituation and recall (scale bars,

20 mm). (D) Mean bouton responses sorted by amplitude, and gray scale rank order bouton ID [vertical
lines indicate FM-sweep onset; in (D) to (H), n = 302 boutons, 10 mice]. (E) Peak z-score of individual bouton
CS+ responses (z-scores > 1.96 are considered responsive). Right: Representative bouton indicated in (C)

(gray, single trials; blue, mean) and fraction of responsive boutons (habituation, 73/302; recall, 129/302). DF/F,
change in fluorescence intensity. (F) Response integral during CS+ over trials (every two binned) in habituation

versus recall (two-way RM ANOVA, P < 0.0001 for session, with Sidak’s multiple comparisons test between
trials). (G) Mean ± SEM bouton response. (H) Response integral during CSs (Friedman test with Dunn’s
multiple comparisons test). (I) Bouton versus pupil response change between habituation and recall in
conditioned mice [Pearson correlation, linear fit and 95% confidence bands, n = 22 mean responses from CS+
blue and CS− green, eight learner mice (dots) and three nonlearner mice (triangles)]. (J) Left: Trial-by-

trial noise correlation between response integrals in a pair of boutons (DF/F*s, Pearson correlation, linear fit,
n = 8 single presentations). Right: Noise correlation between CS responsive and nonresponsive bouton
pairs in learner mice (Kruskal-Wallis test with Dunn’s multiple comparisons test, CS+ habituation n = 2212,
CS+ recall n = 3130, CS− habituation n = 2989, CS− recall n = 3230 correlations). Data shown as mean ± SEM.

n.s. P > 0.05, *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001.

Pardi et al., Science 370, 844–848 (2020) 13 November 2020 2 of 5

RESEARCH | REPORT

77% after learning (Fig. 2, D and E). Second, be predictive of—memory strength (Fig. 2I). A second form of salience that is indepen-
CS+ representation was much more persistent Analyses from each session separately indi- dent of experience derives from the physical
cated a significant correlation between absolute properties of sensory stimuli, such as sound
over trials than in habituation (Fig. 2F). Third, bouton and pupil responses only in recall (fig. intensity (19). To address whether HO-MG af-
CS+ responses became more time-locked to S5E), further underpinning the impact of as- ferents also encode this bottom-up salience,
sociative memory on HO-MG responses. we imaged boutons in response to noise-train
the onset of the CS (Fig. 2, D and G, and fig. stimuli at different intensities (Fig. 3, A and
Finally, we analyzed activity correlations be- B). In contrast to both the changes observed
S6, A to C) and preceded behavioral memory tween boutons, which can limit the content after learning and the typical encoding in first-
of information transfer. Trial-by-trial noise order thalamic afferents (20), bouton responses
expression (fig. S6, D to F). Fourth, the over- correlations between all pairs of boutons dis- during the noise-train decreased with increas-
all CS+ response was strongly potentiated with played a decrease in memory recall relative to ing stimulus intensity (Fig. 3B). This effect was
habituation (fig. S7F). This was mediated spe- generated by a successive increase in the latency
learning (Fig. 2, G and H, and fig. S7), mainly cifically by a decrease in the correlation between of responses, whereas activity still occurred after
CS responsive and nonresponsive boutons (Fig. noise-train offset (fig. S10), indicating a dis-
during the CS but also, to a lesser extent, after 2J and fig. S8D), while correlations among re- engagement of this pathway during stimuli
sponsive and among nonresponsive boutons with increasing bottom-up salience.
CS offset (fig. S7D). were maintained (fig. S7F).
We next tested the requirement of HO-MG
Grouping animals according to the degree Collectively, these data demonstrate that the signaling for associative memory. Synaptic si-
HO-MG conveys information about the behav- lencing of HO-MG neurons with tetanus toxin
of behavioral CS discrimination revealed that ioral relevance of sounds to auditory cortex L1, (TeNT) (3) left basic auditory processing ad-
boutons from mice with low CS− freezing dis- which depends on previous experience both in dressed by startle responses intact (Fig. 3, C
played no response potentiation to the CS−. the short term, during habituation, and in and D) but yielded an impairment of memory
Conversely, potentiation of CS− responses oc- the long term, after associative conditioning. expression to the CS+ along with deficits in
curred in animals with strong CS− threat mem- Whereas remodeling of thalamocortical sys- discrimination after DTC (Fig. 3E, right, and
tems has been described in development and fig. S11). Conversely, behavior during acqui-
ory (fig. S8, A and B). Across all mice, the after sensory deprivation (18), our data high- sition and contextual threat recall was not
light the central role of plasticity in higher-order affected (Fig. 3E, left, and fig. S11, D and F),
observed changes were highly significant for thalamic afferents for associative memory in indicating that the HO-MG is critically re-
the CS+ but not for the CS− (Fig. 2H, figs. S6B the adult. quired for associative auditory memory but
dispensable for foot-shock perception and
and S7, A to D) and were absent in animals nonauditory memory.

that did not learn the CS association (fig. S9). Whereas top-down information is typically
considered to be independent of activity in the
To directly address the relationship between local circuit (1, 2), we hypothesized that these
afferents could be under the functional control
plasticity and memory strength, we measured of a specialized type of neurogliaform IN that
densely innervates L1 and expresses the selec-
the change in pupil dilation from habituation tive marker NDNF (neuron-derived neuro-
trophic factor) (3). Because these neurogliaform
to recall (fig. S5, B to D). Analysis of bouton cells can perform g-aminobutyric acid–mediated
(GABAergic) volume transmission (21, 22), they
response changes per mouse revealed that could control the release probability of HO-MG
boutons through presynaptic GABA type B
interindividual variability of potentiation mag- receptor (GABABR)–mediated inhibition (23)
nitude strongly correlated with—and could thus (Fig. 4A). Such a mechanism would add a layer
of complexity to neuronal processing by en-
60 dB 0.8 0.15 **** 4 abling dynamic and target area–specific con-
1 **** AAV DIO-Tetanus toxin (TeNT) or trol of top-down afferent information.

350 AAV DIO-EYFP (Control) We first asked whether HO-MG boutons in
105 dB 0.10 the cortex can be inhibited through GABABRs.
Recordings in acute brain slices revealed that
1 3 the GABABR agonist baclofen reduced opto-
White DeltaF/F Latency to peak (s) genetically evoked postsynaptic current am-
noise Response (ΔF/F*s) plitude and abolished short-term depression
(Fig. 4, B and C, and fig. S12, A and B), con-
Bouton ID 0.4 0.05 sistent with modulation of release probability
by presynaptic GABABRs. Both effects were re-
2 versed by application of the GABABR antagonist
CGP55845. When axons were stimulated with
0.00 a naturalistic train of light pulses comprising
a range of instantaneous frequencies (Fig. 4D
350 0 -0.05 1 CR-IRES-Cre mice and fig. S12C), baclofen enhanced transmis-
sion of higher frequencies, indicating that
Startle 0246 8 60 75 90 105
response Time (s)
Sound pressure level (dB)
Noise
CS CS
US
Conditioning Memory
recall

TeNT ****

2.0 *** Control 100 **** ****
Control 100
Startle amplitude (a.u.) n.s. Control n.s.
1.5 TeNT 80
n.s. 80 TeNT
60
1.0 n.s. Freezing (%) 60
0.5 n.s. 40 Freezing (%) n.s. n.s.

20 40

n.s.
20

0.0 0 12 20 28 0 CS+
4 Trials Baseline CS-
75 85 95
Sound pressure level (dB)

Fig. 3. The HO-MG is required for auditory associative memory. (A) HO-MG bouton responses to white
noise–trains of 60 and 105 dB sound pressure level (SPL), sorted by amplitude to 60 dB. (B) Response
integral during noise-trains of increasing intensities (Friedman test, n = 350 boutons, 4 mice) and response
latency to peak for responsive boutons (Kruskal-Wallis test, 60 dB n = 139, 75 dB n = 148, 90 dB
n = 149, 105 dB n = 225 boutons). (C) AAV-mediated bilateral expression of TeNT and enhanced yellow
fluorescent protein (EYFP) (control). (D) Startle response (two-way RM ANOVA, P < 0.001 for sound
intensity, Sidak’s multiple comparisons test between animal groups, n.s.). In (D) and (E), n = 8 TeNT and
n = 7 control mice. (E) TeNT inhibition during DTC. Left: Freezing behavior during conditioning (two-way RM
ANOVA, n.s. for animal groups; gray, baseline time; blue, CS+ trials; green, CS− trials). Right: Freezing
behavior during memory recall (two-way RM ANOVA with Tukey’s multiple comparisons test between trial
types and Sidak’s multiple comparisons test between animal groups). Data shown as mean ± SEM.
n.s. P > 0.05, ***P < 0.001, ****P < 0.0001.

Pardi et al., Science 370, 844–848 (2020) 13 November 2020 3 of 5

RESEARCH | REPORT

600 ** n.s. 1.5 identify the HO-MG as a highly experience-
400 *** n.s. dependent source of top-down information
200 that is critical for associative memory, integrates
AAV flex Blue light on layer 1 Amplitude (pA) inputs from a range of brain areas, connects
ChR2 0 widely to cortical circuit elements, and whose
information transfer is in turn modulated by
L1 NDNF IN 1.0 PPR local inhibition in L1. Notably, bottom-up and
HO GABA top-down salience recruit distinct coding mech-
MG AuV HO-MG L1 IN 25 pA anisms in HO-MG afferents to L1: Whereas
ACSF 10 ms CSs elicited potentiated and more time-locked
GABAb-R ChR2 Cesium-IC responses after learning, increased stimulus
Baclofen intensity had the opposite effect, reducing and
ACSF 0.5 delaying the signals. These results suggest that
signaling within L1 may be particularly relevant
Baclofen CR-IRES-Cre mice Baclofen + CGP for perception of stimuli with low bottom-up
salience, in line with the central role of activity
ACSF Bacl. Bacl. in the apical dendrites of L5 Pyrs for percep-
+ CGP tion at the detection threshold (27). Given the
1.0 **** **** described functional differences of HO-MG nu-
Relative amplitude AAV flex AAV fDIO Sound 40% clei (7, 28), future work is needed to address
0.8 n.s. GCaMP6s Chrimson Sound ∆F/F the potentially distinct contribution of their
+ light projections to the auditory cortex.
0.6 4s
1 a.u. In L1, dendrites of Pyrs are a central sub-
0.5 s 0.4 + NDNF strate for integration and compartmentali-
stimulation zation of diverse afferents (12). While our
0.2 ACSF results indicate that these mechanisms will
Bacl. CR-IRES-Cre:: contribute to processing of HO-MG inputs
NDNF-IRES-FlpO mice (4), we also identify a previously unknown
0.0 form of modulation of afferent information
<5 5-10 >10 acting on presynaptic GABABRs, which can
dynamically control synaptic transmission.
Inst. frequency (Hz) The dependence on local inhibition not only
implies that information encoded by long-range
1 0.8 30 ΔF/F*s HO-MG boutons can differ from that encoded at the level
0.6 of the soma but also that collaterals from an
0.4 Sound 0 0.1 s (NHDz)NF Rel. amplitude2 individual axon may convey distinct informa-
0.2 + light 0 Sound 30 1.0 1 tion to different target brain areas. Given that
0 presynaptic GABABRs are present on a wide
Bouton ID 828 DeltaF/F 4 **** Initial release prob. variety of synapses (23, 29), it is likely that ad-
1 3 ditional afferents to cortical L1 are subject to
2 this form of modulation.
Response 1 0.5 10
(ΔF/F*s) Beyond memory expression, top-down af-
Sound Sound 0.0 5 ferents are involved in a range of additional
+ light 0 5 10 functions, such as credit assignment during
0 10 20 30 40 0 learning (30) and predictive coding, in which a
NDNF activity (Hz) 50 generative model of the environment is fed
HO-MG inst. freq. (Hz) back to lower processing stations in order to
828 encode mismatches between expected and
0 5 10 15 actual sensory inputs (1, 2). The present multi-
Time (s) disciplinary dissection of cortical HO-MG af-
ferents in memory-related top-down encoding
Fig. 4. Local modulation by layer 1 NDNF interneurons through presynaptic GABAB receptors. will therefore also enable future work toward
(A) Hypothesis: NDNF INs in L1 modulate HO-MG afferents through presynaptic GABABR-mediated inhibition. a better mechanistic understanding of these
(B) ChR2-expressing HO-MG axon stimulation in AuV L1 in acute slices and recording of L1 INs. (C) Left: diverse cognitive processes.
Representative excitatory postsynaptic currents (EPSCs, 0.5-ms pulses, 20 Hz) in control [artificial
cerebrospinal fluid (ACSF)], baclofen, and baclofen+CGP. Right: First pulse response amplitude (Friedman REFERENCES AND NOTES
test with Dunn’s multiple comparisons test) and paired-pulse ratio (PPR, one-way RM ANOVA with Sidak’s
multiple comparisons test, n = 10 neurons, 7 mice). (D) Left: Representative EPSCs during naturalistic 1. A. M. Bastos et al., Neuron 76, 695–711 (2012).
stimulation, normalized to first response. Right: Normalized response amplitude, grouped according to 2. G. B. Keller, T. D. Mrsic-Flogel, Neuron 100, 424–435 (2018).
instantaneous frequency (two-way RM ANOVA with Sidak’s multiple comparisons between ACSF and 3. E. Abs et al., Neuron 100, 684–699.e6 (2018).
baclofen, n = 10 neurons, 4 mice). (E) HO-MG boutons imaged in AuV L1 (scale bar, 40 mm) in response to a 4. F. Gambino et al., Nature 515, 116–119 (2014).
white noise–train, with optogenetic NDNF L1 IN stimulation in interleaved trials. Traces: Mean ± SEM of 5. M. M. Roth et al., Nat. Neurosci. 19, 299–307 (2016).
boutons [orange bar, light; (E) to (G): n = 828 sound-responsive boutons, 4 mice]. (F) Mean individual bouton 6. N. J. Audette, S. M. Bernhard, A. Ray, L. T. Stewart, A. L. Barth,
responses. (G) Full response integral. Top: Individual boutons [gray, identity line; orange, linear fit; coefficient
of determination (R2) = 0.40, P < 0.001]. Bottom: Mean responses (Wilcoxon test). (H) Computational model of Neuron 103, 277–291.e4 (2019).
NDNF L1 IN modulation of HO-MG afferents. Top left: Representative Poisson spike trains of HO-MG afferents. 7. J. A. Winer, in The Auditory Cortex, J. A. Winer, C. E. Schreiner,
Bottom left: Dependence of HO-MG bouton initial release probability on NDNF L1 IN activity. Insets: Example
EPSCs for constant 25-Hz spike input. Right: Mean EPSC amplitude normalized to the first response as a function Eds. (Springer, 2011), pp. 41–74.
of HO-MG instantaneous spike frequency and NDNF L1 IN activity. Data shown as mean ± SEM. n.s. P > 0.05, 8. T. Dalmay et al., Neuron 104, 1180–1194.e7 (2019).
**P < 0.01, ***P < 0.001, ****P < 0.0001.

presynaptic GABABR activation can robustly NDNF L1 INs exert tight control over HO-MG
modulate the frequency content of afferent afferent information. This interpretation is
information. quantitatively borne out by a model of our
data, which highlights the power of this mech-
While these results demonstrate plausibil- anism for dynamic selection of afferent fre-
ity, a major open question is whether GABA quency content through local NDNF L1 INs
released from endogenous sources in vivo (Fig. 4H and fig. S14). Intriguingly, activity of
can exert similar control. We capitalized on L1 INs depends on both learning (3) and sen-
the genetic tractability of NDNF L1 INs for sory context (24), suggesting that presynaptic
selective in vivo optogenetic stimulation (3) inhibition may itself be subject to modula-
and combined this with imaging of HO-MG tion by behavioral state.
boutons (Fig. 4E). NDNF L1 IN excitation
caused a robust inhibition of HO-MG synaptic Although the higher-order thalamus com-
activity (Fig. 4, E to G), whereas light stimu- prises half of all thalamic nuclei, has selectively
lation in control animals had no effect (fig. expanded in the evolution of higher primates,
S13A). Consistent with GABABR-mediated pre- and has been implicated in attention and other
synaptic modulation, NDNF L1 IN stimulation cognitive capacities from rodents to humans
also affected temporal response dynamics (fig. (25, 26), the function of its connection to the
S13, B to D). Together, these results reveal that cortex remains poorly understood. Our results

Pardi et al., Science 370, 844–848 (2020) 13 November 2020 4 of 5