Science 16.10.2020

RESEARCH | RESEARCH ARTICLE

Fig. 3. Hippo component clustering and aSF formation. (A and B) Jub:GFP intensities in mock ablated aSF (control, blue) and upon aSF ablation (red);
(A), Wts:GFP (B), and MyoII:3xmKate2 [(A) and (B)] distributions at 26 hAPF. n, number of clusters; A.U., arbitrary units. P < 10–3 after time point 30 s.
Arrowheads, aSF tips. (C) Jub:GFP and Wts:GFP ratioin/out of clusters at 18 hAPF (H) Jub:GFP and Wts:CitFP ratioin/out of clusters in wRNAi control and actnRNAi
(low stress) and 26 hAPF (high stress); n, number of cells. Ps < 10–5 for the clones at 26 hAPF; n, number of cells. Ps < 10–5 for Jub:GFP and Wts:CitFP.
comparison between 18 hAPF and 26 hAPF for both Jub:GFP and Wts:GFP. (I) Ban-nls:GFP normalized intensity in wRNAi control and actnRNAi clones at 18 and
26 hAPF; N, number of animals. 18 hAPF, ns; 26 hAPF, P < 10–3. (J) Fraction
(D and D´) Jub:mKate2 and Wts:GFP distributions at 18 hAPF [low stress, (D)]
and 26 hAPF [high stress, (D´)]. Arrowheads, Jub:mKate2 and Wts:GFP co- of cells that divide between 18 and 26 hAPF and between 26 and 34 hAPF in
pnr-G4>wdsRNA control and pnr-G4>actndsRNA tissues; N, number of animals.
clusters. (E) Jub:mKate2 and MyoII:3xGFP distributions during aSF formation.
t = 0 corresponds to aSF nucleation; arrowheads, aSFs. (F) Jub:GFP distribution Prior to 26 hAPF, ns; after 26 hAPF, P < 0.05. Scale bars, 2 mm. *P < 0.05
before and after (t = 90 s) ablation (dashed box) of the MyoII:3xmKate2-labeled
aSF; arrowhead, cluster prior to and after ablation. (G) Jub:GFP cluster [Kruskal-Wallis without Conover post hoc test in (C), (H), and (I) or with Conover

post hoc test in (G); one-tailed Wilcoxon signed-rank test in (J)].

and Wts colocalization increases under high at that aSF’s tip are strongly reduced (Fig. 3, cell apical area. We quantified the intensity of
mechanical stress (Fig. 3, D to D′, and fig. S5G). F and G, fig. S5, L and M, and movie S10). Ban-nls:GFP as a function of cell apical area.
Thus, an increase in mechanical stress corre- Although the loss of Actn function does not Ban-nls:GFP levels increase as cell apical area
lates with the formation of Jub and Wts co- affect the total amount of Jub and Wts at the increases from 10 to 32 mm2 and then reach
clusters at the tips of the aSFs along the AJs. To junction (fig. S5, P and Q), it decreases the a plateau for cells between 32 and 40 mm2
determine whether Jub clustering modulates number of Jub and Wts clusters and their (Fig. 4A and fig. S6A). In contrast, the level
Hippo signaling, we used the Cry2Olig opto- colocalization as well as the ratiosin/out of Jub of nls:GFP under the control of a ubiquitin
genetic clustering system (22) to induce Jub and Wts clusters (Fig. 3H and fig. S5, R and S). promoter (Ubi-nls:GFP) is independent of cell
clustering independent of aSF formation. We Last, we found that the loss of Actn function apical area (Fig. 4A). In agreement with the
found that light-induced clustering of Jub is leads to a decrease of Ban-nls:GFP Yki transcrip-
sufficient to cocluster Wts and to up-regulate tional reporter expression (23) and cell prolif- proposed role of Jub clustering in Yki activity
Yki activity (fig. S5, H to J, and movie S8). Next, eration, specifically under high stress (Fig. 3, I regulation, we found that the Ban-nls:GFP
we explored whether the formation of aSFs and J, and fig. S5, T to V). We conclude that aSFs signal increases with the Jub ratioin/out (Fig. 4B).
and their associated tension modulate Jub promote the coclustering of Jub and Wts at
and Wts clustering, Hippo/Yki activity, and cell their tips, accounting for the down-regulation Furthermore, we observed that under stress
proliferation. High-resolution time-lapse imag- of Wts activity and the up-regulation of Yki both the Jub and Wts ratiosin/out of clusters
ing showed that as an aSF forms or is displaced transcriptional activity.
along the AJ, cortical Jub:mKate2 and Wts: increase with cell apical area, consistent
CitFP flow and accumulate at the aSF tip (Fig. Having characterized the link between aSF
3E, fig. S5, K and K′, and movie S9). Conversely, formation and Yki transcriptional activation with the notion that a larger fraction of Wts
upon ablation of a previously formed aSF, the at 26 hAPF, we explored whether the scaling
Jub:GFP or Wts:CitFP clusters initially present between aSF number and cell apical area is inhibited in larger cells under stress (Fig.
might result in the scaling of Yki activity with
4C and fig. S6, B and C). If aSFs contribute to

the scaling between cell size and Hippo/Yki

pathway activation, the impact of Actn loss

of function on Wts and Jub clustering should

be more pronounced in larger cells. Accordingly,

López-Gay et al., Science 370, eabb2169 (2020) 16 October 2020 4 of 8

RESEARCH | RESEARCH ARTICLE

Fig. 4. Hippo scaling as a function of cell apical area. (A) Ban-nls:GFP and Ubi-nls:GFP intensity versus TCJ’s bisector with respect to the main stress
apical cell area; n, number of cells. P < 10–5. (B) Ban-nls:GFP intensity versus Jub:mKate2 ratioin/out of axis (fig. S7, A to F, fig. S8, L, M, and T, and
cluster; n, number of cells. (C) Wts:CitFP and Jub:GFP ratioin/out of cluster versus apical cell area; n, number supplementary text). Thus, we could classify a
of cells. (D and E) Jub:GFP (D) and Wts:CitFP (E) ratioin/out of cluster versus apical cell size in wRNAi control TCJ as a “nucleating TCJ” or a “breaking TCJ”
and actnRNAi clones; n, number of cells. P < 10–4. (F) Fraction of cells that divide between 22 hAPF and based on this orientation. Although the exact
34 hAPF as a function of apical cell size in pnr-G4>wdsRNA control and pnr-G4>actndsRNA tissues; N, number of mechanical and molecular mechanisms driv-
ing aSF nucleation and breakage at TCJs re-
animals. P < 0.05. *P < 0.05 [ANCOVA for difference in regression slopes in (A), (C), (D), and (E); mixed main to be understood, the TCJs’ role in both
processes suggests a geometric mechanism for
ANOVA in (F)]. the scaling of aSF number with cell apical area
based on the following trends: The larger cells
Jub and Wts clustering is not affected in small origins of the scaling of aSF number with cell in a tissue have more TCJs (27) and thus might
actnRNAi cells, whereas clear differences exist apical area. Stress fiber formation is best have higher rates of aSF nucleation. A typical
between large actnRNAi cells and large control studied in individual cells where the fibers’ nucleating TCJ is also farther from a typical
cells (Fig. 4, D and E, and fig. S6D). Further- spatial organization depends on cell shape breaking TCJ in larger cells, so that aSFs in
more, we observed that the effect of Actn (24, 25). Yet even in this context, the mecha- larger cells could be expected to take longer to
loss of function on cell proliferation mirrors nisms controlling stress fiber number and travel from one to the other. Together, these
the defects observed in Jub and Wts clustering dynamics are not fully understood (24, 25). To two effects would lead to larger cells’ having
in actnRNAi cells (Fig. 4F). Finally, we increased investigate how the scaling between cell size more aSFs as the result of an increase in both
cell apical area by preventing cytokinesis and aSF number is achieved in epithelial tis- aSF nucleation rate and aSF lifetime.
and observed a corresponding increase in sues under anisotropic stress, we analyzed aSF
both Jub ratioin/out of clusters and Ban-nls: dynamics at 26 hAPF. Through live imaging of To test this hypothesis, we analyzed how
GFP level (fig. S6, E and F). Together, these aSFs, we found that aSFs form at curved re- aSF nucleation rates and lifetimes differ be-
findings indicate that the scaling between gions of cell-cell junctions aligned with the tween small and large cells. Large cells on
cell area and aSF number leads to a scaling main axis of tensile stress (Fig. 5, A and B, average have higher aSF nucleation rates and
between cell area and the clustering of Jub and movie S11). In epithelial tissues, tricellular longer lifetimes than small cells, in agreement
and Wts, and thus to the scaling of Hippo/ junctions (TCJs), points where three cells meet, with our proposed geometric mechanism (fig.
Yki signaling activity with apical cell size in are often the most curved regions of the cell S8, N to R). Next, we performed computer
epithelial tissues. apical contour (26). Accordingly, aSFs mainly simulations to quantify how much of the ob-
nucleate at TCJs (Fig. 5, A to C). After nuc- served scaling of aSF number with cell apical
Tricellular junctions promote scaling between leation, aSFs sweep across the cell as they peel area could be explained by differences in the
aSF number and apical area from the cortex, most often breaking as they number of TCJs and their positions. In these
encounter another TCJ (Fig. 5, A to C). A TCJ’s simulations, we considered a simple model
Having established the role of aSFs in scaling propensity to nucleate or to break aSFs varies where only the geometric effects are present;
cells’ mechanical and biochemical responses simply as a function of the orientation of the we did not include any explicit dependence
under mechanical stress, we next explored the on cell size (see supplementary text). We ap-
plied this model to cell shapes and orienta-
tions taken directly from experimental images
at 26 hAPF. We found that our geometric TCJ
model explains ~75% of the observed variation
of aSF number with cell apical area with only
one free parameter. This parameter can be in-
terpreted as essentially equivalent to the junc-
tional cortex thickness, and the fitted value
agrees well with direct thickness measure-
ments (0.50 ± 0.12 mm versus 0.51 ± 0.12 mm;
Fig. 5D, supplementary text, and fig. S8S).

To further show that the number of aSFs
within a cell depends on the properties of its
TCJs, we sought to vary TCJ number and po-
sition independently of cell apical area and in
a population of genetically identical cells. To
accomplish this, we performed two distinct
analyses. First, we took advantage of the fact
that a large cell in contact with smaller cells
typically has more TCJs than a large cell sur-
rounded by other large cells (28). To generate
large cells with different local environments,
we induced clones of large cells by blocking
the G2/M transition through overexpression of
trbl (trblUP clones) and compared trblUP cells
at the boundary of the clone (border cells) to
cells within the clone (bulk cells, Fig. 5E). As
anticipated, trblUP cells at the boundary of the
clone have more nucleating TCJs (Fig. 5F) and

López-Gay et al., Science 370, eabb2169 (2020) 16 October 2020 5 of 8

RESEARCH | RESEARCH ARTICLE

Fig. 5. TCJs contribute to a scaling between cell
area and aSF number. (A) E-Cad:3xmKate2 and
MyoII:3xGFP distributions as an aSF nucleates
(t = 0 min; circles denote TCJs), peels from the cortex
(t = 2 to 14 min), moves toward TCJs (t = 18 to
22 min), briefly stalls at TCJs (t = 34 min), and breaks
(t = 42 min). Red arrowhead, nucleating TCJ; yellow
arrowheads, aSF tips; blue arrowheads, breaking TCJs;
open yellow arrowheads, aSF tip positions at the
time of breakage. (B) Illustration of an aSF nucleating
at a nucleating TCJ (red arrowhead) and peeling
from the cortex until the aSF tips reach breaking TCJs
(blue arrowheads) and of the TCJ opening angle a
and TCJ bisector orientation q with respect to the a-p
axis. Orange arrows indicate the TCJ bisector, which
makes an angle q with the a-p axis (q ≈ 5° for the
nucleating TCJ at right; q ≈ 80° for the breaking TCJ at
bottom). (C) aSF nucleation (red) and breakage
(blue) events at TCJs; n, total number of aSFs.
(D) Experimental (gray) and model-predicted (green)
aSF number per cell versus apical area. The model
explains ~75% of the variance of aSF number with cell
apical area; n, number of cells. (E) Schematic based
on a Jub:GFP image illustrating the positions of the
control (white), border (brown), and bulk (blue) trblUP
cells. (F to I) Nucleating TCJs per cell [(F), P < 10–5],
aSF number per cell [(G), P < 10–2 controlling for
cell apical area differences], predicted aSF lifetime (H),
and predicted aSF number per cell [(I), P < 10–2
controlling for cell apical area differences] in trblUP
border and bulk cells; n, number of cells. (J) Schematic
of elongated (top) and ortho-elongated (bottom)
cells. (K to O) aSF number per cell [(K), P < 10–5],
nucleating TCJs per cell (L), cell apical area (M),
predicted aSF lifetime [(N), P < 10–5], and predicted
aSF number per cell [(O), P < 10–5] in elongated
and ortho-elongated cells; n, number of cells. Scale
bar, 2 mm. *P < 0.05 [Kruskal-Wallis tests in (F)
to (I) and (K) to (O)].

thus might be expected to have more aSFs. cell size between these two populations, cells (hereafter referred to as ortho-elongated) should
at the boundary have on average more aSFs have more aSFs than cells of a similar size
To verify that no other differences in cell than cells within the clone (Fig. 5, G and H, and elongated parallel to the main stress axis (Fig.
fig. S9, A to G). Thus, these results substantiate 5J). By restricting our analysis to such ortho-
shape or TCJ distribution lead to a counter- our hypothesis that an increase in the number elongated cells, we showed that at constant
of TCJs correlates with an increase in the num- nucleating TCJ number and cell size, the num-
vailing difference in predicted aSF lifetime ber of aSFs. ber of aSFs is higher in ortho-elongated cells,
as predicted (Fig. 5, K to O, and fig. S9, H to J).
between border and bulk cells, we simulated We then analyzed whether the spatial dis- On the basis of quantitative analyses of aSF
tribution of TCJs modulates aSF number. If the nucleation and lifetime, modeling, and experi-
our geometric model of aSF nucleation and distance between the nucleating TCJ and the ments, we propose that the number and posi-
breaking TCJ(s) is critical to control aSF num- tions of TCJs are major contributors to the
breakage on cell sizes and shapes taken from ber, we predict that cells that are elongated scaling of aSF number with cell apical area in
our images of trblUP clones; the model in- orthogonal to the uniaxial mechanical stress
deed predicts that trblUP cells at the bound-
ary of the clone should have more aSFs per
cell than trblUP bulk cells (Fig. 5I and fig. S9,
A to G). Further, we found experimentally

that, controlling for any small difference in

López-Gay et al., Science 370, eabb2169 (2020) 16 October 2020 6 of 8

RESEARCH | RESEARCH ARTICLE

tissues under uniaxial tension (fig. S10). TCJs of tissue mechanics, proliferation, and 18. C. Rauskolb, S. Sun, G. Sun, Y. Pan, K. D. Irvine, Cytoskeletal
thus play an unexpected role in the regulation tension inhibits Hippo signaling through an Ajuba-Warts
of cells’ mechanical properties and signaling morphogenesis. complex. Cell 158, 143–156 (2014). doi: 10.1016/
under mechanical stress. j.cell.2014.05.035; pmid: 24995985
Materials and methods
Discussion 19. H. Alégot et al., Recruitment of Jub by a-catenin promotes Yki
Drosophila stocks used in this study are given activity and Drosophila wing growth. J. Cell Sci. 132, jcs222018
Scaling between body size and metabolic or in table S2. Methods for genetics, molecular (2019). doi: 10.1242/jcs.222018; pmid: 30659113
biomechanical properties has been the subject
of intense study for decades (29, 30). More biology, live imaging, laser ablations, and quan- 20. C. Ibar et al., Tension-dependent regulation of mammalian
recently, investigations of the scaling of mor- Hippo signaling through LIMD1. J. Cell Sci. 131, jcs214700
phogen gradients with tissue area or length tifications as well as statistics are described (2018). doi: 10.1242/jcs.214700; pmid: 29440237
have unveiled novel mechanisms of cell fate
specification during development (31, 32). Here, in the supplementary materials. The MatLab 21. R. Sarpal et al., Role of a-Catenin and its mechanosensing
we have uncovered a scaling between the num- properties in regulating Hippo/YAP-dependent tissue growth.
ber of aSFs per cell and cell apical area in an codes, the Python Jupyter Notebook for plotting PLOS Genet. 15, e1008454 (2019). doi: 10.1371/journal.
epithelial tissue. We have furthermore provided pgen.1008454; pmid: 31697683
evidence that this scaling is critical to control and statistical analyses, and the Fiji macros
cell and tissue elongation as well as prolifer- 22. A. Taslimi et al., An optimized optogenetic clustering tool for
ation in response to epithelial morphogenetic used are available upon request. Physical mod- probing protein interaction and function. Nat. Commun. 5,
forces. Existing work has emphasized the roles 4925 (2014). doi: 10.1038/ncomms5925; pmid: 25233328
of stress fibers as anchors to extracellular matrix eling, computer simulations, and associated
(ECM) via focal adhesions and the stress fibers’ 23. H. Matakatsu, S. S. Blair, Separating planar cell polarity and
roles in cell locomotion, cell mechanics and quantifications are described in the supple- Hippo pathway activities of the protocadherins Fat and
Hippo/YAP regulation (24, 33). Our work high- Dachsous. Development 139, 1498–1508 (2012). doi: 10.1242/
lights the contributions of aSFs anchored to mentary text. dev.070367; pmid: 22399682
AJs as regulators of both tissue morphogenesis
and proliferation under physiological morpho- REFERENCES AND NOTES 24. L. Blanchoin, R. Boujemaa-Paterski, C. Sykes, J. Plastino,
genetic mechanical stress. We have also found Actin dynamics, architecture, and mechanics in cell motility.
that aSFs promote the coclustering of Jub and 1. E. Hannezo, C.-P. Heisenberg, Mechanochemical Feedback Physiol. Rev. 94, 235–263 (2014). doi: 10.1152/
Wts in the control of Hippo/YAP signaling. The Loops in Development and Disease. Cell 178, 12–25 (2019). physrev.00018.2013; pmid: 24382887
full characterization of the molecular inter- doi: 10.1016/j.cell.2019.05.052; pmid: 31251912
actions between aSF components and Hippo/ 25. E. Kassianidou et al., Extracellular Matrix Geometry and Initial
YAP regulators at the tips of the aSFs will be 2. J. R. Davis, N. Tapon, Hippo signalling during development. Adhesive Position Determine Stress Fiber Network
critical to decipher how aSFs promote the Development 146, dev167106 (2019). doi: 10.1242/dev.167106; Organization during Cell Spreading. Cell Rep. 27, 1897–1909.e4
recruitment of Jub to the aSF tips in response pmid: 31527062 (2019). doi: 10.1016/j.celrep.2019.04.035; pmid: 31067472
to tissue stress. Interestingly, mammalian Jub
localizes at ECM-integrin focal adhesions 3. D. Pinheiro, Y. Bellaïche, Mechanical Force-Driven Adherens 26. T. Higashi, A. L. Miller, Tricellular junctions: How to build
(34); this recruitment of Jub to focal adhesions Junction Remodeling and Epithelial Dynamics. Dev. Cell junctions at the TRICkiest points of epithelial cells. Mol. Biol.
might provide a complementary mechanism of 47, 3–19 (2018). doi: 10.1016/j.devcel.2018.09.014; Cell 28, 2023–2034 (2017). doi: 10.1091/mbc.e16-10-0697;
Hippo/YAP mechanical regulation in individual pmid: 30300588 pmid: 28705832
cells (2, 15, 33, 35–37). Cell cycle regulation is
well predicted by a combination of cell area 4. F. Bosveld et al., Mechanical control of morphogenesis by 27. M. C. Gibson, A. B. Patel, R. Nagpal, N. Perrimon, The
and tension in multicellular assemblies in Fat/Dachsous/Four-jointed planar cell polarity pathway. emergence of geometric order in proliferating metazoan
culture (38). Therefore, it will be useful to Science 336, 724–727 (2012). doi: 10.1126/science.1221071; epithelia. Nature 442, 1038–1041 (2006). doi: 10.1038/
analyze whether the mechanisms we un- pmid: 22499807 nature05014; pmid: 16900102
covered might generally explain the interplay
between cell area and mechanical forces in 5. R. Levayer, C. Dupont, E. Moreno, Tissue Crowding Induces 28. W. T. Gibson et al., Control of the mitotic cleavage plane
cell proliferation control. Earlier studies have Caspase-Dependent Competition for Space. Curr. Biol. by local epithelial topology. Cell 144, 427–438 (2011).
shown that oriented cell divisions relax uni- 26, 670–677 (2016). doi: 10.1016/j.cub.2015.12.072; doi: 10.1016/j.cell.2010.12.035; pmid: 21295702
axial tissue stress in a manner that depends on pmid: 26898471
cell elongation (39); it will be of interest to 29. T. A. McMahon, J. T. Bonner, On Size and Life (Scientific
explore how aSFs, by limiting elongation while 6. S. Curran et al., Myosin II Controls Junction Fluctuations to American Library, 1983).
promoting the division of larger-area cells, Guide Epithelial Tissue Ordering. Dev. Cell 43, 480–492.e6
might modulate tissue stress or mechanical (2017). doi: 10.1016/j.devcel.2017.09.018; pmid: 29107560 30. A. J. Spence, Scaling in biology. Curr. Biol. 19, R57–R61 (2009).
properties. doi: 10.1016/j.cub.2008.10.042; pmid: 19174137
7. N. Founounou, N. Loyer, R. Le Borgne, Septins regulate the
Our characterization of the aSF dynamics contractility of the actomyosin ring to enable adherens 31. B.-Z. Shilo, N. Barkai, Buffering Global Variability of Morphogen
illustrates that the number and distribution junction remodeling during cytokinesis of epithelial cells. Gradients. Dev. Cell 40, 429–438 (2017). doi: 10.1016/
of TCJs largely account for the scaling be- Dev. Cell 24, 242–255 (2013). doi: 10.1016/ j.devcel.2016.12.012; pmid: 28292422
tween cell apical area and aSF number. Thus, j.devcel.2013.01.008; pmid: 23410939
our work defines a functional link between 32. D. Čapek, P. Müller, Positional information and tissue scaling
TCJs and aSFs. Because TCJs and stress fibers 8. F. Bosveld et al., Epithelial tricellular junctions act as during development and regeneration. Development 146,
are prevalent biological structures (24, 26), interphase cell shape sensors to orient mitosis. Nature 530, dev177709 (2019). doi: 10.1242/dev.177709; pmid: 31862792
the molecular characterization of their in- 495–498 (2016). doi: 10.1038/nature16970; pmid: 26886796
terplay might shed light on numerous aspects 33. K. D. Irvine, B. I. Shraiman, Mechanical control of growth:
9. D. Pinheiro et al., Transmission of cytokinesis forces via Ideas, facts and challenges. Development 144, 4238–4248
E-cadherin dilution and actomyosin flows. Nature 545, (2017). doi: 10.1242/dev.151902; pmid: 29183937
103–107 (2017). doi: 10.1038/nature22041; pmid: 28296858
34. S. J. Pratt et al., The LIM protein Ajuba influences p130Cas
10. B. Guirao et al., Unified quantitative characterization of localization and Rac1 activity during cell migration. J. Cell Biol.
epithelial tissue development. eLife 4, e08519 (2015). 168, 813–824 (2005). doi: 10.1083/jcb.200406083;
doi: 10.7554/eLife.08519; pmid: 26653285 pmid: 15728191

11. V. Vasioukhin, C. Bauer, M. Yin, E. Fuchs, Directed actin 35. M. Aragona et al., A mechanical checkpoint controls
polymerization is the driving force for epithelial cell-cell multicellular growth through YAP/TAZ regulation by actin-
adhesion. Cell 100, 209–219 (2000). doi: 10.1016/S0092-8674 processing factors. Cell 154, 1047–1059 (2013). doi: 10.1016/
(00)81559-7; pmid: 10660044 j.cell.2013.07.042; pmid: 23954413

12. J. Millán et al., Adherens junctions connect stress fibres 36. K. Wada, K. Itoga, T. Okano, S. Yonemura, H. Sasaki, Hippo
between adjacent endothelial cells. BMC Biol. 8, 11 (2010). pathway regulation by cell morphology and stress fibers.
doi: 10.1186/1741-7007-8-11; pmid: 20122254 Development 138, 3907–3914 (2011). doi: 10.1242/
dev.070987; pmid: 21831922
13. M. K. L. Han, J. de Rooij, Converging and Unique Mechanisms
of Mechanotransduction at Adhesion Sites. Trends Cell Biol. 37. Z. Meng et al., RAP2 mediates mechanoresponses of the
26, 612–623 (2016). doi: 10.1016/j.tcb.2016.03.005; Hippo pathway. Nature 560, 655–660 (2018). doi: 10.1038/
pmid: 27036655 s41586-018-0444-0; pmid: 30135582

14. S. Huveneers et al., Vinculin associates with endothelial 38. M. Uroz et al., Regulation of cell cycle progression by cell-cell
VE-cadherin junctions to control force-dependent remodeling. and cell-matrix forces. Nat. Cell Biol. 20, 646–654 (2018).
J. Cell Biol. 196, 641–652 (2012). doi: 10.1083/jcb.201108120; doi: 10.1038/s41556-018-0107-2; pmid: 29802405
pmid: 22391038
39. B. G. Godard, C. P. Heisenberg, Cell division and tissue
15. A. Elosegui-Artola et al., Force Triggers YAP Nuclear Entry mechanics. Curr. Opin. Cell Biol. 60, 114–120 (2019).
by Regulating Transport across Nuclear Pores. Cell 171, doi: 10.1016/j.ceb.2019.05.007; pmid: 31288206
1397–1410.e14 (2017). doi: 10.1016/j.cell.2017.10.008;
pmid: 29107331 ACKNOWLEDGMENTS

16. M. Das Thakur et al., Ajuba LIM proteins are negative regulators We thank S. Blair, N. Brown, J. Jiang, G. Struhl, and the
of the Hippo signaling pathway. Curr. Biol. 20, 657–662 (2010). Bloomington, Vienna, Harvard Medical School, and Kyoto Stock
doi: 10.1016/j.cub.2010.02.035; pmid: 20303269 Centres for reagents; the PICT-IBiSA@BDD imaging facility
(ANR-10-INBS-04); and A. Bardin, F. Graner, E. Hannezo,
17. Y. Abe, M. Ohsugi, K. Haraguchi, J. Fujimoto, T. Yamamoto, M. O’Connor, P. Leopold, J.-L. Maitre, N. Minc, A. Maugarny-Cales,
LATS2-Ajuba complex regulates g-tubulin recruitment to L. Alpar, and E. van Leen for comments. Funding: ANR-MaxForce,
centrosomes and spindle organization during mitosis. ERC Advanced (340784), ARC (SL220130607097), ANR Labex
FEBS Lett. 580, 782–788 (2006). doi: 10.1016/
j.febslet.2005.12.096; pmid: 16413547

López-Gay et al., Science 370, eabb2169 (2020) 16 October 2020 7 of 8

RESEARCH | RESEARCH ARTICLE

DEEP (11-LBX-0044, ANR-10-IDEX-0001-02), NSF IOS1353914 and M.S. developed theoretical models and performed simulations; and Figs. S1 to S10
DMR1056456. J.M.L.-G. and F.d.P. acknowledge ARC and FRM J.M.L.-G., D.K.L., H.N., and Y.B. wrote the manuscript. Competing Tables S1 and S2
fellowships, respectively. H.N. and M.S. were supported by NSF interests: The authors declare no competing financial interests. References (40–130)
Graduate Research Fellowships under grant DGE1256260. D.K.L. Data and materials availability: All data are available in the main Movies S1 to S11
was supported by Curie Mayent-Rothschild and ICAM senior text or the supplementary materials. MDAR Reproducibility Checklist
fellowships. Author contributions: J.M.L.-G., D.K.L., H.N., and Y.B.
designed the project; I.G. and S.P. produced reagents; J.M.L.-G., SUPPLEMENTARY MATERIALS View/request a protocol for this paper from Bio-protocol.
F.d.P., F.B., and O.M. performed experiments; J.M.L.-G., H.N., M.S.,
and B.G. developed methods and data analysis scripts; J.M.L.-G., science.sciencemag.org/content/370/6514/eabb2169/suppl/DC1 18 February 2020; accepted 13 August 2020
H.N., M.S., B.G., F.d.P., and F.B. analyzed the data; D.K.L., H.N., and Materials and Methods 10.1126/science.abb2169
Supplementary Text

López-Gay et al., Science 370, eabb2169 (2020) 16 October 2020 8 of 8

RESEARCH

◥ cell types may show unrelated activity patterns
and be irrelevant to behavioral state coding, and
RESEARCH ARTICLE SUMMARY (iii) molecularly defined neurons may respond
similarly within a type, but behavioral state may
NEUROSCIENCE be encoded by combinations of cell types. To
evaluate the role of molecularly defined cell types
Behavioral state coding by molecularly defined in the neural ensemble, it is important to monitor
paraventricular hypothalamic cell type ensembles activity in many individual neurons with sub-
second temporal resolution along with quantita-
Shengjin Xu*, Hui Yang, Vilas Menon, Andrew L. Lemire, Lihua Wang, Fredrick E. Henry, tive gene expression information about each cell.
Srinivas C. Turaga, Scott M. Sternson* For this, we developed the CaRMA (calcium and
RNA multiplexed activity) imaging platform in
INTRODUCTION: Brain function is often com- brain region that is important for behavior which deep-brain two-photon calcium imaging
pared to an orchestral ensemble, where sub- states such as hunger, thirst, and stress. Past of neuron activity is performed in mice during
groups of neurons that have similar activity are work has emphasized specialized behavioral multiple behavioral tasks. This is followed by ex
analogous to different types of instruments state–setting roles for different PVH cell types, vivo multiplexed RNA fluorescent in situ hybrid-
playing a musical score. Brains are composed but it is not clear whether the dynamics of the ization to measure gene expression information
of specialized neuronal subtypes that can be PVH ensemble support this view. in the in vivo–imaged neurons.
efficiently classified by gene expression pro-
files measured by single-cell RNA sequencing RATIONALE: We considered three possibilities RESULTS: We simultaneously imaged calcium
(scRNA-seq). Are these molecularly defined for how PVH neurons could be involved in en- activity in hundreds of PVH neurons from 10
cell types the “instruments” in the neural en- coding behavioral states: (i) PVH neurons of a cell types across 11 behavioral states. Within a
semble? To address this question, we examined molecularly defined cell type may respond molecularly defined cell type, neurons often
the neural ensemble dynamics of the hypo- similarly and be specialized for a behavioral showed similar activity patterns such that we
thalamic paraventricular nucleus (PVH), a small state as a “labeled-line,” (ii) molecularly defined could predict functional responses of individ-
ual neurons solely from their quantitative gene
Calcium dynamics of PVH neurons across multiple behavioral states Gene expression profile expression information. Behavioral states could
be decoded with high accuracy based on com-
state 1 state 2 binatorial assemblies of PVH cell types, which
we called “grouped-ensemble coding.” Labeled-
10x line coding was not observed. The neuromod-
ulatory receptor gene neuropeptide receptor
t post hoc neuropeptide Y receptor type 1 (Npy1r) was
usually the most predictive gene for neuron
state 3 state n functional response and was expressed in mul-
tiple cell types, analogous to the “conductor” of
the PVH neural ensemble.

Neuronal dynamics of molecularly defined cell types CONCLUSION: Our results validated molecularly
defined neurons as important information
Combinatorial cell type responses Ensemble hierarchy Grouped-ensemble coding processing units in the PVH. We found cor-
Cell types respondence between the gene expression
Response PVH cell types Output hierarchies used for molecularly defined cell
behavioral type classification and functional activity hi-
Inhibition Activation erarchies involving coordination by neuromod-
state ulation. CaRMA imaging offers a solution to
1 the problem of how to rapidly evaluate the
function of the panoply of cell types being
2 Type 1 B1 uncovered with scRNA-seq. CaRMA imaging
bridges a gap between the abstract digital
3 elements typically described in systems neuro-
4 B2 science with the “wetware” associated with tra-
Cell type ditional molecular neuroscience. Merging these
5 Npy1r –Npy1r +Type 2 B3 two areas is essential to understanding the re-
6
▪lationships of gene expression, brain function,
7
8 Type 3 behavior, and ultimately neurological diseases.
9 Bn
The list of author affiliations is available in the full article online.
10 Type m Neuromodulator *Corresponding author. Email: [email protected] (S.X.);
1 2 3 4 5 6 7 8 9 10 11 Receptor (Npy1r) [email protected] (S.M.S.)
Behavioral state Cite this article as S. Xu et al., Science 370, eabb2494
(2020). DOI: 10.1126/science.abb2494
CaRMA imaging reveals combinatorial cell type coding of behavior states. CaRMA imaging records
calcium dynamics of PVH neurons across multiple behavioral states followed by gene expression profiling. READ THE FULL ARTICLE AT
Combinatorial assemblies of PVH cell types encoded behavioral states. The PVH neural activity ensemble https://doi.org/10.1126/science.abb2494
was split by Npy1r expression into two main cell classes that were subdivided into cell types. Thus,
neuromodulation coordinates cell types for grouped-ensemble coding to represent different survival
behaviors such as eating, drinking, and stress.

Xu et al., Science 370, 313 (2020) 16 October 2020 1 of 1

RESEARCH

◥ It has been extremely difficult to distinguish
these models because there has not been a
RESEARCH ARTICLE suitable approach to simultaneously examine
the functional dynamics of individual neurons
NEUROSCIENCE having distinct molecular identities across
many behavioral states within a single animal.
Behavioral state coding by molecularly defined Several methods have been developed to map
paraventricular hypothalamic cell type ensembles gene expression onto neuronal function (29–33).
However, these methods have limitations of
Shengjin Xu1*, Hui Yang1,2, Vilas Menon1†, Andrew L. Lemire1, Lihua Wang1, Fredrick E. Henry1‡, temporal sensitivity, dynamic range, molecular
Srinivas C. Turaga1, Scott M. Sternson1* diversity, or restrictions to transparent organ-
isms that have prevented evaluation of this type
Brains encode behaviors using neurons amenable to systematic classification by gene expression. The of problem. Thus, an unbiased and systematic
contribution of molecular identity to neural coding is not understood because of the challenges involved method to relate multigene expression of indi-
with measuring neural dynamics and molecular information from the same cells. We developed CaRMA vidual neurons to their activity patterns in vivo
(calcium and RNA multiplexed activity) imaging based on recording in vivo single-neuron calcium on a subsecond time scale is needed to combine
dynamics followed by gene expression analysis. We simultaneously monitored activity in hundreds of molecular and systems neuroscience.
neurons in mouse paraventricular hypothalamus (PVH). Combinations of cell-type marker genes had
predictive power for neuronal responses across 11 behavioral states. The PVH uses combinatorial To address this problem, we developed the
assemblies of molecularly defined neuron populations for grouped-ensemble coding of survival CaRMA (calcium and RNA multiplexed activity)
behaviors. The neuropeptide receptor neuropeptide Y receptor type 1 (Npy1r) amalgamated multiple cell imaging platform, a comprehensive, three-
types with similar responses. Our results show that molecularly defined neurons are important part strategy compatible with recording neu-
processing units for brain function. ron dynamics deep in the mouse brain (Fig. 1D).
In the first step, we used single-cell RNA se-
N euron activity and neuronal gene ex- tional response types of the neuronal ensemble quencing (scRNA-seq) to classify PVH cells into
pression are considered distinct facets from the mouse paraventricular hypothalamus molecular types (Fig. 1D, a and b). In the sec-
of brain function. Large-scale neuron (PVH). The PVH mediates appetite, stress, thirst, ond step, we performed volumetric, deep-brain,
activity recordings reveal discrete groups and autonomic functions. This brain region is pan-neuronal, two-photon calcium imaging
of neurons that respond similarly (1–4), well known for its diverse peptide-expressing in awake, behaving animals without regard
and gene expression brain atlases show differ- cell types (18, 19), and there are distinct be- to PVH neuronal subtypes to measure the
ential expression patterns across the brain (5). havioral and physiological consequences of response dynamics of hundreds of neurons
Single-cell transcriptomic studies reveal that PVH neuropeptide pharmacology (20) as well during multiple behavioral states (Fig. 1D, c).
the brain is composed of hundreds of molec- as cell type–selective optogenetic and chemo- Subsequently, the brain was removed, sectioned,
ularly defined neuronal subtypes (6–8). Molec- genetic perturbations (21–23). Molecular and and the neurons in the ex vivo tissue sections
ular markers facilitate repeatedly returning to cellular perturbation studies have indicated were registered to those from the in vivo im-
the same type of neurons and also serve as that the PVH acts as a behavioral state “switch- aging volume (Fig. 1D, d and e). In the final
genetic elements that can be used for target- board” in which each molecularly defined pop- step, molecular identity of the neurons imaged
ing functionally similar neurons for causality ulation influences a distinct behavioral state as in vivo was determined post hoc by performing
testing (9, 10) or potentially therapeutics (11). a labeled-line (24). However, it is not known if multiple rounds of three-plex RNA-fluorescent
These applications can only be justified if mo- this is consistent with the natural dynamics of in situ hybridization (FISH) on the ex vivo tissue
lecular markers categorize neurons into groups PVH neurons during these behaviors. sections, guided by marker genes from scRNA-
with similar neuronal dynamics (12) that to- seq in step 1 (Fig. 1D, f to i). This approach
gether encode distinct behavioral states. How- We considered three models for the role of enables monitoring the dynamics of many
ever, there is limited knowledge about the molecularly defined cell types to encode be- molecularly defined neurons in parallel, with-
relationship between activity patterns of cells havioral states. The labeled-line model encodes in a deep-brain region, and across multiple
in the neural ensemble and their underlying different behavioral states based on the activity behaviors.
gene expression profiles. Therefore, the merit pattern of distinct, molecularly defined cell
of using neuronal gene expression as a proxy types (Fig. 1A). Labeled-line coding is fre- RESULTS
for neurons with similar activity patterns is quently assumed for information processing Molecularly defined cell types within PVH
controversial (12–17). in hard-wired neural functions involving es- scRNA-seq of PVH neurons
sential survival behaviors (25–27). At the other
We set out to determine (i) the role of mo- end of the spectrum, the full neural ensemble We used scRNA-seq to transcriptionally pro-
lecularly defined cell types for encoding dif- encodes behavioral state irrespective of cell file 706 manually picked PVH cells from 10
ferent behavioral states and (ii) the association type identity and is primarily a product of animals. We applied iterative unsupervised
of neuronal marker genes with distinct func- neural plasticity (Fig. 1B) (28). In this case, clustering (8) to group cells into 12 molec-
molecular identity is a poor predictor of be- ularly defined cell types within the PVH (fig.
1Janelia Research Campus, Howard Hughes Medical Institute, havioral state, and functional ensembles are S1A). Differentially expressed marker genes for
Ashburn, VA 20147, USA. 2Dominick P. Purpura Department formed using rules that are independent of these clusters included canonical PVH neuro-
of Neuroscience, Albert Einstein College of Medicine, Bronx, cell type. An intermediate model is grouped- peptides: oxytocin (Oxt), vasopressin (Avp),
NY 10461, USA. ensemble coding, in which the members of corticotropin-releasing hormone (Crh), thyrotropin-
*Corresponding author. Email: [email protected] (S.X.); some molecularly defined cell types show releasing hormone (Trh), somatostatin (Sst),
[email protected] (S.M.S.) coordinated responses and encode behavioral proenkephalin (Penk), and prodynorphin (Pdyn).
†Present address: Center for Translational and Computational states based on specific combinations of cell Within these broad classes, clusters were fur-
Neuroimmunology, Department of Neurology, Columbia University types (Fig. 1C). ther divided into subgroups based on the
Medical Center, New York, NY 10032, USA. expression of additional marker genes: gluta-
‡Present address: Innate Pharma, Rockville, MD 20850, USA. mic acid decarboxylase 2 (Gad2), reelin (Reln),

Xu et al., Science 370, eabb2494 (2020) 16 October 2020 1 of 15

RESEARCH | RESEARCH ARTICLE

Fig. 1. CaRMA imaging for investigating A B C
models of behavioral state coding by
molecularly defined cell types. (A to Labeled-line coding Full-ensemble coding Grouped-ensemble coding
C) Models of multiple molecularly defined PVH neurons PVH neurons
cell types encoding multiple behavioral PVH neurons
states after processing diverse internal and Behavioral state Behavioral Behavioral
external inputs (I). (A) Labeled-line B1 state state
coding uses a specialized cell type for a B1 B1
behavioral state in which individual mem- Input Type 1 Input Input Type 1
bers respond similarly, where the number of I1 B2 I1 Type 2 B2
encoding cell types (m) is equal to the I1 B2
number of distinctly encoded behavioral I2 I2
states (n). (B) In a full-ensemble-coding I2 Type 2
model, molecularly defined cell types do not
respond similarly, and behavioral state I3 B3 I3 B3 I3 B3
coding is independent of cell type. Type 3
(C) Grouped-ensemble coding uses Il Type 3
combinations of molecularly defined cell Il Bn
types in which individual molecularly Il
defined cell types act as a coherent Bn
functional unit. (D) Schematic of the
CaRMA imaging platform. l=m=n Bn m< l, n
Type m Type m
Type 1 Type 2 Type 3 Type m

D (a) Single-cell RNA-seq (b) Gene expression profile
Tissue dissection Single-cell suspension
Cell Type
Classification (i) Assign
(scRNA-Seq)
expression
Sequencing profiles cDNA and library Manual picking Cell
Genesynthesis

(f) Marker genes(c) Functional imaging(d) Confocal imaging

GRIN 10x (e) Ex vivo to in vivo registration
lens
20x
Perfusion Section Slice

CaRMA PVH 3-plex
Imaging FISH

(h) multiple rounds (..gS3)-tprilpexPFroISbeHs
Image registration
across all rounds

20x 20x 20x

Multiplexed RNA FISH

netrin G1 (Ntng1), and neuropeptide Y receptor (3D) segmentation of cell boundaries (figs. S2 types was largely intermingled but some well-
type 1 (Npy1r) (fig. S1B). All PVH clusters ex- to S4; see the materials and methods). organized patterns were apparent (Fig. 2G;
pressed Sim1, contained the excitatory neuron fig. S6, A and B; and table S1). Expression levels
marker Vglut2 (Slc17a6), and lacked the inhib- We observed marker gene enrichment in of the marker genes were distributed continu-
itory neuron marker Vgat (Slc32a1). This tar- different PVH subregions (Fig. 2B): Two genes ously, with strong rightward skew and high
geted, deep-sequencing approach provides a (Gad2 and Ntng1) were significantly enriched variance (Fig. 2H). Clustering implicitly thresh-
combinatorial set of marker genes for molec- in aPVH, three genes (Vglut2, Crh, and Avp) olded gene coexpression relationships, and
ularly defined cell type assignment. were significantly enriched in mPVH, and nine most FISH clusters (12/13) were dominated
genes (Vglut2, Gad2, Npy1r, Crh, Reln, Ntng1, by high expression of one gene (Fig. 2E). FISH
12-plex FISH in PVH Pdyn, Oxt, and Avp) were significantly enriched clusters were correlated with the scRNA-seq
in pPVH. Pairwise gene coexpression also dif- dataset (fig. S6, C and D). Likewise, scRNA-
We mapped the spatial and coexpression pat- fered based on anterior-posterior position (fig. seq clusters were represented in the FISH dataset
tern of these 11 differentially expressed genes S5). For example, a majority of Crh cells coex- (fig. S6, E to G) and mapped to different regions
plus Vglut2 by performing four rounds of three- pressed Npy1r in pPVH (75%) but considerably of the PVH (fig. S6G).
plex FISH from two mice in PVH subregions less in mPVH (30%) and aPVH (35%).
separated by 280 mm: the anterior PVH (aPVH, CaRMA imaging
1394 cells), middle PVH (mPVH, 1855 cells), and We determined the proportions of PVH cells
posterior PVH (pPVH, 1263 cells) (Fig. 2A). This in situ that expressed 0 to 12 of the marker To record PVH neuronal dynamics, we ex-
was achieved by developing new methods for genes (Fig. 2C). Coexpression (more than two pressed GCaMP6m in PVH neurons without
stripping fluorescently labeled probes, align- genes) was common (81% of cells), and the regard to neuronal subtypes. A thin gradient
ing individual cells in three dimensions across mode was three genes. Unsupervised hierar- refractive index (GRIN) lens was implanted
multiple rounds of FISH using 4′,6-diamidino- chical clustering of gene expression profiles into the PVH (Fig. 3A). Volumetric two-photon
2-phenylindole (DAPI) fluorescence, as well as from these 4512 PVH cells resulted in 13 tran- calcium imaging was acquired from PVH neu-
generalizable methods for three-dimensional scriptional clusters (Fig. 2, D to F). The spatial rons under the GRIN lens (range: 80 to 240 mm)
distribution of these molecularly defined cell

Xu et al., Science 370, eabb2494 (2020) 16 October 2020 2 of 15

RESEARCH | RESEARCH ARTICLE

AB C

Vglut2 90 aPVH
Gad2 mPVH
Npy1r pPVH Percentage of cells
Crh
Distance 20
Percentage of cells
Reln 60

Ntng1 10
Pdyn

Penk 30
Trh

100 µm Oxt 0 0
Avp Vglut2 Gad2 Npy1r Crh Reln Ntng1 Pdyn Penk Trh Oxt Avp Sst 0 3 6 9 12
Sst
Number of coexpressed genes

D Cluster #

1 23 4 5 67 8 9 10 11 12 13
1.0 505
0.3 Normalized Expression
0.5
1602 194 189 630 136 199 308 114 322 0
0
Oxt
Pdyn
Avp
Gad2
Sst
Trh
Vglut2
Crh
Npy1r
Ntng1
Penk
Reln

Neuron 147 56110

E G 50 µm pPVH H
dp
Oxt 3v Vglut2
Pdyn pv mpv

Avp

Gad2 Gad2
Sst
Trh 0.3
tSNE2
Vglut2 Normalized Expression Npy1r

Crh Crh
Npy1r
Ntng1 Reln
Penk

Reln

1 2 3 4 5 6 7 8 9 10111213 0 Ntng1

Cluster#

F Pdyn

C12-Penk C13-Reln Penk
C5-Low C1-Oxt
lp
C2-Pdyn
Trh

C7-Trh C8-Vglut2 C1-Oxt C8-Vglut2 Oxt
C2-Pdyn C9-Crh
C11-Ntng1 C3-Avp C10-Npy1r Avp
C4-Gad2 C11-Ntng1
C9-Crh C10-Npy1r C5-Low C12-Penk Sst
C6-Sst C13-Reln
C4-Gad2 C7-Trh 0 0.5 1
Normalized expression level
C6-Sst C3-Avp
in positive neurons
tSNE1

Fig. 2. Molecular and spatial characterization of PVH neurons by genes. (E) Mean expression pattern of marker genes in 13 cell types. (F) Cell
multiplexed FISH with 12 marker genes. (A) Maximum intensity projection types in (D) plotted by t-distributed stochastic neighbor embedding (tSNE).
image of multiplexed FISH with 12 marker genes in pPVH. Right, mRNA puncta “Cell types” are transcriptional clusters denoted as Ci-xxx, where i is the
pseudocolor legend. Solid white line is the contour of pPVH; dashed white line cluster number in (D) and xxx is a highly expressed gene or “Low” (low
is the region in fig. S4B. (B) Percentages of cells expressing each marker expression for all probed genes). (G) Spatial organization of 13 molecularly
gene in aPVH, mPVH, and pPVH. Error bars indicate mean ± SEM. Gray circles defined cell types in pPVH (data are from four samples). Each symbol
are sample data. *P < 0.05, **P < 0.01, ***P < 0.001. Statistics are provided represents one neuron. PVH subregion boundaries are from reference (55).
in table S2. (C) Histogram of PVH cells coexpressing various number of pv, periventricular; dp, dorsal parvicellular; lp, lateral parvicellular; mpv,
marker genes. (D) Gene expression profiles of 13 molecularly defined PVH cell medial parvicellular ventral zone; 3V, 3rd ventricle. (H) Normalized expression
types from hierarchical clustering of normalized expression of 12 marker levels of marker genes from PVH neurons.

Xu et al., Science 370, eabb2494 (2020) 16 October 2020 3 of 15

RESEARCH | RESEARCH ARTICLE

AB C

77 µm 88 µm 101 µm * 114 µm * uncorrected * corrected

GRIN lens 129 µm 145 µm 163 µm 184 µm

GCaMP E Round 1 Round 2 Round 3 Round 4
Crh Pdyn Oxt
D Vglut2
Reln
21 40
11 22 Ntng1
35 23
20
626542524218412371283731352
In vivo 38
39

9 1 16
18 29
19 2
3 17 36 37
30
31

34

10

21 40 Gad2 Penk Avp
22
11 23

6 5 2035
264252421842173 8 15
Ex vivo 7 38 39
12 3332

3 9 1 16
191829 2
17
31 30 3637
34

10

Npy1r Trh Sst

11

Bottom slice 65 8
4 7

91

32
37

10

F Hunger Eating Hedonic Eating/Fear Retrieval Ghrelin Gene Expression 0.3

1 0.2 Normalized expression
2
3 0.1
4
5 Sst 0
6 Avp
7 Oxt
8 Trh
9 Penk
10 Pdyn
Ntng1
ID Reln
Crh
Npy1r
Gad2
Vglut2

ttsshooppennoodeeuuottoonoinfincfuet ating

sshppuoonuugtteiornuetating
30 s 30 s 1 t-auROC ITI ITI ITI
(inject) 30 s

0
-1 t-auROC = 2×auROC -1

Fig. 3. Ex vivo ↔ in vivo registration for CaRMA imaging. (A) Top, head- a substack of the in vivo image volume. Ex vivo image is overlay of z-projected
fixed mouse during two-photon calcium imaging of PVH neurons through a confocal stacks from three consecutive 14-mm brain slices (pseudocolors:
GRIN lens. Bottom, schematic of GRIN lens targeting GCaMP-expressing PVH cyan, green, and red indicate top, middle, and bottom slices). White dashed
neurons. (B) Eight planes from a two-photon imaging volume during behavior. line is the resolvable in vivo FOV. Dashed, thin, and thick contours are
Upper left, distances between imaging planes and GRIN lens. Arrowheads neurons from the top, middle, and bottom slices, respectively. (E) Four
mark the corresponding neurons in (C). Red or green asterisks indicate the rounds of three-plex FISH from neurons in the bottom slice in (D). (F) Calcium
imaging plane in Fig. 3C or fig. S9A, respectively. (C) Computational dynamics of neurons in (E) across multiple behavioral states and their
correction of optical aberrations from in vivo imaging. Arrowheads indicate 12-plex gene expression profiles. t-auROC, transformed auROC (see the
neurons from deeper imaging planes in (B) because of field-of-view (FOV) materials and methods); ID, neuron number from (D); ITI, intertrial interval
curvature correction. (D) Example neurons showing the ex vivo registration to (2 min). Scale bars in (D) and (E), 15 mm.

Xu et al., Science 370, eabb2494 (2020) 16 October 2020 4 of 15

RESEARCH | RESEARCH ARTICLE

A Distance Threshold C Activated D Activated 30 Inhibited 2
Max labeled-line neurons (%)45Inhibited 45 3
1 Max labeled-line neurons (%)230 30 15 4
3 5
0.5 4 15 15 0 6
5 4 6 8 10 2 4 6 8 7
0 6 8
7 Number of behavioral state classes 9
B 8 10
9 11
10
11 10

Neuron number00
2 4 6 8 10 2
GhrelinNumber of behavioral states
Leptin
Saline
Eat-Aft
Drink-Aft
Spout-Aft
Tone-Aft
Fear
Eat-Hed
Eat-Hng
Drink-T
Leptin
Saline
Ghrelin
Spout-Aft
Tone-Aft
Fear
Eat-Hed
Drink-Aft
Drink-T
Eat-Aft
Eat-Hng
Leptin
Saline
Ghrelin
Spout-Aft
Tone-Aft
Fear
Eat-Hed
Drink-Aft
Eat-Aft
Eat-Hng
Leptin
Saline
Ghrelin
Spout-Aft
Tone-Aft
Fear
Eat-Hed
Drink-Aft
Eat-Hng
Leptin
Spout-Aft
Tone-Aft
Fear
Drink-Aft
Drink-T
Eat-Aft
Eat-Hng
Leptin
Spout-Aft
Fear
Eat-Hed
Drink-Aft
Drink-T
Eat-Hng
Saline
Ghrelin
Fear
Eat-Hed
Drink-T
Eat-Hng
Ghrelin
Fear
Eat-Hed
Drink-T
Eat-Hng
Ghrelin
Fear
Drink-T
Eat-Hng
Ghrelin
Fear
Eat-Hng
Ghrelin
Eat-Aft

k=11 2* k=10 2* k=9 5* k=8 5* k=7 5
10
k=2 5*k2=* 6 *
5*5* 20

5*3k*=5* 3*
120 40

5*k5=*24*4*

100 Inhibited 60

80 k=3 Search for maximum number of labeled-line neurons 5k*5=*35*
5*5* in all 11 choose k combination behavioral states k5=*52*
Neuron number
Mean response 1
0
(t-auROC) -1
60
k=11
Activated

40 k=4
4* 5*

k=5

20 * 2* * 2* k=6
* * k=7
10 * k=8 k=9
2*
5 k=10

Leptin
Saline
Ghrelin
Spout-Aft
Tone-Aft
Fear
Eat-Hed
Drink-Aft
Drink-T
Eat-Aft
Eat-Hng
Saline
Ghrelin
Spout-Aft
Tone-Aft
Fear
Eat-Hed
Drink-Aft
Drink-T
Eat-Aft
Eat-Hng
Saline
Spout-Aft
Tone-Aft
Fear
Eat-Hed
Drink-Aft
Drink-T
Eat-Aft
Eat-Hng
Spout-Aft
Tone-Aft
Fear
Eat-Hed
Drink-Aft
Drink-T
Eat-Aft
Eat-Hng
Spout-Aft
Tone-Aft
Eat-Hed
Drink-Aft
Drink-T
Eat-Aft
Eat-Hng
Spout-Aft
Tone-Aft
Eat-Hed
Drink-T
Eat-Aft
Eat-Hng
Tone-Aft
Eat-Hed
Drink-T
Eat-Aft
Eat-Hng
Fear
Drink-Aft
Eat-Aft
Eat-Hng
Saline
Ghrelin
Fear
Ghrelin
Fear

Fig. 4. Screening for labeled-line neurons encoding multiple behavioral states. (A) Dendrogram of neuron ensemble response similarities across 11 behavioral
states. Dashed lines are thresholds for grouping state classes. (B) Mean response maps of the maximum number of labeled-line neurons in all 11-choose-k
combinations of behavioral states. Bottom left, Activated labeled-line neuron sets. Top right, inhibited labeled-line neuron sets. Fisher’s exact test was used to evaluate
whether neurons are significantly specialized for a behavioral state. (C) Number of labeled-line neurons depends on the number of behavioral states. (D) Number of
labeled-line neurons depends on the number of behavioral state-classes [see (A)]. *, P < 0.05; 2*, P < 0.01; 3*, P < 0.001; 4*, P < 0.0001; 5*, P < 0.00001.

in a head-fixed behaving mouse (Fig. 3B). We lens were imaged by confocal microscopy. depth-dependent magnification changes (fig.
performed a set of calcium-imaging experiments Because the GRIN lens introduced multiple S8, C, D and I). After correcting for these
involving eating during hunger, drinking water optical distortions, we characterized these aberrations (Fig. 3C and fig. S9A), we found
during thirst, hedonic eating (ad libitum fed, aberrations and transformed the in vivo neurons with distinctive shapes in the two-
palatable food), fear retrieval (movie S1), and image volume coordinates to the view obtained photon–imaging volume (neurons 1 to 3 in
hormone-induced hunger (ghrelin) or energy by confocal imaging in ex vivo tissue sections Fig. 3D; neurons 1, 2, 4, and 35 in fig. S9B)
surfeit (leptin) sequentially over 10 days (fig. (fig. S8). For a successful in vivo ↔ ex vivo that could be recognized in the ex vivo con-
S7). Subsequently, the brain was removed and alignment, it was crucial to correct field cur- focal images. We used these neurons as start-
sectioned. vature introduced by the GRIN lens (fig. S8, ing points to match the surrounding neurons
C, D, F, and J), as well as the nonlinear imaged in vivo and ex vivo with 96% corre-
To register the GCaMP-expressing neurons relationship between sample position and spondence between two independent observers.
in the ex vivo–sectioned brain to the in vivo objective lens position (fig. S8, H and J) and Most of the neurons (89.3%, 334/374) imaged
image volume, the sections below the GRIN

Xu et al., Science 370, eabb2494 (2020) 16 October 2020 5 of 15

RESEARCH | RESEARCH ARTICLE Ghrelin Saline Leptin B Gene expression

A Hunger eating Thirst drinking Fear retrieval

1011 9 8
14 2 34 17

Avp Cluster #
Oxt 7 65 4
Sst
Trh 32
Gad2
Penk 1
Pdyn
Crh
Npy1r
Vglut2
Reln
Ntng1
92 14 9 38 29 30 40 Neuron
tshstooppennoodeeuuottoonoifnincfuet ating1 minITI ITI ITI ITI ITI ITI ITI ITI ITI 01
dssprpionoukutintoginut (inject) (inject) (inject) Distance
sshppuoonuugtteiornuetating
-1 1 0 0.3
Response (t-auROC) Normalized expression

C D E

MC8-Penk 50 µm MC1-Ntng1 50 µm A
MC2-Reln L
MC5-Crh MC7-Low MC10-Trh A MC3-Vglut2
L MC4-Npy1r P
MC4-Npy1r MC5-Crh
MC3-Vglut2tSNE2 P MC6-Pdyn
MC2-Reln MC7-Low
3V MC8-Penk
MC6-Pdyn MC9-Gad2
MC1-Ntng1 MC9-Gad2 MC10-Trh
MC11-Sst
tSNE1
3V
F
1 G Eat-Hng Drink-T Eat-Hed
1 11
0 Consistent-response (t-auROC)
MC5-Crh Response (t-auROC) 0.5 0 0 Response
-1 (t-auROC)
1
0 -1 -1Purity 0 10 20 30 0 10 20 30 0 10 20 30
0 Time (s) Time (s) Time (s)
1 1 min 1
-1
MC8-Penk 0.5 0 Cumulative prob 1 P < 10-8 P < 10-8 P< 10-6

0.5

0 -1 MC5-Crh

sstthooppennoodeeuuottonooinifncfuet ating thsstooppennoodeeuuottnoooiifnncfuet ating MC6-Pdyn
sthtsooppennoodeeuuottnooioifnncfuet ating
0
-0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 -0.2 0 0.2 0.4

Slope (t-auROC/s) Slope (t-auROC/s) Slope (t-auROC/s)

H Response Response purity Max consistent-response Gene expression

MC1-Ntng1 1 Response
MC2-Reln 0 (t-auROC)
-1
MC3-Vglut2 1 Purity
MC4-Npy1r 0.5
0 Normalized
MC5-Crh 0.3 expression
MC6-Pdyn
MC7-Low 0
MC8-Penk
MC9-Gad2
MC10-Trh

Trh
Gad2
Penk
Pdyn
Crh
Npy1r
Vglut2
Reln
Ntng1

Leptin
Saline
Ghrelin
Spout-Aft
Tone-Aft
Fear
Eat-Hed
Drink-Aft
Drink-T
Eat-Aft
Eat-Hng

Leptin
Saline
Ghrelin
Spout-Aft
Tone-Aft
Fear
Eat-Hed
Drink-Aft
Drink-T
Eat-Aft
Eat-Hng

Leptin
Saline
Ghrelin
Spout-Aft
Tone-Aft
Fear
Eat-Hed
Drink-Aft
Drink-T
Eat-Aft
Eat-Hng

Figure 5 6 of 15
Xu et al., Science 370, eabb2494 (2020) 16 October 2020

RESEARCH | RESEARCH ARTICLE

Fig. 5. Calcium dynamics and gene expression profile of PVH neurons Colored arrowheads indicate the corresponding neurons in (E). (E) Fluorescence
across 11 behavioral states. (A) Calcium response dynamics of the same PVH image of GCaMP-expressing example neurons within a FOV (dashed circle). Blue is
neurons (319 cells) during multiple behaviors from three mice. Neurons are DAPI and green is GCaMP. (F) Responses of MC5-Crh and MC8-Penk neurons during
clustered by gene expression profiles in (B). Temporal scale bars, 1 min. fear retrieval. Left, response traces of individual neurons. Middle, red shaded lines
(B) Gene expression profile (12-plex RNA-FISH) of neurons in (A). Molecularly are the mean responses ± SEM across neurons; gray lines are purities. Right,
defined cell types are clustered based on the expression of the first nine genes. instantaneous consistent-responses. (G) Different response temporal profiles of
Cluster 11 has two neurons with high Sst expression even though Sst was not MC5-Crh and MC6-Pdyn neurons. Top, Mean responses aligned with food or water
used for clustering (excluded for later analysis because of low neuron count; see presentation. Bottom, Cumulative distribution of instantaneous response slopes of
the materials and methods). (C) Cell types in (B) plotted by tSNE. “Molecular individual neurons from these cell types during the light-blue-shaded periods from the
clusters” are denoted as MCi-xxx, where i is the cluster number in (B) and xxx is top panel (two-sample Kolmogorov-Smirnov test). (H) Temporal maximum for
a highly expressed gene or “Low” (low expression for all probed genes). consistent-response and corresponding response and purity of PVH cell types defined
(D) Spatial distribution of these cell types in the three imaging FOVs (dashed circles). by combinatorial gene expression profiles across 11 behavioral states.

in vivo could be found in the brain slices, and were statistically greater than chance for not neurons do not necessarily respond similarly in
95.5% (319/334) of the aligned neurons could more than four states (Fig. 4B). We also con- all behaviors and may be tuned to only some
be tracked across the entire sequence of ex- sidered the possibility that grouping similar behavioral states.
periments (Fig. 3D and fig. S9B). states might make labeled-line coding more
apparent. States were grouped into “state We quantified the similarity of neuronal
After matching neurons between the in vivo classes” either subjectively based on ideas responses within each cell type cluster using
and ex vivo image spaces, we quenched about behavioral similarity (fig. S13A) or sys- linear purity (fig. S17; see the materials and
GCaMP6m with acidic buffer (34) and per- tematically based on PVH neuronal response methods). During fear cue presentation, MC5-
formed four rounds of three-plex RNA-FISH similarity metrics for unsupervised clustering Crh had high purity (approaching 1), whereas
(Fig. 3E and fig. S9C). We developed a non- (Fig. 4A). However, grouping further reduced MC8-Penk had low purity (close to zero; Fig.
rigid 3D registration pipeline for the GCaMP- performance of the labeled-line model (Fig. 5F). Cell types with purity >0.5 can be con-
expressing brain slices across multiple rounds 4D and figs. S13 and S14). Examining a small sidered to respond similarly (fig. S17). Next, we
of processing (fig. S10). Calcium dynamics of number of behavioral states might thus lead defined a metric called “consistent-response”,
individual neurons during multiple behaviors to the erroneous conclusion that a set of neu- which weights cell type activity by the re-
were extracted from the in vivo image volumes rons is involved in highly selective labeled-line sponse purity (Fig. 5F and fig. S17). A high
(fig. S11), along with the corresponding molec- coding. Although the labeled-line configuration consistent-response molecularly defined cell
ular profiles for each neuron acquired from cannot be excluded for some brain functions, type is composed of neurons that respond
ex vivo tissue sections (Fig. 3F and fig. S9D). labeled-line coding is poorly scalable. strongly to a stimulus with similar temporal
These computational and image analysis tools dynamics. Low consistent-response cell types
are available online (see “Data and materials Molecularly defined neural have either low purity (a diversity of responses)
availability” in the Acknowledgments). ensemble responses or low response magnitude in a behavioral
state (fig. S17).
PVH ensemble activity across Next, we used unsupervised clustering of gene
11 behavioral states expression profiles to group the PVH neurons Combinatorial molecularly defined cell
into 11 molecularly defined clusters (MCs) type tuning to behavioral states
We recorded calcium dynamics of 319 neurons (Fig. 5, A to C; see the materials and methods)
from three mice in the same portion of the so that we could evaluate the activity patterns We computed the temporal maximum of
pPVH (fig. S12) during all behaviors. We seg- in molecularly defined cell types across 11 be- consistent-responses and purities of each cell
mented the behavioral tasks into 11 behavioral havioral states. Although GCaMP expression type in 11 behavioral states (Fig. 5H and fig.
states (fig. S7). These behavioral states could was not observed in Oxt- or Avp-expressing S18, B and C). In the consistent-response map,
be well separated by unsupervised clustering neurons because of adeno-associated virus some cell types responded differently to in-
of the recorded PVH neural activity patterns tropism (fig. S15), these 11 MCs were con- dividual behavioral states, whereas others were
(Fig. 4A). sistent with cell types clustered from FISH- similarly tuned across multiple states. MC8-
only pPVH tissue (fig. S16). Spatially, these Penk neurons, which had relatively low response
Labeled-line coding molecularly defined cell types were inter- and low purity during fear, showed higher pu-
mingled, primarily in the medial parvicel- rity activation after injection of the hormone
First, we investigated whether this PVH en- lular and lateral parvicellular subdivisions ghrelin. This demonstrates that molecularly
semble showed evidence for labeled-line coding of the pPVH (Fig. 5, D and E). defined neurons are tuned to respond sim-
of behavioral states irrespective of any gene ilarly in some behaviors but not others. The
expression information. A labeled-line neuron Response purity within molecular clusters purity of the molecularly defined cell types
was defined as specifically activated or inhibited resulting from unsupervised clustering of mul-
in one state but not in other states. We searched If cell type information is important to encode tiple genes produces significantly higher re-
for the maximum number of labeled-line neu- a behavioral state, then we expected that neu- sponse purities than when cells are grouped
rons among all k combinations of 11 behavioral rons within a molecularly defined type, grouped by expression of a single gene (fig. S19, A to
states (Fig. 4B). Although selectively tuned solely using gene expression information, C). The high purity activation response to
neurons were evident with comparison of a should respond similarly in that state. When food consumption of MC5-Crh neurons that
small number of behavioral states, the num- we looked at MC5-Crh neurons in response to highly express Crh and coexpress Vglut2 and
ber of neurons showing apparent labeled-line hedonic eating and fear retrieval, we found Npy1r in the pPVH is opposite to the in-
coding was dependent on the number of be- that neurons within this cluster responded hibitory response reported by low-resolution,
havioral states examined (Fig. 4C). Beyond six similarly. By contrast, MC8-Penk neurons multineuron fiber photometry measurements
behavioral states, there were no selective neu- showed highly heterogeneous responses in of all CRH neurons in the PVH, although both
rons for each state. Moreover, the proportions these states (Fig. 5F). Thus, molecularly defined methods give the same response to fear-inducing
of neurons selective for each behavioral state

Xu et al., Science 370, eabb2494 (2020) 16 October 2020 7 of 15

RESEARCH | RESEARCH ARTICLE

AB CD

neuronal dynamics of 1 cell type *** Eat-Hng 45 22 22 11

60 Eat-Aft 67 22 11 Average Response (t-auROC) 1
0.5
11 behavioral states Drink-T 11 45 11 22 11
0
MCm (m=1...10) or All Accuracy (%) Drink-Aft 11 11 22 22 34 100 -0.5
True 50
40 Eat-Hed 22 22 56 0

123 nm Fear 11 89

Tone-Aft 11 78 11

Choose 1 20 Spout-Aft 33 11 56
neuron 0
SalineGhrelin100
1 Xi Ghrelin
Saline 11 11 11 56 11
-1 Leptin
Decode behaviorial states Spout-Aft
Drink-Aft
by its activity profile Eat-Aft
Drink-T
Eat-Hed
Eat-Hng
Tone-Aft
Fear

Leptin
Saline
Ghrelin
Spout-Aft
Tone-Aft
Fear
Eat-Hed
Drink-Aft
Drink-T
Eat-Aft
Eat-Hng

MC8-Penk
MC2-Reln
MC9-Gad2
MC1-Ntng1
MC3-Vglut2
MC10-Trh
MC7-Low
All
MC4-Npy1r
MC6-Pdyn
MC5-Crh
Leptin 11 11 22 56

-1

Predicted

E FG

combined neuronal dynamics of 10 cell types 11 behavioral states Eat-Hng 89 1 3 51 1 100
MC10 80
MC1 MC2 Eat-Aft 94 1 1 1 11 1 60 p=0
10 types 10 dummy types
12 n1 1 2 n2 12 n10 Drink-T 3 92 1 2 2

Drink-Aft 1 98 1 Accuracy (%)

True Eat-Hed 6 1 2 1 89 1 100
50
Fear 1 1 1 1 95 1 0

1 X1 X2 X10 Tone-Aft 1 2 95 2
-1
Spout-Aft 1 2 2 95

Ghrelin 100

Saline 4 1 2 132 86 1

Leptin 2 1 12 1 1 92

Leptin
Saline
Ghrelin
Spout-Aft
Tone-Aft
Fear
Eat-Hed
Drink-Aft
Drink-T
Eat-Aft
Eat-Hng
Predicted
H I K
J
MC1-Ntng1 Eat-Hng
MC2-Reln Eat-Hed Fear Max consistent-response

MC3-Vglut2 1
MC4-Npy1r 0.5
0.25
MC5-Crh 0.05
MC6-Pdyn -0.05
MC7-Low -0.25
MC8-Penk -0.5
MC9-Gad2 -1
MC10-Trh
Ghrelin

Normalized
decoding weight

1

0.5
0.25
0.1
0.05
<0.01

Fig. 6. Decoding behavioral states with neuron dynamics of molecularly matrix with the combined neuronal dynamics of 10 neurons using the
defined PVH cell types. (A) Schematic procedure for decoding behavioral procedure in (E). (G) Decoding accuracies with the combined neuronal
states with the temporal dynamics from individual cell types or all dynamics of 10 neurons from 10 PVH cell types or from 10 dummy cell types
neurons (All, disregards cell type). (B) Decoding accuracies for all behavioral that scramble cell type information (fig. S21B). Box plots show the median,
states using the temporal dynamics of one cell type or All neurons (Chi-squared interquartile range, and minimum to maximum values of the distributions
test followed by the Marascuillo procedure was used). (C) Normalized of decoding accuracies (Wilcoxon rank-sum test). (H to K) Cell type ensemble
confusion matrix for behavioral state decoding using MC5-Crh neurons. response-decoding diagrams for homeostatic and hedonic eating, fear, and
(D) Response amplitude differences of MC5-Crh neurons across behavioral ghrelin injection. Diagrams with temporal maximum of consistent-response
states. Blue lines are mean responses; error bars are SEM; gray connected (proportional to line width) for each cell type and their decoding weights
circles are individual MC5-Crh neuron responses (one-way repeated-measures (proportional to arrowhead area) for hunger eating (H), hedonic eating (I),
ANOVA followed by Tukey-Kramer test). (E) Schematic of procedure for decoding fear retrieval (J), and ghrelin injection (K). Behavioral state decoding is
behavioral states with the combined temporal response profiles of the for all 11 behavioral states (also see fig. S24). *P < 0.05, **P < 0.01,
PVH cell types using one neuron from each cell type. (F) Average confusion ***P < 0.001. Statistics are provided in table S2.

Xu et al., Science 370, eabb2494 (2020) 16 October 2020 8 of 15

RESEARCH | RESEARCH ARTICLE

AB C Ghrelin
Eat-Hng Gene expression Ghrelin Gene expression Pdyn+
1 Response (t-auROC) 100
Act-FC Cumulative percentage of positive neurons Act-FC
Inh-FC 50 ** Inh-FC

0

-1 0
100 Crh+

**
50

Inh-FC Act-FC 0.3 Normalized expression 0
0.15 100 Npy1r+
0
50 *****

10 Sst 0
Distance Trh 0 0.5 1
Ntng1 Normalized Expression
Vglut2
Gad2
Reln
Penk
Npy1r
Crh
Pdyn

Sst
Trh
Ntng1
Vglut2
Gad2
Reln
Penk
Npy1r
Crh
Pdyn
sshppuoonuugtteionruetating
1 min 1 0 ITI ITI ITI
Distance (inject)

D Eat-Aft Drink-T Drink-Aft Eat-Hed Fear Tone-Aft Spout-Aft Ghrelin Saline Leptin

Eat-Hng

Act-FC
Inh-FC
Vglut2
Gad2
Npy1r
Crh
Reln
Ntng1
Pdyn
Penk

0 2 4 0 2 40 20 50 50 5 0 50 5 0 10 0 2 0 1 2

-log10(p value)

Fig. 7. Functional clustering of PVH neurons and differential enrichment of neurons (middle), and Npy1r in Npy1r+ neurons (bottom) between FCs in (B)
marker genes in 11 behavioral states. (A and B) Hierarchical clustering of PVH after ghrelin injection. (D) Enrichment of marker genes in the FCs across
neurons based on their responses while eating in a hunger state (A) and after 11 behavioral states. Gene enrichments were ranked by –log10(P value). Gene
identity is indicated by bar outline, and the bar fill color indicates the enriched
ghrelin injection (B). Right, gene expression profiles of individual neurons. Blue FC. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 10−4, *****P < 10−5. Statistics
are provided in table S2.
bar marks the behavioral state. Temporal scale bar, 1 min. (C) Comparisons of
expression-level distributions of Pdyn in Pdyn+ neurons (top), Crh in Crh+

stimuli (35, 36). However, low-expressing namics of cell type activity patterns. For ex- decoded from the PVH ensemble dynamics of
Crh+/Vglut2+/Npy1r− neurons in the pPVH ample, MC5-Crh and MC6-Pdyn showed similar molecularly defined cell types. Because cell type
were inhibited in response to food ingestion consistent-responses during hunger eating, encoding of behavioral states has been tradi-
(fig. S20), and this cell type is more prevalent drinking, and hedonic eating (absolute dif- tionally examined either one cell type at a time
in the mPVH (fig. S20, F and G), which is the ferences ≤0.18; Fig. 5H), but their temporal (37–39) or was used without regard to cell type
major target region in the reported photometry response profiles were significantly different (2, 3, 40), we first examined the decoding
experiments (35, 36). Thus, CaRMA imaging (Fig. 5G). MC5-Crh neurons ramped to max- performance of individual molecularly defined
is a powerful platform for systematically imum magnitude, whereas MC6-Pdyn neu- cell types to distinguish 11 behavioral states.
discovering the relationship between cell types rons responded more quickly to the ingested
defined by multiple gene markers and their stimuli (Fig. 5G). Similarly, the return to base- We used multinomial logistic regression
response consistency in different behavioral line after withdrawing food, water, or the offset classification to determine the overall perform-
states. of the fear retrieval cue was slower for MC5-Crh ance of predicting behavioral states from the
(fig. S18A). Although these cell types have sim- temporal dynamics of single neurons with-
The consistent-response map shows the re- ilar consistent-response measures, their dif- out regard to cell type (All) as well as from
lationships between molecularly defined cell ferent temporal dynamics indicate distinct single neurons of known cell types (Fig. 6A;
type activity patterns for behavioral state coding functions for encoding behavioral states, as is see the materials and methods). Single neu-
(Fig. 5H). From these data, it is apparent that also the case for other cell types (fig. S18A). rons from several cell types showed decoding
behavioral state is not represented by a single performance superior to a classifier trained
molecularly defined cell type (fig. S19D). Instead, Decoding behavioral states with molecularly using all neurons without regard to cell type
the consistent-response pattern indicates com- defined cell types (Fig. 6B and fig. S21A). The decoding accuracy
binatorial coding. What is different from pre- Individual cell type coding from the dynamics of MC5-Crh neurons was
vious views of hypothalamic circuits is that the highest (Fig. 6B), and the confusion matrix
behavioral state is encoded by the different CaRMA imaging enabled us to investigate for behavioral state decoded by MC5-Crh
combinations, magnitudes, and temporal dy- whether behavioral states can be quantitatively neurons showed perfect decoding of the

Xu et al., Science 370, eabb2494 (2020) 16 October 2020 9 of 15

RESEARCH | RESEARCH ARTICLE

A B
gene expression
profile of neuron i Predictors Supervised Classification Response
Normalized Features Functional Clusters
Gi (g1i , g2i ... gmi) Logistic regression Functional
Feature Training, Cross-Validation, Clustering
Rescaling Feature Selection Neuronal Dynamics
Voxels, Mean intensity Confusion Matrix,
Sum intensity Accuracy, F1 Score,
Feature auROC
Extraction
FISH Signals

Inh-FC Act-FC

Predict

C Eat-Hng Eat-Aft Accuracy Eat-Hed Fear
60 70 50 60 70 50 Drink-T (%) Drink-Aft
50 50 60 70 50 60 70
All 60 70 50 60 70 Trh
SFFS1
Pdyn Npy1r Npy1r Vglut2 Crh Npy1r Npy1r Penk
0 Penk Crh Ntng1 Npy1r Vglut2 Crh Pdyn
50 Ntng1
All Trh Trh Reln Pdyn Vglut2 Penk Reln
SFFS1 Npy1r Gad2 Crh Penk Ntng1 Crh
Ntng1 Pdyn Npy1r
0 Penk Gad2
Vglut2

0.3 0.6 0 0.3 0.6 0 0.3 0.6 0 0.3 0.6 0 0.3 0.6 0 0.3 0.6 Act-FC
Average coefficient 50
Tone-Aft Spout-Aft Inh-FC
60 70 50 60 70 50 Ghrelin Saline Leptin Improvement
60 70
60 70 50 60 70

Npy1r Crh Npy1r Npy1r Npy1r Npy1r Accuracy (%) 70
Crh Pdyn Pdyn Penk 60
Trh Vglut2
Pdyn Reln Vglut2
Gad2 Gad2
Ntng1

0.3 0.6 0 0.3 0.6 0 0.3 0.6 0 0.3 0.6 0 0.3 0.6 50 SFFS1
All

Average coefficient significance level

DE F G Ghrelin
95th percentile GCaMP Npy1r
real optimal of shuffle all positive expression threshold 0.7 1
Target precision
80 p =0 0 0 0 0 0 0 0 0 0 5e-4 1
Pdyn+
Npy1r+ 2
Npy1r+
Accuracy (%) 3
Crh+
60 0.5 Npy1r+ 0.35 4
5
Npy1r+
Npy1r+ 6
Npy1r+
Npy1r+
Npy1r+
Npy1r+

Response purity

40 7
0
Leptin 08
Saline all expression ITI 1 min 1 t-auROC
Ghrelin 0
Spout-Aft positive threshold ID injection ITI -1
Tone-Aft
Fear
Eat-Hed
Drink-Aft
Drink-T
Eat-Aft
Eat-Hng

Saline
Drink-T
Eat-Aft
Tone-Aft
Leptin
Spout-Aft
Eat-Hng
Drink-Aft
Eat-Hed
Fear
Ghrelin

Fig. 8. Gene expression predicts FCs in multiple behavioral states. for 11 behavioral states (real, red circles) and their corresponding 95th percentile
(A) Schematic of FCi prediction for each neuron solely using its gene of shuffled accuracies (violin plots). Dashed line indicates 50% accuracy.
expression profile (Gi). (B) Supervised learning schematic using logistic (E) Optimal precision for FC prediction based on expression of a single gene in
regression to predict FCs of PVH neurons by their gene expression profiles each behavioral state [first gene in SFFS1 from (C)]. Comparisons between using
(see the materials and methods). (C) Predictive accuracies (blue scale) for all neurons expressing that gene (orange) and using the subset of neurons
functional classes using all marker genes (All) and the optimal marker gene above the optimal expression threshold of that gene (green). (F) Comparison of
set from SFFS1, with stacked bars showing the contribution of individual response purities from all neurons expressing the most predictive gene and the
genes (color legend, right). Bottom subpanels show the average coefficients subset of neurons above the optimal expression threshold. Each datapoint is
(red scale) of the SFFS1 genes (see the materials and methods). Stem plot response purity in one behavioral state. (G) Example responses from predicted
colors indicate to which FC the expression of a gene is positively related. Inh-FC neurons after ghrelin injection using Npy1r+ neurons above the expression
Bottom right panel shows that SFFS1 significantly improved predictive threshold for a subset of cells in fig. S32A. Neuron 8 is an example of
accuracy over using the full marker gene set (All) across all states. misclassification by the model (i.e., it was activated despite high Npy1r expression).
(D) Predictive accuracies of the optimal gene set for FC prediction ranked *P < 0.05, **P < 0.01, ***P < 0.001. Statistics are provided in table S2.

Xu et al., Science 370, eabb2494 (2020) 16 October 2020 10 of 15

RESEARCH | RESEARCH ARTICLE

A SstEat-Aft Drink-AftSpout-Aft GhrelinSalineLeptin
Trh
0.3 Ntng1Trh all
0.2 Vglut2
0.1 Gad2Penkoptimal
Reln
0FDE PenkPdynsignificance level
Npy1rNtng1most important
B9 CrhReln2nd most important
PdynCrh
0
Tone-Aft Npy1r
C Fear
Gad2 FDE:
Eat-Hed
Drink-T Vglut2 fraction of deviance explained
Eat-Hng
Rnds DE F 0.15 0.3 Normalized expression

0

MC11-Sst
MC10-Trh
MC9-Gad2
MC8-Penk
MC7-Low
MC6-Pdyn
MC5-Crh
MC4-Npy1r
MC3-Vglut2
MC2-Reln
MC1-Ntng1

1 min 0 0.5 1
-1 0 1 FDE
Response (t-auROC)

Fig. 9. Gene expression profiles of individual PVH neurons predict their gene expression profiles. (B) Number of SFFS rounds with FDE above
temporal responses in multiple behavioral states. (A) Prediction significance level in each timestamp. (C) Responses of PVH neurons ranked
performance (fraction of deviance explained, FDE) of neuronal response at by FDE across behavioral states. (D) FDE ordered high to low of individual
each timestamp using expression levels of all marker genes, the optimal gene neurons for the entire time series across all behavioral states. (E and F) Gene
set, and the two most important genes measured by mSFFS. The optimal expression profiles (E) and molecularly defined cell types (F) of the
gene set was composed of the genes that provided the highest FDE from mSFFS corresponding neurons in (D). Black or white lines in (C) to (F) indicate the
at each timestamp. Significance level is the 95th percentile FDE after shuffling boundary of the most highly predictive quartile.

ghrelin-induced hunger state and high de- states using 10-fold cross-validation (accuracy: gether, our findings support a grouped-ensemble
coding performance for fear retrieval (Fig. 6C). 93.13 ± 0.02%, Fig. 6, E and F; see the materials coding strategy used by molecularly defined
MC5-Crh neurons showed significantly differ- and methods). With molecularly defined cell PVH neurons (Fig. 1C).
ent response amplitude and sign across several type information, 11 different behavioral states
behavioral states (Fig. 6D), which explained could be accurately decoded with the combi- Cell type ensemble response-decoding
the high decoding performance. However, our natorial responses of only 10 PVH neurons, one diagrams
analysis also showed that individual cell types from each of the 10 cell types. This was sig-
in the PVH are insufficient to decode many nificantly higher decoding accuracy (P = 0; To reveal the relative contributions of differ-
behavioral states. Fig. 6G) than a neural ensemble that ignores ent molecularly defined cell types to grouped-
molecularly defined cell type information by ensemble coding (Fig. 1C) for different behavioral
Grouped-ensemble cell type randomly assigning neurons into dummy cell states in the PVH, we used the multinomial
ensemble coding types (fig. S21B). regression coefficients and the consistent-
response measurements to generate cell type
CaRMA imaging enables the activity of many We also examined the effect of adjusting the response-decoding diagrams for each behav-
cell types to be imaged simultaneously in the hierarchical molecularly defined cell type ioral state (Fig. 6, H to K, and fig. S24). In these
same animal. This permitted us to test models clustering threshold on behavioral state de- diagrams, cell type consistent-responses are
of behavioral state coding using an ensemble coding. Purity and decoding accuracy generally weighted by each cell type’s contribution to
of molecularly defined cell types (Fig. 1C) or, increased with the number of molecular clus- decoding accuracy and quantitatively summar-
alternatively, an ensemble of neurons in which ters (figs. S22 and S23). Although grouped- ized across multiple behavioral states. MC5-Crh
molecularly defined cell type information was ensemble coding was supported with a range and MC6-Pdyn neurons had large consistent-
ignored (Fig. 1B). We trained a classifier on an of clustering thresholds, this approach involves responses in several behavioral states (23, 35, 36).
ensemble of molecularly defined cell types “peeking” at the functional responses to super- These neurons were activated by negative and
model by randomly choosing one cell’s activity vise molecular clustering. Because our objec- positive stimuli, indicating an unexpected gen-
trace from each of 10 molecularly defined cell tive was to evaluate the predictive value of eral role for encoding salience of diverse
types and concatenating them into 10-cell en- molecular information on functional and be- stimuli in a variety of different states. Other
sembles for each behavioral state (repeated havioral responses, we proceeded using the molecularly defined cell types, such as MC8-
100 times with replacement), followed by mul- unbiased determination of 11 MCs in the pPVH Penk and MC10-Trh neurons, contributed sub-
tinomial logistic regression for 11 behavioral ensemble (to avoid circular logic). Taken to- stantial weights for distinguishing behavioral

Xu et al., Science 370, eabb2494 (2020) 16 October 2020 11 of 15

RESEARCH | RESEARCH ARTICLE

A B if Npy1r, Pdyn, and Crh were enriched for a
behavioral state, then they were found in the
Modulated grouped-ensemble coding same FC because of partial coexpression in the
same cells. For example, Npy1r, Pdyn, and Crh
PVH neurons were enriched in the Act-FC for Eat-Hed
Output (hedonic) and Fear states. By contrast, the
Inh-FC was enriched with neurons expressing
behavioral state Penk. We also examined combinations of mo-
lecularly defined cell types within FCs in each
Input Type 1 B1 behavioral state. For most behavioral states
I1 Type 2 B2 (eight of 11 states), activated or inhibited FCs
B3 also showed significant enrichment of cell
Type 4 Type 2 I2 types (fig. S28). The cell types with the highest
FC selectivity were MC5-Crh and MC6-Pdyn,
I3 which also showed high consistent-responses
in most behavioral states (Fig. 5H). These gene
Type 3 and cell type enrichment analyses demon-
Il strated a strong association between neuron
function during behavior and the molecular
Bn profile of individual neurons. However, this
analysis lacks useful metrics for ranking the
Type 1 Type 3 Type m predictive power of gene expression for func-
tional response class, and it does not define
Neuromodulator Receptor Type m Neuromodulator the optimal set of marker genes for predict-
Receptor ing responses across the neural ensemble.

Fig. 10. Cooperative regulation of multiple cell types by neuromodulation. (A) Diagram illustrating Gene expression prediction of FCs
volume diffusion of a neuromodulator to selectively regulate subgroups of cell types expressing its receptor.
(B) Schematic of modulated grouped-ensemble coding model for multiple behavioral states that includes We therefore determined the best gene com-
the role of neuromodulation. Neuromodulator release cooperatively regulates multiple PVH cell types binations for predicting neuron functional re-
expressing its receptor. This coding configuration highlights the relationship between the hierarchical sponse classes in each behavioral state (Fig. 8A).
functional organization of PVH neurons and their molecular hierarchy. We used supervised machine learning to predict
the FCs of individual PVH neurons solely by
states (Fig. 6, H to K, and fig. S24C). For ex- have coordinated activity patterns that com- their gene expression profiles with a logistic
ample, food withdrawal (Eat-Aft) in hunger prise hierarchical functional groupings in the regression classifier (Fig. 8B; see the materials
elicited a transient inhibitory response in neural ensemble. This raises the second ma- and methods). Three normalized features rep-
MC10-Trh after spout retraction that transi- jor question: Do differentially expressed genes resented the expression level of each gene
tions to a prolonged activated response, which map to functional response types in the neural (figs. S29 and S30A). This approach does not
was not observed after removing water in ensemble? binarize gene expression by thresholding.
thirst (Drink-Aft) (fig. S25A). This contributes Instead, it explicitly incorporates the broad
to behavioral state coding (fig. S25B) and is We aimed to identify optimal combina- continuous distribution of gene expression
also consistent with past work indicating that tions of genes that best predict the functional levels that we found (Fig. 2H) as the inde-
TRH neurons project to the arcuate nucleus response classes within the neuronal ensem- pendent variables in the predictive model.
and activate hunger-associated Agouti-related ble. Moreover, neurons have complex dynamics Sequential forward feature selection (SFFS)
peptide (AGRP) neurons (22). By contrast, most during behaviors, so we assessed the temporal determines the most predictive individual gene
MC8-Penk neurons showed prolonged inhibi- window over which molecular markers offer for FC classification and sequentially adds the
tion after food withdrawal in hunger but only useful predictive information about functional next most predictive gene until predictive ac-
transient inhibition after drinking in thirst (fig. responses. curacy no longer increases. This approach iden-
S25, C and D). More generally, these response- tified optimal gene sets that best predicted the
decoding diagrams show a functional orga- Differential gene expression in functional response class while minimizing the
nization in the pPVH, with major branches functional clusters contribution of redundant gene coexpression
associated with the sign and amplitude of information (Fig. 8C and fig. S30, B to D). We
broadly tuned MC5-Crh and MC6-Pdyn neurons For each behavioral state, we grouped all neu- compared the predictive power of the molecu-
and individual twigs associated with a more rons on the basis of their neuronal response lar profiles for functional response class with
selective combinatorial code of cell types that dynamics into functional clusters (FCs) by un- the 95th percentile of predictive accuracies
distinguish behavioral states. supervised hierarchical clustering (Fig. 7, A after shuffling the relationship between gene
and B, and fig. S26). Two FCs emerged for each expression of individual neurons and their
Gene expression prediction of ensemble behavioral state associated with primarily acti- FCs (Fig. 8, C and D, and fig. S30B). The two
functional responses vated (Act-FC) or inhibited (Inh-FC) neurons. behavioral states during which neuronal en-
We first examined how functional clustering semble response could be best predicted by
Grouped-ensemble coding relies on similar re- segregates neuron gene expression profiles. gene expression profiles were during ghrelin-
sponses from the neurons that make up a cell induced hunger [accuracy: 73%, area under
type, thereby simplifying the neural ensemble A subset of gene expression profiles specif- the receiver-operating characteristic (auROC):
to component cell types instead of the much ically segregated with functional response clus- 0.784, P = 0) and fear retrieval (accuracy:
larger total number of cells. Consistent-response ters (Fig. 7, A to C, and fig. S26). We performed 69.9%, auROC: 0.745, P = 0) states. The lowest
maps also indicate higher-level organization differential gene expression analyses between
based on similar responses between groups of the two FCs in each behavioral state (fig. S27A),
cell types. For example, MC5-Crh and MC6- which was measured by either the P value of
Pdyn show similar response tuning to multiple the expression-level difference or by the ratio of
behavioral states. By contrast, MC8-Penk typi- expression level between FCs (Fig. 7D and fig.
cally responds oppositely to these cell types S27B). Npy1r was significantly enriched across
(Figs. 5H and 6, H to K, and fig. S24C). Thus, 11 behavioral states and was ranked as the first
different molecularly defined cell types may or second enriched gene by P value. Crh and
Pdyn were significantly enriched in eight and
nine behavioral states, respectively. Moreover,

Xu et al., Science 370, eabb2494 (2020) 16 October 2020 12 of 15

RESEARCH | RESEARCH ARTICLE

predictive accuracy was for saline injection concatenated across all 11 behavioral states dictions of neural activity, we ranked neu-
(accuracy: 62.1%, auROC: 0.570, P = 0.0005) (fig. S31). rons by the fraction of deviance explained
(Fig. 8D and fig. S30D). with our linear regression models along the
Gene expression levels for improved entire time series across behavioral states
Predictive contribution of individual genes FC prediction (Fig. 9, C and D, fig. S33A). A majority of
neurons (65%) showed statistically signifi-
Combinations of genes always had higher pre- Using the quantitative gene expression level cant predictive power. Neurons within the
dictive power for functional response class to predict functional responses also offers a top quartile of predictive powers showed sig-
than single genes (fig. S30B). In addition, for systematic method for defining the gene ex- nificantly higher expression levels of Vglut2,
all behavioral states, the optimal accuracies pression threshold that targets a functional Gad2, Npy1r, Crh, Reln, Pdyn, and Trh (fig.
from the first round of SFFS (SFFS1) with response class with the best precision for each S33D). Next, we remapped the corresponding
just two to six genes were higher than the behavioral state. We identified optimal cut gene expression profiles onto molecularly de-
accuracies predicted with all genes (Fig. 8C). points for the gene expression threshold using fined cell types (Fig. 9, E and F). MC4-Npy1r,
This showed that a subset of genes provided Youden’s J statistic with the ROC curve (fig. MC5-Crh (exclusively), and MC6-Pdyn neu-
higher prediction performance than the whole S32). Prediction precision with a single gene rons were enriched in the most highly pre-
set of genes (41). Next, we quantitatively as- was up to 40% higher, and response purity was dictive quartile (fig. S33E), consistent with
sessed and ranked the predictive accuracy of significantly improved when the optimal ex- the high predictive accuracy associated with
single genes for FCs across 11 behavioral states. pression threshold was used instead of a lower Npy1r coexpression.
Pdyn, Npy1r, or Crh showed the highest pre- threshold based on expression over back-
dictive power in the first round for all behav- ground (Fig. 8, E to G). Thus, neuronal func- DISCUSSION
ioral states. Npy1r was always in the optimal tional identity is not a binary quality of gene
gene set from SFFS1 in all behavioral states. expression. Neural coding by an ensemble-of-cell-types in-
Moreover, Npy1r expression strongly predicted dicates that many molecularly defined neu-
fear-activated neurons. We determined the Temporal profile of gene → functional rons also form functional groupings that work
roles of individual genes within the optimal response prediction in concert to encode behavioral state. It is
gene set to improve prediction performance unlikely that coding in the pPVH follows the
by measuring the average coefficients of those Next, we investigated how the predictive power labeled-line model because we could not find
genes from logistic regression (Fig. 8C). For of gene expression profiles was related to neu- such neurons in the ensemble irrespective of
ghrelin-induced hunger, SFFS1 showed an ad- ron temporal dynamics. For example, does a molecular identity. Our results validate the use
ditive relationship for Npy1r and Pdyn predic- gene have similar predictive value throughout of molecularly defined cell types to reduce the
tion of the inhibited functional response class a state? For each timestamp (0.4-s interval), dimensionality of the PVH neuronal ensemble
(Fig. 8C). Both genes have been previously we modeled the relationship of gene expres- to understand behavioral state coding. Ensemble-
associated with appetite control by the PVH sion level to neuronal response (Fig. 9, A to C, of-cell-types coding of behavioral states provides
(42). For food consumption during hunger and fig. S33A). Our results complemented those superior efficiency, scalability, and flexibility
(Eat-Hng), Pdyn and Npy1r predicted the ac- obtained from prediction of FCs, including (i) than labeled-line coding and offers superior
tivated response type, whereas the inhibited higher predictive power of combinatorial ex- robustness (due to redundancy) than cell type–
response class was predicted by Trh and Penk pression profile (Fig. 9A), (ii) improvement in independent, full-ensemble coding. Operation-
(Fig. 8C). For fear retrieval, Npy1r and Crh prediction performance by mSFFS (Fig. 9A), ally, grouping by molecularly defined cell types
best predicted activated neurons and Penk pre- (iii) redundant information in molecular pro- is justified by high consistent-responses and is
dicted fear-inhibited neurons (Fig. 8C). files for predicting temporal response (Fig. 9B), analogous to the pooling operation in artificial
and (iv) that Npy1r predictive power was high neural networks, which is used to extract in-
Multiple rounds of SFFS (mSFFS; see the and statistically significant across all behavioral variant information (43–45). This may also be a
materials and methods) revealed genes that states (Fig. 9A). However, predictive power de- key function when molecularly defined circuits
could compensate for the removal of the most pended on cell type–specific neuronal dynam- are used for coding behavioral states, which
predictive gene from the previous rounds (fig. ics. During eating and drinking, the predictive should be invariant to irrelevant external con-
S30B). This indicated that the most predictive power of marker genes increased as consump- ditions. Taken together, these cell types and
genes could be largely compensated by combi- tion progressed. For example, Pdyn neuron the patterns of activity reported here provide
nations of additional genes in subsequent activity rose more quickly than Pdyn predictive a parts list for systematically examining the
rounds. For example, Pdyn was most predic- accuracy for Eat-Hed and Eat-Hng. In this neural coding of behavioral state by the PVH.
tive during Eat-Hng, but this predictive ac- case, the transient activation at onset of food CaRMA imaging offers a quantitative method
curacy was largely replaced by Crh in round 2 presentation was not specific within Pdyn- to evaluate the extent to which other brain
and Npy1r in round 3 of mSFFS, and we found expressing neurons but was also widespread regions rely on molecularly defined cell type
that Npy1r neurons coexpressed Pdyn and Crh in Pdyn-negative neurons (fig. S33B). This ensemble coding.
in a hierarchical relationship in pPVH (fig. S30, shows that cell type–specific responses de-
E and F). Consistent with this, Npy1r showed velop progressively in some behaviors. By Using CaRMA imaging, we demonstrate an
statistically significant predictive power for contrast, fear retrieval showed an immediate important role for gene expression informa-
every behavioral state, but if Npy1r was re- cell type–specific response that was best pre- tion to predict both neuron functional re-
moved from the analysis, its predictive power dicted by Npy1r and gradually became less sponse class as well as temporal dynamics
could be largely replaced by other genes (fig. well predicted by gene expression (fig. S33B). within the PVH neural ensemble. High pre-
S30B). Thus, the hierarchical gene expression dictive power of Npy1r for neuron activity
relationship of these three genes in the pPVH In five of 11 behavioral states (Eat-Hng, was observed with all analysis methods. How-
was associated with similar neuron functional Drink-Thirst, Eat-Hed, Fear, and Ghrelin), a ever, combinations of other cell type marker
responses. Neurons with low or no Npy1r ex- single gene had the greatest predictive accu- genes also showed significant prediction for
pression typically responded oppositely. We racy throughout a behavioral state. However, functional responses. The highest consistent-
also found similar results for the predictive for six other states, the most predictive gene response cell types, MC5-Crh and MC6-Pdyn,
power of marker genes for neuron FCs for the ensemble functional response switched coexpressed Npy1r, and cell types that responded
with time. Moreover, using the optimal pre-

Xu et al., Science 370, eabb2494 (2020) 16 October 2020 13 of 15

RESEARCH | RESEARCH ARTICLE

oppositely typically lacked Npy1r. scRNA-seq expression levels that are optimal for target- 19. R. A. Romanov et al., Molecular interrogation of hypothalamic
analysis revealed hierarchical relationships of organization reveals distinct dopamine neuronal subtypes.
gene expression within neuron populations, ing specific functional perturbations. Nat. Neurosci. 20, 176–188 (2017). doi: 10.1038/nn.4462;
where the smallest subpopulations are des- pmid: 27991900
ignated as “types” and the largest subdivi- CaRMA imaging provides a solution to the
sions as “classes” that encompass many types 20. L. W. Swanson, P. E. Sawchenko, Hypothalamic integration:
(46). We found evidence that this molecular general problem of assessing functional roles Organization of the paraventricular and supraoptic nuclei.
hierarchy extends to functional responses in Annu. Rev. Neurosci. 6, 269–324 (1983). doi: 10.1146/
the neuronal ensemble, in which similarly for the large number of new molecularly de- annurev.ne.06.030183.001413; pmid: 6132586
tuned neuron types form classes that can be
well described by a neuromodulator receptor fined cell types that are being discovered 21. M. M. Li et al., The paraventricular hypothalamus regulates
gene that is expressed across multiple molec- satiety and prevents obesity via two genetically distinct
ularly defined cell types (fig. S30, E and F). routinely using single-cell transcriptomics. circuits. Neuron 102, 653–667.e6 (2019). doi: 10.1016/
This is interesting in the context of the neu- j.neuron.2019.02.028; pmid: 30879785
romodulatory function of Npy1r and its en- Because CaRMA imaging does not rely on
dogenous agonist NPY (47). Because Npy1r is 22. M. J. Krashes et al., An excitatory paraventricular nucleus to
expressed across multiple cell types and NPY genetic modifications, only viral targeting, it AgRP neuron circuit that drives hunger. Nature 507, 238–242
signals through volume transmission (48), this (2014). doi: 10.1038/nature12956; pmid: 24487620
receptor can cooperatively regulate diverse cell is applicable to any organism suitable for
types (Fig. 10A). This may enhance the gener- 23. T. Füzesi, N. Daviu, J. I. Wamsteeker Cusulin, R. P. Bonin,
ality of neural output in the presence of noisy in vivo functional imaging. By producing un- J. S. Bains, Hypothalamic CRH neurons orchestrate complex
neural inputs, which is important to facilitate behaviours after stress. Nat. Commun. 7, 11937 (2016).
behavioral state coding but is also a computa- biased and systematic datasets that integrate doi: 10.1038/ncomms11937; pmid: 27306314
tional role akin to regularization in machine
learning. Nevertheless, despite the response genes, neuron dynamics, and behavior, CaRMA 24. C. B. Saper, B. B. Lowell, The hypothalamus. Curr. Biol. 24,
similarity between neurons that express Npy1r, R1111–R1116 (2014). doi: 10.1016/j.cub.2014.10.023;
there are reliable temporal differences that imaging offers a path to discoveries about the pmid: 25465326
map onto cells with specific gene expression
profiles. These temporal responses are impor- organization of brain function at both the 25. K. K. Ishii et al., A labeled-line neural circuit for pheromone-
tant for distinguishing behavioral states and mediated sexual behaviors in mice. Neuron 95, 123–137.e8
functional response types in the neuronal molecular level and the systems level. (2017). doi: 10.1016/j.neuron.2017.05.038; pmid: 28648498
ensemble. In the analogy of an orchestral
ensemble, Npy1r is a conductor selectively REFERENCES AND NOTES 26. A. K. Graebner, M. Iyer, M. E. Carter, Understanding how
coordinating the responses of multiple cell discrete populations of hypothalamic neurons orchestrate
types (Fig. 10B) that are like instruments 1. C. Stringer et al., Spontaneous behaviors drive complicated behavioral states. Front. Syst. Neurosci. 9, 111
that play together but maintain distinctive multidimensional, brainwide activity. Science 364, 255 (2019). (2015). doi: 10.3389/fnsys.2015.00111; pmid: 26300745
contributions to a score. doi: 10.1126/science.aav7893; pmid: 31000656
27. S. M. Sternson, Hypothalamic survival circuits: Blueprints for
Marker-gene expression in the PVH can be 2. J. E. Allen, T. E. Sutherland, Crystal-clear treatment for allergic purposive behaviors. Neuron 77, 810–824 (2013). doi: 10.1016/
dynamic (49, 50), but to the extent that this disease. Science 364, 738–739 (2019). doi: 10.1126/ j.neuron.2013.02.018; pmid: 23473313
affects our results, it may degrade the pre- science.aax6175; pmid: 31123126
dictive accuracy of gene expression level to a 28. D. O. Hebb, The Organization of Behavior (Erlbaum, 1949/
functional or behavioral response. Additional 3. J. Gründemann et al., Amygdala ensembles encode behavioral 2002).
genetic markers (51, 52) may improve these states. Science 364, eaav8736 (2019). doi: 10.1126/
predictions by offering increased coverage of science.aav8736; pmid: 31000636 29. J. R. Moffitt et al., Molecular, spatial, and functional single-cell
all neurons in the ensemble. However, pre- profiling of the hypothalamic preoptic region. Science 362,
dictive accuracies solely from gene expression 4. X. Chen et al., Brain-wide organization of neuronal activity eaau5324 (2018). doi: 10.1126/science.aau5324;
information are not expected to be perfect and convergent sensorimotor transformations in larval pmid: 30385464
because functional responsiveness is also zebrafish. Neuron 100, 876–890.e5 (2018). doi: 10.1016/
influenced by plasticity and connectivity j.neuron.2018.09.042; pmid: 30473013 30. D. Lee, M. Kume, T. E. Holy, Sensory coding mechanisms
mechanisms that are independent of cell revealed by optical tagging of physiologically defined neuronal
type (13, 16, 28). Moreover, our findings also 5. E. S. Lein et al., Genome-wide atlas of gene expression in the types. Science 366, 1384–1389 (2019). doi: 10.1126/science.
highlight potential limitations of site-specific adult mouse brain. Nature 445, 168–176 (2007). doi: 10.1038/ aax8055; pmid: 31831669
recombinase (e.g., Cre) mouse lines that use nature05453; pmid: 17151600
low gene expression thresholds to specify cell 31. M. Lovett-Barron et al., Multiple convergent hypothalamus-
type identity (53, 54). 6. A. Saunders et al., Molecular diversity and specializations brainstem circuits drive defensive behavior. Nat. Neurosci. 23,
among the cells of the adult mouse brain. Cell 174, 959–967 (2020). doi: 10.1038/s41593-020-0655-1;
Although we focused here on decoding be- 1015–1030.e16 (2018). doi: 10.1016/j.cell.2018.07.028; pmid: 32572237
havioral states and predicting functional re- pmid: 30096299
sponses solely from molecular information 32. M. Lovett-Barron et al., Ancestral circuits for the coordinated
about the underlying cells, we found that if 7. A. Zeisel et al., Molecular architecture of the mouse nervous modulation of brain state. Cell 171, 1411–1423.e17 (2017).
functional information is used to supervise system. Cell 174, 999–1014.e22 (2018). doi: 10.1016/ doi: 10.1016/j.cell.2017.10.021; pmid: 29103613
clustering, then molecularly defined cell types j.cell.2018.06.021; pmid: 30096314
can be defined with higher purity (17). This 33. A. M. Kerlin, M. L. Andermann, V. K. Berezovskii, R. C. Reid,
may be useful in the future for better defining 8. B. Tasic et al., Shared and distinct transcriptomic cell types Broadly tuned response properties of diverse inhibitory neuron
combinations of molecular markers and their across neocortical areas. Nature 563, 72–78 (2018). subtypes in mouse visual cortex. Neuron 67, 858–871 (2010).
doi: 10.1038/s41586-018-0654-5; pmid: 30382198 doi: 10.1016/j.neuron.2010.08.002; pmid: 20826316

9. A. R. Adamantidis, F. Zhang, A. M. Aravanis, K. Deisseroth, 34. L. M. Barnett, T. E. Hughes, M. Drobizhev, Deciphering
L. de Lecea, Neural substrates of awakening probed with the molecular mechanism responsible for GCaMP6m’s
optogenetic control of hypocretin neurons. Nature 450, Ca2+-dependent change in fluorescence. PLOS ONE 12, e0170934
420–424 (2007). doi: 10.1038/nature06310; pmid: 17943086 (2017). doi: 10.1371/journal.pone.0170934; pmid: 28182677

10. D. Atasoy, S. M. Sternson, Chemogenetic tools for causal 35. C. Li et al., Defined paraventricular hypothalamic populations
cellular and neuronal biology. Physiol. Rev. 98, 391–418 (2018). exhibit differential responses to food contingent on caloric
doi: 10.1152/physrev.00009.2017; pmid: 29351511 state. Cell Metab. 29, 681–694.e5 (2019). doi: 10.1016/
j.cmet.2018.10.016; pmid: 30472090
11. C. J. Magnus et al., Ultrapotent chemogenetics for research
and potential clinical applications. Science 364, eaav5282 36. J. Kim et al., Rapid, biphasic CRF neuronal responses encode
(2019). pmid: 30872534 positive and negative valence. Nat. Neurosci. 22, 576–585
(2019). doi: 10.1038/s41593-019-0342-2; pmid: 30833699
12. M. Jazayeri, A. Afraz, Navigating the neural space in search
of the neural code. Neuron 93, 1003–1014 (2017). 37. Y. Mandelblat-Cerf et al., Bidirectional anticipation of future
doi: 10.1016/j.neuron.2017.02.019; pmid: 28279349 osmotic challenges by vasopressin neurons. Neuron 93, 57–65
(2017). doi: 10.1016/j.neuron.2016.11.021; pmid: 27989461
13. K. Hardcastle, S. Ganguli, L. M. Giocomo, Cell types for our
sense of location: Where we are and where we are going. 38. J. Y. Cohen, S. Haesler, L. Vong, B. B. Lowell, N. Uchida,
Nat. Neurosci. 20, 1474–1482 (2017). doi: 10.1038/nn.4654; Neuron-type-specific signals for reward and punishment in the
pmid: 29073649 ventral tegmental area. Nature 482, 85–88 (2012).
doi: 10.1038/nature10754; pmid: 22258508
14. C. F. Stevens, Neuronal diversity: Too many cell types for
comfort? Curr. Biol. 8, R708–R710 (1998). doi: 10.1016/ 39. J. N. Betley et al., Neurons for hunger and thirst transmit a
S0960-9822(98)70454-3; pmid: 9778523 negative-valence teaching signal. Nature 521, 180–185 (2015).
doi: 10.1038/nature14416; pmid: 25915020
15. T. E. Holy, “Yes! We’re all individuals!”: Redundancy in neuronal
circuits. Nat. Neurosci. 13, 1306–1307 (2010). doi: 10.1038/ 40. Y. Li et al., Neuronal representation of social information in the
nn1110-1306; pmid: 20975750 medial amygdala of awake behaving mice. Cell 171, 1176–1190.
e17 (2017). doi: 10.1016/j.cell.2017.10.015; pmid: 29107332
16. S. Fusi, E. K. Miller, M. Rigotti, Why neurons mix: High
dimensionality for higher cognition. Curr. Opin. Neurobiol. 37, 41. G. V. Trunk, A problem of dimensionality: A simple example.
66–74 (2016). doi: 10.1016/j.conb.2016.01.010; IEEE Trans. Pattern Anal. Mach. Intell. 1, 306–307 (1979).
pmid: 26851755 doi: 10.1109/TPAMI.1979.4766926; pmid: 21868861

17. A. J. Northcutt et al., Molecular profiling of single neurons of 42. A. Kask, L. Rägo, J. Harro, Evidence for involvement of
known identity in two ganglia from the crab Cancer borealis. neuropeptide Y receptors in the regulation of food intake:
Proc. Natl. Acad. Sci. U.S.A. 116, 26980–26990 (2019). Studies with Y1-selective antagonist BIBP3226. Br. J.
doi: 10.1073/pnas.1911413116; pmid: 31806754

18. W. E. Armstrong, “Hypothalamic supraoptic and
paraventricular nuclei,” in The Rat Nervous System, G. Paxinos,
Ed. (Elsevier, 2004), pp. 369–388.

Xu et al., Science 370, eabb2494 (2020) 16 October 2020 14 of 15

RESEARCH | RESEARCH ARTICLE

Pharmacol. 124, 1507–1515 (1998). doi: 10.1038/ 51. K. H. Chen, A. N. Boettiger, J. R. Moffitt, S. Wang, X. Zhuang, friction treadmill; and N. Spruston, A. Lee, M. Ahrens, and
sj.bjp.0701969; pmid: 9723965 RNA imaging. Spatially resolved, highly multiplexed RNA V. Jayaraman for comments on the manuscript. Funding: Research
43. M. Ranzato, F. J. Huang, Y. Boureau, Y. LeCun, “Unsupervised profiling in single cells. Science 348, aaa6090 (2015). was funded by HHMI (S.M.S.). Author contributions: S.M.S.
learning of invariant feature hierarchies with applications to doi: 10.1126/science.aaa6090; pmid: 25858977 and S.X. conceived the project and designed the experiments
object recognition,” in IEEE Conference on Computer Vision and analyses. S.X. built the two-photon microscope, designed the
and Pattern Recognition (IEEE, 2007), pp. 1–8; 52. S. Shah, E. Lubeck, W. Zhou, L. Cai, In situ transcription behavior control and log system, performed two-photon-GRIN-lens
https://doi.org/10.1109/CVPR.2007.383157. profiling of single cells reveals spatial organization of cells in calcium-imaging experiments, and analyzed the data. H.Y., S.X.,
44. D. H. Hubel, T. N. Wiesel, Receptive fields, binocular interaction the mouse hippocampus. Neuron 92, 342–357 (2016). S.M.S., and F.E.H. developed the multiplex FISH method. H.Y. and
and functional architecture in the cat’s visual cortex. J. Physiol. doi: 10.1016/j.neuron.2016.10.001; pmid: 27764670 S.X. performed the multiplex FISH experiments. A.L.L., L.W., and
160, 106–154 (1962). doi: 10.1113/jphysiol.1962.sp006837; F.E.H. performed scRNA-seq. V.M. analyzed the scRNA-seq data
pmid: 14449617 53. J. A. Harris et al., Anatomical characterization of Cre driver and assisted in the gene and cell type enrichment analysis.
45. Y. LeCun, I. Kanter, S. A. Solla, “Second order properties of mice for neural circuit mapping and manipulation. Front. Neural S.C.T. supervised the machine-learning analyses. S.M.S. and S.X.
error surfaces: Learning time and generalization,” in Advances Circuits 8, 76 (2014). doi: 10.3389/fncir.2014.00076; wrote the manuscript with input from other authors. Competing
in Neural Information Processing Systems 3, R. P. Lippmann, pmid: 25071457 interests: S.M.S., F.E.H., H.Y., and S.X. have filed a patent
J. E. Moody, D. S. Touretzky, Eds. (NIPS, 1990), pp. 396–404. application on multiplexed FISH methods. Data and materials
46. H. Zeng, J. R. Sanes, Neuronal cell-type classification: Challenges, 54. K. Y. Lee et al., Lessons on conditional gene targeting in mouse availability: The GEO submission ID for the single-cell PVH data
opportunities and the path forward. Nat. Rev. Neurosci. 18, adipose tissue. Diabetes 62, 864–874 (2013). doi: 10.2337/ is GSE148568. Code for CaRMA imaging analysis pipelines,
530–546 (2017). doi: 10.1038/nrn.2017.85; pmid: 28775344 db12-1089; pmid: 23321074 example data, preprocessed multiplex FISH data, and
47. H. Herzog, 30 Years of NPY research. Neuropeptides 46, 251 preprocessed CaRMA imaging data are available at GitHub (56).
(2012). doi: 10.1016/j.npep.2012.10.002; pmid: 23141043 55. J. Biag et al., Cyto- and chemoarchitecture of the hypothalamic
paraventricular nucleus in the C57BL/6J male mouse: A study SUPPLEMENTARY MATERIALS
48. S. O. Fetissov, J. Kopp, T. Hökfelt, Distribution of NPY of immunostaining and multiple fluorescent tract tracing.
receptors in the hypothalamus. Neuropeptides 38, 175–188 J. Comp. Neurol. 520, 6–33 (2012). doi: 10.1002/cne.22698; science.sciencemag.org/content/370/6514/eabb2494/suppl/DC1
(2004). doi: 10.1016/j.npep.2004.05.009; pmid: 15337370 pmid: 21674499 Materials and Methods
Figs. S1 to S33
49. R. A. Romanov, A. Alpár, T. Hökfelt, T. Harkany, Molecular 56. S. Xu, H. Yang, V. Menon, A. L. Lemire, L. Wang, F. E. Henry, Tables S1 and S2
diversity of corticotropin-releasing hormone mRNA-containing S. C. Turaga, S. M. Sternson, Code for CaRMA imaging analysis References (57–68)
neurons in the hypothalamus. J. Endocrinol. 232, R161–R172 pipelines, example data, preprocessed multiplex FISH data, Movie S1
(2017). doi: 10.1530/JOE-16-0256; pmid: 28057867 and preprocessed CaRMA imaging data for: Behavioral state MDAR Reproducibility Checklist
coding by molecularly defined paraventricular hypothalamic
50. A. G. Watts, L. W. Swanson, Diurnal variations in the content of cell type ensembles, GitHub (2020); https://github.com/ View/request a protocol for this paper from Bio-protocol.
preprocorticotropin-releasing hormone messenger ribonucleic sternson-lab/CaRMA-imaging.
acids in the hypothalamic paraventricular nucleus of rats of 19 February 2020; accepted 18 August 2020
both sexes as measured by in situ hybridization. Endocrinology ACKNOWLEDGMENTS 10.1126/science.abb2494
125, 1734–1738 (1989). doi: 10.1210/endo-125-3-1734;
pmid: 2788078 We thank V. Goncharov and C. McRaven for assistance with two-
photon microscope construction; M. Barbic and T. Harris for
the one-photon and two-photon excitable grid target; Z. Guo for
GRIN lens testing; A. Hu for brain sectioning; J. Arnold for the low-

Xu et al., Science 370, eabb2494 (2020) 16 October 2020 15 of 15

RESEARCH

◥ in comparison to ileal and colonic iEANs,
express significantly higher levels of tran-
RESEARCH ARTICLE scripts encoding receptors involved in the re-
sponse to proximal EEC- or iEAN-derived signals,
M I C R O B I O TA such as cholecystokinin receptor A (Cckar),
glucagon receptor (Gcgr), and tachykinin re-
Microbiota-modulated CART+ enteric neurons ceptor 3 (Tacr3) (Fig. 1E and fig. S3, A to C).
autonomously regulate blood glucose There are also regional differences in neuro-
peptide transcripts, including neuropeptide
Paul A. Muller1*†‡, Fanny Matheis1*, Marc Schneeberger2*, Zachary Kerner1, Y (Npy) enrichment in duodenum iEANs. Ter-
Veronica Jové3, Daniel Mucida1‡ minal ileum and colon iEANs are enriched in
neuropeptide transcripts, including somato-
The gut microbiota affects tissue physiology, metabolism, and function of both the immune and nervous statin (Sst), cocaine- and amphetamine-regulated
systems. We found that intrinsic enteric-associated neurons (iEANs) in mice are functionally adapted to transcript (CART, encoded by Cartpt), pro-
the intestinal segment they occupy; ileal and colonic neurons are more responsive to microbial enkephalin (Penk), gastrin releasing peptide
colonization than duodenal neurons. Specifically, a microbially responsive subset of viscerofugal CART+ (Grp), agouti-related peptide (Agrp), and tachy-
neurons, enriched in the ileum and colon, modulated feeding and glucose metabolism. These CART+ kinin 1 (Tac1), all of which, besides Agrp, are
neurons send axons to the prevertebral ganglia and are polysynaptically connected to the liver and thought to be involved in controlling intes-
pancreas. Microbiota depletion led to NLRP6- and caspase 11–dependent loss of CART+ neurons and tine motility through the myenteric plexus
impaired glucose regulation. Hence, iEAN subsets appear to be capable of regulating blood glucose levels (8–12) (Fig. 1E and fig. S3, A to C). Immuno-
independently from the central nervous system. fluorescence analysis confirmed a region-
specific compartmentalization of neuropeptides
E nteric-associated neurons (EANs) com- Microbiota-dependent transcriptional changes at the protein level and identified regional dif-
prise a numerous and heterogeneous in enteric neurons are region-specific ferences in neuronal numbers along the intes-
population of neurons innervating the To profile iEANs in an untargeted and region- tine (Fig. 1F and fig. S3D). For instance, NPY,
gastrointestinal (GI) tract that monitor specific manner, we performed translating which is typically involved in intestinal inflam-
and respond to various environmental ribosomal affinity purification (TRAP), a cell mation or inhibition of neurotransmission
cues such as mechanical stretch and luminal type–specific actively translated mRNA profil- (13, 14), was enriched in duodenum iEANs
content (1, 2). The vast majority of luminal ing approach. We interbred pan-neuronal (Fig. 1G and fig. S3E). By contrast, the neuro-
stimuli are derived from the diet and commensal Snap25Cre mice with Rpl22lsl-HA (RiboTag) peptide SST, which is involved in EEC regu-
microbes, which may be sensed directly by EAN mice, allowing hemagglutinin (HA) immuno- lation of several GI hormones (15) and inhibition
fibers positioned along the intestinal epithelium precipitation of actively translated mRNA. of smooth muscle contraction (8, 13), is highly
or indirectly through signals derived from non- Expression of HA-tagged ribosomes was ob- expressed in the ileum and colon but scarcely
neuronal cells inhabiting the same compart- served in neurons in the myenteric plexus of expressed in the duodenum (Fig. 1H and fig.
ment (1, 3). Intrinsic EANs (iEANs) are neural Snap25RiboTag mice (Fig. 1A). RNA sequencing S3F). We also observed increased numbers of
crest–derived and organized in two distinct (RNA-seq) of bound transcripts revealed sig- CART+ neurons, which are thought to play a role
layers: the myenteric or Auerbach’s plexus and nificant enrichment of neuron-specific genes in intestinal nitrous oxide neurotransmission
submucosal or Meissner’s plexus (2). iEANs can and pathways in Cre+ animals compared with and neuroprotection (9, 16), in the ileum and
operate autonomously and are primarily tasked Cre– animals (fig. S1, A to C). Given their deep colon (Fig. 1I and fig. S3, G to I). Finally, we
with modulation of intestinal motility and integration into the intestinal tissue, we sought found that duodenum iEANs are particularly
secretory function (2). Recent studies have to understand how iEANs might differ from enriched in pleiotropic growth factors previ-
demonstrated that the gut microbiota influ- extrinsic EANs (eEANs) (5) innervating the gut. ously shown to be expressed in intestinal
ences the basal activity of intestine-associated TRAP RNA-seq (TRAP-seq) analysis of iEANs epithelium, such as follistatin 1 (Fst1) (17) and
cells, including EANs and immune cells (2, 3), and eEANs (nodose, NG; celiac–superior mes- WNT inhibitory factor 1 (Wif1) (18). Immu-
as well as host metabolism (4). These studies enteric, CG-SMG; and dorsal root ganglion, nofluorescence analysis confirmed prominent
highlight the impact of the gut microbiota DRG) (5) suggested that iEANs possess a distinct FST1+ neurons and nerve fibers in the duode-
on EANs and key mammalian physiological translational profile (Fig. 1B and fig. S2A). We num that were absent in the ileum and sparse
processes; however, the cellular circuits and found that iEANs were primarily defined by within the colon (Fig. 1J and fig. S3J). These
molecular components that mediate gut-EAN enriched transcripts related to neuropeptide data reveal the region-specific translational
communication remain poorly understood. signaling as compared with either NG and profiles of iEANs that likely reflect the function
We sought to determine how the microbiota DRG or CG-SMG, which had increased expres- of distinct intestinal regions.
affects iEANs to better characterize their role sion of transcripts involved in sensory pro-
in host physiology. cesses and catecholamine production, respectively Because the density and diversity of the gut
(Fig. 1C and fig. S2, B to D). microbiota increase from the proximal to
1Laboratory of Mucosal Immunology, The Rockefeller distal intestine, we examined whether region-
University, New York, NY, USA. 2Laboratory of Molecular Comparison between translational profiles ally distinct iEAN translational programs
Genetics, Howard Hughes Medical Institute, The Rockefeller of myenteric iEANs isolated from the duode- are partially influenced by the microbiota. To
University, New York, NY, USA. 3Laboratory of Neurogenetics num, ileum, and colon indicated that iEANs determine if microbial stimuli influence iEAN
and Behavior, Howard Hughes Medical Institute, The segregate based on their anatomical location morphology, we first performed AdipoClear
Rockefeller University, New York, NY, USA. (6) (Fig. 1D and fig. S3, A to C). The proximal on whole-mount intestinal tissue followed by
*These authors contributed equally to this work. small intestine is highly absorptive and enriched light-sheet microscopy of iEANs in the ileum and
†Present address: Kallyope, New York, NY, USA. with enteroendocrine cell (EEC) subsets asso- colon of germ-free (GF) or specific pathogen–
‡Corresponding author. Email: [email protected] (P.A.M.); ciated with lipid and nutrient detection (7). free (SPF) mice. In both GF and SPF mice,
[email protected] (D.M.) Consistently, we found that duodenal iEANs, iEANs were organized into distinct plexuses,
and we observed vast mucosal innervation

Muller et al., Science 370, 314–321 (2020) 16 October 2020 1 of 8

RESEARCH | RESEARCH ARTICLE

A HA Overlay B C Neuropeptide Signaling Ileum
ANNA-1 Eating Behavior NG
DRG Glucose Homeostasis
Duodenum NG Secretion
CG-SMG Temperature Detection
Duodenum Myelin Assembly
Ileum Neuron Action Potential
100 Colon Potassium Ion Transport

Dim2 (17.5%) 50 10 5 0 5 10

-log Fisher P-value
10 elim

Ileum 0 Positive regulation of Calcium Ileum
Neuropeptide Signaling

−50 Dopamine Synthesis CG-SMG
Norepinephrine Synthesis

−100 GPCR Signaling 84 048
Chemical Stimuli Detection
Ileum -log Fisher P-value
Neuropeptide Signaling DRG 10 elim
Action Potential Propagation
Colon −150 Regulation of Neuron Projection

−100 0 100 TCA Cycle

Dim1 (28%)

20 0 20

-log Fisher P-value
10 elim

D Duodenum E Ileum (5) Duodenum (3) Colon (3) Duodenum (3) Ileum (5) Colon (3) F

Ileum −log (p ) −log (p ) −log (p ) Krt19
Colon 10 adj 10 adj 10 adj
Cckar Duodenum
40 20 40 60 80 Fst 20 40 60 80 20 40 60 80 Ileum
Colon
Pou3f3 Wif1 1400 ****

Dim2 (21.7%) Sst 1200
1000
0 Pou3f3 Neurons/mm2
800
Cck Fst 600
Cckbr 400
Penk Gpr88 200
Tacr3 Sst
−40 Gal Npy Cckar Calca Ffar3 0
Agrp Rarb Htr7 Htr7
Gcgr Nmur2 Casr 25
Grp Chrnb4 Agrp Npy Gcgr Rbp4
Htr3a Tac1 Tacr3 Htr2b ***
−100 −50 0 50 100 0 Calca Cartpt Htr3b 0 0
20
Dim1 (43.8%) −10 −5 0 5 10 −10 −5 0 5 10 −10 −5 0 5 10 15
10
log2 Fold Change log2 Fold Change log2 Fold Change
5
G Duodenum H Duodenum I Duodenum J Duodenum 0
Ileum Ileum Ileum Ileum
Colon Colon
Colon Colon **
60 ***
NPY+ Neurons/mm225 * 8 100 ****SST+ Neurons/mm2 10 **** ** % CART+ Neurons 10 50 FST+ Neurons/mm2
% NPY+ Neurons**** % SST+ Neurons 30 **** % FST+ Neurons
**** 80 CART+ Neurons/mm28 40
20 ** 25 8 *** **
6 30
15 4 60 6 20 6
2 15 20 *
10 0 40 4
4 10
5 20 2 10
0
52

0 00 00

Fig. 1. TRAP-seq profiling of iEANs reveals anatomical region–specific DEGs between myenteric iEANs. Pink dots represent all intestine neuronal
differences. Snap25RiboTag SPF mice were analyzed in (A) to (E). (A) Whole-
mount immunofluorescence (IF) of duodenum, ileum, and colon myenteric immunoprecipitation (IP)–enriched transcripts; colored dots represent
plexus using anti-neuronal nuclear (ANNA-1, green) and anti-HA (red) DEGs of interest between each pair of intestine segments. Sample numbers
antibodies. Scale bars are 50 mm. (B) Principal components analysis (PCA)
of neuronal translatomes from DRG, NG, CG-SMG, duodenum, ileum, and are indicated in parentheses. (F) Number of total iEANs in different gut
regions. (G to J) Numbers and percentages of (G) NPY+, (H) SST+, (I)
colon. (C) Gene ontology (GO) pathways of differentially expressed genes CART+, and (J) FST+ myenteric iEANs in different gut regions. *P < 0.05,
(DEGs) [log2 fold change (FC) > 1, and adjusted P value (Padj) < 0.05] **P < 0.01, ***P < 0.001, and ****P < 0.0001; for numbers of iEANs in (F)
enriched in ileum versus indicated ganglia. GPCR, G protein–coupled to (J), Brown-Forsythe and Welch analysis of variance (ANOVA) with
receptor; TCA cycle, tricarboxylic acid cycle. (D) PCA of neuronal
translatomes from duodenum, ileum, and colon. (E) Volcano plots of Dunnett’s T3 multiple comparisons test was performed; for percentages of
iEANs in (G) to (J), Kruskal-Wallis test with Dunn’s multiple comparisons test
was performed. Error bars indicate SD.

in the small and large intestines that extended though no transcripts were significantly up- sion profile, we generated a list of transcripts
into individual villi with fibers adjacent to the regulated in the duodenum in SPF mice com- enriched in SPF duodenum as compared with
epithelium (Fig. 2A and movies S1 to S4). We pared with GF mice, 750 and 117 transcripts SPF ileum and colon (table S1). A subset of
noted that ileum villi are thin and blunted in were significantly up-regulated in the ileum duodenum-enriched transcripts was up-regulated
GF animals, inherently leading to different nerve and colon, respectively (Fig. 2C). The absence in the ileum, but not colon, of GF animals as
fiber structure (Fig. 2A and movies S1 to S4). of significant microbiota-dependent changes compared with SPF animals (Fig. 2D). The
in the duodenum could be related to the de- third principal component also showed segre-
To determine whether the microbiota af- creased microbial density and diversity in this gation of colon samples from the small intestine,
fects iEAN gene expression profiles along the region. Furthermore, principal components which may reflect the presence of iEANs derived
intestine, we rederived Snap25RiboTag mice analysis showed that duodenum, ileum, and from sacral progenitors in the large intestine, or
under GF conditions (fig. S4A). Analysis of colon samples from GF mice all clustered functional differences between colon and ileum
TRAP-seq from duodenum, ileum, and colon together with duodenum samples of SPF mice (fig. S4B). Direct comparison between regions
myenteric iEANs of GF Snap25RiboTag mice (Fig. 2B). To determine whether these distal in GF mice also indicated segregation, sug-
revealed microbiota-dependent transcriptional regions gain a “duodenum-like” gene expres- gesting that certain features of region-specific
changes in each region (Fig. 2, B and C). Al-

Muller et al., Science 370, 314–321 (2020) 16 October 2020 2 of 8

RESEARCH | RESEARCH ARTICLE

iEAN programming are microbiota-independent transcription factors for the ileum and colon GF diet (exGF). Colonization of 8-week-old
(fig. S4C). in SPF mice (fig. S4I). Because the level of GF animals with SPF feces was sufficient to
phosphorylated CREB (pCREB) in neurons is increase the number of SST+ and CART+ neurons
In the ileum and colon, we found microbiota- often used as an indirect measure of activation in the colon and ileum to levels similar to SPF
dependent transcripts encoding neuropeptides and it is a mediator of neuropeptide transcrip- animals after 2 weeks, as well as a notable in-
associated with neuro-immune cross-talk, such tion (20), we used immunofluorescence anal- crease in the density of SST+ and CART+ nerve
as Nmu (19); EAN physiological function, such ysis of pCREB at serine-133, which is a key fibers (Fig. 2, F and G, and fig. S4, L to O). We
as Sst or Cartpt; and functions outside of the modification to induce gene transcription. We also noted that recolonization restored iEAN
intestine, like Agrp (colon only) (Fig. 2, C and found a significant reduction of pCREB in the numbers in the ileum, whereas the colon re-
E). SST and CART protein expression changes ileum myenteric plexus of GF mice compared mained unaffected by colonization (Fig. 2H).
were confirmed by quantification of immuno- with SPF mice (fig. S4, J and K), demonstrat- These results establish both regional differ-
fluorescence images from GF and SPF mice ing that iEANs may be hypoexcitable under ences and the microbial influence on iEAN
(Fig. 2, F and G, and fig. S4, D to G). Quan- gnotobiotic conditions, as previously proposed numbers and gene expression profiles, partic-
tification of iEANs in the myenteric plexus of (21), and providing a possible explanation ularly on neuropeptidergic coding.
GF and SPF mice also revealed a significant for the reduction in neuropeptide transcripts
reduction in iEAN numbers in the duodenum observed. Microbiota modulate iEAN numbers through
and ileum, but not colon, of GF mice (Fig. 2H NLRP6 and caspase 11
and fig. S4H). Finally, analysis of GF and SPF To address whether altered neuropeptide
datasets using PASTAA (predicting associated levels and iEAN numbers in GF mice are the We next asked if microbiota-dependent changes
transcription factors from annotated affinities) result of a developmental defect, we provided in iEANs can be observed in SPF mice after
identified cAMP response element–binding adult C57BL/6 GF mice with age- and sex- administration of broad-spectrum antibiotics
protein (CREB) among the most enriched matched feces from SPF mice on a matched or if a microbiome must be absent from birth.

A SPF GF SPF B
GF
Dim2 (17.1%) 40
Ileum
SPF Duodenum
C GF Duodenum
GF (3) Duodenum SPF (3)
20TUJ1 0 SPF Ileum
GF Ileum

SPF Colon
GF Colon

−40

Colon

−40 0 40 80 120

Dim1 (43.3%)

GF (3) Ileum SPF (5) GF (2) D GF (3) Ileum SPF (5) GF (2) Colon SPF (3)
Gm10359 Duodenum Signature Duodenum Signature
Colon SPF (3)
Spon1 10
−log10(padj) −log10(padj) 477 750 −log (p ) 4 117 -log (p ) -log10(padj)
10 adj 10 adj 14 1
20 40 60 80 100 40 60 80 100 0 2 4 6 8 10 12
20 40 60 80 100 0 2 4 6 8 10 12 Gpr149
Capn6

Sulf2 Pde2a
Gabbr2
Tmem132b
Cartpt Dock10 Cartpt Dlx2 Pkp1

20 Npcd Sst Grp Gm14327 Gal Areg Tacr3 Hcn4 Pmaip1 Krt19
Nmu Lars2 Cbarp Nrgn
Gm14327 Ucn3 Cckar −2 −1 0 1 2
Gm5884 Calcb
Calca Slc4a11 log Fold Change
2
0 0 0 Pcdh8 Agrp

-10 -5 0 5 10 -10 -5 0 5 10 -10 -5 0 5 10 −2 −1 0 1 2

log2 Fold Change log2 Fold Change log2 Fold Change log Fold Change
2

Microbial Density/Diversity

E F exGF C57BL/6J G exGF C57BL/6J H exGF C57BL/6J
GF GF
exGF FC 600 GF
25 25 exGF FC 500 exGF FC
Enriched GO in SPF vs GF SST+ Neurons/mm2 200 ** 50 400 1400
CART+ Neurons/mm2 300
Translation 20 *** 200 *** 1200
Neuropeptide Signaling SST+ Neurons/mm2 100 ** 1000
CART+ Neurons/mm2****
15 **** 0
Neurons/mm2
Neurons/mm2
Ileum 20 150 40 800

**** 30 ***
15 **** 100

Hormone Secretion Colon 10 10 20 600
Neuropeptide Signaling
55 50 10 400

200

024 68 0 0 0 0 0

-log Fisher P-value Ileum Ileum Colon Colon Ileum Colon
10 elim

Fig. 2. Microbiota affect iEAN translatomes in a compartmentalized manner. numbers are indicated in parentheses. (E) GO pathways of DEGs enriched in SPF
(A) AdipoClear light-sheet images of ileum and colon of GF or SPF C57BL/6J mice versus GF samples. In (B) to (E), Snap25RiboTag mice were analyzed. (F and G) Num-
using anti-TUJ1 antibody. Scale bars for ileum and GF colon are 200 mm; scale bers of SST+ and CART+ iEANs in (F) ileum and (G) colon of GF mice, exGF mice,
bar for SPF colon is 100 mm. (B) PCA of GF and SPF iEANs from duodenum, ileum,
and colon. (C and D) Volcano plots of (C) duodenum, ileum, and colon iEAN DEGs and GF mice colonized with exGF feces (exGF FC). (H) Total iEANs in ileum (left) and
between GF and SPF mice and (D) differentially expressed “duodenum signature”
transcripts in ileum and colon iEANs of GF and SPF mice. Pink dots represent all colon (right) of GF, exGF, and exGF FC mice. **P < 0.01, ***P < 0.001, and ****P <
intestine neuronal IP-enriched transcripts; colored dots and numbers represent 0.0001; for numbers of SST+ iEANs in (F) and (G), CART+ iEANs in (F), and total
significantly differentially expressed transcripts (log2 FC > 1, and Padj < 0.05). Sample
iEANs in (H), one-way ANOVA with Tukey’s multiple comparisons test was performed;
for numbers of CART+ iEANs in (G), Brown-Forsythe and Welch ANOVA with

Dunnett’s T3 multiple comparisons test was performed. Error bars indicate SD.

Muller et al., Science 370, 314–321 (2020) 16 October 2020 3 of 8

RESEARCH | RESEARCH ARTICLE

We administered antibiotics (vancomycin, am- cin, also induced a reduction in total neuro- some sensor NLRP6 in microbiota-mediated
picillin, metronidazole, and neomycin) through nal numbers, suggesting a possible role for iEAN regulation.
drinking water to SPF mice for 2 weeks and specific bacteria in the physiological iEAN
detected a significant decrease in the number maintenance (Fig. 3E and fig. S5, H and I). Microbiota-modulated CART+ iEANs are
of iEANs in the ileum and colon but not in the Quantification of microbiota load after oral viscerofugal and glucoregulatory
duodenum (Fig. 3A and fig. S5A). We recently or intraperitoneal antibiotic administration,
described an inflammasome-dependent path- and subsequent neuronal quantification, further To test possible functional roles for microbiota-
way whereby caspase 11 (caspase 4 in humans) suggested that neuronal loss was induced by modulated iEANs in intestinal physiology, we
and NOD-like receptor family pyrin domain microbial depletion and not by direct antibiotic focused on CART+, NPY+, and AGRP+ neuronal
containing 6 (NLRP6) are key mediators of neurotoxicity or dysbiosis per se (fig. S5, J to M). populations because of their distinct features.
neuronal death after infection (22). To evalu- We next examined the microbiota-modulated CART+ neurons are enriched in the ileum and
ate whether the iEAN reduction observed after neuropeptide pathways that we identified in colon and are bidirectionally modulated by
microbiota depletion was also dependent on GF mice and analogously observed a significant the microbiota, and unlike SST, CART is not
this pathway, Casp1–/–Casp11–/– (ICE–/–) or decrease in number and percentage of SST+ expressed by EECs in these gut regions (23).
Casp11–/– mice were exposed to Splenda or and CART+ neurons in the ileum and colon, Meanwhile, AGRP+ neurons are particularly
antibiotics in the drinking water. Quantifi- but not duodenum, after antibiotic treatment enriched in the colon and reduced in GF mice,
cation of iEANs in the ileum of antibiotic- (Fig. 3F). Consistently, short-term microbial de- and NPY+ neurons are enriched in the duode-
treated mice did not reveal iEAN loss in ICE–/– pletion by single-dose streptomycin administered num and not affected by the microbiota. These
or Casp11–/– mice, suggesting an additional to wild-type (WT) mice or continuous broad- three neuropeptides are also expressed by
role for caspase 11 in the reduction of iEANs spectrum antibiotic treatment of Casp1/11–/–, neuronal populations in the hypothalamus
during dysbiosis (Fig. 3A and fig. S5, A to D). Casp11–/–, Snap25DNlrp6, and Snap25DCasp11 mice that work in concert to regulate energy balance
We confirmed that these changes in neuronal did not affect neuropeptide-specific iEAN (24) and, as such, could potentially play a
numbers were not the effect of morphological numbers in the ileum and colon (Fig. 3, F to similar role in gut-specific circuits to influence
differences in the intestine induced by anti- H, and fig. S6, A to O). Given that neuropep- feeding behavior. Whole-mount analysis of in-
biotics or genotype (fig. S5, E to G). Further- tide expression can vary with drastic changes testinal muscularis externa using RNA in situ
more, neuronal- and neuroendocrine-specific in nutrient availability during fasting (fig. hybridization confirmed the expression of Npy
deletion (Snap25) of Nlrp6 and Casp11 may S7, A to C), we confirmed that antibiotic treat- and Cartpt in the ileum and colon and Agrp in
prevent loss of enteric neurons after antibiotic ment did not result in body weight change the midcolon (fig. S8, A to C). We obtained Cre
administration (Fig. 3, B and C). Importantly, at the time of analysis, indicating that food lines corresponding to the three neuropep-
this neuronal reduction was not permanent, intake was comparable between groups (fig. tides and validated Cre, Cartpt, Npy, and Agrp
because antibiotic withdrawal for 2 weeks re- S7D). The above data establish that iEANs, expression in the periphery using in situ hy-
sulted in the recovery of neuronal numbers to including SST+ and CART+ neuropeptide sub- bridization (fig. S8, D to G). Because these
SPF levels (Fig. 3D). Treatment with vancomy- sets, can be tuned by the microbiota. Ad- neuropeptides are known to be expressed in
cin, ampicillin, or metronidazole alone, but ditionally, these analyses define a role for both the periphery and central nervous system
not neomycin or a single dose of streptomy- caspase 11 and the noncanonical inflamma- (CNS) (13, 24–26), we used a local retrograde
viral delivery approach into the duodenum,

Neurons/mm2A C57BL/6J Neurons/mm2Splenda 1400 * Neurons/mm2 Snap25Cre- Snap25Cre-/Casp11+/+Neurons/mm2 D Neurons/mm2 Splenda
Casp11-/- Neurons/mm2Abx1200 Colon Abx
600 600 **** 1000 B Snap25Nlrp6fl/+ C Snap25Casp11fl/+ 500 Switch
500 400 ** *
500 800 Snap25∆Nlrp6 Snap25∆Casp11 300
400 600 600 Abx 600 Abx 200 Ileum
400 400 500 **** 100
300 200 500
300 400 0
200 0 400
200 300
100 300
100 200
0 200
0 100
Duodenum 100
Ileum 0
0
Ileum
Ileum

E **** Splenda F CART+ Neurons/mm2 Splenda CART+ Neurons/mm2 100 G Snap25Cre- CART+ Neurons/mm2 H Snap25Cre-/Casp11+/+CART+ Neurons/mm2
Metro CART+ Neurons/mm2 40 Abx
Neurons/mm2 500 Amp C57BL/6J 30 **** 80 * Snap25Nlrp6fl/+ Snap25Casp11fl/+
400 Vanco 10 Casp11-/- Snap25∆Nlrp6 Snap25∆Casp11
300 Neo 20 60 25 * 15 Abx
200 8 Abx
100 10 10
6 20

4 15

2 40 10 5
20 5

0 0 0 0 0 0

Ileum Duodenum Ileum Colon Ileum Ileum

Fig. 3. iEAN loss after antibiotic treatment is mediated by NLRP6 and metronidazole (Metro), ampicillin (Amp), vancomycin (Vanco), neomycin (Neo),
caspase 11. (A) Myenteric iEAN numbers in duodenum, ileum, and colon of or Splenda. (F) CART+ myenteric iEAN numbers in duodenum, ileum, and
C57BL/6J and Casp11–/– mice treated with antibiotics (vancomycin, ampicillin, colon of C57BL/6J and Casp11–/– mice on Abx or Splenda. (G and H) CART+ ileal
metronidazole, and neomycin, referred to collectively as Abx) or Splenda. myenteric iEAN numbers for (G) Snap25DNlrp6 and control or (H) Snap25DCasp11
(B and C) Ileal myenteric iEAN numbers for (B) Snap25DNlrp6 or (C) Snap25DCasp11
mice on Abx. (D) Ileal myenteric iEAN numbers for C57BL/6J SPF mice after and control mice on Abx. *P < 0.05, **P < 0.01, and ****P < 0.0001; for (A)
4 weeks on Abx or Splenda, or Abx followed by Splenda for 2 weeks each
and (F), two-way ANOVA with Tukey’s multiple comparisons test was performed;
(Switch). (E) Ileal myenteric iEAN numbers for C57BL/6J SPF mice on
for (B), (D), (E), and (G), one-way ANOVA with Tukey’s multiple comparisons

test was performed. Error bars indicate SD.

Muller et al., Science 370, 314–321 (2020) 16 October 2020 4 of 8

RESEARCH | RESEARCH ARTICLE

A B C D cFos Overlay
tdTomato tdTomato CART
mCherry mCherry

Ileum CG-SMG Ileum Ileum

E F Control G H
hM3Dq
Cart Cre- 300 Fed Fed Fed
1.0 Cart Cre+ Food Consumption (g) *** 20 Cart Cre-
Food Consumption (g)2.0 ** 8 * 1.0 Cart Cre- 1.0 ****
* 6 Cart Cre+
Glucose (mg/dL)1.54 0.8 Cart Cre+ 0.8
0.8 AUC (x1000)2 15
Insulin (ng/mL)1.02000
0.6 Insulin (ng/mL) 10
0.5
0.4 Glucagon (pmol/L) 5
0.0 0.6 0.6
0.2
4 Hours 100 0.4 ** 0.4
0.0
Fed 0.2 0.2
2 Hours
0 0.0 0.0 0 0 30 90 0 30 90
0 20 40 60 80 100 120 0 30 60 90 0 30 90 0 30 90
Time (minutes)
Time (minutes) Time (minutes) Time (minutes)

Fig. 4. CART+ iEANs in the distal intestine are viscerofugal and glucoregu- levels after C21 administration (left) and area under curve (AUC) analysis (right)
in fed CartCre+ and CartCre– (control) mice injected with AAV9-hSyn-DIO-hM3Dq-
latory. (A and B) Whole-mount IF image of the (A) ileum myenteric plexus (MP)
and (B) CG-SMG of CartCre+ injected with AAVrg-FLEX-tdTomato into (A) ileum mCherry or control AAV9-hSyn-DIO-mCherry into ileum and colon. (G and
H) Plasma (G) insulin and (H) glucagon levels after C21 administration in CartCre+
and (B) duodenum, ileum, and colon. (C and D) Whole-mount IF image of the and CartCre– mice injected with AAV9-hSyn-DIO-hM3Dq-mCherry into ileum
ileum MP of CartCre+ mice injected with AAV9-hSyn-DIO-hM3Dq-mCherry into
and colon. Scale bars in (A) to (D) are 50 mm. *P < 0.05, **P < 0.01,
the ileum, (C) stained for CART (green) and mCherry (red) and (D) 3 hours after
***P < 0.001, and ****P < 0.0001; for (E) and (F), two-tailed unpaired Student’s
C21 administration, stained for mCherry (red) and cFos (green). (E) Food
t test was performed; for (G), two-way (left) or one-way ANOVA (right) with
consumption at night 2 hours (left) and 4 hours (right) after C21 administration
in CartCre+ and CartCre– mice injected with AAV9 as in (C). (F) Blood glucose Tukey’s multiple comparisons test was performed. Error bars indicate SD.

ileum, and colon to guide us on subsequent chemogenetic approach by injecting excitatory not change blood glucose levels (Fig. 4F and
DREADD (designer receptor exclusively acti- fig. S9, H and I). On measurement of canonical
gut-restricted adeno-associated virus (AAV) glucoregulatory hormones, we found a signif-
vated by designer drugs) virus (AAV9-FLEX- icant decrease in insulin levels at 30 and 90 min
approaches. Injection of retrograde AAV Syn-hM3Dq-mCherry) into the distal ileum after C21 administration to CartptEAN-hM3Dq
and proximal colon of CartptCre and NpyCre mice, whereas glucagon levels were not sig-
(AAVrg)-FLEX-tdTomato into the intestine of mice or into the midcolon of AgrpCre mice (Fig. nificantly altered (Fig. 4, G and H). These data
CartptCre, NpyCre, and AgrpCre mice (generat- 4, C and D, and fig. S9A). We found no change indicate that stimulation of ileum and colon
ing CartptEAN-tdTomato, NpyEAN-tdTomato, and in total intestinal transit time for any of the CART+ iEANs results in increased blood glucose
AgrpEAN-tdTomato, respectively) revealed a and decreased insulin levels, with a subsequent
prominent population of tdTomato+ neurons three neuropeptide lines tested after admin- reduction in food consumption.
in the myenteric plexus of CartptEAN-tdTomato istration of the DREADD ligand, compound
and NpyEAN-tdTomato mice (Fig. 4A and fig. S8, H We next asked how CART+ iEANs can exert
and I). NpyEAN-tdTomato and CartptEAN-tdTomato 21 (C21), although changes in either small or their glucoregulatory function. One possible
neurons displayed considerable innervation of large intestine motility separately cannot be route could be through direct detection of sig-
nals coming from the epithelium. However,
the circular and longitudinal smooth muscle ruled out (fig. S9, B to D). However, we ob- imaging analyses confirmed that CART+ neu-
rons are not present in the submucosal plexus
within these segments of the intestine, with served a significant decrease in food consump- nor do they project to the epithelium (fig. S10,
CartptEAN-tdTomato also exhibiting dense inter- tion during day feeding at 1 and 2 hours, as A and B, and movie S5). We confirmed their
viscerofugal nature with viral anatomical and
ganglionic patterning (Fig. 4A and fig. S8, well as during night feeding at 2 and 4 hours, cholera toxin subunit B (CTB) tracing (Fig. 5, A
after C21 injection in CartptEAN-hM3Dq but not and B, and movie S5). We also noted that some
H and I). We found a sparse population of NpyEAN-hM3Dq or AgrpEAN-hM3Dq mice (Fig. 4E CART+ neurons appear to directly interact
tdTomato+ neurons in the midcolon of and fig. S9, E to G). We found no clear evidence with other CART– viscerofugal neurons (fig.
AgrpEAN-tdTomato, which exhibited muscular S10C). These CART+ viscerofugal neurons send
for hM3Dq (mCherry) expression outside of the axonal projections to the CG-SMG, which in
and interganglionic innervation (fig. S8J). We distal ileum and proximal colon, indicating that turn provides sympathetic innervation to
several visceral organs, including the pan-
confirmed a lack of tdTomato expression in the the observed changes in feeding behavior are creas and liver (29, 30). Sympathetic inner-
dependent on iEAN-specific neuronal stimula- vation of the pancreatic islets can stimulate
submucosal plexus, NG, DRG, and CG-SMG in glucagon release and inhibit insulin secretion
CartptEAN-tdTomato and AgrpEAN-tdTomato mice. tion. We next evaluated whether the reduction through adrenergic receptor engagement on
We also observed a population of tdTomato+ in feeding was accompanied by acute changes alpha and beta cells, respectively (29). Sym-
pathetic stimulation of the liver can drive
neurons in the submucosal plexus and gut- in blood glucose or by glucoregulatory hor- gluconeogenesis and glycogenolysis through
projecting CG-SMG neurons in NpyEAN-tdTomato mone levels that can regulate the activity of hepatocyte adrenergic receptor activation
mice (fig. S8K). Additionally, a significant num- CNS nuclei that control feeding behavior (28).
ber of tdTomato+ fibers were detected in the We assessed the effects of either excitatory
CG-SMG of CartptEAN-tdTomato mice (Fig. 4B and
fig. S8L), suggesting that some CART+ neurons (hM3Dq) or inhibitory (hM4Di) DREADD
viruses in CartptCre mice, again injected into
are viscerofugal (iEAN projecting axons outside their distal ileum and proximal colon. Ad-
ministration of C21 led to significantly
of the intestine and previously described as higher blood glucose levels in CartptEAN-hM3Dq
mice than in control mice, whereas inhibition
mechanosensitive) (27). of these neurons in CartptEAN-hM4Di mice did

To directly modulate neuronal activity and

assess the function of these intestine neuro-

peptide populations, we used a gut-restricted

Muller et al., Science 370, 314–321 (2020) 16 October 2020 5 of 8

RESEARCH | RESEARCH ARTICLE

A B CG-SMG C D

mCherry mCherry PRV-GFP cFos
ANNA-1 mCherry
PRV-GFP CartCre-
PRV-RFP CartCre+
300 *
Mesentery CTB+ CART+ Neurons/mm2
cFos+ cells/CG-SMG 200
Myenteric
Plexus 100

Ileum 0

Colon Ileum CG-SMG

CG-SMG Splenda
Abx
E F CTB+ CART+ Neurons/mm2 G H

CTB+ Neurons/mm2 C57BL/6J Splenda C57BL/6J Casp11-/- CTB+ Neurons/mm2 Splenda Casp11-/- Splenda
Abx 15 **** Abx Abx
15 5 10
**
** 4 10 ** 8

10 *** 3 5 6
**
2* 4 p = 0.1013
5 2
1
** 0 0
0
0 DI C Ileum
Ileum
DIC

Fig. 5. CART+ viscerofugal neurons are polysynaptically connected to the and colon (labeled D, I, and C, respectively) of (E) C57BL/6J or (G) Casp11−/−
mice treated with Splenda or antibiotics for 2 weeks after CTB injection into the
liver and pancreas through the CG-SMG. (A and B) Whole-mount IF image of CG-SMG. (F and H) CTB-AF647+ CART+ neuron numbers in the ileum of (F)
(A) CG-SMG and (B) ileum MP of CartCre+ mice injected with AAV9-hSyn-DIO- C57BL/6J or (H) Casp11−/− mice treated with Splenda or antibiotics for 2 weeks
after CTB injection into the CG-SMG. *P < 0.05, **P < 0.01, ***P < 0.001,
hM3Dq-mCherry into ileum and colon. Scale bars are 100 mm. (C) Whole-mount and ****P < 0.0001; for (D) and comparisons between Splenda and Abx [(E) to
(H)], two-tailed Student’s unpaired t test was performed; for Splenda group
IF image of colon (left) and ileum MP (right) of SPF mice injected with PRV-GFP comparisons between gut segments [(E) and (G)], one-way ANOVA with Tukey’s
multiple comparisons test was performed. Error bars indicate SD.
(pancreas) and PRV-RFP (liver). Scale bars are 50 mm. (D) (Left) cFos (green)
and mCherry (red) expression and (right) number of cFos+ neurons in the CG-
SMG of CartCre+ mice injected with AAV as in (A), 3 hours after C21 injection.
Scale bar is 50 mm. (E and G) CTB-AF647+ neuron numbers in duodenum, ileum,

(30). To determine if a synaptically connected activation (5). As expected, we observed a sig- Casp11–/– mice did not result in loss of vis-
circuit exists between the gut, sympathetic nificant increase in cFos expression in C21- cerofugal neurons (Fig. 5, G and H, and fig.
ganglia, and the pancreas or liver, we per- injected CartptEAN-hM3Dq mice as compared with S11, K to Q), demonstrating that their loss is
formed polysynaptic retrograde tracing using control mice (Fig. 5D). Inhibition of catechol- also dependent on this noncanonical inflam-
pseudo-rabies virus (PRV). We injected en- amine release by guanethidine prevented masome effector. We conclude that intestinal
hanced green fluorescent protein (EGFP)– the increase in glucose levels in C21-treated CART+ neurons that can modulate blood glucose
expressing PRV into the pancreas and monomeric CartptEAN-hM3Dq mice, further suggesting the are viscerofugal and microbiota-dependent.
red fluorescent protein (mRFP)–expressing PRV involvement of sympathetic activation (fig. S11A).
into the parenchyma of the liver and assessed However, guanethidine administration did Modulation of glucose by the microbiota is
their synaptic connections to the CG-SMG and not prevent the reduction in blood glucose dependent on CART+ iEANs
the intestine (fig. S10D). We detected viral spread induced by antibiotics (fig. S11B). To deter-
or CTB labeling from both organs to the CG- mine whether neuropeptide release affects To confirm whether microbiota depletion af-
SMG as early as 1 day after injection (fig. S10E). glucose regulation, we exogenously admin- fects glucose regulation, we analyzed antibiotic-
We observed GFP+ neurons in the myenteric istered CART peptide in antibiotic- or control- treated SPF and GF mice and found a significant
plexuses of the duodenum, ileum, and colon treated mice but observed no change in blood reduction in blood glucose, irrespective of diet
4 days after injection, with the highest con- glucose (fig. S11C). However, we cannot defin- or feeding state (Fig. 6, A and B, and fig. S12,
centration of neurons occurring in the colon itively rule out a direct effect of CART peptide A, B, and H), corroborating previous studies
and ileum, whereas RFP+ neurons were only owing to incomplete knowledge of its role in (31–33). Consistent with neuronal loss, am-
observed in the ileum (Fig. 5C). We did not the periphery and lack of identified receptor(s). picillin treatment alone specifically led to a
observe dual RPF and GFP labeling of ileal Next, we investigated whether viscerofugal reduction in blood glucose levels (Fig. 6C).
viscerofugal neurons, suggesting that the pan- populations would be affected after microbial Fasting blood glucose levels could be res-
creas and liver are connected by two separate depletion. Indeed, retrograde fluorescent CTB cued after colonization with the microbiota
circuits. Together, we found that glucoregula- tracing from the CG-SMG revealed a loss of from SPF animals, irrespective of genetic back-
tory organs are polysynaptically connected to CTB+ neurons, including CTB+ CART+ neurons, ground (Fig. 6D and fig. S12, C to G). To de-
the gut through viscerofugal neurons. specifically in the ileum of antibiotic-treated termine whether microbiota-mediated changes
mice, whereas only a minor reduction was in glucose levels are associated with loss of
To investigate whether CART+ viscerofugal observed in the colon and no changes were iEANs, we measured blood glucose levels in
neuron activation could directly modulate sym- found in the sparsely retrograde-labeled duo- global or conditional knockout mice targeting
pathetic neuronal activity, we dissected the CG- denum (Fig. 5, E and F, and fig. S11, D to J). NLRP6 and its downstream effector caspase
SMG after C21 administration and measured As expected, administration of antibiotics to 11, because these genotypes did not lose iEANs
cFos expression as an indicator of sympathetic or CART+ iEANs after antibiotic treatment.

Muller et al., Science 370, 314–321 (2020) 16 October 2020 6 of 8

RESEARCH | RESEARCH ARTICLE

A B GF C Splenda D GF E Snap25Cre- F Snap25Cre-/Casp11+/+
exGF FC
C57BL/6J exGF B6 Ampicillin
Casp11-/- Casp11-/- Neomycin Casp11-/- FC Snap25Nlrp6fl/+ Snap25Casp11fl/+
Casp1/11-/- 120 ** 120 ****
Glucose (mg/dL) Glucose (mg/dL) Glucose (mg/dL) Glucose (mg/dL) * Glucose (mg/dL) Snap25∆Nlrp6 Glucose (mg/dL) Snap25∆Casp11
150 100 100 150 ****
150 120
**** 80 80
** exGF * *** 80
100 60 60 Casp11-/-
100 100
50 40 40
50 50 40
0 20 20
0 0 0
0 0
Abx Abx

G H GF I J Snap25Cre- K Snap25Cre-/Casp11+/+
Snap25Nlrp6fl/+ Snap25Casp11fl/+
C57BL/6J Splenda Splenda GF FC C57BL/6J Splenda Snap25∆Nlrp6 15 Snap25∆Casp11
Abx Casp11-/- Abx 10
300 Abx C57BL/6J 30
Glucose (mg/dL) 15
20 20
AUC (x1000) AUC (x1000) AUC (x1000) AUC (x1000) AUC (x1000)
*** **

200 15 15 10 **** 20

10 10

100 55 5 10 5

Pyruvate 00 000

0 Abx Abx
0 20 40 60 80 100 120

Minutes

Glucose (mg/dL)L CartControl Glucose (mg/dL) 240 P = 0.1868 MInsulin (ng/mL) CartControl Insulin (ng/mL) 1.5 ** NGlucose (mg/dL) CartControl AUC (x1000) 8 P = 0.0511
CartDTA 200 CartDTA CartDTA 6
120 160 1.0 1.0 200 4
100 ** 120 0.8 ** 0.5 2
0.0 150 0
80
80 40 0.6 Fed 100

60 0

40 Fed 0.4 50

20 0.2 Pyruvate

0 0.0 0
0 20 40 60 80 100 120
Fasted Fasted
Minutes

Fig. 6. Control of blood glucose is microbiota- and CART+ iEAN– mice on Abx or Splenda. (J and K) IP-PTT AUC analysis of fasted (J) Snap25DNlrp6
dependent. (A to G) Blood glucose levels of fasted (A) C57BL/6J, Casp11–/–, and control or (K) Snap25DCasp11 and control mice on Abx. (L to N) CartCre+
and Casp1/11–/– mice treated with Abx or Splenda; (B) GF, exGF B6, and mice injected with AAV5-DTA into ileum and colon. Shown are (L) fasted (left) or
Casp11–/– mice; (C) C57BL/6J mice on ampicillin, neomycin, or Splenda; fed (right) blood glucose levels, (M) fasted (left) or fed (right) plasma insulin
(D) GF, exGF FC, and Casp11–/– FC mice; (E) Snap25DNlrp6 mice on Abx; and
(F) Snap25DCasp11 mice on Abx. In (G), an intraperitoneal pyruvate tolerance levels, and (N) fasted IP-PTT blood glucose curves (left) and AUC analysis (right).
*P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001; for (G) and (L) to
test (IP-PTT) of fasted C57BL/6J mice on Abx or Splenda is presented, (N), two-tailed unpaired Student’s t test was performed; for (B) to (E), (H), and
(I), one-way ANOVA with Tukey’s multiple comparisons test was performed;
with blood glucose curves shown on the left and AUC analysis shown on the for (A) and (J), two-way ANOVA with Tukey’s multiple comparisons test was
performed. Error bars indicate SD.
right. (H and I) IP-PTT AUC analysis of fasted (H) GF, GF FC, and C57BL/6J
mice housed in bioexclusion isolator cages, and (I) C57BL/6J and Casp11–/–

We found that blood glucose levels in Casp1/11–/–, compared with control SPF mice after pyru- in blood glucose levels in fasted animals and a
Casp11–/–, Snap25DNlrp6, and Snap25DCasp11 mice significant increase in insulin levels in fasted
were higher after antibiotic treatment as com- vate administration (Fig. 6, G to I, and fig. S12, and fed animals compared with CartptCre mice
pared with that of WT and heterozygous con- injected with control AAV5 (Fig. 6, L and M, and
trols (Fig. 6, A, E, and F, and fig. S12H). L and M). This effect was rescued by micro- fig. S13, A and B). Similar to what was observed in
biota reconstitution in GF mice (Fig. 6H and GF and antibiotic-treated mice, we found a trend
We sought to determine which glucose- toward decreased gluconeogenic capacity (Fig.
modulating pathways may be regulated by fig. S12L) or global loss of caspase 11 in SPF 6N). Thus, loss of CART+ viscerofugal iEANs de-
changes in the microbiota and, more specifically, mice (Fig. 6I and fig. S12N); additionally, it was creases blood glucose levels, presumably owing
whether the NLRP6–caspase 11 inflammasome to the lack of pancreas- and liver-specific sym-
pathway was involved. We first confirmed that partially rescued in SPF mice with neuronal- pathetic regulation. Gut CART+ neurons are
glucagon-like peptide 1 (GLP-1) is increased specific deletion of NLRP6 and caspase 11 (Fig. therefore both sufficient and necessary to mod-
in antibiotic-treated mice (31) but found that ulate blood glucose through glucoregulatory
this increase was independent of caspase 11 6, J and K, and fig. S12, O and P). Of note, be- organs. Together, these experiments establish a
(fig. S12I). Furthermore, administration of cause we noticed antibiotic-resistant bacteria microbiota-sensitive, polysynaptic glucoregu-
the GLP-1 receptor (GLP-1R)–blocking peptide in global Casp11–/– mice, these experiments latory circuit connecting the gut, sympathetic
Exendin-9-39 did not change fasting blood were performed after antibiotic depletion ganglia, and the liver and pancreas (fig. S14).
glucose levels of antibiotic-treated WT mice
(fig. S12J). Additionally, insulin levels did followed by microbiota reconstitution from Discussion
not change after antibiotic treatment in WT, “antibiotic-resistant” feces from Casp11–/– mice
Casp1/11–/–, and Casp11–/– mice (fig. S12K). We (fig. S12Q). These data suggest that iEANs can The gut microbiota influences several physio-
next investigated whether pyruvate-induced regulate liver gluconeogenesis independently logical and pathological processes, including
gluconeogenesis (33) was affected by micro- local nutrient absorption and lipid metabo-
bial manipulation. We observed significantly of pancreatic insulin production or intestinal lism (4, 31, 32, 34), as well as activation of the
blunted temporal changes in blood glucose GLP-1 release in a microbiota- and inflammasome- gut-associated and systemic immune sys-
levels in GF and antibiotic-treated SPF mice tems (35). Dysbiosis or depletion of commensal
dependent manner.
To directly test the necessity of gut CART+

neurons in glucose regulation, we injected AAV5-
mCherry-FLEX-DTA into the ileum and colon
of CartptCre mice to selectively delete intestinal
CART+ neurons. Two weeks after CART+ neu-

ron ablation, we observed a significant reduction

Muller et al., Science 370, 314–321 (2020) 16 October 2020 7 of 8

RESEARCH | RESEARCH ARTICLE

bacteria has also been shown to affect iEAN coronavirus disease 2019 (COVID-19) lab clo- 28. J. Havrankova, J. Roth, M. Brownstein, Nature 272, 827–829
excitability and neurochemical code (21, 36), (1978).
behavioral, or cognitive disorders (37). Our data sures, and thus we acknowledge that our initial
revealed microbial- and region-dependent iEAN sample sizes for Snap25DCasp11 were below the 29. J. A. Love, E. Yi, T. G. Smith, Auton. Neurosci. 133, 19–34
functional specialization with the potential standard for an animal experiment, despite two (2007).
to perform metabolic control independent
of the CNS. Adding to our recent findings (22), independent experiments having been run. In 30. K. Mizuno, Y. Ueno, Hepatol. Res. 47, 160–165 (2017).
here we describe a distinct role for the non- 31. A. Zarrinpar et al., Nat. Commun. 9, 2872 (2018).
canonical inflammasome components NLRP6 the time between the First Release and print 32. A. M. Martin et al., Proc. Natl. Acad. Sci. U.S.A. 116,
and caspase 11 in controlling iEAN numbers and
subsequent glucose regulation in response to publication, we were able to complete a final 19802–19804 (2019).
microbiota levels. Although we do not know antibiotic treatment experiment in Snap25DCasp11 33. T. I. Krisko et al., Cell Metab. 31, 592–604.e9 (2020).
the ligand(s) that activate the noncanonical mice. We present these findings with increased 34. T. Korem et al., Cell Metab. 25, 1243–1253.e5 (2017).
inflammasome pathway, bile acids may rep- n in fig. S15. 35. K. Honda, D. R. Littman, Nature 535, 75–84 (2016).
resent an interesting mediator (38) because we 36. J. M. Yano et al., Cell 161, 264–276 (2015).
did not observe significant iEAN expression of REFERENCES AND NOTES 37. S. Kim et al., Nature 549, 528–532 (2017).
direct pathogen recognition receptors, such as 38. M. Levy et al., Cell 163, 1428–1443 (2015).
Toll-like receptors. It remains to be determined 1. E. A. Mayer, Nat. Rev. Neurosci. 12, 453–466 (2011). 39. M. M. Kaelberer et al., Science 361, eaat5236 (2018).
whether CART+ viscerofugal neurons respond 2. J. B. Furness, L. R. Rivera, H. J. Cho, D. M. Bravo, B. Callaghan, 40. W. Han et al., Cell 175, 665–678.e23 (2018).
to the presence of glucose in the intestinal lu-
men, release of neuropeptides, or the move- Nat. Rev. Gastroenterol. Hepatol. 10, 729–740 (2013). ACKNOWLEDGMENTS
ment of fecal matter. It will also be important 3. H. Veiga-Fernandes, D. Mucida, Cell 165, 801–811 (2016).
to determine how CART+ viscerofugal neurons 4. V. K. Ridaura et al., Science 341, 1241214 (2013). We thank all Mucida Lab members, the Bio-imaging and
are functionally connected to potential sensors 5. P. A. Muller et al., Nature 583, 441–446 (2020). Genomics Centers, and Rockefeller University employees for
such as intrinsic primary afferents, EECs, or 6. Q. Sang, H. M. Young, Cell Tissue Res. 284, 39–53 their continuous assistance; and A. Rogoz and S. Gonzalez for
mechanosensitive populations and the down- the maintenance of mice. We thank J. Friedman for the
stream neuronal populations required to per- (1996). generous use of lab equipment and resources. We also thank
form glucoregulatory functions (39, 40). Finally, 7. F. M. Gribble, F. Reimann, Annu. Rev. Physiol. 78, 277–299 A. Lockhart, G. Donaldson, and the Nussenzweig, Victora,
the ability of viscerofugal neurons in the dis- and Lafaille labs for fruitful discussions. Funding: This work was
tal ileum and proximal colon to increase blood (2016). supported by NIH P40 OD010996 (Core), Gilliam HHMI (V.J.),
glucose through a peripheral circuit warrants 8. D. H. Teitelbaum, T. M. O’Dorisio, W. E. Perkins, T. S. Gaginella, NIH UL1TR001866 (P.A.M. and D.M.), and NIH F31 DK112601;
additional investigations into CNS-independent Philip M. Levine (P.A.M.), Anderson Graduate (F.M.), and
iEAN circuits. Because we focused on the func- Am. J. Physiol. 246, G509–G514 (1984). Kavli fellowships (P.A.M. and M.S.); NIDDK grant K99 DK120869;
tional characterization of selected neuropeptides 9. E. Ekblad, M. Kuhar, N. Wierup, F. Sundler, Neurogastroenterol. the Robertson Therapeutic Development Fund (M.S.); the
in this study, it will be important to further Burroughs Wellcome Fund; the Kenneth Rainin Foundation; the
explore if additional microbiota-modulated Motil. 15, 545–557 (2003). Food Allergy FARE/FASI Consortium; and NIH R01DK126407
and/or microbiota-independent iEAN neu- 10. A. Lecci, M. Altamura, A. Capriati, C. A. Maggi, Eur. Rev. Med. and Transformative R01DK116646 (D.M.). Author contributions:
ropeptide pathways play complementary or P.A.M. initiated, designed, performed, and helped supervise
redundant roles in GI physiology, including Pharmacol. Sci. 12 (suppl. 1), 69–80 (2008). the research and wrote the manuscript. F.M. and M.S. designed
feeding behavior (3, 32, 36, 39, 40). Targeting 11. P. Holzer, Regul. Pept. 155, 11–17 (2009). and performed experiments. Z.K. performed experiments. V.J.
peripheral-restricted circuits, such as the one 12. L. P. Degen et al., Gastroenterology 120, 361–368 (2001). performed part of the RNA-seq analysis. D.M. initiated, designed,
uncovered here, could bypass undesirable CNS 13. J. R. Grider, J. Pharmacol. Exp. Ther. 307, 460–467 and supervised the research and wrote the manuscript. All authors
effects for the treatment of metabolic disor- revised and edited the manuscript and figures. Competing
ders, such as type 2 diabetes. (2003). interests: The authors declare no competing financial interests.
14. K. N. Browning, G. M. Lees, Neurogastroenterol. Motil. 12, Data and materials availability: TRAP-seq data for extrinsic
Note added in proof: Our final mouse experi- ganglia and intestine segments were previously deposited in the
ments were substantially affected by Rockefeller 33–41 (2000). Gene Expression Omnibus under accession numbers GSE145986
15. O. J. Mace, B. Tehan, F. Marshall, Pharmacol. Res. Perspect. 3, and GSE156142, respectively.

e00155 (2015). SUPPLEMENTARY MATERIALS
16. K. Makowska et al., C. R. Biol. 341, 325–333 (2018).
17. K. Sonoyama, S. Rutatip, T. Kasai, Am. J. Physiol. Gastrointest. science.sciencemag.org/content/370/6514/314/suppl/DC1
Materials and Methods
Liver Physiol. 278, G89–G97 (2000). Figs. S1 to S15
18. T. Byun et al., J. Clin. Pathol. 58, 515–519 (2005). Table S1
19. V. Cardoso et al., Nature 549, 277–281 (2017). References (41–51)
20. V. Lakhina et al., Neuron 85, 330–345 (2015). MDAR Reproducibility Checklist
21. K. A. McVey Neufeld, A. Perez-Burgos, Y. K. Mao, Movies S1 to S5

J. Bienenstock, W. A. Kunze, Neurogastroenterol. Motil. 27, View/request a protocol for this paper from Bio-protocol.
627–636 (2015).
22. F. Matheis et al., Cell 180, 64–78.e16 (2020). 14 July 2020; accepted 18 August 2020
23. A. R. Gunawardene, B. M. Corfe, C. A. Staton, Int. J. Exp. Published online 27 August 2020
Pathol. 92, 219–231 (2011). 10.1126/science.abd6176
24. M. J. Waterson, T. L. Horvath, Cell Metab. 22, 962–970
(2015).
25. R. Gupta, Y. Ma, M. Wang, M. D. Whim, Endocrinology 158,
2572–2584 (2017).
26. X. Yuan, Y. Huang, S. Shah, H. Wu, L. Gautron, eNeuro 3,
ENEURO.0174-16.2016 (2016).
27. T. J. Hibberd, V. P. Zagorodnyuk, N. J. Spencer, S. J. Brookes,
Neuroscience 225, 118–129 (2012).

Muller et al., Science 370, 314–321 (2020) 16 October 2020 8 of 8

RESEARCH

MORPHOGENS terized anti-GFP nanobodies (14), such as GBP1
(also known as vhhGFP4), which binds GFP
Patterning and growth control in vivo by an with a dissociation constant (Kd) of 0.23 nM
engineered GFP gradient (15, 16); we refer to this antibody as Nb1high.
DNA encoding this fusion protein (Nb1highCD8)
Kristina S. Stapornwongkul1, Marc de Gennes1, Luca Cocconi1,2, was knocked into the hedgehog (hh) locus so
Guillaume Salbreux1*†, Jean-Paul Vincent1* that it could be expressed at a physiological
level in a domain that abuts the ptc expression
Morphogen gradients provide positional information during development. To uncover the minimal domain (Fig. 1A, gray shading), where SecGFP
requirements for morphogen gradient formation, we have engineered a synthetic morphogen in is produced. In the presence of both genetic
Drosophila wing primordia. We show that an inert protein, green fluorescent protein (GFP), can form a modifications, a gradient of GFP fluores-
detectable diffusion-based gradient in the presence of surface-associated anti-GFP nanobodies, cence was readily detectable (Fig. 1, B and C,
which modulate the gradient by trapping the ligand and limiting leakage from the tissue. We next fused hh-Nb1highCD8) in both the basolateral and
anti-GFP nanobodies to the receptors of Dpp, a natural morphogen, to render them responsive to apical regions. Here, we focused on the baso-
extracellular GFP. In the presence of these engineered receptors, GFP could replace Dpp to organize lateral gradient; a discussion of the apical gra-
patterning and growth in vivo. Concomitant expression of glycosylphosphatidylinositol (GPI)–anchored dient can be found in fig. S3. The basolateral
nonsignaling receptors further improved patterning, to near–wild-type quality. Theoretical arguments GFP profile (Fig. 1D, green curve) differed
suggest that GPI anchorage could be important for these receptors to expand the gradient length scale while somewhat from a classic exponential, with a
at the same time reducing leakage. shoulder near the source and a nonzero tail
far from the source (length scales and nonzero
D uring development, morphogens pro- approach to investigate whether an inert pro- tail values are listed in table S1). Because GFP
vide positional information by forming tein can form a diffusion-based gradient in can diffuse in and out of imaginal discs, we
long-range concentration gradients. De- the basolateral space of a developing pseudo- considered the possibility that the nonzero
spite the importance of morphogens, stratified epithelium and specify positional tail could arise from GFP that escaped into the
there is still no consensus on how they information. hemolymph (GFPhemo). This was tested by trap-
spread within tissues (1). The most parsimo- ping GFP in the hemolymph with Nb1highCD8
nious view is that morphogens travel by dif- Extracellular binders reveal a expressed at the surface of the fat body, a
fusion (1–3). However, epithelia, monolayered diffusion-based gradient sprawling organ that lines the body cavity (fig.
sheets of cells, present a particular challenge S2A). In the resulting imaginal discs, the GFP
for diffusion-based mechanisms, as ligand leak- Synthetic approaches have become a power- profile decayed all the way to background level,
age is expected to occur, thus compromising ful tool to uncover the key features of natural showing that the tail indeed originated from
planar gradient formation (4) and possibly processes (12, 13). To assess the ability of an inert the hemolymph (Fig. 1, C and D, purple curve).
affecting the development of other tissues protein to form a gradient in wing imaginal In conclusion, a single extracellular binding
and organs (5). discs of Drosophila, we engineered flies to ex- species reveals the gradient of an inert protein
press, from a localized source, green fluores- in vivo, but leakage in the hemolymph occurs
Much of our knowledge about the forma- cent protein (GFP) appended with a secretion and interferes with the gradient’s shape, most
tion and interpretation of morphogen gra- targeting signal (SecGFP). This was achieved obviously far from the source, at the tail end
dients in epithelia comes from studies of the by integrating SecGFP coding DNA into the of the gradient.
bone morphogenetic protein (BMP) homolog patched (ptc) locus (fig. S1), a gene which, like
Decapentaplegic (Dpp) in wing imaginal discs dpp, is expressed along the A/P boundary (Fig. Key parameters of gradient formation
of Drosophila. In these epithelial pouches, Dpp 1A). GFP was detectable, albeit weakly, in the
is produced by a stripe of cells and spreads to expression domain (Fig. 1, B and C, no bind- Having established that an inert protein can
form a gradient that organizes growth and ers). GFP fluorescence was also present uni- form a gradient in a developing epithelium,
patterning along the anterior–posterior (A/P) formly in the peripodial space, an enclosed we set out to investigate the importance of
axis (6). It has been suggested that Dpp spreads lumen on the epithelium’s apical side. By con- the surface binders’ affinity for GFP. To ask
by planar transcytosis or on specialized filopo- trast, the basolateral space was devoid of de- if the high affinity of Nb1 for GFP (0.23 nM)
dia called cytonemes (7, 8). Both mechanisms tectable GFP, most likely because it is exposed is needed for a detectable gradient to form,
would ensure planar transport; however, so far, to the larval circulation, which could provide this parameter was changed by using, as an
direct functional evidence remains scant. By an escape route (fig. S2A). Indeed, GFP can extracellular binder, LaG3, which binds GFP
contrast, there is extensive genetic evidence for cross the basal lamina to and from the hemo- with a Kd of 25 nM (17); we refer to this pro-
the requirement of glypicans in morphogen lymph (fig. S2), and leakage could therefore tein as Nblow). In imaginal discs carrying ptc-
transport (9–11). It has been suggested that prevent locally expressed GFP from forming SecGFP and hh-NblowCD8, GFP fluorescence
morphogens can piggyback on laterally diffus- a detectable gradient in the basolateral space. was above background but not detectably
ing glypicans and pass from cell to cell through Natural morphogens, which form gradients, graded (Fig. 1, C and D, compare blue and
cycles of dissociation and reassociation, thus are known to bind various receptors and extra- black curves;fig. S4 informs a discussion of
remaining within the plane of the epithelium. cellular components (11). We therefore asked Nb-mediated GFP fluorescence boosting). This
Here, we have taken a forward engineering whether adding GFP-binding species in the indicated that a low-affinity binder can trap
extracellular space would reduce leakage and extracellular GFP, but also that sufficiently
1The Francis Crick Institute, 1 Midland Road, London NW1 1AT, enable the formation of a detectable gradient high affinity is needed for a meaningful gra-
UK. 2Imperial College, Department of Mathematics, London, UK. in SecGFP-expressing wing imaginal discs. dient of surface-associated GFP to form.
*Corresponding author. Email: [email protected] (J.-P.V.);
[email protected] (G.S.) Extracellular GFP-binding proteins are readily To formalize the role of extracellular binders
†Present address: Department of Genetics and Evolution, University of engineered by fusing a transmembrane protein and leakage in GFP gradient formation, we
Geneva, Quai Ernest-Ansermet 30, 1205 Geneva, Switzerland. (e.g., human CD8) to one of the many charac- devised a diffusion–degradation–leakage math-
ematical model (Fig. 2A and supplementary
text), building on previous work (18, 19). Free

Stapornwongkul et al., Science 370, 321–327 (2020) 16 October 2020 1 of 6

RESEARCH | RESEARCH ARTICLE

GFP was assumed to diffuse in the inter- the concentration of GFP in the hemolymph. level of Nb1highCD8 ~20-fold, with hh-Gal4 and
cellular space with a diffusion constant D and The model also replicated the effect of the fat UAS-Nb1highCD8) (fig. S7A illustrates the quan-
to bind and unbind at rates kon and koff to body trap by increasing the hemolymph degra- tification) would lead to a steepening of the
receptors internalized and degraded at rate k: dation rate ~20-fold (Fig. 2C, purple curve). gradient and an increase in GFP level near
The flux between hemolymph and the epi- the source, although the latter was not as
The model could also be used for de novo marked in the experiment as in the model
thelium was assumed to be driven by the con- predictions. With the parameters determined (Fig. 3, compare A and C). Both model and
above, it predicted that increasing ligand pro- experiment showed a reduced nonzero tail
centration difference between them, with a duction at the source should lead to gradient in this condition, confirming that receptor-
proportionality coefficient k. At steady state, extension as well as flattening near the source mediated internalization contributes to limit-
in the posterior compartment (where the re- because of saturation (fig. S6A). This was ing leakage into the hemolymph. Therefore,
indeed found experimentally in imaginal discs increasing ligand–receptor avidity could con-
ceptors are expressed), the concentrations of overexpressing SecGFP under the control of tribute to reducing the amount of GFPhemo
free (c) and receptor-bound ðnbÞ GFP follow ptc-Gal4 (fig. S6, B to D). Also confirmed ex-
these equations: perimentally was the prediction that increasing flowing back in the tissue, although this could
receptor expression (achieved by boosting the be at the cost of a reduced range.
0 ¼ D@x2c À k nb À kðc À cHÞ ð1Þ
h

nb ¼ nT koff konc þ k ð2Þ
þ konc
A C
ptc domain SecGFP source no binders
where nT refers to the density of receptors at
the cell surface, h is the intercellular distance, hh domain binders
and cH is the free GFP concentration in the
hemolymph. Analytical exploration of the A/P boundary
model showed that it recapitulated the essen-
tial features of the bound GFP gradient profile wing pouch
(Fig. 2B and fig. S5): (i) close to the source,
receptor saturation leads to a shoulder; (ii) peripodial apical
further away from the source, the profile decays
on a length scale determined by the diffusion space ROI
constant, the degradation of receptors, and
leakage to the hemolymph; (iii) far from the ROI hh-Nb1highCD8
source, the concentration of GFP remains at
a constant nonzero value that depends on the B no binders basal
hemolymph GFP concentration.
hh-Nb1highCD8
We then tested whether the model, and its
consideration of leakage in particular, could hh-Nb1highCD8 + “fat body trap”
quantitatively account for observed experi-
mental profiles. To derive the concentration D no binders
cH of GFP in the hemolymph, we surmised Intensity norm. to source
that it is set by the balance between input from hh-Nb1highCD8
tissue leakage and loss by degradation in the
hemolymph (kH) (supplementary text, where hh-Nb1highCD8 hh-NblowCD8
we also discuss the contribution of other larval +“fat body trap”
tissues that produce and degrade the ligand).
Parameters were chosen from reasonable esti- hh-NblowCD8
mates or published data, with the remaining
unknown parameters obtained from a fit to A/P Distance from source [µm]
experimental curves, as described in table
S2. Our fitting procedure indicated a substan- Fig. 1. Establishment of a GFP gradient in a developing epithelium. (A) Schematic representation of a
tial leakage rate, k, of ~1/(13 s). Comparison of wing imaginal disc of Drosophila. SecGFP is expressed under the control of the ptc promoter (brown), and a
Figs. 2C and 1D shows that the model provides membrane-tethered anti-GFP nanobody is expressed under the control of the hh promoter (gray). The region
a suitable framework to rationalize experi- of interest (ROI) indicates the areas depicted in (B) and (C). Blue shading indicates the region used to
mental observations. The effect of reducing generate the profiles shown in (D). (B) In the absence of binders, SecGFP can be seen at the source but is
affinity was recapitulated by setting this pa- not detectable in basolateral focal planes. Upon expression of high-affinity binders in the posterior
rameter to that measured for Nblow (Fig. 2C, compartment (hh-Nb1highCD8), a gradient is readily seen. (C) Cross sections of imaginal discs expressing
blue curve). Being a poor binder, NblowCD8 is SecGFP in the ptc domain show that, in the absence of binders, GFP is detectable in the peripodial space but
unable to trap much GFP at the cell surface, not in the basolateral space. In the presence of binders (hh-Nb1highCD8), a gradient can be seen in the
reducing the amplitude of the gradient near basolateral space but with a nonzero tail, which is largely abrogated by concomitant activation of UAS-
the source. Consequently, NblowCD8 takes up Nb1highCD8 in the fat body (+ fat body trap). Only a shallow basolateral gradient is detected when a low-
and degrades GFP at a relatively low rate, affinity binder is expressed (hh-NblowCD8). (D) Fluorescence intensity profiles derived from preparations like
leading to increased leakage (supplementary those shown in (C). The vertical dotted line marks the estimated posterior edge of the source. The numbers
text, section 1.5). Thus, lowering ligand–binder of discs analyzed are as follows: no binders, n = 10; hh-Nb1highCD8, n = 11; hh-Nb1highCD8 + fat body trap,
affinity adversely affects the gradient both by n = 7; hh-NblowCD8, n = 10. Scale bars, 20 mm.
reducing gradient amplitude and by increasing

Stapornwongkul et al., Science 370, 321–327 (2020) 16 October 2020 2 of 6

RESEARCH | RESEARCH ARTICLE

Engineering a GFP-dependent signaling next asked if this gradient could provide posi- Bound GFP [normalized]A 2.5 High affinity binder
gradient in vivo tional information. The best-characterized mor- High affinity binder x20
In the previous section, we identified minimal phogen in wing imaginal discs is Dpp, which 2
conditions for an inert protein to form a gradient promotes growth and specifies the position of
along the plane of a developing epithelium. We veins along the A/P axis. As a first step toward 1.5
asking if GFP could substitute for Dpp in vivo,
Distance from source [µm] we engineered the Dpp receptors Thickveins 1
(Tkv) and Punt (Put) to render them respon-
Fig. 2. A diffusion–degradation–leakage model for sive to GFP. Normally, Dpp dimers bind to 0.5
GFP gradient formation. (A) Schematic representa- two pairs of Tkv and Put, leading to phospho-
tion of the model for gradient formation. Parameters rylation of Mad (20) and transcriptional re- 0 20 40 60 80
are described in the main text and supplementary pression of the brinker gene (brk) (21). The 0
text. (B) Main features of bound GFP profiles predicted resulting inverse Brk gradient in turn controls
by the model using parameters listed in table S2 the nested expression of target genes such as Distance from source [µm]
(unless specified otherwise). The yellow-green curve spalt (sal) and optomotor-blind (omb) (22, 23)
shows the profile exhibiting receptor saturation near the (Fig. 4A). We reasoned that GFP dimers might B hh-Gal4, UAS-Nb1highCD8
source and a nonzero tail due to GFPhemo (ligand initiate the same signaling cascade if Tkv and
production rate j ¼ 0:5nM=s; nT ¼ 100nM:mm). The Put were fused to anti-GFP nanobodies [GBP1 CIntensity norm. to source
blue curve shows that without saturation, the gradient is (referred to here as Nb1high) and GBP6 (here
an exponential with a nonpzeffiffirffioffiffiffiffitffiaffiffiiffilffi,ffiffinffiffib∞ffiffiffi≤ffiffi nTcHkon=koff, called Nb2high)] that recognize nonoverlapping hh-Nb1highCD8
and a decay length [l ¼ D=ðkr þ kÞ] that depends epitopes (16, 24) (fig. S8A). We created plasmids hh-Gal4, UAS-Nb1highCD8
on diffusion, effective degradation with rate kr, to express Nb2highTkv and Nb1highPut and co-
and leakage with rate k ( j ¼ 3Á10À4nM=s, transfected them in S2 cells. Addition of GFP A/P Distance from source [µm]
nT ¼ 3Á104nM:mm). The red curve shows that the dimers (or monomers) to the culture medium
ligand concentration set to zero in the hemolymph led to accumulation of phospho-Mad (pMad), Fig. 3. Predicted and experimental effects of
abolishes the nonzero tail (blue curve with cH ¼ 0). The suggesting that the chimeric receptors can be increasing binder expression. (A) Predicted GFP
supplementary text, section 1.3, provides full parameter activated by GFP (fig. S8B), although we cannot profile after a 20-fold increase in binder expression
definitions. (C) Bound GFP profiles normalized to be sure that the signaling kinetics normally (orange; compare to the green curve, which is
the total concentration of receptors. The blue and achieved by the natural ligand were entirely reproduced from Fig. 2C). Note the steep gradient
green curves were obtained with the known on- and recapitulated. and the lower nonzero tail. Bound GFP concen-
off-rates for the low- and high-affinity receptors, trations are normalized to the lower value of total
respectively. The purple curve was obtained by On the basis of these encouraging results receptor concentration. (B) Cross section of a ptc-
increasing degradation in the hemolymph. Compare with cultured cells, we created a transgene SecGFP imaginal disc overexpressing the high-
to corresponding experimental curves in Fig. 1D. that expresses, in a Flippase (Flp)-dependent affinity binder (hh-Gal4, UAS-Nb1highCD8).
manner, both engineered receptors under the (C) GFP profiles in hh-Gal4, UAS-Nb1highCD8
control of the ubiquitin (ubi) promoter (Fig. (orange curve, n = 8), and hh-Nb1highCD8 discs
4B). This transgene, ubi-[>STOP>Nb2highTkv (green curve from Fig. 1D). Scale bars, 20 mm.
2A Nb1highPut], where > indicates Flp recom-
bination targets, is referred to here as SR (for (26). In addition, pMad immunoreactivity was
signaling receptors). We also developed a dpp also present in a salt-and-pepper manner
allele that can be inactivated but at the same throughout the whole pouch, as if residual
time be made to express secreted GFP dimers signaling activity persisted far from the source.
upon Flp expression (dpp-[>Dpp>SecGFP:GFP]) In agreement, brk was repressed over a wider
(Fig. 4C; validated in fig. S9). First, we used the range than in control discs. Both sal and omb
previously described dpp-[>Dpp>] allele (25) to were expressed in GFP-rescued discs, although
confirm that inactivation of Dpp throughout the in a range that did not recapitulate the wild-
wing primordium with rotund-Gal4 (rn-Gal4) type situation; in the posterior compartment,
and UAS-Flp abrogated growth and patterning, the Sal domain boundary was fuzzy, whereas
even in the presence of GFP-responsive recep- the domain of Omb was too broad. Normally,
tors (Fig. 4D, column 2). Crucially, with dpp- Sal and Omb ensure the patterned down-
[>Dpp>SecGFP:GFP], which produces GFP regulation of the Drosophila serum response
upon Dpp inactivation, recognizably patterned factor (DSRF), which is required for vein fate
wings developed (Fig. 4D, column 3). Note that specification (27). However, in the “rescue”
no GFP gradient was detectable in this genetic condition, an oversized domain of DSRF down-
background, perhaps because of rapid inter- regulation could be seen in the posterior
nalization and degradation of GFP by the compartment, along with a corresponding
signaling receptors. The rescuing activity of
secreted GFP dimers was further assessed in
imaginal discs by staining for various markers
of Dpp signaling. pMad immunoreactivity was
unexpectedly low in the GFP-producing cells
(fig. S10). Most relevant to this study, however,
signaling activity was graded on either side of
the source, including in the posterior compart-
ment, which relies entirely on ligand diffusion

Stapornwongkul et al., Science 370, 321–327 (2020) 16 October 2020 3 of 6

RESEARCH | RESEARCH ARTICLE

Fig. 4. Rescue of growth and patterning by GFP. A B ubi 2A

(A) Target genes of Dpp signaling in the pouch, which Dpp STOP Nb2highTkv Nb1highPut

give rise to the wing. (B) Schematic representation of Flp SR
SR, the transgene for conditional expression of
engineered receptors. (C) The dpp locus engineered Brk C endo. dpp
to allow Flp-mediated replacement of an essential
region by sequences encoding secreted GFP dimers. Omb Sal 3. SR HA:Dpp SecGFP:GFP
Throughout the study, rn-Gal4 and UAS-Flp were used GFP Flp 5. SR + NR
to inactivate Dpp and/or trigger SR expression D dimer
specifically in the pouch. (D) Phenotypes of wing 4. SR + SR
imaginal discs and adult wings of various genotypes 1. positive control 2. negative control GFP only
(columns). The positive control (column 1) shows
Dpp dimer Nb2high Nb1high

Tkv Put no Dpp, no GFP GFP only NR
Dpp only
GPI
GFP only

imaginal discs and wings from a dpp-[>Dpp>SecGFP:GFP] Norm. intensity [%] pMad Brinker

homozygous larva. Flp is absent and Dpp is therefore

expressed as in wild-type discs. For the negative

control (column 2), we used larvae homozygous for a

different conditional allele (dpp-[>Dpp>]) (25) and pMad A P AP AP AP AP

carrying the SR transgene. Here, Flp expression Brinker

inactivates Dpp in the pouch without triggering

GFP:GFP production while at the same time

activating expression of the engineered receptors. Distance [µm] Distance [µm] Distance [µm] Distance [µm] Distance [µm]
The resulting phenotypes recapitulated those

of classical dpp mutants (e.g., brk derepression and

growth impairment). Abrogation of Dpp activity Omb

shows that the engineered receptors do not trigger

signaling in the absence of GFP. If, in combination

with the SR transgene, the dpp-[>Dpp>SecGFP:GFP]

allele is used (column 3, SR), signaling activity

(e.g., pMad immunoreactivity near the source) and Sal

growth are restored, albeit imperfectly. Note the

occasional spots of pMad throughout the pouch, the

expanded zone of brk repression, the fuzzy boundary of

sal expression in the posterior compartment, and the DSRF

disrupted vein pattern. Adding a second SR transgene

(column 4, same genotype as in column 3 with one

additional SR transgene, SR+SR) led to enhanced adult wing

pMad at the source and a narrowing of the signaling

gradient (relative to SR alone). Addition of non-

signaling receptors (column 5, same genotype as in
column 3 with addition of dally-NblowGPI, SR+NR)

extended the signaling gradient (relative to SR alone). Note the absence of background pMad far from the source and the wild-type–like expression of target genes.

Note, however, that vein L4 was often disrupted and vein L5 was slightly broadened in the distal part. Scale bars, 50 mm (for wing discs) or 0.25 mm (for adult wings).

sprawling vein L5 in surviving adult wings. SR), background pMad immunoreactivity was (fig. S5, F and G); at higher receptor density,
Ectopic vein material was also seen through- largely abrogated (Fig. 4D, column 4). The hemolymph concentration drops but the gra-
out the wings, which was probably a result of target genes omb and sal were still expressed dient scale shortens because of increased
global ectopic Dpp signaling. Despite these in a nested fashion; however, the width of degradation in the tissue. It appears, there-
limitations, the above results suggest that a these domains, as well as that of the zone of fore, that long-range GFP gradients with low
GFP gradient can stimulate growth and pro- brk repression, were narrower than in the residual signaling far from the source may only
vide substantial patterning information through wild type (Fig. 4D, compare column 1 with col- be achievable within a narrow range of pa-
engineered receptors. umn 4). This suggests that a twofold increase in rameters (fig. S5, F and G).
receptor expression had the beneficial effect
Signaling activity far from the GFP source, of reducing the adverse effect of leakage on Beneficial effects of GPI-anchored
e.g., in the form of ectopic pMad, suggests the signaling activity far from the source, but at nonsignaling receptors
presence of GFP dimers throughout the disc, the expense of a reduced range. These results
perhaps as a result of re-entry from the hemo- can be understood qualitatively in the context The above analysis suggests that, by solely mod-
lymph. During our initial analysis of the GFP of our gradient model (supplementary text ulating the expression of signaling receptors, it
gradient, we found that leakage could be re- section 1.3.2): at low receptor density, recep- is difficult to reduce leakage without shorten-
duced by increasing the level of extracellular tor activation is too low to trigger target gene ing the gradient. Natural morphogens bind not
binders (Fig. 3, B and C). We therefore asked if activation; at intermediate receptor density, only to signaling receptors but also to nonsig-
a similar strategy could be used to reduce leakage can lead to a high ligand concentra- naling extracellular proteins such as glypicans,
nonzero tail signaling in rescued discs. Indeed, tion in the hemolymph, triggering signaling glycosylphosphatidylinositol (GPI)–anchored
with two copies of the transgene expressing and target gene activation far from the source heparan sulfate proteoglycans (11). We there-
engineered GFP-responsive receptors (SR + fore set out to investigate whether low-affinity,

Stapornwongkul et al., Science 370, 321–327 (2020) 16 October 2020 4 of 6

RESEARCH | RESEARCH ARTICLE

A GFP C Fig. 5. Modeling the effect of GPI-anchored non-
SR signaling receptors on a gradient length scale.
NR High target activation domain (HT) (A) Schematic representation of the molecular inter-
actions considered by our model, including GFP
NR level NA RA [µm] dimer handover and NR hopping. Signal transduction
(yellow lightning bolt) is activated by GFP-bound
Signaling complex [nM µm]B NR level SR level [µm] signaling receptor dimers (SR). See supplementary
No activation (NA) Restricted activation (RA) text for details. (B) Predicted profiles of signaling
high target complexes in three conditions: a reference case with
HT signaling receptors only (SR; red), doubling SR
low target levels (SR + SR; green), and adding nonsignaling
Low target activation domain (LT) receptors (SR + NR; blue). As observed experimen-
Distance from source [µm] tally, doubling SR leads to a steeper gradient,
SR RA whereas adding NR reduces GFPhemo signaling and
SR+SR FA extends the gradient, due to nonsignaling receptor
SR+NR effective diffusion. For illustration, arbitrary thresholds
SR level were chosen to indicate the position where high- and
0 20 40 60 80 100 Full activation (FA) Restricted activation (RA) low-level target genes would be activated (tables S2
Distance from source [µm] and S3 report the parameter values). (C) Width of the
LT high (top) and low (bottom) target activation domains
[arbitrary threshold shown in (B)], as a function of
normalized levels of SR and NR. Warmer colors
indicate a wider target activation domain. Colored dots
show parameter combinations used in (B). (Top) For
the normalized SR value of 1, increasing NR initially
lengthens the high target domain, while a further
increase shortens it by preventing access of GFP to
SR [as observed experimentally (fig. S11)]. (Bottom)
For the normalized SR value of 1 and in the absence of
NR, GFPhemo signaling dominates and low target gene
is activated throughout (bright yellow region).
Increasing SR or NR production both lead to a
reduction in the low target domain size.

GPI-anchored, extracellular binders (nonsig- signaling and nonsignaling receptors, target selves nearby, a process we call hopping (Fig.
gene expression (controlled entirely by GFP 5A). Simulations introducing a tissue-scale effec-
naling receptors, NR) would improve the per- dimers) was comparable to that in wild-type tive diffusion constant,Dr ¼ 0:1mm2=s, for GFP-
imaginal discs, and the resulting wings were NblowGPI (representing lateral diffusion in the
formance of the signaling gradient. As a first notably well patterned, proportioned, and con- cell membrane and intercellular hopping) can
sistently sized (fig. S12). explain the extension of the signaling gradient
step, we created a DNA fragment encoding by Nblow-GPI (Fig. 5B). In fact, we find that,
Nblow [Kd of 25 nM, chosen to mimic the To rationalize how nonsignaling receptors in the absence of NR diffusion (or with a low
affinity of BMP for heparin (28)] tethered to could improve the signaling gradient’s char- diffusion constant), NRs can only shorten the
the extracellular face of the plasma mem- acteristics, we devised a formal description of gradient, as nondiffusing NRs provide an addi-
brane by a GPI anchor, NblowGPI. This fragment the relevant molecular interactions (Fig. 5A, fig. tional route for ligand degradation without
was expressed with rn-Gal4, which concomi- S13C, supplementary text and table S3). In this contributing to ligand spread (fig. S13, D and
tantly triggered expression of Flp to inactivate framework, GFP dimers bind signaling and E, and supplementary text). At a high NR con-
Dpp and initiated expression of SecGFP and nonsignaling receptors and transit from one centration, competition for the ligand with SR
the engineered receptors. In the resulting configuration to another. Signaling receptors inhibits signaling (Fig. 5C), as observed exper-
typically undergo rapid endocytosis upon bind- imentally (fig. S11). For intermediate concen-
imaginal discs, the signaling activity of SR ing to their ligands. By contrast, GPI-anchored trations of the NR, however, NR diffusion
binders could have a longer lifetime (29), allow- enables the gradient range to increase while
was suppressed (fig. S11), perhaps because ing them to hand over ligands to signaling preventing uniform activation of low-target
excess NblowGPI prevented SecGFP from ac- receptors (30). Simulations showed, however, genes by leaked ligand.
that addition of membrane-tethered nonrecep-
cessing the signaling receptors. To achieve a tors does not extend the signaling gradient, al- Conclusion
though they can alter its shape near the source
more reasonable expression level, we inserted (supplementary text and fig. S13D). We next We have shown that, in the presence of ex-
the NblowGPI-encoding fragment in the dally considered the relevance of the labile nature tracellular binders, GFP can form a gradient
locus, one of the two glypican-encoding genes of GPI anchors (31). Locally expressed GFP-GPI in an epithelial tissue. Because GFP is inert in
of Drosophila. This allele (dally-NblowGPI) spreads within wing imaginal discs (32, 33), wing imaginal discs, it is unlikely to spread by
was then combined with all the previously suggesting that GPI-anchored proteins can a specialized transport mechanism, such as
detach from cells and possibly reinsert them- planar transcytosis. The low off-rate of Nb1high
described genetic elements needed for a GFP

signaling gradient to form. Addition of this

nonsignaling receptor extended the pMad gra-
dient and narrowed the domain of brk repres-
sion, an indication of reduced GFPhemo signaling

far from the source (Fig. 4D, compare columns

3 and 5). Indeed, with this combination of

Stapornwongkul et al., Science 370, 321–327 (2020) 16 October 2020 5 of 6

RESEARCH | RESEARCH ARTICLE

(koff = 1.7 × 10−4 s−1) also limits the contribu- 5. T. B. Kornberg, A. Guha, Curr. Opin. Genet. Dev. 17, 264–271 ACKNOWLEDGMENTS
tion of ligands passing from one receptor to (2007).
another. We therefore suggest that the GFP We thank C. Alexandre for generating vg-Gal4 and dally-attP and
gradient forms by free diffusion, even though 6. M. Affolter, K. Basler, Nat. Rev. Genet. 8, 663–674 (2007). advice on genome engineering and the Crick Fly Facility for DNA
the readily detectable gradient is largely made 7. E. V. Entchev, A. Schwabedissen, M. González-Gaitán, Cell 103, injections. We also acknowledge the technical and intellectual
up of bound GFP (Fig. 1C and fig. S5, H to K). contributions of S. Crossman and I. McGough. J. Briscoe provided
In the presence of engineered GFP-responsive 981–992 (2000). comments on the manuscript. We are also grateful to A. Lander for
receptors (SR), diffusing GFP can act as a 8. S. Roy, H. Huang, S. Liu, T. B. Kornberg, Science 343, pointing out the possible relevance of GPI’s loose anchorage to
morphogen. One limitation of free diffusion membranes. We thank H. Ashe, G. Pflugfelder, and A. Salzberg for
is that it allows leakage into the circulation, 1244624–1244624 (2014). the generous gift of antibodies. The Developmental Studies
a potential threat to positional information. 9. U. Häcker, K. Nybakken, N. Perrimon, Nat. Rev. Mol. Cell Biol. 6, Hybridoma Bank also provided antibodies. Drosophila stocks
As we have shown, signaling from leaked GFP obtained from the Bloomington Drosophila Stock Center (NIH
can be reduced by increasing the level of SR. 530–541 (2005). P40OD018537) were used in this study. Funding: This work was
However, this was at the expense of a reduced 10. H. Nakato, J. P. Li, Int. Rev. Cell Mol. Biol. 325, 275–293 (2016). supported by core funding from the Francis Crick Institute
range. Leakage can be reduced without a con- 11. D. Yan, X. Lin, Cold Spring Harb. Perspect. Biol. 1, (FC001204 to J.-P.V. and FC001317 to G.S.) and a Wellcome Trust
comitant decrease in gradient range by adding Investigator Award to J.-P.V. (206341/Z/17/Z). K.S.S. was the
GPI-anchored nonsignaling receptor (NR). We a002493–a002493 (2009). recipient of a PhD Studentship from the Wellcome Trust (109054/
suggest that, by virtue of their labile associa- 12. P. Li et al., Science 360, 543–548 (2018). Z/15/Z). Author contributions: This project was conceived by
tion with cell membranes, GPI-anchored non- 13. S. Toda et al., Science 370, 327–331 (2020). K.S.S., J.-P.V., and G.S. K.S.S. designed and performed all the
signaling receptors can undergo tissue-level 14. S. Harmansa, I. Alborelli, D. Bieli, E. Caussinus, M. Affolter, experiments. The results were analyzed by all authors. The
diffusion and thus extend the gradient. Al- model was conceived by G.S., M.d.G., and L.C., and numerical
though this hypothesis remains to be dem- eLife 6, e22549 (2017). simulations were performed by M.d.G. and L.C. The main text was
onstrated experimentally, our results so far 15. U. Rothbauer et al., Nat. Methods 3, 887–889 (2006). written by K.S.S., J.-P.V., and G.S. with comments from M.d.G.
show that a combination of free and NR- 16. A. Kirchhofer et al., Nat. Struct. Mol. Biol. 17, 133–138 (2010). and L.C. Competing interests: The authors declare no competing
assisted diffusion suffices to emulate the 17. P. C. Fridy et al., Nat. Methods 11, 1253–1260 (2014). or financial interests. All data are described in the main text or
range and activity of a natural morphogen. 18. A. D. Lander, Q. Nie, F. Y. M. Wan, Dev. Cell 2, 785–796 (2002). supplementary materials. Data and materials availability: All
19. T. Bollenbach, K. Kruse, P. Pantazis, M. González-Gaitán, materials are available upon request. A link to the computer code
REFERENCES AND NOTES can be found in the supplementary text.
F. Jülicher, Phys. Rev. E Stat. Nonlin. Soft Matter Phys. 75,
1. P. Müller, K. W. Rogers, S. R. Yu, M. Brand, A. F. Schier, 011901 (2007). SUPPLEMENTARY MATERIALS
Development 140, 1621–1638 (2013). 20. S. J. Newfeld et al., Development 124, 3167–3176 (1997).
21. B. Müller, B. Hartmann, G. Pyrowolakis, M. Affolter, K. Basler, science.sciencemag.org/content/370/6514/321/suppl/DC1
2. S. Zhou et al., Curr. Biol. 22, 668–675 (2012). Materials and Methods
3. S. R. Yu et al., Nature 461, 533–536 (2009). Cell 113, 221–233 (2003). Supplementary Text
4. A. D. Lander, Q. Nie, B. Vargas, F. Y. Wan, J. Mech. Mater. 22. R. Barrio, J. F. de Celis, Proc. Natl. Acad. Sci. U.S.A. 101, Figs. S1 to S13
Tables S1 to S3
Struct. 6, 321–350 (2011). 6021–6026 (2004). References (34–61)
23. R. Sivasankaran, M. A. Vigano, B. Müller, M. Affolter, K. Basler, MDAR Reproducibility Checklist

EMBO J. 19, 6162–6172 (2000). View/request a protocol for this paper from Bio-protocol.
24. J. C. Tang et al., Cell 154, 928–939 (2013).
25. P. S. Bosch, R. Ziukaite, C. Alexandre, K. Basler, J.-P. Vincent, 21 March 2020; accepted 21 August 2020
10.1126/science.abb8205
eLife 6, e22546 (2017).
26. C. J. Evans et al., Nat. Methods 6, 603–605 (2009).
27. J. F. De Celis, BioEssays 25, 443–451 (2003).
28. R. Ruppert, E. Hoffmann, W. Sebald, Eur. J. Biochem. 237,

295–302 (1996).
29. S. Mayor, H. Riezman, Nat. Rev. Mol. Cell Biol. 5, 110–120 (2004).
30. J. Schlessinger, I. Lax, M. Lemmon, Cell 83, 357–360 (1995).
31. G. A. Müller, Arch. Biochem. Biophys. 656, 1–18 (2018).
32. V. Greco, M. Hannus, S. Eaton, Cell 106, 633–645 (2001).

33. C. Tempesta, A. Hijazi, B. Moussian, F. Roch, PLOS ONE 12,
e0185897 (2017).

Stapornwongkul et al., Science 370, 321–327 (2020) 16 October 2020 6 of 6

RESEARCH

◥ S2B). We observed the formation of a long-
range gradient (~4 mm) of both secreted GFP
REPORT and synNotch reporter gene expression (mCherry)
in contrast to the narrow band of activation
MORPHOGENS observed when pole cells expressed membrane-
tethered GFP (Fig. 1D and fig. S2C). The GFP
Engineering synthetic morphogen systems that can gradient formed within ~24 hours, whereas
program multicellular patterning induced mCherry reporter expression reached
steady state in ~96 hours (fig. S2, C and D, and
Satoshi Toda1*†, Wesley L. McKeithan1, Teemu J. Hakkinen2, Pilar Lopez1, movie S1). We found that another bivalently
Ophir D. Klein2,3, Wendell A. Lim1* recognized protein—a fusion protein of mCherry
with a PNE peptide (mCherry-PNE)—could also
In metazoan tissues, cells decide their fates by sensing positional information provided by specialized be recognized by an analogous anchor-receptor
morphogen proteins. To explore what features are sufficient for positional encoding, we asked system (fig. S3 and supplementary text), which
whether arbitrary molecules (e.g., green fluorescent protein or mCherry) could be converted into further supports the idea that arbitrary proteins
synthetic morphogens. Synthetic morphogens expressed from a localized source formed a gradient can be converted into morphogens (Fig. 1B).
when trapped by surface-anchoring proteins, and they could be sensed by synthetic receptors. Despite
their simplicity, these morphogen systems yielded patterns reminiscent of those observed in vivo. To investigate how the shape of a synthetic
Gradients could be reshaped by altering anchor density or by providing a source of competing inhibitor. morphogen gradient could be modulated, we
Gradient interpretation could be altered by adding feedback loops or morphogen cascades to receiver systematically perturbed different interaction
cell response circuits. Orthogonal cell-cell communication systems provide insight into morphogen parameters. Using the LaG2 anti-GFP nanobody
evolution and a platform for engineering tissues. as an anchor protein, we tested how anchor
protein density can regulate gradient shape.
D evelopment of multicellular organisms ognition domain [e.g., nanobody or single We generated four different levels of anchor
requires precise spatial control of cell chain antibody (scFv)], the Notch core reg- protein density by varying both the anchor
fate. Morphogens are molecules that ulatory domain, and an intracellular tran- protein expression level and the anchor cell
provide positional information. They scriptional domain (e.g., TetR-VP64). When an number in the body. More specifically, we
diffuse from a source to form a concen- anti–green fluorescent protein (GFP) synNotch constructed high- or low-expressing anchor
tration gradient that is interpreted by neigh- receptor recognizes membrane-tethered GFP cells and mixed them at 50% (a 1:1 ratio) with
boring cells (1–4). In metazoans, a small set of on a neighboring sender cell, the synNotch receiver cells in the body (Fig. 2A and fig. S4).
specialized molecules, including sonic hedgehog core undergoes cleavage, which releases its We also created a combination anchor-receiver
(Shh), Wnt, fibroblast growth factor (FGF), and intracellular domain to enter the nucleus to L929 cell, which was used at 100% in the body,
bone morphogenetic protein (BMP)–transforming activate target gene expression. Soluble GFP yielding an approximately twofold higher den-
growth factor–b (TGFb) family members, serve does not activate synNotch because exposure sity of anchor proteins for a given expression
as morphogens (5). Reconstitution of morpho- of the cleavage site requires the mechanical level (fig. S4). These studies showed that when
gen signaling in vitro is a powerful approach force of a membrane-tethered ligand (8). Here, the overall density of anchor protein was re-
to understand how morphogens encode posi- we reengineered the synNotch system to detect duced, the maximal signal amplitude close
tional information (6, 7). To define the mini- soluble molecules by tethering the diffusible to the pole decreased, but the signal range
mal requirements for a functional morphogen, ligand to a complementary engineered anchor extended a longer distance as morphogen dif-
we asked whether it is possible to construct a cell (Fig. 1A). In the case of GFP, we used two fused further before being trapped. The impor-
synthetic morphogen signaling system that func- noncompetitive anti-GFP nanobodies: One tance of anchor density as a determinant of
tions orthogonally to endogenous morphogens. serves as the anchor binding domain, and morphogen gradient shape is consistent with
Orthogonal morphogen signaling would en- the other serves as the receptor binding do- trends predicted by a simple computational
able the systematic exploration of patterning main (Fig. 1B). We designed this diffusible model (fig. S5B).
circuits, free from confounding cross-talk with synNotch system with anti-GFP LaG2 nano-
endogenous systems. body fused to a transmembrane domain as the The spatial distribution of morphogens can
anchor molecule and anti-GFP LaG17 nanobody also be regulated by antimorphogen inhibitors.
To create synthetic morphogens, we modi- as the recognition domain of the synNotch re- For example, during Xenopus embryogenesis,
fied the recently developed synthetic Notch ceptor (fig. S1A). These nanobodies recognize BMP is secreted from a ventrally located pole,
(synNotch) system to detect user-defined solu- different sites on GFP (9). SynNotch receiver whereas Chordin—a BMP-binding inhibitor—
ble factors. SynNotch receptors are a modular cells were only activated by soluble GFP in is secreted from an opposing pole (10, 11). Such
platform for engineering orthogonal juxtacrine the presence of anchor cells (Fig. 1C and fig. antagonism by an inhibitor is a common theme
signaling (8), composed of an extracellular rec- S1, B and C). observed in development. Inspired by activator-
inhibitor opposition, we designed a three-region
1Cell Design Institute, Department of Cellular and Molecular To test whether this soluble synNotch sys- configuration on the culture dish: a central body
Pharmacology, and Howard Hughes Medical Institute, tem allowed GFP to function as a morphogen, surrounded by a morphogen-secreting pole
University of California San Francisco, San Francisco, CA we reconstituted an in vitro model of L929 and an inhibitor-secreting pole on opposite
94158, USA. 2Program in Craniofacial Biology and fibroblast cells organized into a pole and a sides (Fig. 2B; see fig. S6, A and B, and mate-
Department of Orofacial Sciences, University of California body (fig. S2A) (culture well divided by insert rials and methods for development of mor-
San Francisco, San Francisco, CA 94143, USA. 3Department wall; see materials and methods). The pole phogen inhibitor). We tested both morphogen
of Pediatrics and Institute for Human Genetics, University of was composed of GFP secretor cells and the and inhibitor sets (GFP and mCherry-PNE) in
California San Francisco, San Francisco, CA 94143, USA. body of mixed anchor and receiver cells. To the double-pole system. When the GFP inhib-
*Corresponding author. Email: [email protected] (W.A.L.); minimize disruptive convective flow, we used itor was secreted from the opposing pole, we
[email protected] (S.T.) a solidified media containing 1% agarose (fig. observed a reduced amplitude of the activation
†Present address: WPI Nano Life Science Institute, Kanazawa gradient generated by the GFP morphogen
University, Kanazawa 920-1192, Japan. (fig. S6C). The mCherry-PNE inhibitor showed

Toda et al., Science 370, 327–331 (2020) 16 October 2020 1 of 5

RESEARCH | REPORT

Fig. 1. Turning arbitrary proteins A B
into synthetic morphogens.
(A) SynNotch receptors detect synNotch: Diffusible synNotch: Kd=16 nM Kd=0.04nM
juxtacrine signals (e.g., membrane- Juxtacrine Trapping and Signaling
tethered GFP). In the diffusible ANCHOR
synNotch system, soluble
GFP is produced from a secretor SENDER SECRETOR ANCHOR anti-GFP Anchor anti-PNE
cell, then trapped by anti-GFP CELL
anchor protein, and finally CELL CELL (LaG2) anchor Anchor
presented to anti-GFP synNotch membrane
on a receiver cell. (B) Multiple GFP Anchor peptide
arbitrary proteins with two recog- anti-GFP tag (PNE)
nition sites could be converted anti-GFP synNotch
synNotch diffusion GFP Protein X

soluble mCherry
GFP
anti-GFP anti-mCherry Receptor
synNotch (LaG17) synNotch

into synthetic morphogens (see mCherry mCherry RECEPTOR

fig. S3 for construction of RECEIVER CELL RECEIVER CELL Kd=50 nM Kd=0.18nM
mCherry-PNE peptide morpho-

gen). Kd, dissociation constant. C D Juxtacrine ligand Diffusible ligand

(C) Testing diffusible GFP GFP Morphogen memb. BODY GFP BODY
synNotch system in L929 mouse Receiver Cell
GFP 4 mm 4 mm
fibroblasts. Anchor cell expresses
SECRETOR ON
anti-GFP LaG2 anchor protein. +

Receiver cell expresses anti-GFP ANCHOR L929

LaG17 synNotch (induces mCherry SECRETOR OFF mCherry Output
reporter). 1 × 104 GFP-secreting + mCherry Output
cells, 0.5 × 104 anchor cells, and 500µm 500µm
0.5 × 104 receiver cells were parental L929 10 OUTPUT 50

cultured overnight, and mCherry No SECRETOR OFF 96hr OUTPUT

induction in receiver cells was 102 103 104 105 0 96hr
0 1000 2000 3000 4000
measured by flow cytometry. mCherry Distance (µm) 0
(OUTPUT) 0 1000 2000 3000 4000
(D) Juxtacrine versus diffusible Distance (µm)
GFP signaling gradient. Left pole
has 3 × 104 sender cells, and right
body has 1.5 × 104 cells (100%

receiver cells for juxtacrine; 50:50

anchor:receiver cells for diffusible GFP; see fig. S2, A and B). Images were taken by incucyte system over 4 days when system reached steady state (movie S1). Individual

lines show mCherry intensity every 12 hours. memb., membrane.

a somewhat different pattern, reducing the anchor cells when engineered with sufficiently pathways involving self-induced morphogen
signaling range of the activation gradient gen- low background expression of GFP (fig. S7A). degradation, endocytosis, or inhibition (6, 12).
erated by mCherry-PNE morphogen with smaller By contrast, when a body of positive feedback In these cases, the common principle of
effects on amplitude (Fig. 2B and supplemen- receiver cells (with anchor cells) was placed morphogen-induced negative feedback is
tary text). Together, these results show that the next to a pole of GFP-producing cells, the thought to lead to higher levels of negative
signaling range and amplitude of a synthetic positive feedback circuit caused rapid, high- regulation in the vicinity of the morphogen
morphogen gradient can be tuned to gen- amplitude spatial propagation of morphogen source but lower levels of negative regulation
erate a variety of shapes similar to those seen signal and reporter gene activation across at a greater distance from the source, which
in embryos. the entire body (Fig. 3, B and C; fig. S7C; and yields a more extended and stable gradient.
movie S1).
In natural morphogen systems, receiver cells We have created a modular toolkit of in-
can interpret a morphogen gradient in diverse, We also constructed intercellular negative tercellular signaling components that can
higher-order ways. We explored several mech- feedback receiver cells, in which the receiver flexibly reshape morphogen production and
anisms to reshape how the body cells interpret cells sense the GFP morphogen and, in re- interpretation. We wanted to test whether
the synthetic morphogen gradient, taking ad- sponse, produce the soluble GFP inhibitor these modules could be combined to program
vantage of the flexibility with which synNotch (Fig. 3A and fig. S7B). When negative feed- higher-order pattern formation, such as the
signaling can be engineered to drive any ge- back receiver cells (mixed with anchor cells) formation of multiple distinct segments with-
netically encoded payload. We particularly fo- were placed in the body, the maximal am- in a body plan (Fig. 4). We first asked whether
cused on engineering intercellular positive and plitude of the response gradient was notably we could create a body plan with two distinct
negative feedback loops among receiver cells dampened. The negative feedback activity gra- domains (activated and unactivated). To do
(Fig. 3). dient reached a stable steady state far more so, we combined the positive feedback mor-
rapidly (~30 versus 70 hours for negative phogen circuit with a counteracting inhibitor
We first designed an intercellular positive feedback versus nonfeedback cells) (Fig. 3D pole on the opposite side of the body (Fig. 4A).
feedback circuit in which the receiver cells and fig. S7D). Such stable gradient forma- This circuit yielded two distinct domains: re-
sense the GFP morphogen and, in response, tion is similar to that which is observed in ceiver cells close to morphogen source triggered
induce expression of more GFP (Fig. 3A). response to natural morphogens (e.g., wing- positive feedback to drive strong synNotch ac-
Cells with this positive feedback circuit did less or hedgehog), with negative feedback tivation, whereas signal activation close to the
not trigger spontaneous activation with the

Toda et al., Science 370, 327–331 (2020) 16 October 2020 2 of 5

RESEARCH | REPORT B

A

Fig. 2. Systematic control over distance range of synthetic morphogen gradient. morphogen pole, body, and inhibitor pole. The morphogen pole has a mixture of
(A) Anchor density can tune synthetic GFP morphogen gradient shape and signaling 1 × 104 mCherry-PNE–secreting cells and 2 × 104 parental L929 cells. The body has
range. We constructed bodies with four different densities of LaG2-anchor by a 1.5 × 104 mixture of anti-PNE anchor cells and anti-mCherry synNotch receiver
using two types of variations: anchor expression level (high or low) and fraction
of cells in the body that express anchor (100 or 50%). See materials and cells (50:50 ratio). The inhibitor pole has cells expressing anti–mCherry-PNE inhibitor
methods for details. Norm., normalized. (B) Controlling signaling range of mCherry- (total cell number: 3 × 104 of cells; varying number of inhibitor cells: 0, 0.1 × 104,
PNE morphogen with inhibitor. We used a three-well insert wall to build three regions: 0.3 × 104, or 1.0 × 104; remaining cells were parental L929). BFP output was quantified

by In Cell Analyzer 6000 at day 4 (see fig. S6 for GFP inhibitor analysis).

Fig. 3. Reshaping morphogen interpretation A B No feedback
with positive or negative feedback. (A) In a 96 hr
positive feedback circuit, GFP morphogen activates morphogen feedback circuit C
receiver cells to induce the secretion of more 500µm
GFP. In a negative feedback circuit, GFP morphogen anti-GFP 80
induces the expression of antimorphogen inhibitor GFP inhibitor Diffuse 40 Positive feedback
by receiver cells. TF, transcription factor. 96 hr
(B) Comparison of mCherry output in the body or
with and without positive feedback at 96 hours No feedback (n=6)
(see fig. S7C for time course). The pole has Input aGFP Positive feedback (n=4)
3 × 104 GFP-secreting cells; the body has a synNotch
1.5 × 104 mixture of anchor cells and receiver Secretion 96hr
cells engineered with a positive feedback circuit
(50:50 ratio). Images were taken by incucyte Anchor TF
system for 4 days (movie S1). (C) Activity
gradient profiles at 96 hours, with and without target gene
positive feedback. Shaded area shows SD from /mCherry
multiple experiments. (D) The mCherry-positive
area (integral of top plots) plotted over time Positive feedback circuit mCherry intensity
shows that the body with negative feedback
reaches steady state faster than it does without Anchor cell
feedback. AU, arbitrary units.
GFP

spatial propagation 0mCherry-positive area (AU)
0 1000 2000 3000 4000
Negative feedback circuit Distance (µm)

Anchor cell D

GFP 5 No feedback
4 (n=6)

3
2 + Negative feedback

(n=3)
1

0
0 20 40 60 80 100
Time (hr)

Toda et al., Science 370, 327–331 (2020) 16 October 2020 3 of 5

RESEARCH | REPORT

Fig. 4. Combining synthetic A B
morphogen interpretation C D
circuits to engineer multido-
main spatial patterns. (A and
B) Programming two-domain pat-
tern by combining positive
feedback circuit with opposing
morphogen and inhibitor poles.
Morphogen pole (X) has 3 × 104
GFP-secreting cells; the body has
a 50:50 mixture of anchor cells
and receiver cells with positive
feedback (used in Fig. 3B) (1.5 ×
104 total cells); and the inhibitor
pole (Y) has 2 × 104 anti-GFP
inhibitor–secreting cells. Images
were taken by incucyte at
120 hours (movie S1). See
fig. S8B for variant circuits.
(C and D) Programming three-
domain pattern. We combined the
two-morphogen cascade and
positive feedback circuit with
opposing morphogen and inhibitor
poles. The body contains two types
of cells: Cell A expresses anti-
mCherry synNotch that induces
BFP reporter and GFP morphogen
(mCherry-PNE→GFP cascade),
and cell B expresses anti-GFP LaG17
synNotch that induces expression
of GFP morphogen (GFP→GFP
positive feedback). See materials
and methods for details. Image was
taken at 96 hours by In Cell
Analyzer 6000 (movie S2). IFP,
infrared fluorescent protein.

inhibitor source was blocked, which lead to a hibitor to sharpen the initial morphogen gra- ceptor activation. Whereas natural morphogens
nonlinear, switch-like transition from active to dient. With this composite circuit, we observe often function autonomously, most participate
inactive domains (Fig. 4B). By changing the the robust formation of three distinct domains in weak tethering interactions that are analo-
number of GFP inhibitor–secreting cells, we (Fig. 4D): a BFP+GFP+ domain closest to the gous to anchoring—whether interacting with
could tune the widths of the domains (fig. S8, mCherry-PNE morphogen pole; a BFP–GFP+ cell surface proteoglycans, extracellular matrix,
A and B). middle domain (where the GFP+ region is ex- or cell membranes through lipid modifications
tended by GFP→GFP positive feedback); and, (fig. S9A) (13–15). These tethering interactions
Finally, we set out to engineer a circuit that furthest (closest to the inhibitor pole), a BFP–GFP– are proposed, in many cases, to constrain sig-
produces a three-domain body, akin to Wolpert’s domain (fig. S8D and movie S2). This toolkit naling range and to prevent the leakage of
classic French Flag pattern (2). We designed a of synthetic cell-cell communication compo- morphogens (16, 17) (supplementary text and
circuit that incorporates a two-morphogen nents can be used to write spatial programs fig. S10).
cascade, positive feedback, and opposing pole capable of encoding multiple, distinct body
inhibition (Fig. 4C). In this circuit, the left domains. These synthetic morphogen platforms can
pole secretes mCherry-PNE morphogen. The program positional information without cross-
body contains two types of uniformly mixed Although evolution has relied on a relatively talk to endogenous signaling pathways. Thus, it
receiver cells (fig. S8C): receiver A cells that small set of specialized morphogen families, may be possible to deploy them in vivo as inert
sense mCherry-PNE to induce the expression we find that arbitrary proteins with no known tools to probe or redirect development. Related
of a blue fluorescent protein (BFP) reporter history of functioning as morphogens can be studies (18) have shown that an analogous
and GFP morphogen (mCherry-PNE→GFP converted into effective morphogens if they synthetic GFP morphogen can function in vivo
two-morphogen cascade) and receiver B cells are deployed with a complementary system of in Drosophila. Thus, these synthetic morpho-
that sense GFP to induce the secretion of GFP receptors, anchoring interactions, and inhib- gen systems could be used to facilitate con-
(GFP→GFP positive feedback) (cells A and B itors. These synthetic morphogens differ from trolled forward engineering of tissues and
are engineered to serve as anchors for each natural morphogens in that they explicitly re- organs, both in a native-like or a modified
other). To oppose morphogen signaling, the quire a distinct anchoring protein to constrain fashion. We found that the diffusible synNotch
right pole secretes the anti–mCherry-PNE in- their distribution and mediate synNotch re- system functioned in other cell types, including

Toda et al., Science 370, 327–331 (2020) 16 October 2020 4 of 5

RESEARCH | REPORT

immune cells (fig. S11), which provides further 15. Y. Wang, X. Wang, T. Wohland, K. Sampath, eLife 5, e13879 manuscript. T.J.H. designed and tested simulation works,
possibilities as to how such synthetic signaling (2016). wrote methods for simulations, and edited the manuscript.
systems could be deployed to shape spatially P.L. helped with material construction. O.D.K. oversaw research
controlled functions in vivo. 16. P. Müller, K. W. Rogers, S. R. Yu, M. Brand, A. F. Schier, and edited the manuscript. W.A.L. developed, planned, and
Development 140, 1621–1638 (2013). oversaw research and wrote and edited the manuscript.
REFERENCES AND NOTES Competing interests: W.A.L. and S.T. have a financial interest
17. T. B. Kornberg, A. Guha, Curr. Opin. Genet. Dev. 17, 264–271 (2007). in Gilead Biosciences. W.A.L. and S.T. are inventors on a patent
1. A. M. Turing, Phil. Trans. R. Soc. Lond. B 237, 37–72 18. K. S. Stapornwongkul, M. de Gennes, L. Cocconi, G. Salbreux, application (PCT/US2016/019188) held by the Regents of
(1952). the University of California that covers binding synthetic Notch
J.-P. Vincent, Science 370, 321–327 (2020). receptors. Data and materials availability: All data are
2. L. Wolpert, J. Theor. Biol. 25, 1–47 (1969). available in the main manuscript and supplementary materials.
3. K. W. Rogers, A. F. Schier, Annu. Rev. Cell Dev. Biol. 27, ACKNOWLEDGMENTS All plasmids developed in this study will be deposited at
Addgene (www.addgene.org/Wendell_Lim/), where they will
377–407 (2011). We thank A. McMahon for Shh-GFP and L. Morsut, K. Roybal, be publicly available.
4. A. D. Lander, Science 339, 923–927 (2013). J. Brunger, N. Frankel, and members of the Lim laboratory for
5. T. Tabata, Y. Takei, Development 131, 703–712 (2004). discussion and assistance. We also thank members of the SUPPLEMENTARY MATERIALS
6. P. Li et al., Science 360, 543–548 (2018). University of California San Francisco Center for Systems and
7. R. Sekine, T. Shibata, M. Ebisuya, Nat. Commun. 9, 5456 Synthetic Biology and the NSF Center for Cellular Construction. science.sciencemag.org/content/370/6514/327/suppl/DC1
Funding: This work was supported by the Human Frontiers Materials and Methods
(2018). of Science Program (HFSP); the Senri Life Science Foundation; Supplementary Text
8. L. Morsut et al., Cell 164, 780–791 (2016). the Kato Memorial Bioscience Foundation; JSPS KAKENHI Figs. S1 to S11
9. P. C. Fridy et al., Nat. Methods 11, 1253–1260 (2014). grant no. 20K15828 and World Premier International Research Table S1
10. E. M. De Robertis, H. Kuroda, Annu. Rev. Cell Dev. Biol. 20, Center Initiative (WPI), MEXT, Japan (to S.T.); a Ruth L. Kirschstein References (19–27)
National Research Service Award (NRSA) Individual Postdoctoral MDAR Reproducibility Checklist
285–308 (2004). Fellowship F32DK123939 (to W.L.M.); the NSF DBI-1548297 Movies S1 and S2
11. E. Bier, E. M. De Robertis, Science 348, aaa5838 (2015). Center for Cellular Construction; the DARPA Engineered
12. A. Eldar, D. Rosin, B.-Z. Shilo, N. Barkai, Dev. Cell 5, 635–646 Living Materials program; and the Howard Hughes Medical View/request a protocol for this paper from Bio-protocol.
Institute (to W.A.L.). T.J.H. and O.D.K. were supported by NIH
(2003). R01-DE028496 and R35-DE026602. Author contributions: 31 March 2020; accepted 24 August 2020
13. D. Yan, X. Lin, Cold Spring Harb. Perspect. Biol. 1, a002493 S.T. developed and planned research; carried out design, 10.1126/science.abc0033
construction, and testing of multicellular patterning circuits;
(2009). and wrote the manuscript. W.L.M. planned research, carried out
14. A. Parchure, N. Vyas, S. Mayor, Trends Cell Biol. 28, 157–170 testing of multicellular patterning circuits, and edited the

(2018).

Toda et al., Science 370, 327–331 (2020) 16 October 2020 5 of 5

RESEARCH

ULTRACOLD CHEMISTRY of the spatial wave function (da), as illustrated
in Fig. 1A. Thus, the atomic motion and the
Coherently forming a single molecule in
an optical trap spin are coupled with each other, that is, there
exists an SMC (34). Consequently, the wave
Xiaodong He1,2*†, Kunpeng Wang1,2,3*, Jun Zhuang1,2,3, Peng Xu1,2, Xiang Gao4,5, Ruijun Guo1,2,3, function overlap jhn j n′ij ðn ≠ n′Þ between
Cheng Sheng1,2, Min Liu1,2, Jin Wang1,2, Jiaming Li6,7,8, G. V. Shlyapnikov9,10,11, Mingsheng Zhan1,2† different motional states can become notice-
able, where n′ and n are vibrational quantum
Ultracold single molecules have wide-ranging potential applications, such as ultracold chemistry, numbers in two spin states. SMC in the optical
precision measurements, quantum simulation, and quantum computation. However, given the difficulty
of achieving full control of a complex atom-molecule system, the coherent formation of single molecules trap leads to the appearance of sidebands in
remains a challenge. Here, we report an alternative route to coherently bind two atoms into a weakly
bound molecule at megahertz levels by coupling atomic spins to their two-body relative motion in a the microwave transition, similar to the case
strongly focused laser with inherent polarization gradients. The coherent nature is demonstrated by of a state-dependent optical lattice (35, 36). For
long-lived atom-molecule Rabi oscillations. We further manipulate the motional levels of the molecules two isolated colliding atoms in a tight OT, the
and measure the binding energy precisely. This work opens the door to full control of all degrees of spatial displacement di of a given atom with
freedom in atom-molecule systems. mass mi and position →r i (i = 1 or 2) is
straightforwardly transferred to the center-of-
M any fundamental research topics re- ecules out of atoms and the full control over
lated to ultracold molecules, such as atom-molecule systems remain a challenge. →

ultracold collisions and chemistry Here, we report an alternative approach mass (c.m.) and relative (rel.) coordinates, R ¼
(1–3), strongly correlated quantum that allows us to bypass the strong dephasing ðm1→r 1 þ m2→r 2Þ=ðm1 þ m2Þ and →r ¼ →r 1 À →r 2 ,
systems (4, 5), quantum degenerate encountered (without Feshbach resonances) respectively. This indicates that a microwave
gases (6, 7), precision measurements (8), and in the photoassociation of atoms. The idea is
quantum information processing (9, 10), are to couple the spin of one of the interacting spin-flip transition of this atom in the presence
rapidly developing and attracting broad atten- atoms to the two-body relative motion. In this
way, a noticeable displacement can be induced of SMC also enables a possibility of manipulat-
tion. After the work on producing coherent on the wave function of the relative motion
under the action of a microwave spin-flip ing the two-atom relative motion and even
coupling between atoms and molecules in a transition. The induced displacement could
85Rb Bose-Einstein condensate near a Feshbach enhance the atom-molecule wave function tuning the atom-molecule overlaps so as to
resonance (11), various types of ultracold di- overlap and thus the corresponding Franck-
atomic molecules in the gas phase (12–17) and Condon (FC) factor in the molecule formation. reliably bind the targeted atom to the colliding
in an optical lattice (18–20) have been formed Such spin-motion coupling (SMC) is mediated
via Feshbach resonances or photoassociation, by the inherent polarization gradients in a partner, as schematically depicted in Fig. 1B.
strongly focused trapping laser, namely in an Specifically, for the 85Rb87Rb bound states with
leading to the observation of ultracold chem- OT. We experimentally demonstrated the
istry (21) and to the simulation of quantum idea of engineering the quantized motion binding energy at the megahertz level, SMC
spin models (22). At the same time, because of of a two-atom system and coherently binding
recent developments in high-level individual two atoms, one 85Rb atom and one 87Rb atom, could increase the atom-molecule overlap by a
particle control and detection (23–27), bottom- into a single molecule. We obtained long-lived factor of 2 to 3 (37).
up assembly of single molecules in optical atom-molecule Rabi oscillations and further
manipulated the motional levels of the formed The starting point of the experiments was
tweezers (OTs) has attracted attention for molecules. We measured their binding energy the preparation of one 85Rb atom and one 87Rb
ultracold molecule applications (28–30). How- in free space as well as the differential light atom in internal hyperfine states j↑i85 ≡ j3; À3i85
ever, the coherent formation of single mol- shifts between scattering and molecular states. and j↓i87 ≡ j1; À1i87, respectively, in which in-
elastic collisions between these two atoms are
1State Key Laboratory of Magnetic Resonance and Atomic Optical tweezers are typically realized with
and Molecular Physics, Wuhan Institute of Physics and strongly focused Gaussian beams to provide energetically forbidden. The microwave radia-
Mathematics, APM, Chinese Academy of Sciences, Wuhan strong spatial confinements of laser-cooled tion can drive transitions between j↑i85 and
430071, China. 2Center for Cold Atom Physics, Chinese atoms. When the beam waists wOT are compa- j↓i85 ≡ j2; À2i85 for the 85Rb atom. Both atoms
Academy of Sciences, Wuhan 430071, China. 3School of rable to the wavelength l (here, wOT ≈ 0.75 mm occupy the ground state in a linearly polarized
Physical Sciences, University of Chinese Academy of and l = 0.852 mm), the tightly focused beams
Sciences, Beijing 100049, China. 4Institute for Theoretical exhibit longitudinal polarization components OT. The details for preparing such a two-atom
Physics, Vienna University of Technology, A-1040 Vienna, (31), which give rise to spatially varying ellip- reservoir and the atomic state–resolved detec-
Austria. 5Beijing Computational Science Research Center, tical polarizations, even for linearly polar- tion can be found in our previous work (38).
Beijing 100193, China. 6Department of Physics and Center for ized input fields. The resulting polarizations After simultaneous Raman sideband cool-
Atomic and Molecular Nanosciences, Tsinghua University, function as magnetic field gradients (32, 33), ing of 85Rb and 87Rb in two separate traps, the
Beijing 100084, China. 7Key Laboratory for Laser Plasmas and therefore atoms with different spin projec- two atoms in spin-stretched states of j↑i85 and
(Ministry of Education), and Department of Physics and tions mF of hyperfine state F have different j↓i87 were merged into a single OT via species-
Astronomy, Shanghai Jiao Tong University, Shanghai 200240, equilibrium positions. Because of this effect, a dependent transport, by which these two atoms
China. 8Collaborative Innovation Center of Quantum Matter, spin-flip microwave transition between dif-
Beijing 100084, China. 9LPTMS, CNRS, Univ. Paris-Sud, ferent mF states could induce a displacement were split for imaging.
Université Paris-Saclay, 91405 Orsay, France. 10Russian
Quantum Center, Skolkovo, Moscow 121025, Russia. 11Van der In the first set of experiments, we used SMC
Waals-Zeeman Institute, Institute of Physics, University of
Amsterdam, 1098 XH Amsterdam, Netherlands. to probe the quantized motion of two atoms
*These authors contributed equally to this work.
†Corresponding author. Email: [email protected] (X.H.); and deduce the displacements of the c.m. and
[email protected] (M.Z.)
rel. motions. To induce SMC, the axis of the

guiding magnetic field of 2.0 G was switched

on perpendicular to both the propagation di-
rection of the OT laser (z axis) and the linear
polarization vector (x axis). This arrangement
initiated displacement predominantly along
the x axis, which was measured to be 21.6(1) nm
(fig. S1) (37, 39). The trap depth is set at 1.6 mK,
corresponding to a radial trap frequency wx =
2p × 165 kHz and axial trap frequency wz =
2p × 27 kHz. We then used conventional Rabi
spectroscopy to measure the two-atom vibra-
tional transitions jysijϕ0i → jys′ijϕN x i (jyi
and jϕi are the rel. and c.m. motional states;
Nx = {0,1,…} denotes the quantum number of

He et al., Science 370, 331–335 (2020) 16 October 2020 1 of 4

RESEARCH | REPORT

AB (Fig. 2A). The peak f1, closely resembling the
single-atom transition f0, was identified as the
|ψm〉 spin-flip transition together with the motional
transition jϕ0i → jϕNx¼1i in the c.m. motion.
|↑〉85 600 |ψm〉 |↑〉85|↓〉87 The peak f2 corresponds to the transition
|0〉 jysi ¼ jjmrj ¼ 0i → jys′i ¼ jjmrj ¼ 1i in the
|φ1〉 rel. motion, where mr is the angular momen-
tum quantum number of a two-dimensional
-600 600
y (a0) harmonic oscillator describing the relative
MW |ψs0〉 1500 |ψs〉
motion. For the contact interaction in the
+ SMC
ultracold regime, the interaction potential was
0 85 85Rb 87Rb 85Rb87Rb
characterized by a pseudopotential that can
-1500 x (a0)
1500 only shift the quantized energy of states with
|ψs〉 mr = 0 and even k, where k is the quantum
|ψs0〉 number of the one-dimensional harmonic os-
cillator in the axial direction (z axis). The colli-
|1’〉 drel. |↓〉85|↓〉87 |φNx=0〉 sional shift Ddð→r Þ for all mr = 0 states has been
|↓〉85 |0’〉 Relative dc.m. found analytically (40). In contrast, the wave
functions of the states with mr ≠ 0 vanished
d Center of mass at r = 0, and these states did not feel the s-
a wave interaction. Thus, the splitting between
85Rb-87Rb f1 and f2 was equal to the interaction shift of
85Rb the initial two-atom state.

Fig. 1. Schemes of SMC and molecular association in a tight OT. (A) SMC of a single 85Rb atom. The Next, we observed the coherent driving of

trapping potentials in the spin states of j↑i85 and j↓i85 are displaced in space by da, leading to the coupling the two-atom quantized motion, from which
between j0i and j1′i with a strength of W0h85 (W0 is the bare spin-flip coupling, and h85 is an additional the displacements dc.m. and drel. of the c.m. and
factor resulting from the wave function overlap). (B) Making one molecule via SMC. The wave functions are shown rel. motions, respectively, were extracted. The

with two-dimensional density plots and the corresponding cross sections in the y = 0 plane, and the potential profiles coherent Rabi oscillations for these motions

include the trapping potential and the molecular potential. In an external trapping field without SMC (dashed were obtained with the corresponding Rabi

red curve), the background overlap of the atomic rel. wave function jys0i (red wave packet) and the molecular wave frequencies W1 and W2, respectively. The re-
function jymi is relatively low. While in the presence of SMC, a displacement (drel.) between the atomic rel. sulting oscillations are shown in Fig. 2B. The
wave function jysi (purple wave packet) and jymi along the x direction shows up, as well as a displacement dc.m.
in the c.m. motion. The drel. could enhance the FC factor for binding 85Rb and 87Rb into one molecule by lower contrast of the rel. motion resulted from
implementing the microwave (MW) spin-flip transition j↓i85j↓i87→j↑i85j↓i87. The dc.m. allows us to make the
resulting molecule in the motional ground state |ϕ0i (Nx = 0) or in the excited state jϕ1i (Nx = 1). a finite ground-state fidelity. Given the mea-
sured W1 and W2, the displacements dc.m. and
A B 0.8 drel. were estimated to be about 9 and 25 nm,
1.0 respectively (37).

0.8Survival probability 0.6 Having demonstrated the manipulation of
Survival probability
f0 quantized motion of two atoms with micro-

0.6 wave radiation, we now describe the forma-

0.4 f1 f2 85Rb SB transition 0.4 tion of a single least-bound molecule in the
0.2 85Rb survival prob.
87Rb survival prob. 0.2 c.m. motion presence of SMC. We first considered the least-
-0.21 0.0 rel. motion bound molecule for the channel j↑i85j↓i87 ,
where the background scattering length was
-0.18 -0.15 -0.12 -0.09 0.1 0.2 0.3 0.4
MW detuning (MHz) MW pulse (ms) the largest among all collisionally stable chan-
nels (41). From coupled-channel calculations,
Fig. 2. Observation of sideband transitions of two atoms via SMC. (A) Microwave spectra of sideband the scattering length for this channel was
(SB) transitions of 85Rb in the presence (filled squares and circles) and absence (filled triangles) of 87Rb. 314.4 a0 (where a0 is the Bohr radius) at a mag-
The detuning is given with respect to the carrier transition frequency of a single 85Rb atom. The corresponding netic field of 2.0 G, and the calculated binding
energy was −0.901 MHz (37).
spin-flip coupling W0 = 2p × 17.4(1) kHz. All of the data points are averaged over 100 runs. The solid curves
are Gaussian fits of the data, yielding resonant detunings of f1 = −167.9(5) kHz and f2 = −144.5(8) kHz To drive the molecule-formation transition
for the two-atom case, and f0 = −167.7(5) kHz for the one-atom case. (B) The Rabi oscillations at detunings jysi → jymi, the corresponding spin-flip cou-
f1 (filled squares) and f2 (filled circles). The data points are averaged over 100 runs, and the error bars pling was increased to W0 = 2p × 64.5(4) kHz
denote the standard deviation of the mean. The solid curves are damped sinusoidal fits of the oscillations, by using high-power microwave radiation.

yielding the Rabi frequencies W1 = 2p × 6.1(1) kHz and W2 = 2p × 6.5(2) kHz. Then we applied a sequence of three pulses:

the c.m. motion in the x direction) together with after the species-dependent transport, leading the first and third were resonant p pulses driv-
the spin-flip transition j↑i85j↓i87 → j↓i85j↓i87. to the disappearance of the atomic fluorescent ing only the transitions j↑i85j↓i87 → j↓i85j↓i87
When the 85Rb spin was flipped, the resulting signals. Compared with the Dnx = 1 sideband and j↓i85j↓i87 → j↑i85j↓i87 while leaving atoms
transition of a single 85Rb atom, the two-atom in the scattering ground statejysi unchanged.
two atoms, j↓i85 and j↓i87 , had vector light spectrum exhibited a double-peak structure The frequency fm of the middle pulse was
shifts of the same sign and moved together scanned in order to search for the molecular
transition. When fm was resonant with the
molecular transition, the sequence resulted
in the formation of a 85Rb87Rb molecule.

Otherwise, the sequence would implement

He et al., Science 370, 331–335 (2020) 16 October 2020 2 of 4

RESEARCH | REPORT

A 1.0 B 85Rb survival prob. resulting single molecules were in the ground
0.8 0.8 87Rb survival prob.
Survival probability state of OT, and for the latter, the resulting
Survival probability 0.7
single molecules were in the first excited motional
0.6 0.6 statesjϕNx¼1i). The difference in strength of these
resonances was due to an additional FC factor
0.4 Molecular sideband δfm1 0.5
Molecular carrier 0.4 of 0.35 for the quantized motion of the molecule.
Only 85Rb present δfm0
0.0 Having confirmed the formation of a single
-1.25 -1.20 -1.10 -1.05 -0.90 -0.85 0.2 0.4 0.6 0.8
MW detuning (MHz) Pulse duration (ms) molecule, we subsequently observed coher-

Fig. 3. Atom-molecule transitions and coherent Rabi oscillations. (A) Microwave spectra of the ent atom-molecule Rabi oscillations. The time-
jysij↓i85j↓i87 → jymij↑i85j↓i87 transition at the trap frequency of wz = 2p × 27 kHz. The survival probability
of 85Rb atoms is measured as a function of detuning from the carrier transition of a single 85Rb atom resolved Rabi oscillations allowed full control

and the background signals are shown with filled gray circles. The pulse duration is 0.11 (0.22) ms for the of the final superposition state. Figure 3B shows

carrier (sideband) transition. All the data points are averaged over 100 experimental runs, and the error the recorded Rabi oscillation for the transition

bars denote the corresponding standard deviations. The solid curves are Gaussian fits of the peaks, starting from the two-atom ground state, at
the resonant detuning of dfm0. Slowly damped
yielding dfm0 = −1.048(1) MHz and dfm1 = −0.8825(5) MHz. (B) The atom-molecule Rabi oscillation for the oscillating survival probabilities of atoms were
molecular carrier transition at the trap depth of wz = 2p × 18.5 kHz. The detected 85Rb signals oscillate obtained with technical noise–limited decay
similarly to 87Rb ones, with decay times of 1.1 ± 0.6 ms and 1.7 ± 1.5 ms, respectively. The average Rabi time of about 1 ms. We subsequently mea-

frequency is Wm = 2p × 3.95(5) kHz. The data points are averaged over 150 experimental runs. sured the lifetime of the resulting molecules,
which was about 20(7) ms (37). To illustrate
A 12 |↑〉85|↓〉87 with SMC B the range of adjustable coupling strengths,
9 |↑〉85|↓〉87 background 1.10
6 |↓〉85|↓〉87 with SMC 1.05 we measured the Rabi frequencies versus the
|↓〉85|↓〉87 background 1.00
Cs,m (1 × 10−2) 0.95 Molecular transition detunings trapping frequencies wz and deduced the asso-
Transition detuning (MHz) 0.90 After correction ciated FC factors for the atopm-ffiffimffiffiffiffioffiffilecule tran-
sition Cs,m. The values of Cs;m ¼ Wm=W0 ,
3 which amounts to the wave function overlap

0 10 15 20 25 30 0.85 5 10 15 20 25 30 between the scattering and the molecular
5 ωz (2π × kHz) 0 ωz (2π × kHz)
state, are shown in Fig. 4A. The wave function
Fig. 4. Atom-molecule wave function povffieffiffirffiffilffiaffiffip and molecular binding energy. (A) The dependence of the
atom-molecule wave function overlap ð Cs;mÞ on the axial trapping frequency wz. The purple (red) filled overlap and thus the FC factor increased as

circles are of measured overlap integrals for making j↑i85j↓i87 ðj↓i85j↓i87Þ molecules in the presence of SMC, the confinement got stronger. To measure the
and the corresponding fitting with the pseudopotential model is shown with the solid purple (red) line. In the
background atom-molecule overlap, the mag-
absence of SMC, the measured background overlap of j↑i85j↓i87 ðj↓i85j↓i87Þ molecules is denoted via the netic field was set along the x axis. The back-
purple (red) filled squares and the corresponding calculation with coupled-channel method is shown with ground overlap was measured as the Rabi

the purple (red) dashed line. The error bars denote standard deviations. (B) Measurement of the binding frequency ratio of the atom-molecule transi-

energy Eb of the j↑i85j↓i87 bound state. The transition detuning dfm0 (filled squares) is plotted as a function tion and single-atom transition. As compared
of the axial trapping frequency wz, and the error bars include the fitting errors (standard deviations) and
fluctuations (±3 kHz) induced by the magnetic field. After taking into account the zero-point energy En=0, the with the measured and calculated background
interaction shift Ddð→r Þ, and the energy shift due to confinement DEb, the resulting dependence of transition
resonances (filled circles) on wz is obtained. The solid line is a linear fit, yielding −Eb = 0.890(3) MHz. overlap, SMC could enhance the overlap by a

an effective 2p rotation for 85Rb. The species- spacing of 165(1) kHz. This spacing, which factor of more than two (Fig. 4A). The cor-
dependent transport was unable to dissociate
amounted to wx/2p of the constituent atoms responding fitting with the pseudopotential
such a molecule that has binding energy on in the OT, originated from the resolved quantized
model leads to a shift of the wave function of
the order of megahertz. Thus, the molecule motion of the formed molecules. Under the about 9 nm (37).

formation would manifest itself as the dis- same spin-flip transition,j↓i85j↓i87 → j↑i85j↓i87, To deduce the binding energy in free space,
appearance of the atomic 87Rb and 85Rb fluo- the recorded resonant vibrational transi- Eb, we measured the molecular transition
detuning dfm0 as functions of wz (Fig. 4B).
rescence signals. As shown in Fig. 3A, we tions at detunings of dfm0 and dfm1 corre- The molecule formation was accompanied by
observed two molecular resonances dfm0 and sponded tojysijϕ0i → jymijϕ0i andjysijϕ0i →
dfm1 at detunings of about −1 MHz with a jymijϕNx¼1i, respectively (for the former, the a change in the energy of the relative motion,
Erel. For the initial state, the associated Erel
consisted of the zero-point energy En=0 and
interaction shift Ddð→r Þ. As for the molecular
state, it included Eb and the energy shift due to
the confinement, DEb. The calculated DEb and
Ddð→r Þ as functions of wz can be found in fig. S2.
After subtracting the contributions of trap-
dependent energy values fEn¼0; Ddð→r Þ; DEbg,
we obtained the resulting dependence of the

binding energy on wz. With this correction,
Eb was determined to be −0.890(3) MHz by
linearly extrapolating wz→0 (Fig. 4B). The
evidently linear dependence has a strong con-

nection with the separated external trapping

potentials seen by the constituent atoms, which
lifts the energy of the bound state (42). From
the binding energy Eb measured with a kilo-
hertz uncertainty, the scattering length can be
calculated to be 316.9(4) a0 with an analytical

He et al., Science 370, 331–335 (2020) 16 October 2020 3 of 4

RESEARCH | REPORT

formula in the framework of multichannel control the molecule-photon interactions in 37. See supplementary materials for more details.
quantum defect theory (43), which is slightly 38. K. P. Wang et al., Phys. Rev. A 100, 063429 (2019).
larger than the value 314.4 a0 from coupled- these systems. 39. K. P. Wang et al., Chin. Phys. Lett. 37, 044209 (2020).
channel calculations (37). The slight difference 40. Z. Idziaszek, T. Calarco, Phys. Rev. A 74, 022712 (2006).
may arise from a finite precision of the ana- REFERENCES AND NOTES 41. K. P. Wang et al., arxiv:1907.12722 [quant-ph] (30 July 2019).
lytical formula (37, 41). 42. M. Krych, Z. Idziaszek, Phys. Rev. A 80, 022710 (2009).
1. H. Yang et al., Science 363, 261–264 (2019). 43. B. Gao, J. Phys. At. Mol. Opt. Phys. 37, 4273–4279
Aside from the production of the j↑i85j↓i87 2. R. Chapurin et al., Phys. Rev. Lett. 123, 233402 (2019).
least-bound molecule, we also created the 3. M.-G. Hu et al., Science 366, 1111–1115 (2019). (2004).
j↓i85j↓i87 molecule. We then observed the cor- 4. M. A. Baranov, M. Dalmonte, G. Pupillo, P. Zoller, Chem. Rev. 44. L. Caldwell, M. R. Tarbutt, Phys. Rev. Res. 2, 013251 (2020).
responding coherent Rabi oscillations and mea- 45. J. Wolf et al., Science 358, 921–924 (2017).
sured the binding energy in free space, Eb (see 112, 5012–5061 (2012). 46. J. Wolf, M. Deiß, J. Hecker Denschlag, Phys. Rev. Lett. 123,
fig. S5). With the determined value of Eb = 5. J. A. Blackmore et al., Quantum Sci. Technol. 4, 014010 (2018).
−2.224(6) MHz, we deduced the correspond- 6. L. De Marco et al., Science 363, 853–856 (2019). 253401 (2019).
ing scattering length to be 230.5(2) a0, which 7. W. G. Tobias et al., Phys. Rev. Lett. 124, 033401 (2020). 47. P. S. Żuchowski, J. Aldegunde, J. M. Hutson, Phys. Rev. Lett.
is also slightly larger than the value 227.9 a0 8. S. S. Kondov et al., Nat. Phys. 15, 1118–1122 (2019).
from coupled-channel calculations (37). Com- 9. J. W. Park, Z. Z. Yan, H. Loh, S. A. Will, M. W. Zwierlein, 105, 153201 (2010).
pared with the case of the j↑i85j↓i87 bound 48. D. A. Brue, J. M. Hutson, Phys. Rev. A 87, 052709 (2013).
state, the atom-molecule wave function over- Science 357, 372–375 (2017). 49. D. E. Chang, J. S. Douglas, A. González-Tudela, C.-L. Hung,
10. K.-K. Ni, T. Rosenband, D. D. Grimes, Chem. Sci. 9, 6830–6838
lap was smaller and amounted to 0.064(5) at H. J. Kimble, Rev. Mod. Phys. 90, 031002 (2018).
(2018). 50. X. He et al., Data for figures: Coherently forminga single
the trap frequency of wz = 2p × 27 kHz (Fig. 4A), 11. E. A. Donley, N. R. Claussen, S. T. Thompson, C. E. Wieman,
while, without SMC, the background overlap molecule in an optical trap, Version 1, Zenodo (2020);
was only 0.032 (37). The fitting with the pseu- Nature 417, 529–533 (2002). https://doi.org/10.5281/zenodo.3941379.
dopotential model leads to a wave function 12. K.-K. Ni et al., Science 322, 231–235 (2008).
shift of about 23 nm (37). 13. J. G. Danzl et al., Science 321, 1062–1066 (2008). ACKNOWLEDGMENTS
14. T. Takekoshi et al., Phys. Rev. Lett. 113, 205301 (2014).
The least-bound molecule can be further 15. P. K. Molony et al., Phys. Rev. Lett. 113, 255301 (2014). We thank P. Zhang for insightful discussions. Funding: This work
16. J. W. Park, S. A. Will, M. W. Zwierlein, Phys. Rev. Lett. 114, was supported by the National Key R&D Program of China
transferred to the rovibrational ground state (grants 2017YFA0304501, 2016YFA0302800, 2016YFA0302002,
205302 (2015). and 2016YFA0302104), the Key Research Program of Frontier
via the two-photon stimulated Raman adiabatic 17. M. Guo et al., Phys. Rev. Lett. 116, 205303 (2016). Science of the Chinese Academy of Sciences (CAS) (grant ZDBS-
18. T. Rom et al., Phys. Rev. Lett. 93, 073002 (2004). LY-SLH012), the National Natural Science Foundation of China
passage. Our work opens the door to the studies 19. F. Lang, K. Winkler, C. Strauss, R. Grimm, J. H. Denschlag, (grants 11774389, 11774023, and U1930402), the Strategic Priority
of motional state controlled (44) state-to-state Research Program of CAS (grant XDB 21010100), and the Youth
cold chemistry (45, 46) and few-body molecu- Phys. Rev. Lett. 101, 133005 (2008). Innovation Promotion Association of CAS (grants 2019325 and
lar dynamics. The SMC method demonstrated 20. A. Chotia et al., Phys. Rev. Lett. 108, 080405 (2012). 2017378). The work of G.V.S. was supported by the Russian
21. S. Ospelkaus et al., Science 327, 853–857 (2010). Science Foundation (grant 20-42-05002). Author contributions:
here is especially important for atom-molecule 22. B. Yan et al., Nature 501, 521–525 (2013). X.H. and M.Z. conceived of the experiments. K.W., X.H., J.Z.,
23. P. Xu et al., Nat. Commun. 6, 7803 (2015). R.G., and C.S. established the setup and performed the experiment.
transfer between states of near-equal and large 24. D. Barredo, S. de Léséleuc, V. Lienhard, T. Lahaye, A. Browaeys, X.G., X.H., K.W., and J.L. performed theoretical calculations. K.W.,
X.H., P.X., X.G., M.L., J.W., G.V.S., and M.Z. analyzed the data and
scattering lengths, where direct microwave tran- Science 354, 1021–1023 (2016). discussed the results. X.H., K.W., G.V.S., and M.Z. wrote the
25. M. Endres et al., Science 354, 1024–1027 (2016). manuscript. All authors reviewed the manuscript. X.H. and M.Z.
sitions are weak, thus SMC enhancement plays 26. T.-Y. Wu, A. Kumar, F. Giraldo, D. S. Weiss, Nat. Phys. 15, supervised the project. Competing interests: None declared.
Data and materials availability: All data needed to evaluate the
a key role. This method may find applications 538–542 (2019). conclusions in the paper are present in the paper or the
in interesting systems such as alkali atom– 27. J. P. Covey, I. S. Madjarov, A. Cooper, M. Endres, Phys. Rev. supplementary materials and are deposited in (50).
alkaline-earth atom pairs (47, 48). Further-
more, SMC can be realized with optical near Lett. 122, 173201 (2019). SUPPLEMENTARY MATERIALS
28. L. R. Liu et al., Science 360, 900–903 (2018).
fields of nanoscale waveguides and nanopho- 29. L. R. Liu et al., Phys. Rev. X 9, 021039 (2019). science.sciencemag.org/content/370/6514/331/suppl/DC1
tonic structures (49), and our work may be 30. L. Anderegg et al., Science 365, 1156–1158 (2019). Materials and Methods
directly applied to form molecules and to 31. B. Richards, E. Wolf, Proc. R. Soc. London Ser. A 253, 358–379 Figs. S1 to S6
Tables S1 and S2
(1959). References (51–58)
32. A. M. Kaufman, B. J. Lester, C. A. Regal, Phys. Rev. X 2,
1 January 2020; accepted 27 August 2020
041014 (2012). Published online 24 September 2020
33. J. D. Thompson, T. G. Tiecke, A. S. Zibrov, V. Vuletić, 10.1126/science.aba7468

M. D. Lukin, Phys. Rev. Lett. 110, 133001 (2013).
34. A. Dareau, Y. Meng, P. Schneeweiss, A. Rauschenbeutel,

Phys. Rev. Lett. 121, 253603 (2018).
35. L. Förster et al., Phys. Rev. Lett. 103, 233001 (2009).
36. X. Li, T. A. Corcovilos, Y. Wang, D. S. Weiss, Phys. Rev. Lett.

108, 103001 (2012).

He et al., Science 370, 331–335 (2020) 16 October 2020 4 of 4

RESEARCH

HYDROGELS The sliding friction Fs between the hydrogel
and a polished stainless steel surface, as well
Cartilage-inspired, lipid-based as surfaces of other materials (supplemen-
boundary-lubricated hydrogels tary text section 7), was examined over a range
of loads Fn, corresponding to different mean
Weifeng Lin1, Monika Kluzek1, Noa Iuster1, Eyal Shimoni2, Nir Kampf1, Ronit Goldberg1*†, Jacob Klein1* contact stresses P and sliding velocities vs
(materials and methods section 7 and supple-
The lubrication of hydrogels arises from fluid or solvated surface phases. By contrast, the lubricity mentary text section 12), yielding the coeffi-
of articular cartilage, a complex biohydrogel, has been at least partially attributed to nonfluid, cient of sliding friction m = Fs/Fn. Figure 3A
lipid-exposing boundary layers. We emulated this behavior in synthetic hydrogels by incorporating shows the tribometer configuration, with rep-
trace lipid concentrations to create a molecularly thin, lipid-based boundary layer that renews resentative directly recorded traces from
continuously. We observed a 80% to 99.3% reduction in friction and wear relative to the lipid-free gel, which Fs, and thus m, is determined, whereas
over a wide range of conditions. This effect persists when the gels are dried and then rehydrated. panels B and C in Fig. 3 reveal and quantify
Our approach may provide a method for sustained, extreme lubrication of hydrogels in applications the transfer of lipids between gel and steel
from tissue engineering to clinical diagnostics. surface during contact and sliding.

S ynthetic hydrogels are widely used in reduction in friction and wear is attributed Figure 3, D and E, shows the variation of m
biomedical and other applications (1–3), to a lipid-based boundary layer at the hydrogel with load for the two lipids used, at room
and their lubricity is crucial for their ef- surface, which is continually reconstructed (25°C) and at physiological temperature (37°C).
ficient function whenever surfaces slide as it wears, through progressive release of A reduction in friction is seen for the lipid-
past each other. The lubrication is attrib- lipids, as indicated schematically in Fig. 1. incorporating gels relative to the lipid-free
uted to fluid interfacial layers intrinsic to the We used this approach to create several differ- gels. For the former, m can range from ~0.02
gels, such as exuded liquid films or solvated ent self-lubricating hydrogels of both biological at lower loads to 0.005 at higher loads; for
flexible polymers at their surfaces (4–7). By and synthetic polymers (supplementary text the latter, at low loads and contact stresses
contrast, biological materials such as articu- section 1); here we focus on the widely ex- 0:5 ≲ m ≲ 1 was measured [similar to earlier
lar cartilage remain well lubricated over a life- ploited (1) poly(hydroxyethylmethacrylate) measurements of high friction for steel sliding
time of sliding and wear. The low friction of (pHEMA) hydrogel. against pHEMA hydrogels (24)], whereas for
cartilage has been attributed to fluid pressur- P higher than ~0.5 MPa the metal surface
ization supporting much of the load [an effect Hydrogels were prepared either without lip- would deform or tear the gel, without sliding
that does not apply in synthetic hydrogels (8)], ids or with low concentrations of the PC lip- (Fig. 3A; see materials and methods section 7).
whereas its boundary lubrication has been ids dimyristoylphosphatidylcholine (DMPC) The reduction in friction (Fig. 3, D and E)
attributed to nonfluid boundary layers at its or hydrogenated soy phosphatidylcholine arising from the lipid incorporation thus ranges
surface (9–13). These layers expose phosphati- (HSPC) added in the form of multilamellar from 95% to 99.3% at the higher loads and
dylcholine (PC) lipids whose highly hydrated vesicles (MLVs) (though smaller unilamel- contact pressures. Although both lipids reduce
phosphocholine headgroups may reduce fric- lar vesicles may also be present; materials the friction relative to that of the lipid-free gel,
tion via the hydration lubrication mechanism and methods section 2). These lipids were at room temperature DMPC is slightly more
(14, 15). The maintenance of such boundary chosen because they are respectively above lubricious at low loads, whereas at 37°C the
layers after frictional wear occurs through cel- and below their gel-to-liquid transition—
lular replenishment and self-assembly of their known to affect lubrication (23)—at room A Counter - Surface Newly-exposed gel
components (14, 16, 17), including hyaluronan, and physiological temperatures. Their dis- surface
lubricin, and especially PC lipids (18–22), which tribution is revealed by freeze-fracture cryo– Frictional
reduce friction at the slip plane via their hy- scanning electron microscopy (cryo-SEM) and
drated phosphocholine headgroups (12). These confocal fluorescence microscopy (materials wear
components are ubiquitous in both cartilage and methods section 6), as depicted in Fig. 2
and the surrounding synovial fluid (16, 20, 21) (see also supplementary text section 4). The B C
and thus are readily available to maintain the lipid-free hydrogel (Fig. 2A) displays a fea-
lubricating layer at the articular surface. tureless internal surface, as expected at this Counter - Surface
resolution. Figure 2, B to F, shows the incor-
We adapted this mechanism to lubricate porated DMPC and HSPC MLVs within the Fig. 1. Schematic illustrating the self-lubrication
synthetic hydrogels via the incorporation of hydrogel bulk. These are sequestered in clus- of lipid-incorporating hydrogels. As the surface
small amounts of PC lipids to form micro- ters, either as spherical microreservoirs filled of the hydrogel, incorporating lipids as vesicles in
reservoirs throughout the gel bulk, by mixing with roughly spherical vesicles (DMPC; Fig. 2, microreservoirs (A), wears away because of friction,
a low concentration of PC lipids with the de- B to D), whose size is consistent with dynamic additional microreservoirs of lipid are exposed. This
sired monomer solution, then polymerizing light scattering measurements on the lipo- enables boundary layers of lipids to form on the
and cross-linking to form the hydrogel. The some dispersions (materials and methods sec- surfaces (B and C), leading to friction reduction via
tion 2), or in less regular clusters (HSPC; Fig. 2, the hydration lubrication mechanism at the slip
1Department of Materials and Interfaces, Weizmann Institute E and F). Rheometrically determined mechan- plane between the highly hydrated lipid headgroups.
of Science, Rehovot 76100, Israel. 2Department of Chemical ical properties (materials and methods sec-
Research Support, Weizmann Institute of Science, Rehovot tion 4) reveal that the hydrogel storage modulus
76100, Israel. G′ (>>G′′, the loss modulus) varies, over a fre-
*Corresponding author. Email: [email protected] quency ( f ) range of 0.1 to 10 Hz, by ~30% or less
(R.G.); [email protected] (J.K.) between lipid-free and PC-incorporating hydro-
†Present address: Liposphere Ltd., Pinhas Sapir Street 3, Nes gels (Fig. 2G).
Ziona 7403626, Israel.

Lin et al., Science 370, 335–338 (2020) 16 October 2020 1 of 4

RESEARCH | REPORT B CD

A

5 µm 5 µm 500 nm 20 µm

EF G

5 µm 1 µm

Fig. 2. Characterization of lipid-free and lipid-incorporating hydrogels. (A) Freeze microscopy section of the hydrogel incorporating fluorescently labeled DMPC vesicles,
fracture surface of lipid-free pHEMA hydrogel. (B) Freeze-fracture surface of the showing the lipid microreservoir distribution. (E) Freeze-fracture surface of the gel
gel incorporating DMPC vesicles, showing the microreservoirs transected by the incorporating HSPC vesicles. (F) Microreservoir from (E) at larger magnification.
surface. (C) A single microreservoir from (B) at larger magnification. (D) Confocal (G) Storage and loss moduli of lipid-free and lipid-incorporating pHEMA gels.

HSPC lipids reduce friction more effectively. When the lipid-incorporating gels are fully gel (Fig. 4D, left, for which m ≈ 0.01 and ss =
We attribute this to the interplay between head- dried and then rehydrated (materials and mP ≈ 104 N/m2 is much less than G′) is barely
group hydration and bilayer robustness for the methods section 9), the friction returns to its affected after a full hour of sliding under this
two lipids, arising from their different phase low value, and the lubrication is once again load. Notably, after 1 hour of sliding at this
states (23). Figure 3F shows the near-constant self-sustaining (Fig. 4B). This robustness to high load and contact stress (1.53 MPa), the
value of m with sliding velocity over some three drying and rehydration has particular implica- surface of the liposome-containing gel has
orders of magnitude in vs, a clear signature of tions for hydrogel coating and storage of hydro- worn by 9 ± 3 mm (materials and methods
boundary as opposed to fluid-film lubrication. gels. Finally, friction as well as wear and surface section 7 and supplementary text section 11),
damage were reduced by the incorporated yet despite this removal-by-wear of the origi-
Incorporating lipids within the bulk hydro- lipids. Figure 4C compares wear of the lipid- nal gel surface, the frictional force remains
gel resulted in much lower friction than when free hydrogel with that of a DMPC-MLV– unchanged, with m ≈ 0.01 throughout. Because
the gels were exposed to lipids externally, as incorporating hydrogel after 2 hours of sliding. this extent of surface wear far exceeds the
shown in Fig. 4A. For lipid-incorporating hy- In these conditions, the wear of the lipid-free thickness of any boundary-lubricating layer or
drogels sliding under water, m ≈ 0.01, as com- pHEMA gel surface was 57 ± 3 mm for a 1-N load, size of the microreservoirs (Fig. 2), it is clear
pared with m ≈ 0.06 to 0.08 for lipid-free gels whereas the wear of the lipid-incorporating that such layers are continuously renewing
sliding immersed in a PC-MLV dispersion, and gel (where m ≈ 0.01 throughout the 2 hours of as friction abrades the surface, as indicated
m ≈ 0.08 to 0.15 for lipid-free hydrogels in- sliding) was below the detection limit (±3 mm) in Fig. 1. Hydrogels conform affinely when
cubated overnight in PC-MLV dispersions, of the tribometer, even for a load 10 times as compressed by a countersurface already at very
then measured in water. In the latter case, the large. This implies that such gels could resist low contact stresses (higher than a fraction
friction rises sharply with sliding time (sup- substantial wear over many sliding cycles of an atmosphere; see supplementary text
plementary text section 2). Thus, lubrication (supplementary text section 13). The effect of section 14). Thus the lubricating lipids are ex-
by such external application of lipids is far less low surface wear and low damage manifests at pected to spread over the entire hydrogel con-
effective than when the lipids are incorporated a higher load, as shown in Fig. 4D. The lipid- tact area, as long as the sliding amplitude
in the bulk hydrogel. For the case of sliding in free gel is damaged and torn after just a few exceeds the mean inter-microreservoir spacing
a liposome dispersion, this observation is at- seconds of back-and-forth motion of the steel of a few micrometers.
tributed to the lipids having poor access to the countersurface. This phenomenon is attributed
intersurface region, which arises from the very to the high friction strongly shearing the gel Sliding takes place through hydration-
large distortion energy required for liposomes and arises because the shear stress ss at the gel lubricated slip between the exposed hydrated
to enter the intersurface gap (supplementary surface—given by ss = mP ≈ 106 N/m2, where headgroups of the lipid bilayers or liposomes
text section 8). For the case of sliding after the friction coefficient m ≈ 0.5 to 1 and the (supplementary text section 6). These are ex-
overnight lipid adsorption, the lubrication de- mean pressure P ≈ 1.5 MPa—greatly exceeds tracted by interfacial shear, due to the sliding,
teriorates rapidly through wear once the ex- the gel shear modulus G′ ≈ 6 × 104 N/m2. At from the surface-exposed microreservoirs (Figs.
ternal PC source is removed (supplementary the same time, the lipid-incorporating hydro- 1 and 2) and are thereby spread to coat the op-
text section 2). posing (gel and metal) surfaces (supplementary

Lin et al., Science 370, 335–338 (2020) 16 October 2020 2 of 4

RESEARCH | REPORT

A BFn attached bilayers, as indicated in Fig. 1. Similar
hydration lubrication by PC boundary layers
R 2a on model substrates (14, 28) yields even lower
a friction (m ≈ 0.001 or less) than observed with
our hydrogels (for which m ≈ 0.005 to 0.02).
i) Hydrogel 2a This finding reflects the softer and rougher na-
2a ture of our substrates, with consequent addi-
tional pathways for frictional dissipation, such
Petri dish iii) as viscoelastic losses, relative to these earlier
studies in which rigid and extremely smooth
Scanning substrates were used (23, 28, 29). Evidence for
laser beam this phenomenon is provided by the varia-
tion of m for pHEMA gels of different moduli,
ii) induced by varying the cross-linker density
within the lipid-incorporating gels, where softer
CD gels have higher sliding friction (supplemen-
tary text section 9).
EF
The self-renewal of the lubricating bound-
Fig. 3. Sliding between a steel sphere and pHEMA hydrogels. (A) UMT (Universal Mechanical Tester) ary layer, as the hydrogel abrades under fric-
tribometer and sliding configuration (materials and methods section 7) and typical friction-versus-time tion, is attributed to its continuous healing
traces for lipid-free and lipid-incorporating gels, where Fs is taken as half of the amplitude between through availability of lipids at the surface,
sliding plateaus. The middle trace (no sliding plateau) indicates that sliding is not achieved and as microreservoirs of the PC vesicles within
that Fs must be larger than half of the amplitude between peaks. (B) Transfer of fluorescently labeled the gel become progressively exposed and
lipids incorporated in the gel to the contact area (radius a) with the sliding steel sphere, as in (i), is sheared by the countersurface (Fig. 1). A more
monitored as follows: After sliding, the steel sphere is placed in a Petri dish and the dye in the detailed consideration (materials and meth-
transferred lipid layer is excited with a scanning laser beam, as in (ii), and imaged with a photomultiplier ods section 10) shows that the total volume of
tube to yield images, as in (iii), that show the lipid transfer after 5 min of sliding. The amount of lipids available at the gel surface from res-
transferred lipids is then quantified as in (C) (see also materials and methods section 11). a.u., arbitrary ervoirs transected by it would be sufficient to
units. (D) and (E) show friction coefficient values m at room and physiological temperatures, respectively, form a lipid film of thickness d given by
at a series of loads and corresponding contact stresses. “Wavy-topped” columns indicate that no
sliding was achieved [see middle trace in (A)], so m must be larger than the column height shown. d ≈ 2fR0
Error bars indicate SD from at least six independent measurements. (F) Variation of m with sliding
velocity vs for lipid-free (blue) and lipid-incorporating (HSPC, orange; DMPC, green) pHEMA gels. Error where f is the mean volume fraction of lipids
bars indicate SD from at least three measurements. incorporated in the bulk hydrogel, and R0 is
the radius of a microreservoir. R0 could be
text section 5). The zwitterionic phosphocho- fer of fluorescently labeled gel-incorporated varied by changing the concentration of lipids
line headgroups adhere to both the nega- lipids to the sliding steel sphere counter- incorporated within the gel (supplementary
tively charged (25) pHEMA surface, which is surface. The area covered by the transferred text section 10). When R0 ≈ 1.5 mm (Fig. 2, B
rich in dipolar hydroxyl groups (26) (such lipids can be imaged (Fig. 3B) and its thick- and C) and f ¼ 0:012 (materials and methods
dipole-charge interactions may also help ness evaluated from their total calibrated fluo- section 2), as in our experiments, d ≈ 36 nm,
localize the vesicles within the microreser- rescence intensity (Fig. 3C and materials and which is equivalent to the thickness of some
voirs), and the negatively charged stainless methods section 11). This shows that the lipid seven to eight bilayers of DMPC. This is more
steel countersurface (27), as seen by atomic layer on the metal surface is 1.5 ± 0.5 bilayers than sufficient for hydration lubrication between
force microscopy imaging (supplementary text thick, consistent with hydration lubrication at a bilayer or even a compressed vesicle layer
sections 3 and 4). More directly, we see trans- the slip plane between metal-attached and gel- attached on each of the opposing surfaces (23, 29),
indicating that the mechanism shown in Fig.
1 can amply account for the self-sustaining
lubricating boundary layers as the gel surface
wears away. It is also consistent with the
thickness of the lipid layer transferred to the
metal surface, as indicated above (Fig. 3, B and C).

We show that trace incorporation of PC lip-
ids provides a simple route to create hydro-
gels that can continuously lubricate themselves
as they wear, via the hydration lubrication
mechanism attributed to PC-exposing bound-
ary layers on articular cartilage. Such gels main-
tain very low friction and wear—up to contact
stresses of several megapascals and sliding
velocities up to the centimeter-per-second
scale—while minimally perturbing their bulk
mechanical properties. Our approach may pro-
vide a platform for creating self-lubricating
hydrogels wherever low friction and low wear
are required.

Lin et al., Science 370, 335–338 (2020) 16 October 2020 3 of 4

RESEARCH | REPORT

AB 15. W. H. Briscoe et al., Nature 444, 191–194 (2006).
16. J. A. Buckwalter, H. J. Mankin, A. J. Grodzinsky, Instr. Course
C D 50 N load
Lect. 54, 465–480 (2005).
After 60 mins, MLV-DMPC After 12 seconds, Lipid-free 17. J. Klein, Science 323, 47–48 (2009).
18. D. P. Chang et al., Soft Matter 5, 3438–3445 (2009).
Fig. 4. Friction and wear of lipid-treated pHEMA gels. (A) Sliding friction between lipid-free gels (blue), between 19. A. Singh et al., Nat. Mater. 13, 988–995 (2014).
lipid-free gels after adsorption of HSPC (olive green) and DMPC (violet) followed by washing, after 30 min of sliding, 20. M. K. Kosinska et al., Arthritis Rheum. 65, 2323–2333
and between lipid-free gels immersed in lipid dispersions (purple and red), compared with lipid-incorporating gels
immersed in water (orange and green). Error bars indicate SD from at least three independent measurements. (2013).
(B) Effect of rehydration after dehydration of lipid-free and HSPC-incorporating gels, showing retention of the 21. A. V. Sarma, G. L. Powell, M. LaBerge, J. Orthop. Res. 19,
characteristic self-lubricating ability for the latter. Wavy-topped columns represent a lower bound on m (see also
Fig. 3, D and E, and materials and methods section 7). Error bars indicate SD from at least three independent 671–676 (2001).
measurements. (C) Wear of lipid-free and DMPC-incorporating gels, showing typical friction traces as well as the visual 22. B. A. Hills, B. D. Butler, Ann. Rheum. Dis. 43, 641–648
appearance of the gel samples, after 2 hours of sliding of the steel sphere on the gels, at 1-N load for the lipid-free gel
and 10-N load for the lipid-incorporating gel. (D) Appearance of gel samples under 50-N load (1.53-MPa contact (1984).
stress) on the steel sphere. After 1 hour of sliding (corresponding to 9 mm of wear), the appearance of the DMPC- 23. R. Sorkin, N. Kampf, Y. Dror, E. Shimoni, J. Klein, Biomaterials
incorporated gel was unchanged, whereas the lipid-free gel was torn after 12 s of back-and-forth shear under the load.
34, 5465–5475 (2013).
REFERENCES AND NOTES 8. E. Porte, P. Cann, M. Masen, J. Mech. Behav. Biomed. Mater. 24. M. E. Freeman, M. J. Furey, B. J. Love, J. M. Hampton, Wear
90, 284–294 (2019).
1. N. A. Peppas, J. Z. Hilt, A. Khademhosseini, R. Langer, 241, 129–135 (2000).
Adv. Mater. 18, 1345–1360 (2006). 9. C. W. McCutchen, Nature 184, 1284–1285 (1959). 25. A. H. Hogt et al., J. Colloid Interface Sci. 106, 289–298
10. V. C. Mow, G. A. Ateshian, in Basic Orthopaedic Biomechanics,
2. J. J. Green, J. H. Elisseeff, Nature 540, 386–394 (2016). (1985).
3. M. Liu et al., Nature 517, 68–72 (2015). V. C. Mow, W. Hayes, Eds. (Lippencott-Raven, 1997), pp. 275–315. 26. Q. Chen, D. Zhang, G. Somorjai, C. R. Bertozzi, J. Am. Chem.
4. J. P. Gong, Soft Matter 2, 544–552 (2006). 11. H. Forster, J. Fisher, Proc. Inst. Mech. Eng. H 210, 109–119
5. A. A. Pitenis, J. Manuel Urueña, A. C. Cooper, T. E. Angelini, Soc. 121, 446–447 (1999).
(1996). 27. L. Boulangé-Petermann, A. Doren, B. Baroux, M.-N. Bellon-Fontaine,
W. Gregory Sawyer, J. Tribol. 138, 042103 (2016). 12. S. Jahn, J. Seror, J. Klein, Annu. Rev. Biomed. Eng. 18, 235–258
6. X. Zhang, J. Wang, H. Jin, S. Wang, W. Song, J. Am. Chem. Soc. J. Colloid Interface Sci. 171, 179–186 (1995).
(2016). 28. A.-M. Trunfio-Sfarghiu, Y. Berthier, M.-H. Meurisse, J.-P. Rieu,
140, 3186–3189 (2018). 13. S. Lee, N. D. Spencer, Science 319, 575–576 (2008).
7. J. Cui, D. Daniel, A. Grinthal, K. Lin, J. Aizenberg, Nat. Mater. 14. J. Seror, L. Zhu, R. Goldberg, A. J. Day, J. Klein, Nat. Commun. Langmuir 24, 8765–8771 (2008).
29. R. Goldberg et al., Adv. Mater. 23, 3517–3521 (2011).
14, 790–795 (2015). 6, 6497 (2015).
ACKNOWLEDGMENTS

We thank S. Safran and A. Butcher for useful discussions and
M. Urbakh for comments on the manuscript. The electron microscopy
studies were conducted at the Irving and Cherna Moskowitz Center for
Nano and Bio-Nano Imaging at the Weizmann Institute of Science.
Funding: This project has received funding from the European
Research Council (ERC) under the European Union’s Horizon 2020
research and innovation program (grant 743016). We thank the
McCutchen Foundation, the Israel Science Foundation (grant 1715/
2014), the Israel Science Foundation–National Science Foundation
China (grant 2577/17), the Israel Ministry of Science and Technology
(grant 713272), and the Weizmann-EPFL Collaboration Program funded
by the Rothschild Caesarea Foundation for support of this work.
This work was made possible in part through the historic generosity of
the Harold Perlman family. Author contributions: R.G. and J.K.
conceived the project; R.G., W.L., M.K., N.I., and N.K. carried out
experiments; E.S. and R.G. carried out cryo-SEM imaging; J.K., R.G.,
W.L., and M.K. wrote the manuscript and analyzed the data; and
all authors commented on the manuscript. Competing interests: The
Weizmann Institute has a patent on low-friction hydrogels
(US20160175488A1). Data and materials availability: All data are
available in the manuscript or the supplementary materials.

SUPPLEMENTARY MATERIALS

science.sciencemag.org/content/370/6514/335/suppl/DC1
Materials and Methods
Supplementary Text
Figs. S1 to S16
References (30–39)

23 July 2019; resubmitted 10 May 2020
Accepted 6 August 2020
10.1126/science.aay8276

Lin et al., Science 370, 335–338 (2020) 16 October 2020 4 of 4

RESEARCH

SPECTROSCOPY response. Notably, the delays expected from

Zeptosecond birth time delay in that travel time of a photon across molecular
orbitals are one to two orders of magnitude

molecular photoionization shorter than the Wigner delay—i.e., they man-
ifest in the zeptosecond (1 zs = 10−21 s) domain.

In the description of light-matter interactions,

Sven Grundmann1*, Daniel Trabert1, Kilian Fehre1, Nico Strenger1, Andreas Pier1, Leon Kaiser1, such ultrashort time differences are often ig-
Max Kircher1, Miriam Weller1, Sebastian Eckart1, Lothar Ph. H. Schmidt1, Florian Trinter1,2,3, nored, and the dipole approximation is invoked.
Till Jahnke1*, Markus S. Schöffler1, Reinhard Dörner1* Doing so corresponds to neglecting the spatial

dependence of the light wave, and the light’s

Photoionization is one of the fundamental light-matter interaction processes in which the absorption electric field is approximated to be present

of a photon launches the escape of an electron. The time scale of this process poses many open instantaneously with the same phase over
questions. Experiments have found time delays in the attosecond (10−18 seconds) domain between the whole relevant region of space. Beyond

electron ejection from different orbitals, from different electronic bands, or in different directions. this approximation, however, the birth time

Here, we demonstrate that, across a molecular orbital, the electron is not launched at the same time. delay leads to a phase shift between the corre-

Rather, the birth time depends on the travel time of the photon across the molecule, which is sponding contributions to the overall photo-

247 zeptoseconds (1 zeptosecond = 10−21 seconds) for the average bond length of molecular hydrogen. electron wave. Such relative phases—and

Using an electron interferometric technique, we resolve this birth time delay between electron hence the birth time delays across a molecular

emission from the two centers of the hydrogen molecule. orbital—are accessible by experiments that ex-

ploit the interference between the different

parts of the wave function. Figure 1 outlines

P hotoionization is a fundamental quan- ingly, the birth time delay quantifies to what our metrology to measure tb.
tum process that has become a powerful extent a delocalized molecular orbital reacts Our approach builds on the close analogy
tool to study atoms, molecules, liquids, simultaneously as a single unit on being hit by
and solids. Facilitated by the advent of a photon. For example, it shows whether the between a plane wave behind a double slit
attosecond technology, it can now even part of the orbital facing toward an approach- (Fig. 1, A and B) and the photoelectron wave
emitted from a homonuclear diatomic molecule

be addressed in the time domain. Timing in ing photon reacts first and whether the part (Fig. 1, C and D). This analogy was proposed

photoionization usually refers to the time it downstream of the photon beam has a retarded by Cohen and Fano (8) and is well established

takes for an electron to escape to the con-

tinuum after absorption of the photon. This

time depends, for example, on the electronic

orbital (1, 2), on the energy band in solids (3, 4),

or on the orientation (5) and handedness (6)

of the target molecule. The escape time differ-

ence manifests as a phase shift of the electron

wave in the far field. This relation is based on

the concept of the Wigner delay, which is the

energy derivative of the phase of a photoelec-

tron wave at an asymptotic distance from the

source (7). Typical numbers are, for example,
20 attoseconds (1 as = 10−18 s) for the Wigner

delay between emission from the 2s and 2p

shells in neon (1).

However, the Wigner delay does not cover

another notable question about the timing in

photoionization of extended systems—namely

the temporal buildup of the photoelectron wave

across its spatially extended source. Hereafter,

we refer to these variations in the temporal

structure of the electron wave as the birth time

delay tb. Although the Wigner delay is caused
during the travel of the electron to the contin-

uum after its birth, tb relies on different birth
times of the contributions to the total photo-

electron wave along a molecular orbital. Accord- Fig. 1. Concept of birth time delay measurement. (A) Intensity distribution on a screen in the far field

behind the double slit in (B). (B) A plane wave impinges on a double slit. The phase shift ½DfŠ in the

1Institut für Kernphysik, Goethe-Universität, Max-von-Laue- right slit causes a tilt of the interference pattern. (C and D) Emission of a photoelectron wave from
Strasse 1, 60438 Frankfurt, Germany. 2Photon Science, two indistinguishable atoms of a homonuclear diatomic molecule mimics the double-slit setup in (B).
Deutsches Elektronen-Synchrotron (DESY), Notkestrasse 85, Here, the angle a is enclosed by the electron momentum vector and the molecular axis. A time delay (Dt)
22607 Hamburg, Germany. 3Molecular Physics, Fritz-Haber- between the emission from one of the two centers—e.g., originating from the travel time of the photon
Institut der Max-Planck-Gesellschaft, Faradayweg 4-6, 14195 impinging from the left side in (D)—leads to a shift of the interference fringes in (C). The ratio of slit distance
Berlin, Germany. (molecular bond length R, respectively) to wavelength is 1.65 in both cases [(B) and (D)]. In (B) the
right-hand part of the wave is delayed by Df ¼ p=2, whereas in (D) a birth time delay of 247 zs causes
*Corresponding author. Email: [email protected] Df ≈ p=11 for R = 0.74 Å. light dir., light direction; mol. dir., molecular direction.

(S.G.); [email protected] (T.J.); doerner@atom.

uni-frankfurt.de (R.D.)

Grundmann et al., Science 370, 339–341 (2020) 16 October 2020 1 of 3

RESEARCH | REPORT

today (9, 10). It has been used, for example, to Fig. 2. Electron emission from H2 mimics
study the onset of decoherence of a quantum the double-slit experiment. (A and
system (11) and entanglement in an electron B) Interference pattern of fast electrons
pair (12). The angular emission pattern from a (Ee = 735 ± 15 eV) from one-photon
gerade orbital has a maximum that is perpen- double ionization of H2 by 800 eV circularly
dicular to the molecular axis, which corresponds polarized photons for the average inter-
to the zeroth-order interference maximum be- nuclear distance of R = 0.74 ± 0.02 Å
hind the double slit. If a constant phase shift [purple line in (A)] and as function of R (B).
Df is introduced to one of the slits, the inter- The blue line in (A) models a double-slit
ference pattern behind the double slit be- interference pattern for a slit distance
comes asymmetric, and the angular position R = 0.74 Å and l = 0.45 Å, which is the
of the interference fringes moves as a function average de Broglie wavelength of the fast
of Df on an intensity screen in the far field. electron. The subset S of the data
In Fig. 1, A and B, we illustrate this relation. are used for (A) and for the subsequent
A corresponding shift may also occur when analysis of the birth time delay.
two interfering electron waves are emerging
from the two indistinguishable centers of a Fig. 3. Results of birth time delay measurement. Birth time delay of fast electrons (Ee = 735 ± 15 eV) from
homonuclear diatomic molecule upon photon one-photon double ionization of H2 by 800 eV circularly polarized photons for the average internuclear
absorption (Fig. 1, C and D). If the contribu- distance of R = 0.74 ± 0.02 Å (selected subset S as shown in Fig. 2B). (A) Electron angular distribution with
tions to the photoelectron wave are launched respect to the molecular axis, which is aligned parallel to the light propagation direction [cos(b) > 0.87,
simultaneously across an orbital of a diatomic corresponding to the top row of bins in (B)]. Red curve is the Gaussian fit used to obtain the angular
molecule, the electron emission pattern in the position of the zeroth-order maximum cos(a0). (B) Electron angular distribution in the molecular frame of
molecular frame of reference is symmetric reference as a function of cos(b). Dashed line is perpendicular to the molecular axis—i.e., the location
with respect to the normal of the bond axis. of the zeroth-order maximum in the absence of birth time delays. (C) Location of the maxima of the
However, any initial phase shift between the zeroth-order interference fringe as a function of cos(b). The maxima are obtained with Gaussian fits, as
waves emerging from one or the other center indicated by the red line in (A). The error bars include statistical and systematic errors, and the purple-
leads to an angular shift of the diffraction shaded error range indicates the systematical error (see supplementary materials for further details).
pattern—just as in the double-slit case. Hence, Left axis: cos(a0); right axis: birth time delay calculated with Eq. 2. The blue line resembles a birth time delay
inspecting the angular emission distribu- given by the travel time of light across the molecule (Eq. 3). The red line is a prediction combining atomic
tion of photoelectrons emitted from a homo- nondipole effects and the travel time of the photon (see text for details). elec., electron.
nuclear diatomic molecule for such angular
shifts offers a way to measure the birth time
delay tb (13).

For an electron of kinetic energy Ee and mo-
mentum pe, tb can be obtained from the mea-
sured angle a0—the angle to which the central
interference maximum is shifted—in the follow-
ing way: An electron wave is emitted and prop-
agates with the phase velocity vph = Ee/pe for
the time tb before a second electron wave is
born at a distance R. The zeroth-order inter-
ference maximum occurs under the emission
angle a0 for which the path length difference
between both electron waves vanishes


cosða0Þ Á R tb Á vph
2p l À l ¼0 ð1Þ

where l is the electron’s de Broglie wave- dimensional momenta of both protons in (KER) from the relative momentum of the
length. Accordingly, the birth time delay tb can coincidence with one electron. From the sum protons (15, 16). In the reflection approxima-
be inferred from the angular shift a0, and an momentum of the three measured particles, tion, the internuclear distance R at the moment
angle of a0 = 90° corresponds to zero birth time we inferred the missing electron momentum of photoabsorption is related to the kinetic
delay (simultaneous emission) vector by means of momentum conservation energy release through KER = e2/(4pe0R),
including the photon momentum. The mol- where e is the elementary charge and e0 is the
tb ¼ cosða0Þ R ð2Þ ecules in the target gas jet were randomly vacuum permittivity.
vph oriented with respect to the light propagation.
After ejection of the two electrons, the two One-photon double ionization typically pro-
We implemented this scheme by studying protons were driven apart by their Coulomb
one-photon double ionization of H2 using left- repulsion. We obtained the orientation of the ceeds by means of one of two sequential pro-
handed circularly polarized photons with an molecular axis and the kinetic energy release
energy of 800 eV. Using a COLTRIMS reac- cesses. A primary photoelectron is set free by
tion microscope (14), we measured the three-
the absorption of the photon, and the second

electron is either shaken off or knocked out to

Grundmann et al., Science 370, 339–341 (2020) 16 October 2020 2 of 3