RESEARCH ARTICLE SUMMARY◥

NEUROSCIENCE

Thirst regulates motivated behaviorthrough modulation of brainwideneural population dynamicsWilliam E. Allen, Michael Z. Chen, Nandini Pichamoorthy, Rebecca H. Tien,Marius Pachitariu, Liqun Luo*, Karl Deisseroth*

INTRODUCTION:Motivational drives are in-ternal states that can explainwhy animals adapt-ively display distinct behaviors in response tothe same external stimulus. Physiological needs(e.g., for food and water) are thought to pro-duce specific drives that engage particular goal-directed behaviors to promote survival. Thesedrives are established through the activity ofpopulations of hypothalamicneurons that sensephysiological variables and relay this informa-tion to other parts of the brain. However, it isunclear how the activity of these hypothalamicneurons activates and coordinates other brainsystems involved in sensation, decision-making,and action in order to produce the appropriategoal-directed behavior. The cellular dynamicsof brainwide states underlying basic survivaldrives have heretofore remained inaccessible.

RATIONALE:Understanding the neural basisof motivated behavior would require measur-ing activity within distributed neural circuitsin the brain and studying how this activity isregulated by the animal’s motivational state.Recent advances in extracellular electrophysio-logical recording technology have enabledlarge-scale measurements of neuronal activitythroughout the brains of behaving animals.This technology could lead to a spatiotemporalmap of activity flow across the brain as ananimal senses and responds to stimuli underdifferent conditions. By examining the animalin different motivational states, we could de-termine how these states are represented indiverse brain regions and how these statesinfluence neuronal activity across the brainto alter behavior.

RESULTS: We recorded the activity of ~24,000neurons throughout 34 brain regions duringthirst-motivated choice behavior; these record-ings were performed across 87 sessions from21 mice as they gradually consumedwater andbecame sated. We found that more than halfof recorded neurons were modulated by thistask, with a rapid and widespread response toa water-predicting olfactory cue precedingsustained activity related to water acquisi-

tion. The animal’s satietystate gated this wave ofactivity: In satiated mice,the same sensory cue pro-duced only a transientchange in activity and nobehavioral response. Sur-

prisingly, the spontaneous baseline activity ofneurons across many brain regions was alsomodulated by the animal’s satiety state.We found diverse correlates of the task at

the single-neuron level, with many neurons’activity correlating with specific sensory cuesor behavior, and some of them highly modu-lated by satiety state. Neurons from each ofthese groups were distributed throughout thebrain, with each brain area differing in therelative proportions of different types of neurons.We separated the high-dimensional populationactivity dynamics into lower-dimensional state-,cue-, and behavior-related modes. Whereas thecue- and behavior-related modes were pri-marily modulated by the task in a satiety state–dependent manner, activity along the statemode was persistent throughout the trial, in-cluding the period before odor onset. Thus,there existed a brainwide representation ofsatiety state that appeared to gate the flow ofactivity in response to sensory inputs.Finally, as a causal test of the role of this

widespread encoding of satiety state, we op-togenetically stimulated hypothalamic thirstneurons and caused mice to perform moretrials after reaching satiety. We found thatthis focal stimulation returned the behavior,brainwide single-neuronwithin-trial dynamics,and brainwide single-neuron correlates ofsatiety state back to the state corresponding tothirst. Stimulation also specifically modulatedactivity along the satiety-related mode in asubset of brain areas.

CONCLUSION: Our experiments revealed aglobal representation of the thirst motiva-tional state throughout the brain, which ap-pears to gate the brainwide propagation ofsensory information and its subsequent trans-formation into behavioral output.▪

RESEARCH

Allen et al., Science 364, 253 (2019) 19 April 2019 1 of 1

The list of author affiliations is available in the full article online.*Corresponding author. Email: lluo@stanford.edu (L.L.);deissero@stanford.edu (K.D.)Cite this article as W. E. Allen et al., Science 364, eaav3932(2019). DOI: 10.1126/science.aav3932

Thirsty Sated

State-dependent brainwide

activity dynamics

Brai

n re

gion

s

ThirstySated

Large-scale activity recordingduring head-fixed

thirst-motivated behavior

vs.

...Sensory Motor

Thirst

...Sensory Motor

How does thirst motivation regulate distributed brain activity?

Satiety state specifies initial conditions of

low-dimensional dynamics

Selective

Broadcast

State-related

activity

Cue-related

activity

Beha

vior

-rel

ated

ac

tivity

Widespread encoding of satiety state in

spontaneous activityThirsty Sated

Neu

rons

Brainwide dynamics of thirst motivational drive.Top left: Potential models for widespreador selective regulation of neural dynamics by thirst. Bottom left: Tracks of Neuropixels electrodesused to record activity from different brain regions during thirst-motivated behavior. Center:Brainwide activity dynamics of individual neurons in response to a sensory cue while a mouse wasthirsty and sated.Top right: Neural activity correlated with satiety state. Bottom right:Low-dimensional dynamics of task-related activity.

ON OUR WEBSITE◥

Read the full articleat http://dx.doi.org/10.1126/science.aav3932..................................................

on April 23, 2019

http://science.sciencem

ag.org/D

ownloaded from

RESEARCH ARTICLE◥

NEUROSCIENCE

Thirst regulates motivated behaviorthrough modulation of brainwideneural population dynamicsWilliam E. Allen1,2,3, Michael Z. Chen1,2, Nandini Pichamoorthy1, Rebecca H. Tien1,Marius Pachitariu4, Liqun Luo2,5*, Karl Deisseroth1,5,6*

Physiological needs produce motivational drives, such as thirst and hunger, that regulatebehaviors essential to survival. Hypothalamic neurons sense these needs and mustcoordinate relevant brainwide neuronal activity to produce the appropriate behavior.We studied dynamics from ~24,000 neurons in 34 brain regions during thirst-motivatedchoice behavior in 21 mice as they consumed water and became sated. Water-predictingsensory cues elicited activity that rapidly spread throughout the brain of thirsty animals.These dynamics were gated by a brainwide mode of population activity that encodedmotivational state. After satiation, focal optogenetic activation of hypothalamicthirst-sensing neurons returned global activity to the pre-satiation state. Thus,motivational states specify initial conditions that determine how a brainwide dynamicalsystem transforms sensory input into behavioral output.

Motivational drives are internal states thatcan explain why animals adaptivelychange their behavior in response tothe same external stimuli (1–4). Phys-iological needs (e.g., for food and water)

are thought to produce specific drives that en-gage particular goal-directed behaviors to main-tain homeostasis (1, 4, 5). These drives areestablished through the activity of populationsof hypothalamic neurons that sense physiologicalvariables (6, 7) and relay this information to otherparts of the brain.The nature of these internal states has been

debated for many decades (1). A prevalent modelis that motivational states regulate behavior byamplifying, in a state-dependent manner, thesalience of specific sensory signals that predictreward (4). Neural recordings in hungry animalsfound “gating” of responses to food-predictingstimuli, such that responses to these stimuli weregreatly diminished when animals were satedversus hungry. This modulation was found in avariety of brain regions, including the hypo-thalamus (8–14). However, the mechanisms reg-ulating this sensory modulation remain unclear,as does the neural representation of the motiva-

tional drive itself. Studies suggest the presenceof persistent neural encoding of the food satietystate in spontaneous activity (i.e., in the absenceof sensory stimuli) in several brain regions(15, 16), but the broader importance of thesefindings remains to be explored.We investigated these questions in the con-

text of thirst, a motivational drive that is es-sential for survival, is modulated on a rapid timescale, and is implemented by well-characterizedneural circuitry. Osmolarity- and angiotensin-sensitive neurons in the subfornical organ (SFO)and vascular organ of the lamina terminalis pro-ject to the median preoptic nucleus (MnPO),which integrates these physiological signals andrelays that information to multiple downstreamnuclei (17) (Fig. 1A). Activity in this circuit is aver-sive (18–20); water consumption reduces SFOand MnPO activity (19, 20) via activation oflocal MnPO interneurons (21), implementing amechanism akin to the classical “drive reduc-tion” model for regulation of water-seekingbehavior (1).However, it has remained unclear how thirst

motivation is represented in circuits downstreamof these dedicated thirst neurons, and how thisstate regulates activity relating to sensation andaction to implement goal-directed behavior. Oneimpediment to research in this area is the broadanatomical distribution of downstream targetsof thirst neurons; another is that goal-directedbehavior is thought to involve the coordinatedactivity of multiple interconnected circuits dis-tributed throughout the brain (6). Together theseconsiderations necessitate an approach to re-cording neural activity with single-cell resolutionon a brainwide scale.

Large-scale activity recording duringhead-fixed behaviorWe used a head-fixed thirst-motivated olfactoryGo/No-Go decision-making task (22, 23) to studyhow thirst motivation regulates sensory process-ing and behavior in mice (Fig. 1, B and C). Thistask consisted of a Go cue (67% of trials), afterwhich mice could instrumentally lick to triggerdelivery of a water reward, and a neutral No-Gocue (33% of trials) that yielded no reward uponlicking.Water-restrictedmice performedhundredsof trials per session (226 ± 7 trials, mean ± SEM),licking specifically on the Go cue with a near 100%hit rate, until they eventually became sated andstopped responding to the cue (Fig. 1, D and E,and fig. S1). At this point, we additionally re-corded several dozen trials (78 ± 4 trials, mean ±SEM) while the mouse was receiving the samesensory cues but not responding behaviorally(Fig. 1, D and E, and fig. S1).While mice performed this task, we used high-

density Neuropixels microelectrode arrays (24)to acutely record single units from many brainregions (listed in table S1) distributed acrossthe mouse forebrain and midbrain (Fig. 1, B andF). To ensure that single units were stably spike-sorted throughout the session, we used a recentlydeveloped spike-sorting framework with explicitdrift tracking and correction (25). We targetedthese recordings to include several thalamicand hypothalamic nuclei directly downstream ofthe MnPO [e.g., polymodal association thalamus(DORpm) and lateral hypothalamus (LZ)], di-verse second-order regions [e.g., prelimbic cortex(PL), insular cortex (INS)] that were anatomicallyconnected by axonal projection pathways to thesedirectly downstream regions, andmultiple regionsinvolved in olfactory sensation [e.g., piriform cor-tex (PIR)] and the control of licking [e.g., sec-ondary motor cortex (MOs)]. We developed amethod to fluorescently label the electrode trackand reconstruct its three-dimensional locationwithin a common atlas space at the end of eachexperiment (Fig. 1G, fig. S2, and supplemen-tary materials). We recorded a total of 23,881single units (255± 26units simultaneously,mean±SEM; fig. S3, A and B) from 34 brain regions;these recordings were performed across 87 ses-sions from 21 mice (fig. S3, C and D). A singleprobe was used within each recording session;each region was recorded at most in one sessionper mouse to allow for unambiguous identifi-cation of tracks to sessions, and up to eight elec-trode penetrations were made in the same mouseacross multiple days to record different regions.

Brainwide correlates of task and stateduring thirst-motivated behavior

There were extremely widespread changes infiring rates during task performance. The Go cueproduced a wave of activity in thirsty animalsthat began in olfactory regions within ~150ms ofodor onset and rapidly (within ~300 ms) spreadto neurons in every region we recorded from(Fig. 2A; fig. S4, A and B; andmovie S1). Both theinitial onset and the brainwide propagation ofthis activity wave substantially preceded the

RESEARCH

Allen et al., Science 364, eaav3932 (2019) 19 April 2019 1 of 9

1Department of Bioengineering, Stanford University, Stanford,CA 94305, USA. 2Department of Biology, Stanford University,Stanford, CA 94305, USA. 3Neurosciences GraduateProgram, Stanford University, Stanford, CA 94305, USA.4Janelia Research Campus, Ashburn, VA 20147, USA.5Howard Hughes Medical Institute, Stanford University,Stanford, CA 94305, USA. 6Department of Psychiatry andBehavioral Sciences, Stanford University, Stanford, CA94305, USA.*Corresponding author. Email: lluo@stanford.edu (L.L.);deissero@stanford.edu (K.D.)

on April 23, 2019

http://science.sciencem

ag.org/D

ownloaded from

onset of licking (first licks were measured at0.86 ± 0.04 s after odor onset, mean ± SEMacross sessions). A large fraction of these units inevery brain region (54% of total; 12,986) weremodulated by the task (Fig. 2B). For many re-gions, the change in firing rate (relative to base-line) peaked within 500ms of cue onset; however,a subset of regions developed and maintainedpersistent activity that peaked instead just afterthe time of the reward (2 s after odor onset; Fig.2C). The latter included many known sites im-plicated in motor output, such as primary motorcortex (MOp), MOs, and somatosensory thalamus(DORsm). Many of these neurons exhibited in-creases in activity upon licking onset (fig. S4C)as well as further increases upon reward delivery(fig. S4D), which was behaviorally associated withincreased intensity of licking (Fig. 1E). The dis-

tributions of both baseline firing rates and Goodor-evoked changes in firing rates exhibitedsubstantial variability across brain regions (fig.S4, E and F). The No-Go cue produced a sim-ilarly rapid response to cue onset within manyregions that also responded rapidly to the Gocue, but then failed to produce persistent changesin activity past this initial transient period, cor-responding with the lack of behavioral engage-ment relative to Go trials (Fig. 2, A and C; fig. S4,A and B; and movie S1). When mice were sated,there was still a diminished transient responseto the Go and No-Go cues in many regions, butthe stimuli failed to produce persistent activitypast the sensory cue period (Fig. 2, A and C, andmovie S1). Satiety state–dependent gating ormodulation of those sensory cue–evoked re-sponses appeared to be similarly widespread

(compare Go thirsty and sated conditions inFig. 2, A and C), although widespread buttransient odor responses were observed evenin the sated state.We also observed large and widespread mod-

ulations in the spontaneous firing of neurons asa function of the animal’s satiety state in the pre-stimulus baseline period.Many neurons exhibitedrapid changes in their spontaneous activity overthe course of a few trials as the animal transi-tioned from being thirsty to being sated thatremained persistent for many minutes (Fig. 2D).To characterize thismodulation,we foundneuronsthroughout the brain with activity during thebaseline period that was significantly correlatedwith the animal’s satiety state over the courseof a session (P < 0.01, absolute R value comparedwith shuffled data, Bonferroni correction) (Fig.2E). These neurons constituted a large fractionof neurons in each brain region (N = 8395, 35%of total; Fig. 2F). Although roughly as manyneurons increased or decreased baseline firingrate going from thirsty to sated (3811 versus 4584),there was a significant net change in firing ratein a variety of brain regions (Fig. 2G).

Single-cell correlates of differenttask variables

How are these state-dependent dynamics im-plemented across the brain at the level of singleneurons? The activity patterns of many task-modulated neurons were highly correlated withthe state, cue, or behavior aspects of the task(Fig. 3A). Unsupervised clustering of the averageactivity of task-modulated neurons (N = 12,986)across the thirsty and sated Go and No-Go trialsrevealed clusters of neurons with activity thatqualitatively appeared to correlate with each ofthese state, cue, and behavior variables (Fig. 3B),as well as discrete subclusters within each group.The state-related clusters (N = 3484 units, 27%of total) exhibited only slight modulation uponcue presentation but were active throughout thetrial selectively on either thirsty or sated trials(Fig. 3B, blue clusters). The cue-related clusters(N = 5007 units, 38% of total) were more variedbut were primarily activated or suppressed dur-ing the odor period (Fig. 3B, green clusters); cer-tain cue clusters were also state-modulated in amanner superimposed on the odor modulation.The behavior-related clusters (N = 4495 units,35% of total) were active only during thirsty Gotrials when the mice licked for water; they wereselective for a particular period during the taskbefore, during, or after reward consumption (Fig.3B, red clusters) and exhibited activity that ex-tended beyond the odor period. Neurons fromnearly every cluster were represented in everybrain region, with differing relative proportionsof neurons from each cluster among all task-modulated neurons in different regions (Fig. 3C).Across both the odor and response epochs,

many more clusters were selective for Go thanfor No-Go trials (Fig. 3, B and D), and severalclusters exhibited selective activation on thirstyversus sated trials (Fig. 3, B and E). The Go cuerapidly produced increased firing in neurons

Allen et al., Science 364, eaav3932 (2019) 19 April 2019 2 of 9

Neuropixels array

Odor

Waterspout

A C

Ethyl acetate

2-pentanone

1 2

ø

E

Lick

øø

No lick

Lick

No lick

Hit

Miss

CR

FA

Odor ResponseBaseline

0

ITI

B

D

Sated0123456

Lick

s/s

GoNo-Go

−1 0 1 2 3 4Time (s)

0123456

Lick

s/s

Thirsty

0 2 4Time (s)

0

50

100

150

200

250

Tria

l num

ber

0 2 4Time (s)

0

20

40

60

80

100

120

Go No-Go

PLMOsORB

OLF STRv

ACA LS STRd

P-mot

SC

BLA

PALv

DORpm

LZ MEZ

SSpHPC

RSP

DORsmMBmot

PIRMEZ

AI

SSs

IL

PVZ PVR PAsAMY BMA

MOp

MBsta

BST

PALd

G

Go (67%)

No-Go (33%)

thalamus

hypo.midbrain

VOLT

SFO

MnPO

Time relative to cue (s)

F

PALm

Fig. 1. Brainwide neuronal recording during thirst-motivated behavior. (A) Diagram of theoutput of the hypothalamic thirst circuit to other brain regions. SFO, subfornical organ;VOLT, vascular organ of the lamina terminalis; MnPO, median preoptic nucleus; hypo., hypothalamus.(B) Schematic drawing of the experimental setup. (C) Diagram of olfactory Go/No-Go taskstructure and different task epochs: baseline epoch (–1.1 to 0 s before odor period), odor epoch(0 to 0.5 s after odor onset), and response epoch (1.5 to 2.5 s after odor onset). FA, false alarm;CR, correct rejection; ITI, intertrial interval. (D) Example behavioral data from a single behavioralsession, showing individual licks on Go and No-Go trials over the course of the session. (E) Averagelicking behavior on Go and No-Go trials while thirsty and sated (mean ± SEM, N = 87 sessions).(F) Electrode tracks from all 87 recording sessions, co-registered into the Allen common referencespace. Mouse brain is outlined in wire mesh. (G) Locations of all recorded neurons in AllenBrain Atlas space, colored by brain region. See table S1 for brain region abbreviations (fromAllen Brain Atlas ontology) used here and in subsequent figures.

RESEARCH | RESEARCH ARTICLEon A

pril 23, 2019

http://science.sciencemag.org/

Dow

nloaded from

throughout the brain in thirsty mice, and thesame Go cue in the sated state or the No-Go cuefailed to produce this large-scale activation.Satiety state–dependent modulation prior to theonset of behavior is therefore not manifestedas a broad change in responsiveness across allcue-responsive neurons, but in fact it specifi-cally modulates the activity of certain subclassesof those neurons (Fig. 3E, arrows).

We modeled neuronal activity using a set ofstate-, cue-, and behavior-related regressors todetermine the extent to which single-neuronactivity correlated with specific task elementson a trial-by-trial basis (fig. S5, A to C). Task-modulated neurons exhibited substantially betterfits to the model than nonmodulated neurons(mean 0.17 ratio of deviance explained for task-modulated versus 0.06 for non–task-modulated,

P < 0.001, Wilcoxon rank-sum test; fig. S5D),and these features modeled neurons in differentclusters to different extents (fig. S5E). The state,cue, and behavior regressors tended to explainthe most deviance in clusters that qualitativelyappeared to encode those features (fig. S5, Fand I). Different brain regions had differentfractions of neurons that encoded state-, cue-,or behavior-related features (or combinations

Allen et al., Science 364, eaav3932 (2019) 19 April 2019 3 of 9

Fig. 2. Widespread task- andstate-related activity dynamics.(A) Baseline subtracted, Z-scoredtrial-averaged firing rates for alltask-modulated recorded neurons(N = 12,986 of 23,881) from 21 mice,averaged from all thirsty Go andNo-Go trials. Neurons are sorted firstby brain region, and then by time topeak within Go trial; s.d., standarddeviation. (B) Fraction of task-modulated neurons per brain region.(C) Per-region population firingrate during thirsty and sated Go andNo-Go trials. Regions are sorted byaverage time to peak change in pop-ulation firing rate during the thirstyGo condition, relative to baseline,across all neurons from all mice.Dots indicate times of maximalchange in firing rate for each regionin each trial type. (D) Example state-modulated neurons (baseline firingrate significantly correlated withstate) showing baseline firing rateover trials across session, with brainregion and correlation with state(R value) indicated. (E) All state-modulated neurons, Z-scored andsorted by correlation with state,showing the last 20 consecutivetrials of thirsty and sated blocks(N = 8395). (F) Fraction of state-modulated neurons per brain region.(G) Average change in baselinefiring rate per brain region betweenthirsty and sated conditions.*P < 0.05 (Wilcoxon signed-ranktest, Bonferroni correction). Data aremeans ± 95% confidence interval.

Fra

ctio

n st

ate

mod

ulat

ed

863

268 26 51 41 99 169

1382 50

3 33 316 77 158

765

321

214

442

372

648

212 95 126 39 116

196

215 70 359

144

fir

ing

rate

(sa

ted

- th

irsty

)(s

pike

/s)

Go

-3.5 3.5

No-GoBA

Olfactory

Somatosensory cortex

Amygdala

Thalamus

Hypothalamus

Hippocampus

Striatum

Pallidum

Association cortex

Motor cortex

Midbrain

Hindbrain

Go No-GoOLF

PIR

SSp

SSs

BLA

BMA

sAMY

PA

DORpm

DORsm

PVR

PVZ

LZ

MEZ

HPC

STRd

STRv

LS

PL

ILA

ORB

ACA

AI

RSP

PALd

PALm

PALv

BST

MOp

MOs

SC

MBsta

P-mot

Thirsty Sated

EC

FDNeuron 1 (PL, R=-0.83)

Neuron 3 (PALv, R=0.75)

0

8000

G

Neuron 2 (DORsm, R=-0.78)

Neuron 4 (sAMY, R=0.74)

Thirsty Sated

OLF PIR

SS

pS

Ss

BLA

BM

AsA

MY

PA

DO

Rpm

DO

Rsm

PV

RP

VZ LZ

ME

ZH

PC

ST

Rd

ST

Rv

LS PL

ILA

OR

BA

CA AI

RS

PP

ALd

PA

LmP

ALv

BS

TM

Op

MO

sS

CM

Bm

otM

Bst

aP

-mot

n =

Firi

ng r

ate

(spi

ke/s

)

Thirsty Sated

Neu

ron

num

ber

0 1Fraction taskmodulatedFiring rate (s.d.)

Firing rate (s.d.)

1.5

-1.5

MBmot

PIRPVROLF

STRvMBstaPALv

LSPLLZ

sAMYILASC

PVZHPCACAMEZ

MBmotP-motBMA

PALmRSP

DORpmSTRd

SSpORBMOsMOpSSs

AIBST

DORsmBLA

PALdPA

0Time (s)

2 Trials

Thirsty SatedGo No-Go Go No-Go

-5 5Firing rate (spike/s)

1

Baseline firing rate

0

1

OLF PIR

SS

pS

Ss

BLA

BM

AsA

MY

PA

DO

Rpm

DO

Rsm

PV

RP

VZ LZ

ME

ZH

PC

ST

Rd

ST

Rv

LS PL

ILA

OR

BA

CA AI

RS

PP

ALd

PA

LmP

ALv

BS

TM

Op

MO

sS

CM

Bm

otM

Bst

aP

-mot

0

10

0Time (s)

21

* * * * * * * * * * * * * * * * *

RESEARCH | RESEARCH ARTICLEon A

pril 23, 2019

http://science.sciencemag.org/

Dow

nloaded from

thereof) (fig. S5G), but the majority of neuronsin most regions were selective for only one fea-ture (fig. S5, H and I).

Decomposing brainwidepopulation dynamics

How is this single-neuron activity coordinatedwithin neural populations to implement specificcomputations related to sensory processing, be-havioral output, and internal state? The hetero-

geneous activity patterns of neurons within eachbrain region suggested that we should examinepopulation dynamics–based encoding of internalstate and behavior (26–28). Population single-unit recording has shown that the trajectories inpopulation activity space, of dimension equalto the number of recorded neurons, often existon low-dimensional manifolds within this high-dimensional space (29–32). These results sug-gest that the diversity of activity patterns across

many neurons often reflects a common, lower-dimensional dynamical variable (33).How do these dynamical variables encode the

state-, cue-, and behavior-related aspects of thetask? We took advantage of the fact that theseaspects are temporally separatedwithin differentepochs of single trials.We defined a “Statemode”of activity that maximally separated thirsty andsated states during the baseline epoch, a “Cuemode” that maximally separated activity on Goand No-Go trials during the odor epoch (up to500ms after odor onset, before the mouse beganlicking), and similarly a “Behavior mode” thatseparated activity on Go and No-Go trials fromthe response epoch (1 s around time of the rewardonset) (Fig. 4A) (34). The Cue mode was designedto capture activity related to the sensory responsebut could also include signals related to rewardanticipation, decision-making, and motor prepa-ration. The Behavior mode captured activity po-tentially related to motor output, sensorimotorfeedback, and reward anticipation and consump-tion. Wemade use of this temporal separation todefine orthogonal modes of population activitythat encoded each task element (Fig. 4A), whichtogether captured ~50% of the trial-averagedtask-related variance throughout the brain. Priorto orthogonalization, the angles between the Stateaxis and the Cue and Behavior axes were 82° and84°, respectively, and 65° between the Cue andBehavior axes. The vector defining the Statemodewas stable over the course of the trial, whereas thevectors defining Cue and Behavior modes weremost stable within the epoch from which theywere defined (fig. S6A).We then projected brainwide trial-averaged

neural activity onto these axes to visualize howthe State, Cue, and Behavior variables evolvedover time across different trial conditions (Fig. 4,B and C). The dynamics for eachmode computedfrom all neurons across all animals were verysimilar to those computed using subsampledpopulations of 20 neurons per region (fig. S6B)and to those computed for each animal individ-ually (fig. S7). Finally, the rare thirsty Go trialswhere the animal did not obtain a reward hadCue and State dynamics thatwere similar to thoseof other thirsty Go trials but displayed reducedactivity along the Behavior mode; likewise, thirstyNo-Go trials where the animal incorrectly lickedduring the response period had Cue and Statedynamics that were similar to those of otherthirsty No-Go trials but displayed additionalactivity along the Behavior mode (fig. S6C).These three activity modes displayed distinct

dynamics and sensitivities to satiety state. TheState mode was persistently elevated throughoutboth Go and No-Go trials when the animal wasthirsty and displayed only slight modulation bytask engagement (Fig. 4C, left). By contrast, theCue mode strongly discriminated between GoandNo-Go trials anddisplayed a large reduction inactivity on Go trials upon satiety (Fig. 4C, center).Finally, the Behavior mode was only substantiallyactivated on thirsty Go trials (Fig. 4C, right).Activity along these modes changed over

the course of a session in different ways. Mice

Allen et al., Science 364, eaav3932 (2019) 19 April 2019 4 of 9

-2.5

2.5

A

CB

Go

No Go

Thirsty SatedGo No-Go Go No-Go

Beh

avio

rC

ueS

tate

0 1 2 0 1 2 0 1 2 0 1 2Time (s)

FR(s.d.)

0 25 50 75 100Percent in cluster

PVRPL

PALmILAPIROLFPVZ

MBstaPALv

LZACAHPC

sAMYMEZ

SCSTRv

LSBMA

PAORBMOsRSP

DORpmP-mot

MBmotPALd

BLABST

STRdSSp

DORsmAI

MOpSSs

Spi

kes/

s

0

100

Thirsty Sated

0

30

Neuron 4 (OLF)

Neuron 3 (MOs)

0

75

Neuron 1 (STRv)Thirsty Sated

0

15

Neuron 2 (OLF)

D

0.0

0.4

Sta

te s

elec

tivity

inde

x(T

hirs

ty v

s S

ated

)

Tria

l sel

ectiv

ity in

dex

(G

o vs

No-

Go)

EOdor epoch

Responseepoch

Responseepoch

Baselineepoch

Odorepoch

Cluster Cluster

0

15

Thirsty Sated

Time (s)0 1

Neuron 5 (MOs)

0

Neuron 6 (LZ)

Fig. 3. Representations of sensory processing, behavioral output, and internal state at thesingle-neuron level across the brain. (A) Example neurons correlated with different aspectsof the task or state. Top: Spike raster during Go/No-Go trials while thirsty and sated withina single session. Bottom: Average firing rate during same trials. Neurons 1 and 2 are primarilystate-related, neurons 3 and 4 are primarily cue-related, and neurons 5 and 6 are primarily behavior-related. Brain region is indicated in parentheses for each neuron. (B) Unsupervised clusteringof all 12,986 task-modulated neurons. Behavior-related clusters are red; cue-related clusters aregreen; state-related clusters are blue. Individual neurons are trial-averaged per condition andZ-scored across conditions. FR, firing rate. (C) Hierarchically clustered cluster composition by brainregion, showing fraction of total task-modulated neurons within each region that comes from allclusters. Regions in colored italics: PVZ (periventricular zone of the hypothalamus) containsmostly state-related neurons, PIR mostly cue-related neurons, and MOp mostly behavior-relatedneurons. (D) Trial selectivity index per cluster during odor epoch (0 to 0.5 s after odor onset)and response epoch (1.5 to 2.5 s after odor onset). Trial selectivity index is computed as(hGoi – hNo-Goi)/(hGoi + hNo-Goi) during a particular task epoch, where Go is the average onthirsty Go trials and hNo-Goi is the average on thirsty No-Go trials. (E) State selectivity index percluster during different task epochs. Trial selectivity index is computed as (hThirstyi – hSatedi)/(hThirstyi + hSatedi) during a particular task epoch, where hThirstyi is the average on thirsty Go trialsand hSatedi is the average on sated Go trials. Data in (D) and (E) are means ± 95% confidenceinterval. Arrows indicate sensory clusters with large changes in firing rate during the task as afunction of state.

RESEARCH | RESEARCH ARTICLEon A

pril 23, 2019

http://science.sciencemag.org/

Dow

nloaded from

Allen et al., Science 364, eaav3932 (2019) 19 April 2019 5 of 9

OLF PIR

SS

pS

Ss

BLA

BM

AsA

MY

PA

DO

Rpm

DO

Rsm

PV

RP

VZ LZ

ME

ZH

PC

ST

Rd

ST

Rv

LS PL

ILA

OR

BA

CA AI

RS

PP

ALd

PA

LmP

ALv

BS

TM

Op

MO

sS

CM

Bm

otP

-mot

0.0

0.2

0.4

0.6

0.8

1.0

Thirsty SatedState

Behavior

State

Cue

1. Compute modes from activity in different task epochs

0 1 2

Pro

ject

ion

(AU

)

0 1 2

0

BeginningMiddleEndSated

0 1 2

0

0 1 2

0

400

Pro

ject

ion

(AU

)

Thirsty GoThirsty No-GoSated GoSated No-Go

0 1 2

0

600

0 1 2

0

800

...

A

C State (19.6% EV) Cue (9.2% EV) Behavior (20.1% EV)

Time (s)

Time (s)

D

E

F

800600

Cue

State decoding

OLFPIRSSpSSsBLABMA

sAMYPA

DORpmDORsm

PVRPVZ

LZMEZHPC

STRdSTRv

LSPL

ILAORBACA

AIRSP

PALdPALmPALvBSTMOpMOs

SCMBmotP-mot

Sated Thirsty

R 2

OLF PIR

SS

pS

Ss

BLA

BM

AsA

MY

PA

DO

Rpm

DO

Rsm

PV

RP

VZ LZ

ME

ZH

PC

ST

Rd

ST

Rv

LS PL

ILA

OR

BA

CA AI

RS

PP

ALd

PA

LmP

ALv

BS

TM

Op

MO

sS

CM

Bm

otP

-mot

0.0

0.2

0.4

0.6

0.8

1.0

AU

C

OLF

(27

)

PIR

(11

)

SS

p (3

)

SS

s (3

)

BLA

(3)

BM

A (

5)

sAM

Y (

6)

PA

(3)

DO

Rpm

(22

)

DO

Rsm

(14

)

PV

R (

3)

PV

Z (

2)

LZ (

15)

ME

Z (

6)

HP

C (

8)

ST

Rd

(24)

ST

Rv

(11)

LS (

9)

PL

(9)

ILA

(8)

OR

B (

12)

AC

A (

10)

AI (

2)

RS

P (

1)

PA

Ld (

6)

PA

Lm (

3)

PA

Lv (

5)

BS

T (

1)

MO

p (6

)

MO

s (9

)

SC

(3)

MB

mot

(11

)

P-m

ot (

5)

0.0

0.2

0.4

0.6

0.8

1.0

Sated

Thirsty

AU

C

I

HG

J

K

0 1 2Time (s)

Stimulus decoding

05

101520

Lick

s/tr

ial 25 50 75 100 125 150 175

0 25 50 75 100 125 150 175Trial

SatedThirsty

Thirsty

Sated

Pre

d.

State Decoding (AUC = 0.94) Stimulus Decoding (AUC = 0.71)Lick Decoding (R 2 = 0.82)

Tru

eP

red.

0

400

B

State (A

U)

0

200

400

Cue (AU)0200400

600

Beh

avio

r (A

U)

0

200

400

600

800

Thirsty GoThirsty No-GoSated GoSated No-Go

Odor startOdor stopReward

State thirsty sated

Cue = Go No-Go

Behavior Go No-Go

Trial type A Trial type B

...

Neuron 1

Neuron N

Neuron 2Neuron 3Neuron 4

...

2. Project population activity onto modes

Behavior

0

Max Distance (A

U)

Total licks decoding

Fig. 4. Thirst motivational state regula-tion of brainwide neural populationdynamics. (A) Top: Schematic of howstate-, cue-, and behavior-related modesare defined from population activity indifferent task epochs. Cue and Behaviormodes are only computed from thirstytrials. Bottom: Schematic of how neuralpopulation activity is projected into athree-dimensional space. (B) Brainwideneural population activity projected ontoactivity modes (State, Cue, and Behavior)defined by trial-averaged activity during Goand No-Go trials while thirsty and sated.(C) Same as (B), but showing activityalong each mode individually. EV, totalexplained variance. (D) Brainwide popula-tion dynamics on Go trials projected ontoState, Cue, and Behavior modes overcourse of session (beginning, middle, endof thirsty block and sated). (E) Per-regionaverage projections of simultaneouslyrecorded population activity onto taskaxes. State: average absolute differencebetween thirsty and sated trials projectedonto State mode. Thirsty/sated Cue:average absolute difference between Goand No-Go trials projected onto Cue mode.Thirsty/sated Behavior: average absolutedifference between Go and No-Go trialsprojected onto Behavior mode. Each trace isnormalized to the maximum across allconditions. (F) Example decoding of statefrom baseline epoch activity (1 s beforeodor) of simultaneously recorded neurons inOLF. Prediction on held-out data is shownin dots; observed state values are shown inshaded background. AUC, area under curve.(G) Example decoding of total number oflicks from response epoch activity (1 saround potential reward time), from samepopulation as (F). Black, observed; red,predicted. (H) Example decoding of trialtype from odor epoch activity (first 500 msof odor response) over session, fromsame population as (F). Green, Go trial;magenta, No-Go trial. (I) Per-region averagedecoding from baseline epoch. (J) Per-region lick decoding over whole sessionfrom response epoch. (K) Per-regiondecoding of stimulus from odor epoch, inthirsty and sated states. Regions labeledin red have a statistically significantdecrease in decoding (P < 0.05, two-tailedt test, FDR correction across all regionswith three or more recordings). Regionslabeled in gray have fewer than threerecordings. Numbers of recordings perregion for (I) to (K) are indicated inparentheses in (K). Decoder was trainedonly on thirsty trials. For (I) and (K),chance performance (0.5) is indicatedby dashed line. Data in (I) to (K) aremeans ± 95% confidence interval.

RESEARCH | RESEARCH ARTICLEon A

pril 23, 2019

http://science.sciencemag.org/

Dow

nloaded from

performed the task consistently through the be-ginning and middle of each session but then be-gan to reduce their per-trial licking and toincrease their latency to lick as they approachedsatiety (fig. S1, C and D). Similarly, activity alongthe projection onto the State axis in Go trials washighest at the beginning of the session andslightly lower in themiddle (Fig. 4D). Toward theend of the session, the activity dropped halfwayto the sated state and then dropped further stillupon reaching satiety (Fig. 4D), mirroring thechanges in behavior. By contrast, Go trial projec-tions onto the Cue and Behavior axes had simi-lar activity as long as the mouse was licking forwater but had very little activity when themousewas sated (Fig. 4D).

Regional encoding of taskand state variables

Howdo populations of neurons in different brainregions encode these variables? We defined thesame State, Cue, and Behavior axes using dataonly from simultaneously recorded neurons with-in a single brain region. Projections of single-trialpopulation activity onto these axes, averagedacross multiple sessions, revealed that thesesmaller populations of neurons in nearly everyregion maintained qualitatively similar, state-dependent encodings of the three variables (Fig.4E). Activation along the Cue mode was onlysomewhat diminished with satiety in olfactoryregions OLFC and PIR but was highly reducedin most other brain regions. Similar dynamicswere found using projections onto axes com-puted from trial-averaged data of all neuronsrecorded in each brain region (fig. S8).How well does simultaneously recorded pop-

ulation activity within the different epochs usedto define these modes encode information on atrial-by-trial basis about the animal’s state,sensory input, and behavior? We applied alinear decoding approach to quantify the ex-tent to which activity within each of the threetask epochs (baseline, odor, and response)respectively encoded different task elements(satiety state, trial type, and per-trial licking)(Fig. 4, F to H). Satiety state and odor identityin thirsty mice could be decoded using activityonly in the pre-odor baseline epoch or the first500 ms of odor response (odor epoch), respec-tively, from populations of neurons throughoutthe brain (Fig. 4, I and K). In sated mice, theability to decode odors dropped to near chancelevels throughout much of the brain [18/29 re-gions with >3 recordings per region had sig-nificantly changed decoding accuracy; P < 0.05,two-tailed t test, false discovery rate (FDR) cor-rection] (Fig. 4K). Decoding of per-trial totallicking from the response epoch was much morevariable, with larger differences in decoding ac-curacy between different regions (Fig. 4J); thesame was true for instantaneous decoding of lickrate (fig. S9). Using subsampled populations of20 neurons simultaneously recorded per record-ing yielded reduced, but still substantially abovechance, decoding accuracies from populationsthroughout the brain (fig. S10).

Causal manipulation of thirstmotivational stateAre changes along this State mode causallyrelated to changes in behavior? To answer thisquestion, we examinedwhether focal and targetedstimulation of hypothalamic thirst neurons wouldbe sufficient in sated mice to push brainwideneural population activity back into its thirstystate and restore the subsequent task-relatedglobal neural dynamics. We expressed an ex-citatory channelrhodopsin (35) in SFO Nos1+excitatory neurons (18, 36, 37) (N = 7mice) (Fig.5A). After mice became sated in each session,we stimulated the SFO neurons in a pattern thatcauses water-seeking behavior in sated animals(18, 37). We then had the mice perform a blockof several additional “Stim” trials, then turnedthe stimulation off for a block of “Washout” trials(Fig. 5B). Optogenetic stimulation restored theanimals’ behavior to that during natural thirst.Ending the optogenetic stimulation behaviorallyreturned animals to their sated state (Fig. 5,C to F).This local, targeted optogenetic stimulation

in sated mice globally restored the within-trialactivity patterns and brainwide timing inter-relationships of task-modulated neurons to thenatural dynamics of the thirsty state. Terminat-ing optogenetic stimulation immediately returnedneuronal responses to the sated-state dynamics(Fig. 5G).We next examined the spontaneous activity

during the baseline period of individual neuronsthat were correlated with the animal’s satietystate. The baseline firing rate of many individualneurons during stimulation returned to theiractivity level from when the animal was thirsty(Fig. 5, H to J). The majority of neurons thatdecreased their baseline firing rate as the animaltransitioned from being thirsty to being satedsubsequently increased their firing rate uponstimulation (1195 of 1781, 67%), and vice versa forneurons that increased their firing rate upon thetransition from thirsty to sated (1003 of 1417,71%) (P < 0.05, c2 test) (Fig. 5J). Thus, acute andtargeted optogenetic stimulation not only restoredwithin-trial dynamics related to sensory process-ing and behavior, it globally modulated thespontaneous firing of neurons throughout thebrain to return these persistent representationsof satiety state toward values corresponding towhen the animal was thirsty.

Population response to optogeneticthirst induction

We then projected these brainwide dynamicsonto State, Cue, and Behavior axes defined byactivity in the natural thirsty and sated states.Optogenetic stimulation of the thirst-sensingneurons specifically shifted global activity alongthe naturally occurring State mode toward itsposition when the animal was thirsty, both forneurons pooled across mice (Fig. 6A) and forneurons from individual animals (fig. S11). Afterthe conclusion of optogenetic stimulation, activ-ity along this mode returned to its position priorto the onset of stimulation, when the mouse was

sated (Fig. 6A). Projections of activity from thestimulation period onto the Cue and Behaviormodes also faithfully recovered the naturally oc-curring within-trial dynamics from the thirstystate, with a slight baseline offset detectable onlybefore task initiation (Fig. 6A). Comparing trial-averaged projections from stimulation trials withdata from different points within the session re-vealed that this recovery along the State axis was,in fact, a return to the natural level of activitycorresponding to near the end of the thirsty trials,immediately before the animal became fully sated(Fig. 6B). Single-trial projections from simulta-neously recorded neurons revealed similar effects(Fig. 6C). This stimulation also rapidly and tran-siently modulated the firing rates of neurons inthe expecteddownstreambrain regions (fig. S12).Thus, focal and targeted optogenetic stimulationof the thirst pathway returned both the persistentcellular- and population-level encoding of theanimal’s satiety state across the brain, and thesubsequent task-related global dynamics withina trial, to that when the animal was naturallythirsty earlier in the session.The effect of thirst pathway stimulation on

the State mode—matching the natural level ofactivity toward the end of the thirsty trials, be-fore the animal became fully sated—suggestedthat either (i) a submaximal thirst state was beingrecapitulated throughout the brain, or (ii) thethirst state was being fully recapitulated only in asubset of regions. To discriminate these possibil-ities, we analyzed the predictions of decoders ofsatiety state and trial type, trained on thirsty andsated data from the activity during the baselineepoch and initial odor response epoch, respec-tively, of simultaneously recorded neurons in asingle region. Decoders of satiety state using pre-odor activity predicted “thirsty” in held-out thirstytrials and “sated” in held-out sated trials (Fig. 6D).When decoders were tested on the same taskepoch in stimulation trials (held out from thetraining data), many regions indeed decoded thestate as thirsty (Fig. 6D); in particular, orbito-frontal cortex (ORB), agranular insular cortex(AI), andMOs all decoded the state as being verylikely thirsty. Several of the highest-level corticalregions [corresponding tomedial prefrontal (PL/ILA) and retrosplenial (RSP) cortex] continuedto decode that the animal was sated, consistentwith expected access of these regions to contextualor working-memory cues on recent water con-sumption (38). These other regions may addi-tionally receive visceral or other direct satietycues, and thus the intensity of representationof the induced state in specific regions may betuned by additional influences or informationsources, which may be isolated, defined, andstudied further in this way. These results suggestthat recapitulation of the thirst state in a subsetof regions is sufficient to restore thirst-motivatedsensory processing and behavior, along with theassociated brainwide activity dynamics. Uponending stimulation, most regions again correctlypredicted that the mouse was sated (Fig. 6D).Many of the same regions (e.g., ORB, AI, MOs)

that predicted that the animal was thirsty while

Allen et al., Science 364, eaav3932 (2019) 19 April 2019 6 of 9

RESEARCH | RESEARCH ARTICLEon A

pril 23, 2019

http://science.sciencemag.org/

Dow

nloaded from

being stimulated also immediately returned toaccurately decoding odor identity during thefirst 500 ms after odor onset (Fig. 6E). Olfactorysensory regions (PIR and OLF), while changingin the expected direction in their state decoding,nonetheless continued to decode odor identitywith high accuracy across all four trial blocks(Fig. 6E).

Discussion

Our results comport with the low dimensionalityand pervasiveness of brainwide cellular activity

during behavior in both Caenorhabditis elegans(39) and larval zebrafish (40). In mammals, thesimplicity of the thirst-motivated olfactory Go/No-Go task establishes a baseline of activity towhich brainwide dynamics in other, more in-volved tasks can be compared. For example, dy-namics at least as widespread were observedusing large-scale recordings in a visual decision-making task involving competing sensory stimuli(41). Here, in experiments focused on a basic sur-vival drive behavior, neurons throughout the brainresponded within a few hundred milliseconds

to the odor regardless of the animal’s satietystate or odor identity, and in a state- and odor-dependent manner rapidly came to a determina-tion of whether to propagate the signal furtherto activate a behavioral response. Although thisactivity was extremely widespread, the functionalrelevance of these global activity patterns remainsto be determined. As was shown in a previousstudy of large-scale activity during the sametask in the mouse neocortex (23), it is likely thatonly a subset of brain regions are necessary fortask performance, at least in the simplest task

Allen et al., Science 364, eaav3932 (2019) 19 April 2019 7 of 9

Thirsty Sated Stim Washout

Positively correlated

Thirsty Sated Stim Washout

Negatively correlated

Lick

s/s

3

A

G

SFO

StimThirsty Washout

On3s Off

Sated

Session start Session end

B

FEC D

Thirsty Sated Stim Washout

Ave

rage

lick

s pe

r G

o tr

ial

***

Thirsty Stim

Thirsty

Sated

Stim

Washout

Go

tria

l num

ber Thirsty

Sated

Stim

Washout

Thirsty Sated Stim Washout

I

Neu

ron

Thirsty Sated Stim Washout J

Thirsty Sated Stim Washout

OLF

PIR

SSpSSs

PA

DORpm

DORsmPVRPVZ

LZ

STRd

STRvLS

PL

ILA

ACAAI

RSPPALdPALmPALv

SC

Trial number

H

Trials

Fig. 5. Causal manipulation of hypothalamic thirst-sensitive neuronsrecovers brainwide thirst state and global task-related neuronaldynamics. (A) Expression of ChR2-eYFP in Nos1+ SFO neurons. Scale bar,100 mm. (B) Experimental timeline for optogenetic stimulation sessions. Micefirst perform as in previous figures (“Thirsty” and “Sated”), then areoptogenetically stimulated for a block of trials (“Stim”), then stimulation isturned off for a block of trials (“Washout”). (C) Example licking per trial for allGo trials within a session, with different trial blocks labeled. (D) Averagelicking within each trial block. Data are means ± SEM; N = 36 sessions.(E) Average number of licks per Go trial within each trial block. ***P < 0.001(Wilcoxon signed-rank test, Bonferroni correction); n.s., not significant. TheHit rate was also not significantly different between Thirsty and Stim

conditions (97 ± 1% versus 94 ± 2%, P = 0.67, Wilcoxon signed-rank test).(F) Average latency to first lick on Hit trials in Thirsty and Stim trial blocks(Wilcoxon signed-rank test). (G) Baseline-subtracted, Z-scored trial-averagedfiring rates for all task-modulated neurons (N = 5323 of 10,288 total)from all 36 sessions, on Go trials in each trial block. Neurons are sorted bybrain region, then by time to peak firing rate on thirsty Go trials. (H) Baselinefiring rates across session of example state-modulated neurons in Thirsty,Sated, Stim, and Washout blocks. (I) Z-scored baseline firing rate ofstate-modulated neurons, sorted by each neuron’s correlation with satietystate. The last 20 trials from each block are shown. (J) Average Z-scoredbaseline firing rate from top 25% positively and negatively correlatedstate-modulated neurons. Data are means ± 95% confidence interval.

RESEARCH | RESEARCH ARTICLEon A

pril 23, 2019

http://science.sciencemag.org/

Dow

nloaded from

conditions. Together, these results illustrate thatstudies of how the brain produces behavior re-quire an understanding of the global context inwhich neural computations are occurring.Individual neurons fell into multiple clusters

defined by activity dynamics, with each dynam-ically defined cluster distributed across the en-tire brain and exhibiting task- and state-relatedspecific activity patterns that were highly similaracross multiple regions. Each individual neuronprimarily encoded one aspect of the task, andevery brain region contained a diverse mixtureof neurons encoding different aspects of the task.We further identified a brainwide mode of neuralpopulation activity that existed before odor onset,persisted throughout each trial, correlated withthe behaviorally measured satiety state of themouse, and appeared to determine the patternof odor-evoked neural dynamics and behavior.Specific optogenetic stimulation of thirst-sensingneurons in sated mice was sufficient to recoveractivity along this thirsty state mode in a rapidand reversible manner and to restore the sub-

sequent sensory- and behavior-related activitythroughout the brain. Brain regions that did notfully represent the maximal level of thirst rep-resentation upon optogenetic stimulation couldadditionally be recipients of contextual ormemoryinformation onwater consumption, visceral sig-nals that normally cause animals to terminatedrinking upon satiation [such as those relayedby oxytocin receptor–expressing neurons in theparabrachial nucleus (42)], or other slowly chang-ing variables such as the passage of time (43).The activity of glutamatergic hypothalamic

thirst-related neurons establishes a persistent,aversive motivational drive that animals can re-duce in real time by consuming water (19, 20).The brainwide activity mode we identify appearsto represent the direct effects of this thirst neu-ron activity that is broadcast to downstreamcircuits. This broadcast may proceed throughmultiple levels of fast synaptic transmission and/or via neuromodulators, which feature promi-nently in the regulation of other brainwidearousal states (44). Through imposing this brain-

wide state, thirst establishes persistent activitythat exists prior to the animal sensing a water-predicting cue and can then modulate local cir-cuitry in parallel throughout the brain to gate thesubsequent transformation of sensory input intomotor output in a state-dependent manner. Al-though initial sensory regions continue to respondto stimuli with high fidelity regardless of satietystate, the gating conferred by satiety greatlydiminishes the propagation of sensory informa-tion beyond sensory regions, preventing the ac-tivation of motor circuits.Neural population activity can be well de-

scribed in terms of dynamics that evolve overtime and are subject to internal dynamical rules,initial state settings, and external perturbations(28, 45). Our results suggest that the aversivethirst motivational state sets a landscape of “ini-tial conditions,” spread acrossmany brain regions,that determine subsequent sensory stimuli–evoked brainwide neural dynamics. When thebrain is in the thirsty state, a transient perturba-tion (such as a sensory stimulus that predicts avail-ability of water) elicits activity throughout thebrain that follows internal dynamical rules toproduce a trajectory leading towater-consummatorybehavior. This reward consumption in turnmovesthe animal along an adaptive path toward a newinitial position of less negative valence as thethirst drive is reduced. When the animal is sated,an “energetic barrier” throughout the brain pre-vents the same stimulus from pushing the dynam-ics down the same trajectory that leads to behavior.It remains to be seen whether other physiologicalsurvival-need states modulate different modesof activity, or whether the thirst-related modewe observe represents a more general form ofarousal across multiple types of motivation. Ineither case, specifying initial conditions in thismanner may be a general mechanism for de-fining and regulating motivated behavior.

Methods summary

Mice were implanted with a headbar, allowedto recover, and then water-restricted and trainedon an olfactory Go/No-Go task. Recordings wereperformed using Neuropixels microelectrodearrays (24) inserted into one location per ses-sion. Multiple recordings were made from thesamemice acrossmultiple days, in different loca-tions each day, to allow for unambiguous assign-ment of recording sessions to electrode tracks.During recordings, mice would perform severalhundred trials until they became sated, followedby several dozenmore trialswhile sated. Electrodeswere coated with fixable DiI to allow for trackreconstruction. In a subset of experiments, opto-genetic stimulationwas applied toNos1+ neuronsexpressing ChR2 in the subfornical organ. Brainswere cleared and imaged on a lightsheet micro-scope after killing, then registered to a commonatlas. Putative neurons were identified usingKilosort2 (25). Neurons were assigned to brainarea according to the location of their maximalchannel on the electrode array, with locationtransformed into a common anatomical referencespace. Firing rates were counted in 10-ms bins.

Allen et al., Science 364, eaav3932 (2019) 19 April 2019 8 of 9

−500

150

−50

0

150

0

100

0

100

Thi

rsty

Sat

ed

Stim

Was

hout

P-motMBmot

SCMOsMOp

PALvPALmPALdRSP

AIACAORB

ILAPLLS

STRvSTRdHPCMEZ

LZPVZPVR

DORsmDORpm

PAsAMYBMABLASSsSSp

PIROLF

Thi

rsty

Sat

ed

Stim

Was

hout

Sated

Thirsty

0.5

1

A Stimulus decodingState prediction

AU

C

D E

C0 1 2

Time (s)

Pro

ject

ion

(AU

)

0 1 2

500

00 1 2

0

600BeginningMiddleEndSatedStimWashout

Pro

ject

ion

(AU

)

State Cue Behavior

Thirsty GoThirsty No-Go

Sated GoSated No-Go

Stim GoStim No-Go

Washout GoWashout No-Go

Pro

ject

ion

(AU

)

0 1 2Time (s)

0 1 2 0 1 2

ThirstySated

StimWashout

50

0

-50

20

0

-40

0 1 2

−100

0

200

0 1 2

−100

0

500

0 1 2

0

600

Time (s)

−100

0

200

B

Fig. 6. Optogenetic thirst induction recovers population activity dynamics and decoding inspecific regions. (A) Projections of average activity on Go and No-Go trials from each trialblock onto State, Cue, and Behavior axes (defined using only Thirsty and Sated trials).(B) Projections of average activity on Go trials from trials at different points between Thirstyand Sated trials, and during Stim and Washout blocks, onto different axes. (C) Top: Projections ofsingle trials of simultaneously recorded neurons from ORB onto State, Cue, and Behavior axes.Bottom: Averages of trials in top panels. Data are means ± SEM. (D) Per-region averagepredictions of satiety state (Thirsty versus Sated) in each trial block, using decoder trained onbaseline epoch in Thirsty and Sated trials. (E) Per-region average decoding accuracy of trialtype (Go versus No-Go) in each trial block, using decoder trained on odor epoch in Thirsty and Satedtrials, using last 40 trials from each block. AUC, area under curve; 0.5 = chance performance.Predictions in Thirsty and Sated blocks in (D) and (E) are on held-out test data. Number ofrecordings per region: OLF, 12; PIR, 6; SSp, 1; SSs, 3; BLA, 2; BMA, 3; sAMY, 3; PA, 2; DORpm, 7;DORsm, 4; PVR, 1; PVZ, 1; LZ, 5; MEZ, 3; HPC, 4; STRd, 13; STRv, 3; LS, 2; PL, 2; ILA, 3; ORB, 2; ACA,1; AI, 2; RSP, 1; PALd, 2; PALm, 2; PALv, 1; MOp, 6; MOs, 1; SC, 1; MBmot, 6; P-mot, 2.

RESEARCH | RESEARCH ARTICLEon A

pril 23, 2019

http://science.sciencemag.org/

Dow

nloaded from

REFERENCES AND NOTES

1. C. Hull, Principles of Behavior: An Introduction to BehaviorTheory (Appleton-Century, 1943).

2. R. C. Bolles, Theory of Motivation (Harper & Row, 1967).3. F. Toates, Motivational Systems (Cambridge Univ. Press,

1986).4. K. C. Berridge, Motivation concepts in behavioral neuroscience.

Physiol. Behav. 81, 179–209 (2004). doi: 10.1016/j.physbeh.2004.02.004; pmid: 15159167

5. S. M. Sternson, A.-K. Eiselt, Three Pillars for the NeuralControl of Appetite. Annu. Rev. Physiol. 79, 401–423 (2017).doi: 10.1146/annurev-physiol-021115-104948; pmid: 27912679

6. L. W. Swanson, Cerebral hemisphere regulation of motivatedbehavior. Brain Res. 886, 113–164 (2000). doi: 10.1016/S0006-8993(00)02905-X; pmid: 11119693

7. S. M. Sternson, Hypothalamic survival circuits: Blueprints forpurposive behaviors. Neuron 77, 810–824 (2013). doi: 10.1016/j.neuron.2013.02.018; pmid: 23473313

8. M. Fukuda, T. Ono, H. Nishino, K. Nakamura, Neuronalresponses in monkey lateral hypothalamus during operantfeeding behavior. Brain Res. Bull. 17, 879–883 (1986).doi: 10.1016/0361-9230(86)90102-4; pmid: 3801942

9. M. J. Burton, E. T. Rolls, F. Mora, Effects of hunger on theresponses of neurons in the lateral hypothalamus to the sightand taste of food. Exp. Neurol. 51, 668–677 (1976).doi: 10.1016/0014-4886(76)90189-8; pmid: 819286

10. E. C. O’Connor et al., Accumbal D1R Neurons Projecting toLateral Hypothalamus Authorize Feeding. Neuron 88, 553–564(2015). doi: 10.1016/j.neuron.2015.09.038; pmid: 26593092

11. J. Pager, Ascending olfactory information and centrifugalinfluxes contributing to a nutritional modulation of the ratmitral cell responses. Brain Res. 140, 251–269 (1978).doi: 10.1016/0006-8993(78)90459-6; pmid: 626891

12. H. D. Critchley, E. T. Rolls, Hunger and satiety modify theresponses of olfactory and visual neurons in the primateorbitofrontal cortex. J. Neurophysiol. 75, 1673–1686 (1996).doi: 10.1152/jn.1996.75.4.1673; pmid: 8727405

13. C. R. Burgess et al., Hunger-Dependent Enhancement of FoodCue Responses in Mouse Postrhinal Cortex and LateralAmygdala. Neuron 91, 1154–1169 (2016). doi: 10.1016/j.neuron.2016.07.032; pmid: 27523426

14. Y. Livneh et al., Homeostatic circuits selectively gate food cueresponses in insular cortex. Nature 546, 611–616 (2017).doi: 10.1038/nature22375; pmid: 28614299

15. Y. Nakano et al., Neuronal activity in the medial orbitofrontalcortex of the behaving monkey: Modulation by glucose andsatiety. Brain Res. Bull. 12, 381–385 (1984). doi: 10.1016/0361-9230(84)90109-6; pmid: 6733545

16. I. E. de Araujo et al., Neural ensemble coding of satiety states.Neuron 51, 483–494 (2006). doi: 10.1016/j.neuron.2006.07.009; pmid: 16908413

17. C. Gizowski, C. W. Bourque, The neural basis of homeostaticand anticipatory thirst. Nat. Rev. Nephrol. 14, 11–25 (2018).doi: 10.1038/nrneph.2017.149; pmid: 29129925

18. J. N. Betley et al., Neurons for hunger and thirst transmit anegative-valence teaching signal. Nature 521, 180–185 (2015).doi: 10.1038/nature14416; pmid: 25915020

19. W. E. Allen et al., Thirst-associated preoptic neurons encode anaversive motivational drive. Science 357, 1149–1155 (2017).doi: 10.1126/science.aan6747; pmid: 28912243

20. D. E. Leib et al., The Forebrain Thirst Circuit Drives Drinkingthrough Negative Reinforcement. Neuron 96, 1272–1281.e4(2017). doi: 10.1016/j.neuron.2017.11.041; pmid: 29268095

21. V. Augustine et al., Hierarchical neural architecture underlyingthirst regulation. Nature 555, 204–209 (2018). doi: 10.1038/nature25488; pmid: 29489747

22. T. Komiyama et al., Learning-related fine-scale specificityimaged in motor cortex circuits of behaving mice. Nature 464,1182–1186 (2010). doi: 10.1038/nature08897; pmid: 20376005

23. W. E. Allen et al., Global Representations of Goal-DirectedBehavior in Distinct Cell Types of Mouse Neocortex. Neuron94, 891–907.e6 (2017). doi: 10.1016/j.neuron.2017.04.017;pmid: 28521139

24. J. J. Jun et al., Fully integrated silicon probes for high-densityrecording of neural activity. Nature 551, 232–236 (2017).doi: 10.1038/nature24636; pmid: 29120427

25. C. Stringer et al., Spontaneous behaviors drivemultidimensional, brain-wide population activity. bioRxiv[preprint]. 22 April 2018.doi: 10.1038/nn1233; pmid: 15114356

26. G. Buzsáki, Large-scale recording of neuronal ensembles.Nat. Neurosci. 7, 446–451 (2004). doi: 10.1038/nn1233;pmid: 15114356

27. M. Rabinovich, R. Huerta, G. Laurent, Transient dynamics forneural processing. Science 321, 48–50 (2008). doi: 10.1126/science.1155564; pmid: 18599763

28. K. V. Shenoy, M. Sahani, M. M. Churchland, Cortical control ofarm movements: A dynamical systems perspective.Annu. Rev. Neurosci. 36, 337–359 (2013). doi: 10.1146/annurev-neuro-062111-150509; pmid: 23725001

29. M. Stopfer, V. Jayaraman, G. Laurent, Intensity versus identitycoding in an olfactory system. Neuron 39, 991–1004 (2003).doi: 10.1016/j.neuron.2003.08.011; pmid: 12971898

30. O. Mazor, G. Laurent, Transient dynamics versus fixed points inodor representations by locust antennal lobe projectionneurons. Neuron 48, 661–673 (2005). doi: 10.1016/j.neuron.2005.09.032; pmid: 16301181

31. M. M. Churchland et al., Neural population dynamics duringreaching. Nature 487, 51–56 (2012). doi: 10.1038/nature11129;pmid: 22722855

32. V. Mante, D. Sussillo, K. V. Shenoy, W. T. Newsome,Context-dependent computation by recurrent dynamics inprefrontal cortex. Nature 503, 78–84 (2013). doi: 10.1038/nature12742; pmid: 24201281

33. M. M. Churchland, B. M. Yu, M. Sahani, K. V. Shenoy,Techniques for extracting single-trial activity patterns fromlarge-scale neural recordings. Curr. Opin. Neurobiol. 17, 609–618(2007). doi: 10.1016/j.conb.2007.11.001; pmid: 18093826

34. N. Li, K. Daie, K. Svoboda, S. Druckmann, Robust neuronaldynamics in premotor cortex during motor planning. Nature 532,459–464 (2016). doi: 10.1038/nature17643; pmid: 27074502

35. E. S. Boyden, F. Zhang, E. Bamberg, G. Nagel, K. Deisseroth,Millisecond-timescale, genetically targeted optical control ofneural activity. Nat. Neurosci. 8, 1263–1268 (2005).doi: 10.1038/nn1525; pmid: 16116447

36. Y. Oka, M. Ye, C. S. Zuker, Thirst driving and suppressingsignals encoded by distinct neural populations in thebrain. Nature 520, 349–352 (2015). doi: 10.1038/nature14108;pmid: 25624099

37. C. A. Zimmerman et al., Thirst neurons anticipate thehomeostatic consequences of eating and drinking. Nature 537,680–684 (2016). pmid: 27487211

38. L. Pinto, Y. Dan, Cell-Type-Specific Activity in Prefrontal Cortexduring Goal-Directed Behavior. Neuron 87, 437–450 (2015).doi: 10.1016/j.neuron.2015.06.021; pmid: 26143660

39. S. Kato et al., Global brain dynamics embed the motorcommand sequence of Caenorhabditis elegans. Cell 163,

656–669 (2015). doi: 10.1016/j.cell.2015.09.034;pmid: 26478179

40. M. B. Ahrens et al., Brain-wide neuronal dynamics duringmotor adaptation in zebrafish. Nature 485, 471–477 (2012).doi: 10.1038/nature11057; pmid: 22622571

41. N. A. Steinmetz, P. Zatka-haas, M. Carandini, K. D. Harris,Distributed correlates of visually-guided behavior across themouse brain. bioRxiv [preprint]. 20 November 2018.doi: 10.1038/s41593-017-0014-z; pmid: 29184212

42. P. J. Ryan, S. I. Ross, C. A. Campos, V. A. Derkach,R. D. Palmiter, Oxytocin-receptor-expressing neurons in theparabrachial nucleus regulate fluid intake. Nat. Neurosci.20, 1722–1733 (2017). doi: 10.1038/s41593-017-0014-z;pmid: 29184212

43. A. Tsao et al., Integrating time from experience in the lateralentorhinal cortex. Nature 561, 57–62 (2018). doi: 10.1038/s41586-018-0459-6; pmid: 30158699

44. S. H. Lee, Y. Dan, Neuromodulation of brain states. Neuron76, 209–222 (2012). doi: 10.1016/j.neuron.2012.09.012;pmid: 23040816

45. D. Sussillo, Neural circuits as computational dynamicalsystems. Curr. Opin. Neurobiol. 25, 156–163 (2014).doi: 10.1016/j.conb.2014.01.008; pmid: 24509098

ACKNOWLEDGMENTS

We thank M. Lovett-Barron, C. Kim, T. Machado, D. O’Shea,I. Kauvar, and E. Richman for comments on the manuscript;N. Steinmetz (UCL/University of Washington) and J. Siegle(Allen Institute for Brain Science) for help with using Neuropixels;T. Harris (Janelia Research Campus) for providing Neuropixels;C. Ramakrishnan for assistance with viruses; and D. Sussillo,K. Shenoy, and W. Newsome for useful discussions.Funding: Supported by a Fannie and John Hertz FoundationFellowship (W.E.A.); NINDS (L.L.); the NSF NeuroNex program andHHMI (L.L. and K.D.); and NIMH, NIDA, the Defense AdvancedResearch Projects Agency Neuro-FAST program, the NOMISFoundation, the Wiegers Family Fund, the Nancy and JamesGrosfeld Foundation, the H. L. Snyder Medical Foundation, theSamuel and Betsy Reeves Fund, the Gatsby Foundation,the AE Foundation, and the Fresenius Foundation (K.D.).Author contributions: W.E.A., L.L., and K.D. designed theproject; W.E.A. performed all electrophysiological recordings, spikesorting, and analysis; M.Z.C. performed anatomical tracingexperiments and developed software for decoding, automatingspike sorting, and track reconstruction; N.P. and R.H.T. trainedmice and performed perfusions and genotyping; M.P. contributedKilosort2; and K.D. supervised all aspects of the project.Competing interests: Authors declare no competing interests.Data and materials availability: Data are available athttp://clarityresourcecenter.org.

SUPPLEMENTARY MATERIALS

science.sciencemag.org/content/364/6437/eaav3932/suppl/DC1Materials and MethodsFigs. S1 to S12Table S1Movie S1References (46–55)

12 September 2018; accepted 14 February 2019Published online 4 April 201910.1126/science.aav3932

Allen et al., Science 364, eaav3932 (2019) 19 April 2019 9 of 9

RESEARCH | RESEARCH ARTICLEon A

pril 23, 2019

http://science.sciencemag.org/

Dow

nloaded from

dynamicsThirst regulates motivated behavior through modulation of brainwide neural population

William E. Allen, Michael Z. Chen, Nandini Pichamoorthy, Rebecca H. Tien, Marius Pachitariu, Liqun Luo and Karl Deisseroth

originally published online April 4, 2019DOI: 10.1126/science.aav3932 (6437), eaav3932.364Science 

, this issue p. eaav3932, p. eaav8736, p. eaav7893; see also p. 236Sciencestimuli in the primary visual area is mainly related to arousal and reflects the encoding of latent behavioral states.both visual information and motor activity related to facial movements. The variability of neuronal responses to visual

analyzed spontaneous neuronal firing, finding that neurons in the primary visual cortex encodedet al.areas. Stringer during exploratory and nonexploratory behaviors, possibly reflecting different levels of anxiety experienced in theseamygdala neurons in relation to behavior during different tasks. Two ensembles of neurons showed orthogonal activity

investigated the activity of mouse basalet al.activity across the brain that previously singled thirst. Gründemann response. Optogenetic stimulation of thirst-sensing neurons in one area of the brain reinstated drinking and neuronalIndividual neurons encoded task-specific responses, but every brain area contained neurons with different types of

found that in thirsty mice, there is widespread neural activity related to stimuli that elicit licking and drinking.et al.Allen Hart).papers in this issue present brain-scale studies of neuronal activity and dynamics (see the Perspective by Huk and

How is it that groups of neurons dispersed through the brain interact to generate complex behaviors? ThreeNeuron activity across the brain

ARTICLE TOOLS http://science.sciencemag.org/content/364/6437/eaav3932

MATERIALSSUPPLEMENTARY http://science.sciencemag.org/content/suppl/2019/04/03/science.aav3932.DC1

CONTENTRELATED

http://science.sciencemag.org/content/sci/364/6437/236.fullhttp://stm.sciencemag.org/content/scitransmed/9/401/eaah6733.full

REFERENCES

http://science.sciencemag.org/content/364/6437/eaav3932#BIBLThis article cites 50 articles, 2 of which you can access for free

PERMISSIONS http://www.sciencemag.org/help/reprints-and-permissions

Terms of ServiceUse of this article is subject to the

is a registered trademark of AAAS.Sciencelicensee American Association for the Advancement of Science. No claim to original U.S. Government Works. The title Science, 1200 New York Avenue NW, Washington, DC 20005. 2017 © The Authors, some rights reserved; exclusive

(print ISSN 0036-8075; online ISSN 1095-9203) is published by the American Association for the Advancement ofScience

on April 23, 2019

http://science.sciencem

ag.org/D

ownloaded from