Ventral striatal astrocytes contribute to reinforcement learning

Ventral striatal astrocytes contribute to reinforcement learning | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results Ventral striatal astrocytes contribute to reinforcement learning Julia Pai , Fatih Sogukpinar , Kei Ogasawara , Garrett J Smith , Francesca R Fiocchi , Yanchao Dai , Yifan Wu , Michael J Frank , ShiNung Ching , Federica Lucantonio , Thomas Papouin , Marco Pignatelli , Naoki Hiratani , Ilya E Monosov doi: https://doi.org/10.1101/2025.10.20.683205 Julia Pai 1 Department of Neuroscience, Washington University in St. Louis School of Medicine Find this author on Google Scholar Find this author on PubMed Search for this author on this site Fatih Sogukpinar 2 Department of Electrical and Systems Engineering, Washington University in St. Louis Find this author on Google Scholar Find this author on PubMed Search for this author on this site Kei Ogasawara 6 Departments of Neuroscience, Biomedical Engineering, and Psychiatry, Johns Hopkins University , Baltimore, MD, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Garrett J Smith 6 Departments of Neuroscience, Biomedical Engineering, and Psychiatry, Johns Hopkins University , Baltimore, MD, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Francesca R Fiocchi 3 Department of Psychiatry, Washington University in St. Louis School of Medicine Find this author on Google Scholar Find this author on PubMed Search for this author on this site Yanchao Dai 1 Department of Neuroscience, Washington University in St. Louis School of Medicine Find this author on Google Scholar Find this author on PubMed Search for this author on this site Yifan Wu 1 Department of Neuroscience, Washington University in St. Louis School of Medicine Find this author on Google Scholar Find this author on PubMed Search for this author on this site Michael J Frank 4 Carney Institute for Brain Science, Department of Cognitive and Psychological Sciences, Brown University Find this author on Google Scholar Find this author on PubMed Search for this author on this site ShiNung Ching 2 Department of Electrical and Systems Engineering, Washington University in St. Louis Find this author on Google Scholar Find this author on PubMed Search for this author on this site Federica Lucantonio 3 Department of Psychiatry, Washington University in St. Louis School of Medicine Find this author on Google Scholar Find this author on PubMed Search for this author on this site Thomas Papouin 1 Department of Neuroscience, Washington University in St. Louis School of Medicine Find this author on Google Scholar Find this author on PubMed Search for this author on this site Marco Pignatelli 3 Department of Psychiatry, Washington University in St. Louis School of Medicine 5 Taylor Family Institute for Innovative Psychiatric Research, Washington University in St. Louis School of Medicine, St. Louis, MO, USA. University School of Medicine Find this author on Google Scholar Find this author on PubMed Search for this author on this site Naoki Hiratani 1 Department of Neuroscience, Washington University in St. Louis School of Medicine Find this author on Google Scholar Find this author on PubMed Search for this author on this site Ilya E Monosov 1 Department of Neuroscience, Washington University in St. Louis School of Medicine 2 Department of Electrical and Systems Engineering, Washington University in St. Louis 6 Departments of Neuroscience, Biomedical Engineering, and Psychiatry, Johns Hopkins University , Baltimore, MD, USA 7 Zanvyl Krieger Mind/Brain Institute, Johns Hopkins University , Baltimore, MD, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site For correspondence: ilya.monosov{at}gmail.com Abstract Full Text Info/History Metrics Supplementary material Preview PDF Abstract Astrocytes influence synaptic plasticity and neuronal function through astrocytic calcium dynamics (ACD). However, astrocytes’ contribution to cognitive operations like reinforcement learning (RL) remains unclear. To examine this, we trained mice on a RL dependent probabilistic decision-making task. We attenuated ACD across distinct striatal regions, finding that ACD attenuation specifically in ventral striatum (VS) increased decision noisiness and impaired reward-guided choice. This effect was largely due to a reduction in “win-stay” behavior. Using in-vivo calcium imaging, we found that VS ACD correlated with reward prediction errors (RPEs). Furthermore, these trial-by-trial ACD fluctuations predicted trial-by-trial choice variability. Finally, ex-vivo slice electrophysiology and computational modeling revealed two mechanisms through which astrocytes could regulate behavioral variability: by regulating presynaptic excitatory-inhibitory balance, and by sharing RPE signals across populations of striatal neurons. Together, these results suggest that VS astrocytes contribute to cortico-striatal functions to mediate decision noisiness. Introduction Astrocytes in the central nervous system have well established roles in controlling the neuronal environment. These include the dynamic control of ionic concentrations, reuptake and recycling of transmitter clearance, and mediation of excitation-inhibition balance (e.g., via uptake of glutamate, metabolite regulation, sensing of immune and endocrine signals, and the control of blood flow vascular coupling ( Haydon & Carmignoto, 2006 ; Petzold & Murthy, 2011 )). There is now mounting evidence that, through these multiple functional knobs, astrocytes contribute to synaptic function and ongoing neural activity ( Adamsky et al., 2018 ; Araque et al., 1999 ; A. B. Chen et al., 2024 ; Corkrum et al., 2020 ; De Pittà et al., 2016 ; Guttenplan et al., 2025 ; Henneberger et al., 2010 ; Lefton et al., 2025 ; Murphy-Royal et al., 2023 ; Oliveira et al., 2015 ; Xin et al., 2024 ). However, it is still unclear whether astrocytes contribute to the control of algorithmic computations, such as reinforcement learning. To shed light on this matter, the field needs to generate testable models of astrocytic functions. Astrocytic modulation of neuronal activity and synaptic functions is thought to be dependent on astrocytic calcium dynamics (ACD) and can occur on a spectrum of spatiotemporal scales: relatively fast and locally (sub-second, synapse-level) and relatively slowly and more globally (hours-days, population-level) ( Adamsky et al., 2018 ; Benoit et al., 2025 ; Di Castro et al., 2011 ; Fujii et al., 2017 ; Panatier et al., 2011 ; Scemes & Giuame, 2006 ; Stobart et al., 2018 ). ACD are thought to reflect ongoing circuit activity, including synaptic activity, neuronal firing and neuromodulator release, such dopamine and norepinephrine ( Allen, 2014 ; Bai et al., 2024 ; Bekar et al., 2008 ; Charles et al., 1991 ; Corkrum et al., 2020 ; Cornell-Bell et al., 1990 ; Di Castro et al., 2011 ; Khakh & McCarthy, 2015 ; Lefton et al., 2025 ; Ma et al., 2016 ; Mu et al., 2019 ; Panatier et al., 2011 ; Schipke et al., 2008 ; Sofroniew & Vinters, 2010 ). In addition, astrocyte networks possess unique properties that distinguish them from neurons. For example, astrocytes are thought to tile the brain in non-overlapping domains ( Bushong et al., 2002 ; Halassa et al., 2007 ; Ogata & Kosaka, 2002 ), with each astrocyte closely interacting with hundreds of thousands of synapses ( Oberheim et al., 2009 ), and they are extensively gap-junction coupled into functional networks ( Cooper et al., 2025 ; Dermietzel et al., 1991 ). These abilities might confer the capacity to regulate neuronal computations over a wider spatial and temporal scale in response to local computations. That is, astrocytes could uniquely contribute to neuronal circuits by helping to address a key function required for cognitive operations: adaptively regulating population-level neural variability within local circuits to produce stable representations to guide behavior. This possibility that astrocytes participate in averaging and sharing information across nearby neurons has strong analogs in artificial intelligence, where such properties are crucial for multiple forms of learning algorithms ( Fukushima, 1980 ; Kipf & Welling, 2017 ; Krizhevsky et al., 2012 ; LeCun et al., 1989 , p. 19) and neuromorphic architectures ( Becker et al., 2022 ; Gong et al., 2024 ; Irizarry-Valle & Parker, 2015 ; Kozachkov et al., 2023 ; Nazari et al., 2015 ; G. Tang et al., 2019 ); in such instantiations, input sharing is not an active non-linear component of algorithmic implementation per se, but directly contributes to it. We sought to assess whether and how ACD contributes to reinforcement learning (RL). RL is thought to be dependent on the striatum, with distinct striatal subregions thought to play causal roles in RL, particularly in the learning of values of actions or policies ( Burton et al., 2015 ; Cox & Witten, 2019 ; Ito & Doya, 2011 ; H. Kim et al., 2009 ). We attenuated ACD in the dorsal medial, dorsal lateral, and ventral striatal regions (DMS, DLS, and VS, respectively) of mice performing RL in a probabilistic reversal learning task. This task allowed us to model the animals’ behavior and therefore assess the link between ACD and RL computations. We found that attenuation of ACD in DMS and DLS had relatively little impact on key reinforcement learning parameters, such as on learning rate and inverse temperature (decision noisiness). In contrast, VS ACD attenuation decreased inverse temperature (increasing decision noisiness) and impaired reward-guided choice, specifically by reducing win-stay behavior. Using fiber photometry to monitor ACD during RL revealed that VS ACD reflected model inferred reward prediction errors (RPEs), for many seconds after the outcome, until the start of the next trial. Moreover, trial-by-trial ACD fluctuations predicted trial-by-trial choice variability, particularly in win-stay behavior. Using ex vivo slice electrophysiology, we found that attenuation of VS ACD led to changes in presynaptic excitatory-inhibitory balance. Further in-silico circuit modelling suggested that the behavioral effects could be due to yet an additional mechanism which is akin to astrocytes sharing dopaminergic RPE inputs across MSN populations. These findings highlight a role of ACD in RL and generate novel hypotheses for how astrocytes contribute to value-based decision-making and cognition. Results Attenuating astrocyte calcium dynamics (ACD) in ventral striatum impairs probabilistic decision-making in a reinforcement learning task To begin to assess the role of striatal ACD in reinforcement learning, we first trained mice to perform a non-stationary probabilistic reversal learning task commonly used to study reinforcement learning (RL) algorithms in animals and artificial systems ( Bertsekas & Tsitsiklis, 1996 ; Costa et al., 2015 ; Grossman et al., 2022 ; Ito & Doya, 2011 ; Sutton & Barto, 2018 ). This task allowed us to infer and quantify key RL computations, such as trial-by-trial reward prediction errors and parameters of the decision function. We could then assess how these computations changed after ACD attenuation. We first ran this task on an in-home cage behavioral device (FED3, ( Matikainen-Ankney et al., 2021 )). This allowed us to train large numbers of mice on this task (Methods), and to scale the experiment in order to test the role of ACD in different striatal sub regions known to contribute differentially to RL ( Burton et al., 2015 ; Cox & Witten, 2019 ; Ito & Doya, 2011 ; H. Kim et al., 2009 ; H. Tang et al., 2022 ). At the start of each trial, mice received a visual start cue that informed them that they could make a choice by performing a left or right nose poke. These two ports were associated with two varying reward probabilities: either the left one was 80% reward, and the right one was 20% reward, or the left one was 20% reward, and the right one was 80% reward ( Figure 1A ). The trials were blocked so that the contingency (higher reward probability in left arm or right arm) was stable for 20-30 rewarded trials and then switched. Because these block switches were not explicitly cued to the mice, this task requires animals to maintain and update the value of choosing left or choosing right as a function of prior choice outcomes. Download figure Open in new tab Figure 1. Attenuating astrocyte calcium dynamics (ACD) across striatum subregions leads to performance deficits preferentially in VS. A. Trial diagram of the probabilistic decision-making task. Animals are trained to poke for rewards via an in-home cage operant behavior device, which lights up when trials are available. Animals initiate choices by making a nose poke into the left or right port. Ports have either 80% or 20% probabilities of being rewarded. If rewarded, a reward tone is presented, and a pellet is delivered. If un-rewarded, a different no-reward tone is presented. After each choice, the device enters a 5 second time out, before another trial becomes available. B. Behavior from one example session. Dots on the top and bottom indicate choices the animal made to the left and right and the dashed line shows the rolling average of left choices. C. Experimental timeline. D. Injection regions for VS, DMS, and DLS-injected cohorts. E. Histology of PMCA injection sites in VS, DMS, and DLS. Red shows AAV-gfaABC1D-PMCA-mCherry expression, and green shows the same tissue co-stained for S100β, an astrocyte marker. Yellow overlap shows co-expression of the PMCA with S100β in astrocytes. F. Top: heatmap of astrocyte calcium responses to bath application of norepinephrine (left) or dopamine (right) in control (top heatmap) or PMCA (bottom heatmap) slices. NE or DA application occurs at 100s. Each row represents one slice’s responses to one application; black lines on top indicate the time scale of DA or NE application. Bottom: average traces of control or PMCA ACD responses to NE or DA application. Shaded area is 95% confidence interval. G. Comparison of the peak responses to NE or DA application (maximum response between 100-300 s) between control and PMCA slices. PMCA slices show significantly lower peak amplitude of calcium responses compared to control slices (Wilcoxon rank sum). *= p<0.05, ** = p<0.01, ***=p<0.001. This probabilistic reversal learning task (also known as a ‘bandit’ task) allowed us to test how ACD contributes to learning from outcomes by providing a non-stationary reward structure that has been modeled extensively in computational neuroscience and RL ( Bertsekas & Tsitsiklis, 1996 ; Chakroun et al., 2023 ; Frank et al., 2004 ; Franklin & Frank, 2015 ; Grossman et al., 2022 ; Humphries et al., 2012 ; Sutton & Barto, 2018 ). Similar bandit tasks have been used to study how different neural systems contribute to learning and choice ( Costa et al., 2016 ; Dalton et al., 2014 ; Grossman et al., 2022 ). They have also been used as to assess the algorithmic mechanisms of value-based decision making ( Cohen, 2008 ; Daw et al., 2006 ; Metha et al., 2020 ), including how choice is affected by signed differences between expected and received reward (reward prediction error, RPE) ( Frank et al., 2004 ; Ineichen et al., 2012 ; Montague et al., 1996 ; Sutton & Barto, 2018 ). Behavior from one example session is shown in Figure 1B . This example session shows that the mouse can readily learn the task. Like the example mouse, we found that many mice learned the task, choosing the correct port most of the time just after 5 days of training (mean 67.4% pokes to better port after 5 days of training; SD = 7.6%; Supplemental Figure 1A, left). After the first 5 training sessions, mice sharply decreased the number of non-task period pokes, indicating that they acquired the structure of the task (Supplemental Figure 1A, right). After mice were trained on the task, we attenuated ACD across different subregions of striatum to probe its role in supporting RL and probabilistic decision-making. Rodent striatum is often subdivided into ventral striatum (VS; here we use it to refer to nucleus accumbens), dorsomedial striatum (DMS), and dorsolateral striatum (DLS). Each striatal subregion receives input from cortical, thalamic, and dopaminergic regions ( Burton et al., 2015 ; Haber, 2003 ; Humphries et al., 2012 ; Parent & Hazrati, 1995 ), with some overlap and some differences in anatomical projection patterns across subregions. To our knowledge, few studies have systematically assessed subregion-specific contributions to performance in a probabilistic reversal learning task like ours (but see ( Ito & Doya, 2011 ; Parker et al., 2016 )). However, the anatomical and functional literature suggest that these subregions may play differential roles in motivated behavior, RL, and decision making ( Burton et al., 2015 ; Haber, 2003 ; Ito & Doya, 2011 ; H. Tang et al., 2022 ). We therefore began to screen how ACD contributes to RL by attenuating ACD in VS, DMS, and DLS in task-trained mice and quantifying their behavior before and after ACD attenuation. We injected separate cohorts of mice bilaterally in VS (n = 19), DMS (n = 10), and DLS (n = 11) with AAV5-gfaABC1D-PMCA-mCherry ( Figure 1C , 1D) ( Yu et al., 2018 ). We also injected a cohort of mice with control virus (AAV5-gfaABC1D-tdTomato, or AAV5-GFAP104-mCherry) in VS (n=24). We allowed at least 2 weeks after virus injection for the viral construct to adequately express ( Figure 1C , ( Reimsnider et al., 2007 )). Viral injections resulted in astrocyte-specific expression of plasma membrane Ca2+ ATPase (PMCA), which constitutively extrudes cytosolic Ca2+ from the cell body and functionally attenuates astrocytic calcium transients ( Yu et al., 2018 ). As expected, colocalization with the astrocyte marker S100β but lack of colocalization with NeuN indicated that PMCA was expressed selectively in astrocytes and not in neurons ( Figure 1E ). We used ex vivo calcium imaging to validate that PMCA expression effectively attenuated ACD in ventral striatum in response to bath application of both norepinephrine (which strongly drives astrocyte calcium responses in other brain regions ( Lefton et al., 2025 ; Paukert et al., 2014 ; Reitman et al., 2023 ) and dopamine (a key neuromodulator for RL in the striatum that is also known to drive striatal astrocyte calcium ( Corkrum et al., 2020 ; Haber, 2003 )) ( Figure 1F ). Astrocytes in animals injected with control virus showed strong ACD in response to bath application of norepinephrine and dopamine ( Figure 1F and Supplemental Figure 2). On the other hand, astrocytes in animals injected with PMCA showed attenuated ACD, with significantly smaller peak Ca 2+ response amplitudes ( Figures 1F , 1G). Overall, ACD attenuation is consistent with prior studies in different brain areas, including other areas of the striatum ( Nimitvilai-Roberts et al., 2021 ; Yu et al., 2018 ). To begin to assess whether and how ACD contributes to RL, we compared decision performance (the percent of choices mice made to the better port) before versus after virus injection. Various prior manipulations across different subregions of striatum have been shown to have heterogeneous effects on decision-making behavior ( Beeler et al., 2010 ; Dalton et al., 2014 ; Kwak & Jung, 2019 ). We found that PMCA-induced ACD attenuation in VS, but not in DMS, DLS, or control-injected VS mice, led to a significant decrease in choice performance in the probabilistic reversal task ( Figure 2A ). Download figure Open in new tab Figure 2. Performance deficits in mice with attenuated VS ACD are correlated with decreases in win-stay behavior. A. Change in decision-making performance (quantified as % of choices to the port with better reward probability) pre- vs post-injection for VS-PMCA, DMS-PMCA, DLS-PMCA, and VS-control cohorts. Plotted values are across animal averages; and error bars are S.E.M. across animals. Significance stars within bars = within-animal comparison of pre- vs post-injection, Wilcoxon sign rank test. Significance stars across groups = across group comparison of change, Wilcoxon rank sum test. Bottom row shows pre- and post-injection values for each animal in each cohort. Thick black line is average across animals. Kruskal-Wallis across groups p = 0.086. B. Change in percent win-stay pokes across groups; same convention as A. Kruskal-Wallis across groups p = 0.016. C. Change in percent lose-switch pokes across groups; same convention as A. Kruskal-Wallis across groups p = 0.691. D. Performance aligned to block switches for each cohort (Methods; see ideal observer analyses). Traces were first averaged within animals in sessions before and after virus injection, then averaged across animals. Error bar shows S.E.M. across animals. VS-PMCA animals show a significant decrease in performance in a 25-trial window pre-injection (bottom bar on x-axis) after the injection compared to before. DMS-PMCA, DLS-PMCA, and VS-PMCA animals do not show a significant change in performance in the same time window. All cohorts do not show any pre-post differences in a 5-trial window following block switches. E. Post-win switching (left) and post-lose switching (right) behavior aligned to block switches for VS-PMCA mice. VS-PMCA mice show significantly increased post-win switching but not post-lose switching in a 25-trial window before block switches, consistent with the decrease in win-stay pokes but no change in lose-switch pokes (2B, 2C). Data is averaged over animals; error bars are S.E.M. over animals. F. Pre- vs post-injection change in performance and change in win-stay pokes were correlated across VS-PMCA cohort (Pearson’s correlation), indicating that decreases in performance were associated with decreases in win-stay strategy. *= p<0.05, ** = p<0.01, ***=p<0.001. To visualize and quantify performance around block switches, we aligned the time course of the probability of choosing the higher-valued side to the time of block switches inferred computationally ( Figure 2D ). We used a Bayesian ideal observer model ( Bartolo & Averbeck, 2020 ) to estimate when an agent would detect a switch, correcting for the fact that true switch points can be obscured by stochastic reward delivery (Methods). For instance, if the right port becomes the 80% reward side but the left port continues to deliver rewards by chance, it may take several trials before the change is detectable. The inference-based alignment accounts for such misleading trial sequences and allows for more accurate comparison of behavioral dynamics across blocks. This analysis revealed that, consistent with results averaged across all trials ( Figure 2A ), VS-PMCA but not DMS-PMCA, DLS-PMCA, or VS-control mice displayed a reduction in choice performance. All mice achieved a steady-state of choosing the higher valued option (VS-PMCA: 73.9%, DMS-PMCA: 76.4%, DLS-PMCA: 74.4%, VS-control: 74.5%; average % pokes to better port in 25 trial window preceding block switch). However, only the VS-injected mice showed a significant reduction in choosing the higher valued option (decrease of 3.7%; p = 0.008, measured in a 25-trial window pre-block switch; Figure 2D ). Following block switches, mice switched their choices towards the high valued option. None of the cohorts showed a significant difference in pre- vs post-injection performance in the trials immediately following a block switch (5 trial window following block switch; Figure 2D ), suggesting that the performance deficit in VS-PMCA mice was beyond the specific value-updating processes around block switches. We further validated this by comparing when animals and ideal observers detected block switches. Compared to ideal observers, animals in all cohorts were slower to detect block switches (mean = 3.0 trials, SEM = 0.06 trials across 29,898 blocks). There was no change in this value for any cohort before versus after virus injection (Supplemental Figure 3C), and there was no difference across cohorts (p = 0.92; Kruskal-Wallis test). Next, we assessed whether mice’s win-stay or lose-shift strategies changed due to ACD attenuation. Win-stay/lose-shift strategies are simple, outcome-dependent decision-making heuristics in which subjects tend to repeat a choice after a rewarded outcome (“win-stay”) and switch after a non-rewarded outcome (“lose-shift”). Win-stay/lose-shift strategies have been widely studied to understand the mechanisms of value updating and action selection in reinforcement learning paradigms, both in animals and humans ( Daw et al., 2006 ; Lee et al., 2012 ; Robbins, 1952 ). Because post-reward and post-non-reward strategy can arise from different algorithmic and circuit-level implementations, this analysis could provide deeper understanding of the relationship of ACD and RL performance. For win-stay strategy, we found that mice in the VS-PMCA cohort, but not other cohorts, exhibited a significant decrease in win-stay choices post-injection compared to pre-injection ( Figure 2B ). However, there was no significant difference in lose-switch behavior in any of the cohorts ( Figure 2C ). When we aligned the probability of making switch choices to block switches, we found similarly that mice in the VS-PMCA cohort exhibited significantly more switching behavior following rewarded outcomes and not after unrewarded outcomes (25 trials pre-block switch; Figure 2E ). This effect was not observed in DMS-PMCA, DLS-PMCA mice, or VS-control mice (pre-block switch; Supplemental Figure 3D). Furthermore, the change in VS-PMCA mice correct choice rate was strongly and significantly correlated with the change in their win-stay behavior on a mouse-by-mouse basis ( Figure 2F ). Our data thus far suggests that attenuating ACD in VS impairs performance by altering decision making strategies relative to prior rewards. This shift is consistent with the reduction in the better-poke rate in the pre-block switch period ( Fig. 2D ). This suggests that attenuating ACD in VS influences RL-related computation in the VS circuit. However, VS is known to be highly involved in motivation, appetitive behavior, and reward-seeking ( Haber & Knutson, 2010 ; Kelley, 2004 ). Before delving into changes in RL algorithms of decision-making, we examined the possibility that these decision-making deficits in the VS-PMCA cohort were due to motivational, hedonic, or motor effects brought on by PMCA ACD attenuation. We also aimed to control for surgical injection-related damage to the VS, which could also lead to similar effects. VS ACD attenuation changers were not due to general changes in motivation to initiate trials, motor or impulsivity measures (Supplemental Figures 3 -5 ). For example, VS-PMCA mice did not show a significant change in pokes made outside the task period (e.g., they were not simply poking more at the port outside of the task trials after injection; p = 0.81). In fact, on average, VS-PMCA group initiated more choices (trials) per session after the injection compared to before (Supplemental Figure 3A, mean increase of 21.7 trials per session; SEM = 4.3; p = 0.0007; Wilcoxon sign rank test). Consistent with this notion, VS-PMCA injection also did not impair mice’s motivation to initiate trials (Supplemental Figure 3A). Change in performance was not correlated with change in reward or trial number for VS-PMCA mice (Supplemental Figure 3B), and VS-PMCA group displayed no significant change in side-bias (Supplemental Figure 3A; p = 0.72). We further assessed mouse with several additional independent assays. First, we tested yet more mice injected in VS either with PMCA (n = 12) or control virus (AAV5-gfaABC1D-tdTomato; n = 12) on a progressive ratio task - a commonly used assay related to reward consumption and willingness to expend effort for reward ( Johnson et al., 2022 ; Richardson & Roberts, 1996 ; Stewart, 1975 ). We did not observe significant differences in any measures of willingness to work for reward or reward consumption (Supplemental Figure 4). Second, we tested naive cohorts of VS-PMCA (n = 9) and VS-control (n = 9) mice on a battery of behavioral tests commonly used to assay sensorimotor, cognitive, memory, motivational, and emotion-regulation related phenotypes (Supplemental Figure 5). The only test that revealed a significant difference between cohorts was for the open field test, where VS-PMCA mice spent more time in the center of the arena compared to VS-control mice (uncorrected p value = 0.01; note that with multiple comparison correction considering all assays this effect is not significant). Other measures in this same assay, namely rearing and total movement measures, were not significantly altered. Time in the center of the arena is commonly interpreted as a measure related to less anxiety-like pro-exploratory phenotype ( Seibenhener & Wooten, 2015 ). However, this phenotype may arise due to many sources (e.g., including changes in valuation and expectation that also give rise to win-stay deficits) and therefore must be interpreted with caution. In VS PMCA cohort, there were no significant differences in other assay-measures of sensorimotor coordination and strength, memory (Y-maze and novel object recognition), repetitive behavior (self-grooming), or hedonia (sucrose preference). Altogether, these data suggest that VS-PMCA mice do not show obvious decreases in motivation, memory, or sensorimotor ability. Finally, to double check that the performance deficits in VS-PMCA cohort were indeed due to ACD attenuation and not byproducts of the surgery or other potential environmental factors, we trained additional cohorts of animals on the bandit task (Supplemental Figure 6) and injected them with control virus (AAV5-GFAP104-mCherry or AAV5-gfaABC1D-tdTomato; n=18) or with saline in M1 (n=6). (Note that we pooled these mice in other analyses in the paper where they are collectively referred to as ‘VS-control’ ( Figure 2 , Supplemental Figures 3, 7; see Methods for details). None of these cohorts showed significant changes in correct choice rate, win-stay, or lose-switch behavior. These results suggest that the performance deficit we saw in the VS-PMCA cohort was not due to the injection surgery itself. We also ran a cohort of mice injected in VS with AAV5-gfaABC1D-iβark-mCherry ( Nagai et al., 2021 ) (n=12), which attenuates Gq-GPCR-mediated calcium signaling in astrocytes. We did not see any significant changes in performance, win-stay, or lose-switch behavior (Supplemental Figure 6; far right). This suggests that PMCA-induced behavioral changes were likely not due to Gq-GPCR mediated effects alone, or due to viral infection. In summary, decision-making deficits in VS-PMCA mice were not directly attributable to sensorimotor functions, to memory performance changes, to general motivation, or to the surgery itself. ACD attenuation in VS increases decision noisiness in reinforcement learning To better understand how decision-making algorithms are altered in mice with attenuated astrocytic calcium dynamics (ACD) in the ventral striatum (VS), we fit a reinforcement learning (RL) model to their behavioral data ( Figure 3A ; ( Bertsekas & Tsitsiklis, 1996 ; Sutton & Barto, 2018 )). Download figure Open in new tab Figure 3. RL modelling of behavioral noise following VS ACD attenuation A. Q-learning reinforcement learning (RL) model. Action values for left and right are stored in Q left and Q right for each trial. These values are passed into a softmax function, which outputs a left or right choice that is probabilistically affected by side bias, inverse temperature, and lapse rate parameters. A reward prediction error is calculated based on the choice decision and the current value of the action taken. This is then used to update the value of the action taken; the rate of value updating is weighted by the learning rate. B. Cartoon psychometric functions and how they are affected by shifts in inverse temperature, side bias, and lapse rate. Black indicates ‘baseline’, orange indicates function after shift in specific parameters. X-axis is difference in q-value between left and right options; y-axis is agent’s probability of choosing the left option, which is given as . Note that a change in slope of the psychometric curve corresponds to a change in the β parameter, where a shallower curve (lower slope) corresponds to a more random choice (lower β ), and a steeper curve (higher slope) corresponds to a more deterministic choice (higher β ). On the other hand, a shift of the psychometric curve left or right away from the indifference point (x = 0.0, y = 0.5) results from the side bias term, with an increase in the side bias term shifting the curve towards the right. Finally, a higher maximum value in the psychometric curve indicates a smaller lapse rate (𝜖). C. Behavioral data from one example session (conventions following Figure 1B) and RL model fit. The dashed blue line indicates the rolling average of the mouse’s real left decisions, and the green line indicates the model prediction of probability of left choice. R-squared for this session = 0.66. D. Psychometric curves for VS-PMCA injected mice. Black indicates pre-injection, orange indicates post-injection. Solid lines = mean across animals; error bar = S.E.M. across animals. E. Inverse temperature and learning rate parameters pre- and post-injection for VS-PMCA animals. Each gray line indicates values for one animal; thick black line shows average across animals. VS-PMCA mice show significantly decreased inverse temperature (Wilcoxon sign rank test), signifying noisier behavior, but no change in learning rate. F. Correlations between pre-vs post-injection change in performance and change in inverse temperature (left) and change in learning rate (right) for VS-PMCA mice. Changes in decision noise are associated with changes in performance. Light blue line is best fit line. The RL agent ( Figure 3A ) maintains estimates of the expected value of each action — also called action values, or Q-values —which are updated trial by trial based on experience. To make a choice, these Q-values are passed through a decision function (or a decision process ) that transforms them into action probabilities. The shape of this decision function is influenced by three key parameters: 𝑏 ! – a side bias parameter, capturing any baseline tendency to prefer one side (e.g., left over right); β – an inverse temperature parameter, controlling how sensitive choices are to value differences (i.e., decision “sharpness” or noise), and ε – a lapse rate term, accounting for occasional random choices, even when one option is clearly better. Figure 3B shows cartoon diagrams of how changes in these three parameters – inverse temperature, side bias, and lapse rate – could change the shape of the psychometric function. After each choice and outcome, the model computes a reward prediction error (RPE) — the difference between the received reward and the expected value of the chosen action. This RPE is used to update the value of the chosen action, scaled by the learning rate parameter α, which determines how much recent outcomes influence future behavior. We validated our model fitting procedure by showing that it was able to accurately recover the parameters of simulated datasets where the true parameters were known (parameter recovery analysis; n=1000 simulations, with each parameter drawn randomly from its range). We obtained near-perfect results, where the correlation coefficients between the known and fitted parameters were higher than 0.9, with p < 0.01 for all parameters (Supplemental Figure 8). The model was also able to recreate the behavioral results we saw, with decreased performance and win-stay behavior for the VS-PMCA cohort (Supplemental Figure 9). We further verified that our model was able to fit real mice’s behavioral data. An example session with model fits overlaid on the mouse’s behavior is shown in Figure 3C (r-square = 0.66). We began our model-based examination of VS-PMCA mice’s behavior by utilizing the model to generate psychometric curves for their behavior before and after virus injection. The shape of these curves, which here visualize agents’ choice patterns against the underlying inferred action values, can be used to infer the decision process used to transform internal estimates of value into actual choices ( Ballesta et al., 2020 ; Britten et al., 1996 ; Gold & Ding, 2013 ). VS-PMCA mice’s model-simulated psychometric functions revealed a post-injection shift in the function that resembled a decrease in inverse temperature (data in Figure 3D ; compare with Figure 3B ). Indeed, detailed analyses of the model parameters showed that inverse temperature significantly decreased after virus injection ( Figure 3E ), consistent with the flatter psychometric curves in post-injection mice. These data show that VS-PMCA mice’s decision-making became more ‘noisy’ or ‘random’ after virus injection. Note that psychometric functions or inverse temperature model parameters for DMS and DLS groups did not change after PMCA injection (Supplemental Figure 10). In sum, VS ACD attenuated mice are less likely to make deterministic choices based on underlying learned value estimates. We next assessed whether VS ACD attenuation led to specific changes in RL parameters. VS-PMCA injected animals did not show any change in learning rate pre-versus post-injection ( Figure 3E ). This was also the case for DLS and DMS injected cohorts (Supplemental Figure 11). For VS-PMCA mice, pre- vs post-injection change in performance (choice of the higher valued option) was strongly correlated with change in inverse temperature, but not learning rate, on an animal-by-animal basis ( Figure 3F ; the correlations among performance and inverse temperature and performance and learning rate also were significantly different; p<0.01; tested with a bootstrap test; 10,000 permutations). Model fits for all cohorts of animals are shown in Supplemental Figure 11. We did not observe significant changes in learning rate, inverse temperature, or lapse rate in the DMS-PMCA and DLS-PMCA cohorts. Thus far, we show that VS ACD attenuation impacts decision noisiness, and that this change is associated with the decrease in choice performance on an animal-by-animal basis. Choice stochasticity inferred from models itself could be due to noisier learning or directly due to noisier choices ( Nassar et al., 2010 ). To get yet a deeper understanding of the underlying biology and mechanism, we next sought to shed more light on the functions of VS ACD in RL by monitoring it while mice performed probabilistic reversal learning. Ventral striatum ACD is correlated with RPEs and predicts win-stay behavior on a trial-by-trial basis We first aimed to assess how VS ACD relate to model inferred RPEs, because win-stay strategy is strongly related to post-reward positive RPEs (i.e., agents are more likely to repeat previous actions after unexpected reward outcomes). To this end, mice were trained on a head-fixed version of the task with the same reward statistics as the version used in Figures 1 - 3 , except here mice were presented two lick spouts (left and right), and they indicated their choices by licking the left spout (left choice) or the right spout (right choice; Figure 4A ). Download figure Open in new tab Figure 4. Astrocyte calcium dynamics in VS track reward prediction errors and are correlated with win-switch decisions on a trial-by-trial basis. A. Trial diagram for the head-fixed probabilistic decision-making task. On each trial, a go-tone indicated that the animal could make a choice by licking a left or right lick spout; this choice window persisted for 1.5 second. Reward probabilities were either 80% or 20%. Once the animal made a choice lick, the outcome was immediately delivered (reward or no reward). B. Behavioral data from one example session (conventions as in Figure 1B). Dots on the top and bottom indicate choices the animal made to the left and right and the dashed line shows a rolling average of left choices. The mouse made choices that closely follow the true underlying reward probabilities of left and right pokes. C. Behavioral metrics for mice in the head-fixed decision-making task. Dots indicate averages for each animal (n=4) and bars indicate average across animals. D. Schematic of the surgical procedure. AAV5-gfaABC1D-lck-GCaMP6f was injected into the ventral striatum, and a fiber optic was implanted 200um above the injection site. E. Histology of ventral striatum tissue from an experimental animal. The tissue was co-stained for GFAP (astrocyte marker; red) and NeuN (neuron marker; violet). Cyan arrows in the left merged image indicate areas of GFAP and GCaMP overlap. F. Ventral striatum astrocyte calcium in response to model inferred RPEs for one example session. Individual traces show baseline subtracted fluorescence, averaged over trials with different RPEs. Baseline was calculated per trial as the mean fluorescence in the 4-second window before the go cue started. VS astrocyte calcium activity scales with RPE, for example displaying greatest response for most positive RPE and weakest response for the most negative RPE. Error bars are S.E.M. over trials. Gray bar indicates time window used to calculate trial averages for G. G. Correlation between model-estimated RPE and post-decision fluorescence for the example session in 4F. Each dot represents RPE and fluorescence from one trial. Y-axis is the mean value of baseline-subtracted fluorescence taken from a window 1-3.5 seconds after the outcome. There was a significant positive correlation between RPE and astrocyte calcium activity. Colors denote RPE levels (inset). Spearman’s correlation. H. Average astrocytic calcium transients from all sessions from all animals, split out by RPE. Conventions follow F. Error bars are S.E.M. over sessions. Gray bar indicates time window used to calculate trial averages for I. I. Correlation between RPE and fluorescence, across all sessions (each session contributes 4 dots corresponding to 4 RPE levels; conventions are as in G). J. Model residuals of GLM fitted to post-outcome fluorescence (Supplemental Figure 13; calcium activity 1-3.5 seconds after outcome delivery) for post-stay-win trials, split out by trials where the next trial choices were ‘stay’ or ‘switch’. Trials where mice switched on the next trial had significantly lower than predicted ACD, compared to trials where mice stayed on the next trial. This effect was also significant at shorter time windows (p = 0.015 when GLM fit on calcium activity 1-2 seconds after outcome). Download figure Open in new tab Figure 5. Attenuation of astrocyte calcium disrupts striatal excitatory-inhibitory balance. A-B, Bottom, histograms of the means obtained from amplitudes (left) or frequencies (right) of sEPSCs recorded from NAc D1-MSNs (A) or D2-MSNs (B) in brain slices from PMCA (D1: n = 11 cells, 5 mice; D2: n = 19 cells; 6 mice) or control mice (D1: n = 8 cells, 4 mice; D2: n = 18 cells; 6 mice). Top, example sEPSC traces. Scale bar, 200 ms, 20 pA. C-D, Bottom, histograms of the means obtained from amplitudes (left) or frequencies (right) of sIPSCs recorded from NAc D1-MSNs (C) or D2-MSNs (D) in brain slices from PMCA (D1: n = 13 cells, 5 mice; D2: n = 10 cells; 4 mice) or control mice (D1: n = 12 cells, 4 mice; D2: n = 10 cells; 4 mice). Top, example sEPSC traces. Scale bar, 1 s, 40 pA. E-F, Comparison of E/I ratio between PMCA (D1: n = 9 cells, 5 mice; D2: n = 9 cells; 4 mice) and control mice (D1: n = 8 cells, 4 mice; D2: n = 8 cells; 4 mice) obtained from NAc D1-MSNs (E) or D2-MSNs (F). Right, representative traces of sEPSCs at −70 mV and sIPSCs at +10 mV) recorded from the same neurons. Scale bar, 500 ms, 20 pA. Data are mean ± S.E.M. ns, not significant. * P < 0.05, ** P < 0.01. Similarly to the free-moving task ( Figure 1 ), trials started with an explicit go-cue consisting of an auditory tone and signal LED light, and reward contingencies were 80% and 20%. Reward contingencies switched between left and right spouts after 20-35 trials and inter-trial intervals were 5 to 9 seconds long. Outcomes were delivered immediately after a choice was made ( Grossman et al., 2022 ). An example session is shown in Figure 4B . In this session, the mouse’s choices closely aligned to the true underlying reward probability. It quickly switched its choices to the better side after block switches. Overall, mice exhibited high response rates and averaged correct choice rates of 69.5% +/- 0.8% S.E.M (22,249 trials included across 58 sessions, Figure 4C ) on par with their performance in the free moving task (pre-injection correct choice rate across all groups = 67.4%). Their performance (% pokes to the better port) and switching behavior aligned to block switches is shown in Supplemental Figure 12A. Following block switches, mice increased their switching behavior, and their performance quickly recovered as they corrected their choices to the newly better port. To confirm that mice’s choices reflected recent reward and choice history, we fit a generalized linear model to their trial-by-trial decisions (Supplemental Figure 12B). The GLM captured key behavioral patterns observed in the task: mice tended to repeat previous choices, and this tendency was strengthened when the prior choice had been rewarded. These effects decayed over the last three trials. This model achieved high accuracy, correctly predicting the animal’s choice on 94% of trials, consistent with the hypothesis that recent reward–choice history accounted for a large amount of variance in behavior. Having confirmed that mice can accurately perform the head-fixed version of our task, we next investigated the trial-by-trial relationship of ACD, RL RPEs, and win-stay behavior. We injected AAV5-gfaABC1D-Ick-GCaMP6f into the VS to express the membrane-bound calcium indicator GcaMP6f in astrocytes and implanted a fiber to image bulk ACD over the injection site ( Figures 4D , 4E). All 4 mice included in the study were successfully transfected and thereafter learned the task (Methods). We first qualitatively assessed the relationship of RPEs and ACD. ACD responses to trial outcomes for an example session are shown in Figure 4F . To assess the relationship between VS ACD and RPE, we fit the RL model ( Figure 3 ) to mice’s behavior and extracted model-inferred RPEs for each trial. We then split out ACD activity by RPE level. In the example session shown in Figure 4F , VS astrocytes showed strong outcome responses that varied with the RPE level associated with the outcome. ACD was highest when the RPE was highest (strongly positive RPEs; 0.5 <= RPE < 1), followed by relatively more expected rewards (positive RPEs, 0 < RPE < 0.5), expected non-rewards (negative RPEs, -0.5 < RPE < 0), and unexpected non-rewards (strongly negative RPEs, -1 < RPE <= -0.5). ACD was correlated with RPEs on a trial-by-trial basis ( Figure 4G , Supplemental Figure 13). This effect was strong and significant at the population level as well, across all mice and all sessions ( Figures 4H , 4I). These RPEs were reflected in ACD for many seconds ( Figure 4F and 4H), for a relatively longer time than what is typically observed in single neurons’ firing activity (H. Kim et al., 2009 ; Oyama et al., 2010 ) or neurons’ calcium dynamics in response to RPEs (( Zachry et al., 2024 ), Supplemental Figure 14). The data thus far show that VS ACD are correlated with RPEs, a crucial value updating signal in RL, and that attenuating ACD disrupts decision-making performance by reducing win-stay decisions. Together, these data suggest that VS astrocytes contribute to the utilization of RPE signals in VS-guided post-reward decision-making. Based on these observations, we next predicted that on a trial-by-trial basis, higher ACD after outcomes should predict more win-stay decisions, and lower ACD after outcomes should predict more win-switch decisions. Specifically, our hypothesis is that endogenous variations in RPE-related astrocyte Ca 2+ signals could be related to variations in win-stay versus win-switch choice behavior. To test this, we designed an analysis to maximize our power to detect such endogenous effects, by limiting the influence of exogenous variations in RPEs (e.g. due to variations in choice and outcome history), using a two-step approach. First, we restricted this analysis to trials where both the past choice (stay) and past outcome (win) consistently favored the animal staying on the current trial (Supplemental Figure 13) ( Ito & Doya, 2009 ). The animal’s typical behavior in these conditions is to win-stay; but we hypothesized that they would win-switch on a subset of trials due to variations in ACD occasionally producing abnormally low signals, analogous to how win-switch behavior was induced by artificially lowering ACD using PMCA ( Figure 2 ). Second, to further account for potential influences of trial history on ACD, we used a generalized linear model (GLM) to fit ACD responses in a 1-3.5 second post-outcome time window as a weighted linear combination of both past and current choices and outcomes (Supplemental Figure 13). Consistent with our previous findings that ACD responses are correlated with RPEs ( Figures 4F , 4H), the model’s weights showed a key signature of RPEs – a positive weight for current reward outcomes (consistent with coding of current reward value), and negative weights for previous outcomes received from the same choice (consistent with subtracting predicted value, as in RPE ( Bayer & Glimcher, 2005 )). With this model in hand, we computed the residual ACD response on each trial (actual ACD – model predicted ACD). This allowed us to specifically analyze the variations in ACD responses on a trial-by-trial basis, above and beyond those that could be predicted by the model based on past and current choices and outcomes. We then asked whether trial-to-trial variations in ACD residual responses were linked to future switch or stay choice behavior, following stay-win sequences. Indeed, we found that ACD residuals preceding stay choices were aligned with model predicted values ( Figure 4J , top bar). However, ACD residuals preceding switch choices were lower than predicted by the model, and significantly lower than pre-stay residuals ( Figure 4J , bottom bar). Thus, when VS ACD was low, mice were more likely to make win-switch choices – even after controlling for previous choice and outcome history. This is consistent with the results where attenuating VS ACD led to decreased win-stay (increased win-switch) choices and increased decision noisiness. In summary, we found that on average, VS ACD activity correlated with model-inferred outcome RPE during probabilistic decision-making. We used a GLM to further model how the history of choices and outcomes predicted VS ACD responses on a trial-by-trial basis. Even when accounting for the history of choices and outcomes, we found that lower-than-expected VS ACD levels were more likely to be followed by switch decisions. Across our experiments, we find that lower ACD in VS is associated with an increase in post-rewarded outcome switching and increased decision noisiness or choice stochasticity. Slice physiology and computational modeling of astrocyte-neuron interaction led to novel theories of neural-astrocyte computations in RL Our results suggest that astrocyte calcium dynamics in VS contribute to behavioral stability following rewarded outcomes, which is necessary for adaptive behavior in our probabilistic decision-making task. To further understand how ACD supports this function, we asked how astrocytes affect the main output cells of the ventral striatum – medium spiny neurons (MSNs). MSNs comprise 95% of the striatum ( Kreitzer & Malenka, 2008 ) and can be broadly classified into dopamine receptor 1-expressing (D1) and dopamine receptor 2-expressing (D2) MSNs. They serve as a key site for integrating dopaminergic reinforcement signals with glutamatergic inputs that convey information about environmental, interoceptive relevant stimuli. These genetically defined MSN subtypes differ in their distinct roles driving motivated behaviors (( Calipari et al., 2016 ; Collins & Frank, 2014 ; Faust et al., 2025 ; Gerfen & Surmeier, 2011 ; Heiman et al., 2008 ; Keeler et al., 2014 ; Kravitz et al., 2012 ; Planert et al., 2013 ; Shen et al., 2008 ), but see ( Cui et al., 2013 ; Soares-Cunha et al., 2016 )). The behavioral outcome observed when attenuating VS ACD could have arisen through perturbation of information flow specifically at the level of D1 MSNs, D2 MSNs, or both. To better understand how astrocytes support MSN function and thus gain insight into the algorithms of neuro-astro computation, we studied how VS ACD attenuation affected the electrophysiological properties of D1 and D2 MSNs. Mice expressing tdTomato in either D1 (D1-tdTomato line, n = 6 PMCA, n = 4 control) or D2 (A2A-cre x A14 ROSA26-tdTomato line; n = 6 PMCA, n = 4 control) MSNs were injected bilaterally with PMCA or control virus in VS (same as in Figure 1 ; Methods). After 3-8 weeks, we sacrificed the mice and performed whole-cell patch-clamp recordings ( Lucantonio et al., 2025 ) and measured the spontaneous frequency and amplitude of excitatory post-synaptic currents (sEPSCs) and inhibitory post-synaptic currents (sIPSCs) in D1 and D2 MSNs. We found that the frequency of sEPSCs – a measure of presynaptic neurotransmitter release – increased in both D1 and D2 MSNs in PMCA-injected animals compared to control ( Figures 5A , 5C). We also found that the frequency of sIPSCs decreased in D2 MSNs ( Figure 5D ; but not D1 MSNs, Figure 5B ). In contrast, we did not find any differences in the amplitude of sEPSCs or sIPSCs – a common measure of post-synaptic function / efficacy – across D1 and D2 MSNs ( Figures 5A-D ). We further quantified how PMCA-induced ACD attenuation affected D1 and D2 pathways by calculating the excitation/inhibition (E/I) ratio between sEPSCs and sIPSCs. D1 and D2 MSNs in the PMCA condition showed significantly increased E/I frequency ratios ( Figures 5E , 5F, Supplemental Figure 15) – driven by increased excitation in D1 MSNs ( Figure 5A ), and increased excitation and decreased inhibition in D2 MSNs ( Figures 5C-D ). Overall, PMCA-induced ACD attenuation similarly altered D1 and D2 pathways by shifting E/I balance towards increased excitation, likely through a presynaptic mechanism (or locus of action). Though outside of the scope of our paper, we note that these effects could have arisen from several potential mechanisms that are worth examining, such as through changes in local versus long-range presynaptic inputs, glutamate release probability, or presynaptic inhibition. Circuit modeling of neuron-astrocyte computations To date, there are few (but emerging; see ( Gerasimov et al., 2021 ; Meron Asher & Goshen, 2025 ; Serra et al., 2025 )) techniques for manipulating astrocytes while simultaneously assessing large scale neural dynamics. And manipulation of ACD on timescales relevant for trial-by-trial RL remains a challenge. Therefore, to go further in assessing the algorithms and rules through which astrocytes could contribute to RL and generate testable hypotheses regarding their function, we constructed a computational model of VS with two key features: (1) it incorporates several mechanisms through which ACD could impact MSN function constrained by our ( Figures 1 - 4 ) and others’ previous data ( Benoit et al., 2025 ; Bushong et al., 2002 ; Chai et al., 2017 ; Charles et al., 1991 ; Fujii et al., 2017 ; Mahon et al., 2006 ; Ogata & Kosaka, 2002 ; Stobart et al., 2018 ), and (2) it is capable of performing the bandit task, allowing us to probe distinct theories of astrocyte-neuron function and their effect on behavior (e.g., by comparing models’ behavior and the behavior of PMCA mice). In the model, we can then disrupt astrocytes in a relatively specific manner and observe the details of resulting task-behavior, resulting in precise hypotheses of how astrocyte-neuron interactions may support RL. We first describe the general mechanism of the model and then describe how biologically constrained astrocytic contributions are introduced into its computations. Our minimal model incorporated important aspects of RL-related basal ganglia architecture ( Figure 6A ; ( Glimcher, 2011 ; Roesch et al., 2009 ; Samejima et al., 2005 ; Schultz, 1998 ; Sutton & Barto, 2018 )). Namely, it features two populations of MSNs, each representing the values of left or right actions in our task (Methods). Because we found that PMCA induces changes in E/I balance for both D1- and D2-MSNs, we did not model them separately. Both left- and right-action related MSNs received task- and action-related inputs from the cortex and dopaminergic RPE inputs from VTA. On each trial t , the activity of the MSNs encoded the values of the left and right actions ( q L (t) and q R (t) , respectively ( Samejima et al., 2005 )). These values are propagated to downstream circuits to generate the action a(t) ( Sutton & Barto, 2018 ). The model’s decision-making patterns ( Figure 6B-C ) were qualitatively like mice’s behavior in the free-moving bandit task ( Figures 2A , 3D). Download figure Open in new tab Figure 6. Disruption of astrocytic input sharing in a circuit model impairs decision performance and win-stay behavior. A. Schematic of the circuit model. B. Psychometric curves under the four models with various degree of circuit disruptions. C. Ratio of poke to the better port, win-stay ratio, and lose-shift ratio under the three ablation models compared to the control (color code same as B). The error bars represent the S.E.M. over 10,000 simulated sessions In mice, disruption of astrocytic calcium signaling was associated with noisier behavior, consistent with the possibility that astrocytes contribute to reducing noise in the decision-making process. Our results in earlier experiments offer potential mechanisms through which astrocytes may underlie this function. First, attenuation of ACD leads to increased E/I balance of MSNs ( Figure 5 ); this suggests that astrocytes can regulate excitatory drive and overall MSN activity levels (which is thought to also increase neural variability; see below). Second, ACD is correlated with RPE and is linked to future switch decisions ( Figure 4 ); this suggests that astrocytes can respond to RPE signals’ and contribute to VS circuit functions that support RPE-guided decision making. How can we reconcile or investigate these findings in silico and bridge their underlying effects in the circuit, to specific algorithm-level descriptions of behavior? Our model allowed us to selectively disrupt either MSN excitatory drive or astrocyte RPEs and assess the behavioral outputs of these in silico lesions. First, we tested the effect of changing MSN activity. Assuming MSN firing follows a Poisson like process (N. Kim et al., 2019 ; Mahon et al., 2006 ), an increase in excitatory drive ought to lead to increased noise (Supplementary Figure 14A, light blue line). Consistently, when we increased MSN baseline activity in the model, both the better-port poke ratio and win–stay ratio decreased (blue points in Figure 6C ). However, the effect was relatively small when the baseline shift was less than 10 Hz and the number of MSNs was large (here, N = 100; see Supplemental Figure 16 for parameter dependence). Second, we tested the effect of changing astrocyte responses to RPEs. Although each MSN may receive a noisy and variable RPE signals (for example from dopamine neurons in the VTA), astrocytic averaging of these signals across model neurons can produce cleaner inputs for learning. This is conceptually similar to weight-sharing techniques in deep learning ( Fukushima, 1980 ; Kipf & Welling, 2017 ; Krizhevsky et al., 2017 ; LeCun et al., 1989 , p. 19) which are essential for training convolutional and graph neural networks. This input-sharing function is also consistent with known morphological and physiological traits of astrocytes (e.g., responses to dopamine, spatial spread across domains, gap junction coupling, and integration of inputs over slow timescales; see Discussion). Indeed, when we removed the input sharing from the model, the psychometric curve became broader (red line in Figure 6B ). Furthermore, both the better-port poke ratio and win–stay ratio decreased, while the lose–shift ratio remained nearly unchanged, consistent with experimental observations in our task (red points in Figure 6C ). The same trends were also observed even when both baseline adjustment and signal sharing were removed, which is likely what happens under PMCA injection (gold points in Figure 6C ). This work suggests how input sharing across MSNs could be an additional novel mechanism through which astrocytes regulate decision variability, alongside other potential influences such as modulation of E/I balance. Crucially, using biological constraints, our model generated clear hypotheses that will be readily dissociable by future experiments as new techniques allow ACD excitation and suppression with spatial restriction and fast temporal precision. By formalizing astrocyte-neuron interactions in a model that can be used to study the details of animals’ and models’ task behavior, inspired by the biological data, we provide an early algorithmic description of how astrocytes could contribute to decision-making, highlighting the value of computational models for generating hypotheses that extend beyond current experimental reach. Discussion Astrocyte calcium dynamics are known to influence synaptic function and plasticity, shape neuronal activity, and affect behavior, but their specific role in decision-making algorithms to date has been unclear. Here, we provide evidence that astrocytes in ventral striatum play an important role in regulating decision noisiness. ACD reflected reward prediction errors in the VS, and ACD trial-by-trial variability predicted future choices. Informed by these data, we generated a novel model with neuronal astrocyte interactions that suggests a set of testable hypotheses for the circuit-level contribution of ACD, including input sharing – a prominent idea in deep learning – and regulation of MSN excitatory-inhibitory balance. We focused on the striatum because it is a key region for maintaining and updating value estimates for guiding behavior in neuro-RL frameworks ( Averbeck & O’Doherty, 2022 ; Cox & Witten, 2019 ; Frank, 2025 ; Glimcher, 2011 ; Maia & Frank, 2011 ; Schultz, 1998 ). We note that our theories are agnostic to whether astrocytes play an active role in computing RPEs or value in RL. Future studies must probe this issue with novel causal manipulations of astrocytes. Since, the respective role of different striatal regions in RL remains a subject of investigation, we attenuated astrocyte activity in different parts of the striatum and assessed the details of subsequent changes in RL behavior. Our finding that VS astrocyte calcium attenuation disrupted RL behavior is aligned with theories that propose a functional gradient along the ventral-dorsal axis based on properties of neurons: with medial and ventral-medial regions indexing value and value-updating processes and dorsal-lateral regions more associated with the control of habits, skills, and motor functions ( Alexander et al., 1986 ; Balleine & O’Doherty, 2010 ; Burton et al., 2015 ; Haber & Knutson, 2010 ; Ito & Doya, 2011 ; H. F. Kim & Hikosaka, 2013 ; Mohebi et al., 2024 ). Along with our study, a growing body of evidence supports the notion that astrocytic heterogeneity and function follow this dorsal-ventral functional gradient. In the dorsal striatum, astrocyte function has been linked to behavioral perseveration and modulation of behavioral flexibility ( Kang et al., 2020 ; Montalban et al., 2025 ; Yu et al., 2018 ) as well as functions beyond specific value-learning related computations, such as engagement ( Lazaridis et al., 2024 ). Ventral striatal ACD seems to play a role in preference learning and drug-reward self-administration ( Bull et al., 2014 ; Mongrédien et al., 2025 ; Serra et al., 2025 ). This literature and our study together align with the idea that astrocytic and neuronal functions are aligned in functional domains of the striatum. This further supports the notion that astrocytes are embedded in neuronal computations, albeit on a distinct spatial and temporal scales ( Allen, 2014 ; Araque et al., 1999 ; Khakh, 2025 ; Ma et al., 2016 ; Murphy-Royal et al., 2023 ). We show that VS astrocytes contribute to the regulation of decision noisiness, and that attenuation of ACD specifically reduces win-stay behavior. VS ACD correlated with a model inferred RPE ( Figure 4G , 4I) – a key teaching signal in RL ( Glimcher, 2011 ; Schultz, 1998 ). This post-outcome ACD signal predicted future choices, with larger ACD specifically predicting more win-stay behavior – matching the results of PMCA ACD attenuation. Because VS ACD is highly sensitive to dopamine release ( Corkrum et al., 2020 ), future studies must uncover the details of how dopamine modulates astrocyte-neuron interactions ( Allen, 2014 ; Araque et al., 1999 ; De Pittà et al., 2011 ; Favetta & Bubacco, 2025 ; Lefton et al., 2025 ) as well as create further computational models to formalize these interactions in the context of learning theory. Our model suggests that even simple linear input sharing of RPEs across neighboring MSNs may be enough to impact RL. Figure 6 results incorporating neuron-astrocyte interactions suggests that astrocyte units can exert effects on decision performance and decision-noisiness by sharing dopaminergic inputs across neuronal populations. This type of function is well suited to astrocytic morphologic and physical traits (i.e., responses to dopamine ( Corkrum et al., 2020 ), spatial spread across domains ( Bushong et al., 2002 ; Fujii et al., 2017 ; Halassa et al., 2007 ; Ogata & Kosaka, 2002 ), gap junction coupling ( Dermietzel et al., 1991 ), and integration of inputs over slow calcium timescales ( Benoit et al., 2025 ; Charles et al., 1991 ; Fujii et al., 2017 ; Panatier et al., 2011 ; Scemes & Giuame, 2006 ; Stobart et al., 2018 ). This framework for astrocyte function is particularly interesting in light of a recent study showing that VS astrocytes form functional ensembles associated with specific cue-value associations ( Serra et al., 2025 ). Astrocytes may help control the variability of underlying neurons within functional domains by averaging and sharing information across units representing the same cue or value. This proposal is conceptually like weight-sharing techniques in deep learning ( LeCun et al., 1989 ), which are essential for training convolutional and graph neural networks ( Kipf & Welling, 2017 ; Krizhevsky et al., 2012 ). A recent but growing body of work is examining how astrocytes may be involved in normalization of local neural circuit activity ( Gong et al., 2024 ; Kozachkov et al., 2023 ; Miguel-Quesada et al., 2023 ). Within the VS, the function of astro-neuro domains may be at the level of cue-value association; there is some evidence to suggest astrocytes could perform similar normalization functions as in functional columns in sensory cortex ( Philips et al., 2017 ; Schipke et al., 2008 ). This hypothesis could be extended to test how astrocytes stabilize perceptual and cognitive representations, beyond the striatum or RL ( Doron et al., 2022 ; Miguel-Quesada et al., 2023 ). Future studies are required to understand the relationship between this type of linear input-sharing hypotheses generated by our RL study and recently proposed ideas of how astrocytes can actively shape the types of computations performed by neural circuits in state-dependent, non-linear ways ( Murphy-Royal et al., 2023 ). Astrocyte integration of inputs and active shaping of circuit computations need not be mutually exclusive ( Mu et al., 2019 ). What is shared and what differs between these ideas – i.e., molecular mechanisms and computational, spatiotemporal scales – requires careful assessment. Astrocytes may also modulate dopaminergic tone in the striatum. In-vivo striatal astrocyte calcium responses modulate and reduce subsequent dopamine release ( Lazaridis et al., 2024 ), and astrocytes are known to clear extracellular dopamine through DAT ( Segura-Aguilar et al., 2022 ). Blockade of ACD could potentially attenuate astrocyte dopamine clearing functions and lead to increased tonic dopamine tone. Prior modeling and some experimental work point indeed to the notion that hyperdopaminergic tone could decrease inverse temperature, not learning rate ( Beeler et al., 2010 ; Collins & Frank, 2014 ) ( Figure 3E ). Directly assessing whether and how tonic dopamine tone changes due to astrocyte activation or attenuation across different behavioral states is an important next step for understanding astrocytic function in the context of striatal circuitry. Importantly, a change in dopamine tone over long timescales is unlikely to account for all our results, because we observed variations in ACD linked to win-stay behavior unfolding on a shorter timescale (seconds). Astrocytes could participate in the control of synaptic strength and the balance between excitatory and inhibitory inputs (E/I). This was tested by ex vivo slice physiology ( Figure 5 ) and computational modelling ( Figure 6 ). E/I balance has been linked to the noisiness of neural representations and decision-making ( Lam et al., 2022 ; Rubin et al., 2017 ). We found that PMCA attenuation of VS ACD altered shifted E/I balance towards increased excitation in both D1 and D2 pathways, likely through a presynaptic mechanism ( Figure 5 ). Notably, this finding is consistent with literature indicating that astrocytes can directly inhibit synaptic excitatory drive via an adenosinergic, A1R-dependent mechanism that suppresses presynaptic glutamate release (A. B. Chen et al., 2025 ; Corkrum et al., 2020 ; Lefton et al., 2025 ). Some literature suggests that astrocytes selectively mediate synaptic plasticity among D1-D1 and D2-D2 connections ( Martín et al., 2015 ), which may seem to contrast with anatomical work indicating that single striatal astrocytes make close connections with both D1- and D2-MSNs at about equal rates ( Octeau et al., 2018 ). However, because of the unique properties of D1 and D2 MSNs, it is still possible that the net results of ACD are distinct among the neural subtypes. These distinctions, as well as other potential mechanisms (e.g., changes in local versus exogenous presynaptic inputs, presynaptic inhibition) should be disambiguated in future work with experimental and modeling work that that will account for biophysical, electrophysiological and anatomical properties of these projection neurons (i.e. ( Collins & Frank, 2014 )). We cannot distinguish between the effects of PMCA-related attenuation in nucleus accumbens core versus shell, both of which are in VS. Viral spread in nearly all VS injections encompassed both subregions. The core and shell have distinct input and output projections and are thought to contribute differentially to behavior ( Castro & Bruchas, 2019 ; Di Chiara, 2002 ; Zahm, 1999 ), with some evidence that the shell may be more important for choice performance in probabilistic decision-making ( Dalton et al., 2014 ). However, trying to dissociate astrocyte function across the two subregions is complicated by recent evidence showing that astrocytes calcium responses in the nucleus accumbens can be spatially segregated from the source of glutamatergic input ( Serra et al., 2022 ), and that astrocytes form gap junction-coupled networks that span brain regions ( Cooper et al., 2025 ). For example, optogenetically stimulating ventral hippocampal inputs to the nucleus accumbens (which largely target the shell) leads to widespread astrocyte activation in the core ( Serra et al., 2022 ). Thus, astrocytes may have complex interactions with neural circuits across space that are not cleanly delimited by neural subregion boundaries. Although astrocyte functions broadly align with the dorsal–ventral functional axis (see discussion above), they cannot be cleanly distinguished at narrow spatial scales ( Charles et al., 1991 ; Cooper et al., 2025 ; Dermietzel et al., 1991 ; Fujii et al., 2017 ; Murphy-Royal et al., 2023 ; Oberheim et al., 2009 ; Serra et al., 2022 ). In our computational model, this broad spatial influence is a feature—not a bug—that underlies the input-sharing hypotheses we propose. Input-sharing maybe critical to consider when modelling how the brain implements many forms of RL ( Dabney et al., 2020 ; Lowet et al., 2025 ) in part because the spatial scale of dopaminergic influence on the striatum maybe more sharp or precise than previously assumed ( Yee et al., 2025 ). We used a viral approach that broadly and indiscriminately attenuates astrocyte calcium activity via constitutive calcium pump ( Yu et al., 2018 ) as a first-line strategy to investigate how astrocyte calcium transients, the main functional signal of these cells, impact value-guided behavior. It is important to note that while PMCA significantly attenuates calcium signaling, it does not ‘turn off’ astrocytes – PMCA-expressing astrocytes undergo a variety of gene and protein expression changes ( Yu et al., 2018 ). Interestingly, these gene expression changes are not seen when only Gq-GPCR astrocyte signaling is attenuated ( Nagai et al., 2021 ). Thus, it is possible that some of the behavioral effects we see are due to gene expression changes above and beyond downstream effects of calcium attenuation ( Yu et al., 2018 ). Additionally, it is possible that the extrusion of astrocytic calcium may alter local neural activity by altering neural calcium signaling ( Berridge, 1998 ). Astrocyte cytosolic calcium levels (∼100nM ( Zheng et al., 2015 )) are orders of magnitude below what is needed to alter neural calcium conductance; however, this has not yet been explicitly tested. All this said, it remains that specific targeting of astrocytic calcium signaling in specific brain regions without inducing glial or neural depletion ( Meron Asher & Goshen, 2025 ; Yu et al., 2018 ) makes for a powerful and useful method to study astrocyte contributions to behavior in a circuit-specific manner. In addition to dopamine ( Corkrum et al., 2020 ; Lazaridis et al., 2024 ), striatal astrocytes also respond to influence the processing of nearly all signaling molecules present in striatum and elsewhere, such as glutamate ( D’Ascenzo et al., 2007 , p. 201; Martín et al., 2015 ; Serra et al., 2022 ), GABA ( Kang et al., 2020 ; Nagai et al., 2021 ; Roberts et al., 2020 ; Yu et al., 2018 ), acetylcholine ( Papouin et al., 2017 ; Stedehouder et al., 2024 ; Takata et al., 2011 ), opioids ( Corkrum et al., 2019 ; Murlanova et al., 2023 ), endocannabinoids ( Adermark et al., 2024 ; Martín et al., 2015 ; Navarrete & Araque, 2010 ), serotonin ( Miyazaki & Asanuma, 2016 ), and of even energy metabolites ( Molnár et al., 2011 ). As we gain deeper insight into how different modulators together cooperate to shape learning, their impact on astrocytes and astro-neuro circuit function needs further study. Our findings suggest that VS astrocytes influence decision variability. Their ACD covaries with RPE and predicts future win-stay behavior. Computational modelling suggested that their impact on behavior could occur through input sharing across neuronal populations and via modulation excitability over broader spatiotemporal scales than neurons. Our results generate novel theories for how astrocytes cotribute to RL and algorithm level computation. These implications extend beyond biology. Understanding how astrocytes regulate noise, excitability, and information sharing provides a conceptual bridge between molecular mechanisms, circuit-level models of decision-making, and neuromorphic computing (the effort to design hardware to run AI based on the architectures and circuit motifs of the brain). Embedding astrocytic principles into artificial systems may yield architectures that are both more biologically realistic and computationally efficient — for example, by reducing training duration through hardware that enacts context-dependent input sharing, these principles may enable biologically plausible neural networks to learn faster while retaining the advantages of local learning rules over global gradient descent. Our study provides one entry point into this emerging frontier. Materials and methods Subjects 76 C57BL/6 mice (37 male, 39 female) were obtained from Jackson Laboratory and from in-house breeding. Animals were housed individually and maintained on standard 12/12 light dark cycles. All procedures conformed to the Guide for the Care and Use of Laboratory Animals and were approved by the Institutional Animal Care and Use Committee at Washington University. View this table: View inline View popup Download powerpoint Table 1: Behavioral training Mice were trained to perform tasks on a FED3 device ( Matikainen-Ankney et al., 2021 ). These are small, open-source operant devices running off Arduino code with a left and right nose-poke port, pellet delivery system, a strip of Neopixel lights on the front of the device, and a piezo speaker. A FED3 was placed against the back wall of the home cage of each mouse. Custom cut stainless steel cage dividers magnetically attached to the front of the FED3s prevented the mice from accessing the backs of the devices. During behavioral training, mice were food restricted and received all food (Dustless Precision Pellets® Rodent, Grain-Based 20mg, Product Number, F0163, BioServ) from the FED3. Mice were weighed before behavioral training to establish their baseline weight and weighed every week during training to ensure their weight did not drop below 80%. None of the mice used had weights that fluctuated below 80% of their baseline weight. Mice were first acclimated to the FED3s and to receiving pellets from the FEDs during an initial 1-2 day magazine training period. During magazine training, pellets were automatically delivered into the hopper, or magazine, without any action needed from the mouse. Mice that successfully learned how to pick up pellets from the hopper progressed to the next training stage. The few mice that were unsuccessful in learning how to pick up pellets were removed from the study. The next training stage was a time-unrestricted version of the operant non-stationary probabilistic decision-making task, also known as a bandit task. In this task, the left and right nose-poke ports were associated with either 80% or 20% pellet delivery, such that P(left) = 1 – P(right). If the mouse made a rewarded poke, the FED3 delivered a 200ms 4kHz tone and a pellet was delivered into the hopper. If the mouse made an unrewarded poke, a 300Hz error tone was played. After either a rewarded or unrewarded ‘choice’ poke, the FED3 entered a 5 second time out period during which the Neopixel lights turned off and pokes could not trigger any outcome, although they were still logged. After 20-30 delivered pellets, the reward contingencies between the two nose-poke ports switched, uncued to the mouse. During this training stage, mice learned that nose-poking was necessary for pellet dispensal, and they started to learn the structure of the task. In this intermediate training stage, the task was available to the mice between the hours of 1am and 9pm. When the task was active, the strip of Neopixel lights on the front of the FED3 turned blue, indicating that pokes could lead to reward- or no-reward outcomes. Outside of active task hours, the Neopixel lights turned off and pokes could not trigger any outcome, although they were still logged. After 1-2 days of the relatively time-unrestricted bandit task training stage, mice started the final stage of the bandit task, during which the task was only available to the mice from 5pm to 9pm. Animals on average initiated around 200 trials in each 4-hour session (mean = 194.7, SD = 38.2). After 7-10 days from the start of this training state, the performance of each mouse over days was assessed. Mice that could not consistently achieve 60% correct choice rate were removed from the study because studying their learning or adaptation during block switches was not possible (e.g., low performance indicated that they did not learn or switch as the task contingencies changed). Mice were further trained for 3-4 weeks on the bandit task ( Figure 1 ) and then injected with control, iβark, or PMCA virus into different regions of striatum. After virus injection, mice recovered in their home cage on ad lib normal chow with no FED3. After 2 weeks of recovery and virus expression time, mice were again food restricted, the FED3s programmed with the final time-restricted bandit task were placed back in their home cage, and behavioral data was collected for another 3-4 weeks. To capture block switch related learning changes rather than transient changes related to a post-surgery rest period, we let mice acclimate to the task again after surgery (see Behavioral data analysis section). We ran an additional cohort (n = 7) of VS-PMCA injected mice on a version of the task where block switches occurred every 20-40 trials irrespective of rewards earned (reward non-contingent task), to test that these PMCA-induced behavioral effects were not specific to the reward contingency meta-structure of the first task (reward-contingent task). We found similar decreases in performance and win-stay behavior in this smaller cohort (Supplemental Figure 7). Data from both groups were combined and collectively referred to as VS-PMCA results. To obtain more controls, we injected two cohorts in VS with a control virus (AAV5-GFAP104-mCherry or AAV5-gfaABC1D-mCherry), running one cohort on the reward-contingent version of the task (n = 13) and one cohort on the reward non-contingent version of the task (n = 5). We also injected a cohort of mice on the reward-contingent task with saline in the region of motor cortex overlying VS (n = 6). None of these cohorts exhibited significant shifts in correct choice rate, win-stay, or lose-switch behavior after injection compared to before. Data from these three cohorts were pooled and collectively referred to as ‘VS-control’ (Supplemental Figure 6). For the PR1 progressive ratio task, FED3s were programmed to deliver pellets in respond to increasing number of pokes to the left nose poke port. The right nose poke port was inactive but still logged pokes. The task was available to mice continuously over 5 days of testing. After each pellet was dispensed, the number of pokes required to dispense another pellet increased by one. This continued until the mouse did not make any pokes in a 30-minute window. At this point, the number of pokes the mouse had made for that last pellet was recorded as one breakpoint, and the threshold for dispensing the next pellet reset to one poke. Virus injections Mice received intercranial injections of control (AAV5-GFAP104-mCherry, Addgene 58909-AAV5, or AAV5-gfaABC1D-tdTomato, Addgene 44332-AAV5), PMCA (AAV5-pZac2.1-GfaABC1D-mCherry-hPMCA2w/b, Addgene 111568-AAV5), or iβark (AAV5-pZac2.1-gfaABC1D-iβark-mCherry, Addgene 117691) targeted at different regions of striatum. Anesthesia was induced with 3-4% isoflurane and maintained at 1-2%. Buprenorphine SR was injected subcutaneously at 1.0 mg/kg at the start of surgery for pain relief. After induction of anesthesia, mice’s heads were shaved, sterilized with alternating betadine and alcohol swabs, and an incision was made over the skull. Coordinates were marked bilaterally on the skull for ventral striatum (AP: +1.0-1.2, ML: +/- 1.2, DV: 4.5 relative to dura), dorsomedial striatum (AP: +0.7, ML: +/- 1.5, DV: 2.5 relative to dura), and dorsolateral striatum (AP: +0.7, ML: +/-2.5, DV: 3.0 relative to dura) relative to bregma. A 0.5mm burr hole was drilled over the marked coordinates (Fine Science Tools 19007-05, Foredom K.1070 Drill). Virus was loaded into a syringe (Hamilton Neuros 65457-02), slowly lowered into place, and delivered (VS: 0.25 µl/side at 0.05 µl/min; DLS and DMS, 0.45 µl/side at 0.1 µl/min). After virus delivery, the syringe was raised by 100-200 µm. An additional 5 minutes were allowed to pass before the syringe was slowly raised. After injections were made, the skin incision was sutured, and mice were returned to their home cage to recover. Our study required bilateral expression in VS, DMS, or DLS. Animals without strong bilateral expression were excluded. Occasional virus expression in nearby regions adjacent to the injection needle track was noted; a common and difficult to avoid effect of intracranial virus injection. We report percentages of animals with off-target virus expression in nearby brain regions in Supplemental Figure 1B. For fiber photometry, the same surgical procedure was followed. Mice received intracranial injections of AAV5-pZac2.1-gfaABC1D-Ick-GCaMP6f (Addgene 52924-AAV5) in VS (AP: +1.0-1.2, ML: +/- 1.2, DV: 4.5 relative to dura); a fiber optic (Doric; 200um ID, NA0.37, 5mm long) was implanted 200 µm above the virus injection site. Behavioral data analysis All statistical tests of difference wherever p-values are reported were non-parametric and two-sided unless otherwise noted. Correlations were Pearson’s correlations unless otherwise noted. Sessions before the 6th day of training (both pre- and post-injection), or where animals initiated fewer than 70 trials, were excluded to remove training sessions where animals were still learning the task. All other sessions were included in the analyses. For cohort analyses, data from pre-injection and post-injection session were averaged for each animal. Wilcoxon sign rank tests were then run on pre- and post-injection data for each cohort. For within-animal analyses, Wilcoxon rank sum tests were run for pre- and post-injection sessions for each animal. For Bayesian inference of block switches, we followed methods developed and described in ( Bartolo & Averbeck, 2020 ). In essence, the algorithm operates as follows. Around each true change point, where the reward probability shifts, we defined a window spanning half a block length before and after the change. Within this window, we evaluated the likelihood that each trial represented the change point, under the assumption that exactly one change point was present. Both choice and reward information were used to estimate how an ideal observer would classify a trial as occurring before or after the change. The trial with the highest likelihood was then identified as the inferred block switch point. If multiple trials within the window had nearly identical likelihoods (differing by less than 0.01), the corresponding block was excluded from further analysis. The same window and analysis were used to identify changepoints as the animal detected them, based off choice information alone. The same algorithm was used to analyze block switches for both the reward contingent and reward non-contingent versions of the bandit task. Since the changepoint detection algorithm relies solely on the likelihood, not any priors, the algorithm takes the identical form in both cases. Head Fixed Bandit Training - Habituation After recovering from surgery mice were water restricted at 1ml per day. After 1 day of restriction mice began participating in 2-4 sessions of habituation handling daily. For the first 2-3 days, mice were placed in the palm of the experimenter’s hand and offered water via syringe. For 1-3 days after mice were held over the location where they would be head posted and offered water via syringe. Then for 1-2 days mice were placed on the wheel and held to the head post by the experimenter while being offered water via syringe. Finally, for 1-3 days the mice were fully head posted and were periodically rewarded with 4 ul of 15% sucrose and water via the lickers in the behavior rig. Rewards were given manually in 1-3 second intervals, alternating which side was rewarded every 5-10 rewards. Each habituation session lasted between 5-15 minutes or until the mice showed visible agitation, freezing, or refusal to receive reward. Steps were repeated as needed for mice who showed reluctance to participate. Head fixed Bandit Training - Task Training After habituation training, mice began training for the probabilistic decision-making task. While most mice generally followed the listed progression, steps were repeated as needed for mice who showed reluctance to participate or tended to choose only one licker. At every stage, if the mouse made ∼150-200 consecutive choices on the same side, the task was paused and the position of the lickers adjusted to make the unchosen side closer than the other. The task took place in a sound attenuated box, wherein mice were head fixed atop a stationary styrofoam wheel. Two lickers fixed approximately 4mm apart were placed directly in front of the mouse’s mouth such that each licker was within reach. All tasks described used a 0.3s go cue and 4 microliters of 15% sucrose and water for reward. Rewards were administered immediately after selection. Additionally, to signify the length of the response window and to provide visual feedback, a signal LED was placed in the mouse’s left periphery. This light turned on at the beginning of the response window and remained on until the window closed, or the mouse made a choice. The first day of training was on a Pavlovian task. Mice were rewarded for licking from either licker within 10s after the tone. The task ran for 400-450 trials or until the mouse refused to participate. On day two, mice began a task where only one licker was active at a time. Following the go tone the mice had 10s to lick the active licker to receive reward. Licking the inactive licker gave no feedback. The active licker was determined by alternating blocks 5-10 trails in length. Additionally, we varied the inter trial interval (ITI) between 2.5-5.5 seconds. On days 3-4 we used the same task structure but progressively tightened the window to respond (from 5s to 3s) while increasing the ITI (from 3.5-6.5s to 5-9s). On days 5 and 6 we activated both lickers. The mouse was still tasked with finding the licker that guaranteed reward, but a wrong choice would end the current trial. Both days used a 1.5s response window and 5-9s ITI. Day 5 we used a 15-25 trial block length but increased to 20-35 on day 6. However, only correct choices contributed towards the block length. Such that mice who made more errors had longer total block lengths. Day 7 and onwards used the full task. Probabilities were introduced following the same convention as used in the free moving task. Task parameters were held constant with a 1.5s response window, 5-9s ITI, and 20-35 trial block lengths. Additionally block length was no longer reward dependent Fiber photometry recordings Fluorescent calcium signals were acquired with a FPS-S fiber photometry system from Doric Systems, through which we could excite and collect emission light from GCaMP from the same implanted optical fiber. Excitation wavelength was 488nm; we used a 405nm LED for isobestic excitation. Light was filtered and collected through a 4-channel fluorescent MiniCube (Doric) coupled via a low autofluorescence mono fiber optic patch cable (Doric). Photometry signals were first preprocessed by fitting the isosbestic signal (IS) to the GCaMP transporter signal (TS). We subtracted and then divided the GCaMP signal by this fitted isobestic signal (fIS) (df/F = (TS -fIS) / fIS); this signal was then smoothed with a 200ms window. Since individual trials often started at different initial fluorescences, trial by trial analysis used baseline subtracted values. We defined a baseline for each trial as the mean fluorescence in the 4 secs before tone start. Individual trial fluorescence is reported as the difference between df/F and this baseline value. Additionally, due to the slow time course of astrocytic calcium, when analysis dictated a single measure, we defined the post outcome fluorescence as the mean fluorescence between 1 to 2 seconds after outcome. Head-fixed behavioral and photometry analysis We included 58/89 sessions (24,609 trials; 22,249 trials responded) from 4 mice in our data. Days were excluded if mice did not reach either >60% participation or >60% correct choice rate. This criterion was applied to all analyses performed. RPE was estimated using a Q-learning RL model (see below section on RL modeling). Each trial was sorted into buckets based on its RPE: strong positive (0.5 <= RPE < 1), low positive (0 < RPE < 0.5), low negative (−0.5 < RPE < 0), and strong negative (−1 < RPE <= 0.5). These buckets were used for all analyses that tested for RPE related effects. To capture past trial effects on choice, we constructed a GLM to predict log odds of a left choice using a logistic link function for binomial data. For regressors we included past trial choice (Left: 1, Right:-1), outcome (Rewarded: 1, Unrewarded: -1), Choice x Outcome, and a spatial bias term (1). The model was fitted to 19,386 trials pooled from all 4 mice; trials were excluded if one or more of the regressors was missing a value. We included data from 1 to 3 trials prior. We constructed another GLM to predict mean fluorescence 1-2s post outcome using a linear link function and Gaussian noise distribution. We used a similar construction as in the choice GLM; however, we defined choice in terms of whether it was the same or different than the current trial choice. We included regressors for current outcome (Rewarded: 1, Unrewarded: -1), Past (1-3 trials prior) Choice Same (same:1, else: 0) x PastOutcome, Past (1-3 trials prior) Choice Different (different: 1, else: 0) x PastOutcome, and side bias (1). We fitted the model to 19,386 trials pooled from all 4 mice; trials were excluded if one or more of the regressors was missing a value. We included data from 1 to 3 trials prior. We then looked at the residual fluorescence (actual - model predicted) in trials where mice made two consecutive choices on the same side and were rewarded and divided them based on whether the next choice was the same (“stay-win-stay”, n=11,212) or different (“stay-win-switch”, n=177). Perfusion Animals were deeply anesthetized with a 100/10 mg/kg ketamine-xylazine cocktail injected intraperitoneally. Animals were then intracradially perfused, first with 15ml of 1X phosphate buffered solution (PBS), followed by 15ml of 4% paraformaldehyde (PFA). The brains were extracted and stored in 4% PFA overnight. They were then transferred to a 30% sucrose solution where they were left to fully saturate for at least 2 days for cryoprotection. Serial coronal sections were cut at 50 µm on a cryostat (Leica CM1860). Sections were stored in 1X PBS or 0.1% sodium azide in 1X PBS at 4deg C. Histology Free floating sections were rinsed in 1X PBS three times for 5 minutes each. Sections were incubated in 10% normal goat serum (NGS) in 0.3% Triton-X in 1X PBS for 1 hour at room temperature, then incubated in primary antibodies (rabbit anti-mCherry, abcam ab167453; 1:800-1:1000; guinea pig anti-NeuN, Sigma Aldrich ABN90, 1:1000; guinea pig anti-S100B, Synaptic Systems 287-004, 1:1000) overnight at room temperature. Between primary and secondary incubation, sections were thrice rinsed in 1X PBS. Secondary antibodies (goat anti-rabbit AlexaFluor 555; goat anti-guinea pig AlexaFluor 488; goat anti-rabbit AlexaFluor 488) was applied at 1:300 for 4 hours at room temperature. Finally, sections were rinsed twice with PBS for 5 minutes, followed by a 5-minute incubation in 1:1500 DAPI before being mounted on glass slides and cover slipped with Prolong Diamond Anti-fade mounting medium. Mounted sections were imaged with a Leica DM6 B or a Nikon AXR confocal microscope. All animals tested were imaged to ensure that virus injections were bilaterally expressed and localized to the target region. Animals with unilateral, weak, or overly extensive viral expression were excluded from the dataset. In vitro calcium imaging Two-photon imaging GCaMP6f recordings of calcium activity in the striatum astrocytes were performed in acute slices from C57BL/6J mice microinjected with AAV5-gfaABC1D::lck-GCaMP6f (Addgene 52924-AAV5) three to four weeks prior. Two-photon laser scanning microscopy (2-PLSM) recordings of GCaMP6f fluorescence over a field of view of 452 µm × 452 µm at 512 × 512 pixels resolution were captured using Bruker Ultima 2pPlus with Nikon 25X, 1.10NA objective, 1.6x optical zoom, laser power at 37.5 mW, recording rate at 1 Hz, recording depth at 30-40 μm. A 920 nm tunable laser was used to excite the fluorophores. Acute slices were incubated in a chamber with oxygenated artificial cerebrospinal fluid (aCSF) containing 1µM tetrodotoxin to block neuronal firing. After a stable baseline recording, stock norepinephrine (NE, norepinephrine bitartrate, Tocris 5169) or dopamine (DA, dopamine hydrochloride, Tocris 3548) solution is spiked into the aCSF to reach a working concentration of 20 µM for NE and 100 µM for DA and astrocytic calcium response is recorded. Calcium imaging analysis Astrocytic calcium responses in the striatum to NE and DA are performed across the field of view (FOV) for each recording for both the control and CalEx groups. For the control group, cell boundaries are delineated with the aid of static tdTomato fluorescence using Fiji for further cell-based analysis. Average fluorescence intensity across FOV or each cell is extracted from the recordings, and the traces are aligned so perfusion of NE and DA starts around 100 s into the recording. The traces are then processed using STARDUST ( Wu et al., 2024 ) to detect active calcium signals using two standard deviations of the baseline as the threshold. The largest active signal within 100-300 s of the recording is taken as NE/DA response. For FOV analysis, in cases where no active signal is detected within the pre-determined temporal window, the largest fluorescence value between 100-300 s and its corresponding area under the curve are evaluated. Ex vivo slice electrophysiology Whole-cell patch-clamp recordings were performed as previously described ( Lucantonio et al., 2025 ). Briefly, mice were anesthetized with isoflurane and decapitated. Brains were rapidly extracted and placed in ice-cold, NMDG-based cutting solution (in mM: 92 NMDG, 20 HEPES, 25 glucose, 30 NaHCO₃, 2.5 KCl, 1.2 NaH₂PO₄, 5 sodium ascorbate, 3 sodium pyruvate, 2 thiourea; osmolarity: 303–306 mOsm), continuously bubbled with 95% O₂/5% CO₂. Coronal brain slices (300 µm thick) containing the nucleus accumbens (NAc) were prepared using a Leica VT1200 vibratome at a speed of 0.07 mm/s in the same ice-cold cutting solution. Slices were then transferred to artificial cerebrospinal fluid (aCSF; in mM: 92 NaCl, 20 HEPES, 25 glucose, 30 NaHCO₃, 2.5 KCl, 1.2 NaH₂PO₄, 5 sodium ascorbate, 3 sodium pyruvate, 2 thiourea; osmolarity: 303–306 mOsm) at room temperature, continuously bubbled with 95% O₂/5% CO₂, and incubated for at least 1 hour before recordings. Recordings were conducted at 32°C in a perfusion chamber continuously supplied with aCSF (in mM: 126 NaCl, 2.5 KCl, 1.4 NaH₂PO₄, 1.2 MgCl₂, 2.4 CaCl₂, 25 NaHCO₃, 11 glucose; osmolarity: 303–305 mOsm) at a rate of 1.5–2.0 ml/min using a World Precision Instruments pump. Spontaneous excitatory postsynaptic currents (sEPSCs) were recorded using borosilicate glass microelectrodes (resistance: 2–3 MΩ) filled with a cesium-based internal solution (in mM: 117 cesium methanesulfonate, 20 HEPES, 0.4 EGTA, 2.8 NaCl, 5 TEA-Cl, 4 Mg-ATP, 0.4 Na-GTP, 5 QX-314; osmolarity: 280–285 mOsm). To isolate sEPSCs, neurons were voltage-clamped at –70 mV; to isolate inhibitory postsynaptic currents (IPSCs), cells were clamped at +10 mV. Medium spiny neurons (MSNs) were identified via infrared differential interference contrast (IR-DIC) optics using an Olympus BX5iWI inverted microscope. TdTomato-positive cells were classified based on strong fluorescence signal. Recordings were made using a Multiclamp 700B amplifier (Molecular Devices), filtered at 1 kHz, and digitized at 20 kHz using a Digidata 1440A digitizer (Molecular Devices). Synaptic responses were analyzed with ClampFit (Molecular Devices) and Easy Electrophysiology (Easy Electrophysiology Ltd) software. Series resistance (10–20 MΩ) was monitored via a –5 mV voltage step, and cells exhibiting >20% change in series resistance were excluded from analysis. RL modeling We fit a Q-learning model to the mice’s choices in the probabilistic decision-making task. To ensure near-perfect parameter recovery, for each mouse, we concatenated data from all sessions included in other behavioral analyses into pre- and post-injection datasets. The model estimated trial-by-trial action values for the left and right nose poke actions, Q left and Q right , using a single learning rate. At the first trial of each day, we initialized the Q-values to zero. In addition, we modeled potential side biases using an additional value-like term to the estimated Q-values of the left side. Finally, we also included a lapse rate term to account for potential lapses of mice. On each trial, the difference between the left and right action values and the side bias term, was passed to a probabilistic softmax decision function of the form: The choice probability, was bounded by a lapse rate: After carrying out the choice outputted by the decision function, the outcome was delivered (1 for rewarded, 0 for not rewarded). The reward prediction error (RPE(t)) was computed by subtracting the value of the chosen action from the outcome: Finally, the value of the chosen action was updated for the next trial by adding RPE(t) weighted by a learning rate α: We conducted the standard maximum likelihood fitting using the 𝑝 observed , the probabilities of the model matching mice’s action(i.e., 𝑝 observed = 𝑝(𝑐ℎ𝑜𝑖𝑐𝑒 = 𝑙𝑒𝑓𝑡) if the observed choice of the animal is left, and 𝑝 observed = 1 − 𝑝(𝑐ℎ𝑜𝑖𝑐𝑒 = 𝑙𝑒𝑓𝑡) otherwise). We optimized the log likelihood of the data, 𝐿 = ∑ trial 𝑙𝑜𝑔(𝑝 observed ), using a heuristic global optimization algorithm (Matlab GlobalSearch;( Ugray et al., 2007 )). We used GlobalSearch with its default Matlab parameters (NumTrialPoints = 1000, NumStageOnePoints = 200, FunctionTolerance = 1e-6, XTolerance = 1e-6). We initialized the parameters for the initial local optimization process as: The initial values for subsequent optimization attempts were chosen automatically by the GlobalSearch algorithm. To prevent the optimization from finding unreasonable parameter values, we constrained the search for the parameters as: The coefficient of determination, 𝑟 # values, was calculated as: Where, 𝑛 trials is the total number of trials in the concatenated data, and 𝑝(𝑐ℎ𝑎𝑛𝑐𝑒) = 0.5. Note that this 𝑟 # value is an equivalent of the 𝑟 2 of the linear regression in the case of maximum likelihood fitting and quantifies the model’s ability to predict observed behavior above and beyond chance (Daw, 2009). Circuit model We constructed a simplified circuit model of the ventral striatum, shown in Fig. 6 , in which MSNs receive task- and motor-related input from the cortex and RPE signals from the VTA. We model two populations of MSNs, each representing either left or right actions, with each population containing N =100 neurons. Let w L i denote the (total) synaptic weight of cortical input to the i -th MSN representing the left action. The firing rate of this neuron at trial t is given by where b is the baseline firing rate, and r o is a scaling factor. The firing rate of the i -th MSN in the right-action population is similarly given by 𝑟 R i ( 𝑡 ) ∼ 𝑃𝑜𝑖𝑠𝑠𝑜𝑛 ( w R i ( 𝑡 ) 𝑟 0 + b ). The value of the left action is computed in a downstream area by averaging MSN activity across the population: . A decision is then made by comparing the estimated values of the left and right actions, q L (t) and q R (t): Here, β is the inverse temperature parameter controlling choice stochasticity. A reward signal 𝑅(𝑡) ∈ {0,1} is delivered at the end of each trial based on the correctness of the chosen action a(t): R(t) = 1 if correct, R(t) = 0 otherwise. The true RPE signal is given by 𝛿(𝑡) = 𝑅(𝑡) − 𝑞 𝑎 (𝑡), where q a (t) denotes the value of the chosen action. If dopaminergic inputs are noisy, the signal received by the i -th MSN is modeled as where z i (t) is sampled independently from a standard normal distribution, 𝑧 5 (𝑡) ∼ 𝑁(0,1). Although we assume multiplicative noise (consistent with the Poisson nature of neural variability), the results are robust to additive noise as well. The synaptic weight update based on this noisy RPE signal is 𝑤 L i (𝑡 + 1) = 𝑤 L i (𝑡) + 𝛼𝛿 L i (𝑡) if a(t) = L, and 𝑤 L i (𝑡 + 1) = 𝑤 L i (𝑡) otherwise. In the presence of astrocytes, dopaminergic signals can be shared across MSNs. Denoting the averaged RPE signal as we hypothesize that the weight update under functional astrocytic modulation becomes 𝑤 L i (𝑡 + 1) = 𝑤 L i (𝑡) + 𝛼̄𝛿(𝑡) if a(t) = L, and 𝑤 L i (𝑡 + 1) = 𝑤 L i (𝑡) otherwise. We simulated the model using the 80:20 bandit task from the experiments, in which reward contingencies switched every 20–30 rewarded trials. Each session consisted of 250 trials. The learning rate was fixed at 𝛼=0.5 and the inverse temperature at β =2.5, matching the parameters recovered from RL fitting. For the main figure, we set the scaling factor 𝑟 0 =5.0, baseline firing rate b =5.0Hz, and RPE noise 𝜎 d =0.32. The results were robust to variations in b and 𝜎 d as shown in the supplementary figure. Additional behavioral assays Sensorimotor battery The procedures used follow those previously described (J. Chen et al., 2021 ). For each test the experimenter manually recorded, using a stopwatch, the test time in seconds and hundredths of a second. Two trials were conducted for each test and the average of the two yielded a single time in sec and hundredths of a sec, which was used in the analyses. The order of tests was not counterbalanced between animals so that each animal experienced the test under the same conditions. In brief, for the walk initiation test, mice were placed in the middle of a 24x24cm square boundary facing away from the experimenter; the time it took for the animals to reach the outside of the boundary was recorded. For the ledge test, animals were placed on a ledge apparatus for 60 seconds; the amount of time they were able to balance on the ledge was recorded. For the inverted screen, animals were placed on a 90 degree inclined screen and assessed for how long they could hold on (up to 60 seconds) For the platform test, animals were recorded for how long they could balance on a wooden platform (measuring 3.5cm thick, 3.0cm in diameter, elevated 25.5 cm above the base) for up to 60 seconds. For the pole test, animals were placed head up on top of the vertical, textured pole apparatus; the time it took for them to turn downward 180 degrees and climb to the bottom of the pole was recorded (up to 120 seconds). For the rotorod, animals were placed on a rotating rod and assessed for how long they could stay on it at either constant or increasing speed (up to 60 seconds). Y-maze Animals were placed in a Y-maze composed of 3 arms measuring 40cm x 10cm with 20cm high walls, placed at 120 degrees to each other. They were placed in the in the same arm facing away from the center, then allowed to freely explore the arena for 8 minutes. Animal movement was tracked using ANY-maze (Stoelting Co., Wood Dale, IL; http://www.anymaze.co.uk/index.htm ). Number of entries and alternations between arms were recorded. Arenas were cleaned with Nolvasan between animals. Novel object recognition Animals were handled for at least 3 sessions before the NOR was performed. Objects and sides were counterbalanced across animals. Animals were tested in a 40cm x 40cm open field inside sound attenuating boxes. Animals were tested over 3 days for 10 min each. Day 1 consisted of habituation to the environment with no objects; day 2 consisted of habituation to 2 of the same sample objects; day 3 consisted of testing with one sample object and one novel object (2x 10 minute trials with 50 minute ITI). Time animals spent investigating the objects was quantified as time where the animal was oriented towards and within 20mm of an object. NOR index was quantified as the proportion of time spent investigating the novel object relative to all investigation time. Equipment was cleaned with 70% ethanol between animals. Open field Animals were evaluated in clear 47.6 x 25.4 x 20.6 cm enclosures with lids. They were monitored for 1 hour each using Kinder Scientific Motor Monitor, which tracked distance traveled, rearing behavior, and time spent in the center of the arena (middle 50%). Runs were counterbalanced across mice in different experimental groups. Arenas were cleaned with 70% ethanol between animals. Self grooming Mice were tested for 10 minutes in transparent enclosures (29.5 x 18.5 x 13cm). Behavior was tracked and quantified for self-grooming (licking, scratching, and nibbling) behavior using Ethovision ( https://noldus.com/ethovision-xt ). Mice were acclimated to the room for 1 hour prior to testing and enclosures were cleaned with 70% ethanol between animals. Sucrose preference Testing was performed on single housed mice using automated sucrose preference testing devices (Sippers, ( Godynyuk et al., 2019 ). In brief, animals were single housed for 96 hours and presented with two liquid dispensers that provided either plain water or 1% sucrose. The consumption of sucrose vs plain water was tracked by the Sipper devices over time. The location of sucrose vs water was counterbalanced across animals and switched across days. Prior to testing, animals were presented with plain water in both dispensers. Supplemental Figures Supplemental Figure 1. Performance and virus expression in the free-moving probabilistic decision-making task. A. Performance (% of pokes to the port with higher reward probability) and number of pokes made outside of the active task window (from 9pm - 5pm) over training days, averaged over all mice pre-injection (VS-PMCA, DMS-PMCA, DLS-PMCA, VS-control, VS-iβark). Error bar is S.E.M. over mice. Mice show improvement in performance and decrease in the number of extraneous pokes made outside the task window. Analyses throughout the paper nclude data from day 6 onwards (Methods). B. Quantification of virus spread in striatum and adjacent regions for each injection cohort (VS-PMCA, DMS-PMCA, and DLS-PMCA). Numbers in each square indicate the proportion of animals in that cohort that showed bilateral expression of PMCA in each region. Abbreviations: VS = ventral striatum; VP = ventral pallidum; DMS = dorsomedial striatum; DLS = dorsolateral striatum; S1 = primary sensory cortex; M1 = primary motor cortex. Supplemental Figure 2. Dopamine elicits slower and less acute astrocyte calcium response than norepinephrine. A. Heatmap of astrocyte calcium responses to bath application of norepinephrine (left) or dopamine (right) in control slices. Each row represents one cell’s responses to one application (see Methods on cell-based analysis; due to lack of astrocyte response in the PMCA condition cell-based segmentation and analysis was not available for PMCA slices). NE or DA application occurs at 100s. Black lines on top indicate the time scale of DA or NE application. Bottom: average traces of responses to NE or DA application. Shaded area is 95% confidence interval. B. Comparison of the peak responses (maximum responses between 100-300 s) to NE or DA application for control slices, for both field of view (FOV) and cell-based analyses (Wilcoxon rank sum test). C. Comparison of AUC (100-300 s) to NE or DA application; conventions same as B. D. Comparison of rise time (time to maximum response between 100-300 s) to NE or DA application; conventions same as B. E. Comparison of half maximum decay time (time between maximum response and half-max) for NE and DA application; conventions same as B. Supplemental Figure 3. A. Behavioral metrics (% left pokes, number of rewards earned per session, number of trials initiated per session) change pre- vs post-injection for VS-PMCA, DMS-PMCA, DLS-PMCA, and VS-control mice. Attenuating astrocyte calcium signaling in ventral striatum selectively affects task performance, but this is not due to decreased motivation (assayed through the number of choice pokes made per session, and total rewards earned on a session-by-session basis; also see B). DLS-PMCA injected mice show decreased left bias after injection. Since all injections were bilateral, subsequent behavioral experiments will be required to tease out the detailed mechanism for the alteration of the systematic side bias across animals following DLS PMCA injection. For example, future studies ought to explore the relationship of this change in persistent bias and increases in repetitive behavior following ACD attenuation in DLS that have been characterized by Khakh and his colleagues (Xu et al., 2019) B. Correlation between change in performance and change in number of trials initiated per session (top subplot) and change in number of rewards earned per session (bottom subplot), for VS-PMCA mice. There is no correlation between performance vs trial/reward number change pre- vs post-injection, indicating that mice’s decision-making strategy or performance is not directly tied to motivation to initiate trials or reward rate. C. Difference in trials it took for animals vs ideal observers to detect block switches, based on a Bayesian model (Methods). Difference calculated only for block switches where both the ideal observer and the animal detected a switch. As expected, mice consistently were slower to detect block switches compared to ideal observers, and this difference did not change as a function of virus injection. D. Post-win switching (top row) and pre-win switching (bottom row) for DMS-PMCA, DLS-PMCA, and VS-control mice do not show significant differences between pre-injection and post-injection behavior. We note that VS-control mice display a tendency to decrease post-reward switching in the 5-trial window post-block switch, however this cohort’s baseline (pre-injection) post-reward switching was also higher than other cohorts. *= p<0.05, ** = p<0.01, ***=p<0.001. Supplemental Figure 4. VS-PMCA and VS-control mice do not exhibit differences in motivation in a progressive ratio task. A. Pellets earned per hour for VS-PMCA (n = 12) and VS-control (n = 12) mice in a progressive ratio (PR1) task that was run continuously in the home cages of mice (Methods). We ran the PR1 task continuously to assess not only if motivated behavior differed between groups, but also if circadian patterns in motivation/feeding differed. Unsurprisingly, mice were the most active at engaging with the task and earn the most pellets from 6pm-6am during lights off (gray shaded areas), with two peaks in activity right after lights off and before lights on. Overall, there was no difference in pellets earned or circadian pattern of activity between VS-PMCA and VS-control mice. B. Commonly used measures of motivation assayed in progressive ratio tasks, including active pokes and pellets earned by day, active/inactive poke index (ratio of pokes to the active vs inactive port), and breakpoint (number of pokes mice are willing to make for one pellet). There was no significant difference between VS-PMCA and VS-control mice across any of these measures. Supplemental Figure 5. VS-PMCA mice do not show deficits in assays of sensorimotor, cognitive, or motivational function compared to VS-control mice. Naive cohorts of mice were injected bilaterally in VS with either PMCA (n = 9) or control (n = 9) virus and run on a battery of standard behavioral tests commonly used to assay proxies of sensorimotor control, memory, anxiety, cognition, and motivation (see Methods and (J. Chen et al., 2021 )). Note that p-values have not been corrected for multiple comparisons. *= p<0.05, ** = p<0.01, ***=p<0.001. Supplemental Figure 6. Performance and switching aligned to block switches for VS-control and VS-iβark mice. A. Performance (% correct choice rate, top row), switch post-rewards (middle row), and switching post non-rewards (bottom row) aligned to Bayesian-inferred block switches for VS-control and VS-iβark mice. Mice were run on two versions of the task: one where block switches were contingent on rewards earned (reward-contingent), and one where block switches were not contingent on rewards earned (reward non-contingent; see Methods). For the VS-control cohort; mice were injected either with control virus in VS or with saline in the motor cortex overlying VS. These sub-cohorts were collectively analyzed as ‘VS-control’. Lines and error bars are mean and S.E.M. across animals; Wilcoxon sign rank test. B. Pre- vs post-injection change in correct choice rate, win-stay, and lose-switch pokes for VS-control and VS-iβark cohorts. *= p<0.05, ** = p<0.01, ***=p<0.001. Supplemental Figure 7. Performance and switching aligned to block switches for VS-PMCA mice, split out by reward-contingent and reward-non-contingent block switch versions of the task. A. Performance aligned to block switches in mice injected with VS-PMCA, split out by those run on the reward-contingent vs reward non-contingent versions of the task. These sub-cohorts were collectively analyzed as ‘VS-PMCA’ mice. Because splitting out animals by sub-cohort reduced our power to detect effects, to test if there were differences in behavior pre- vs post-injection we tested across all pre- and post-injection trials for each sub-cohort. Attenuating astrocyte calcium in VS reduced performance in mice run on tasks with different reward contingency meta-structure. Lines and error bars are mean and S.E.M. across trials; gray bars indicate 25 trial period pre-block switch where statistical tests (Wilcoxon rank sum test) were run. B. Switching aligned to block switches, for trials following rewarded trials. Attenuating astrocyte calcium in VS affects post-reward switching in mice run on both versions of the task. Conventions follow A. C. Switching aligned to block switches, for trials following non-rewarded trials. There was no significant difference between pre- and post-injection switching for cohorts of mice run on both versions of the task. Conventions follow A. *= p<0.05, ** = p<0.01, ***=p<0.001. Supplemental Figure 8. Parameter recovery for Q-learning model. Parameter recoveries for the Q-learning model using nsim = 1,000 simulations. To align the artificial data with real task statistics, we selected one representative animal and matched both the total number of trials (ntrials = 4,335) and the number of trials per day to that animal’s data (we reinitialized the Q values at the trials corresponding to the first trial of the day). In each simulation, each parameter of the model is drawn randomly from the uniform distribution defined by the range of that parameter, which are the: learning rate 𝛼 ∼ 𝑈(0,1), inverse temperature β ∼ 𝑈(0,20), lapse rate 𝜖 ∼ 𝑈(0, 0.4) and side bias 𝑏 s ∼ 𝑈(−1, 1) . (For the lapse rate term we did not cover the full range (0,1), as the actual fitted values were almost always below 0.4, except for a very small number of the fits). Then the model was simulated with these known parameters, and the fitting procedure was done with the exact same settings as in fitting to the actual data (see Figure 3 and methods). x-axis: the true parameter values, y-axis: recovered parameter values, black dotted line: identity line. A. Top, left: Recovery results for the learning rate 𝛼 (𝑟 = 0.976, 𝑝 = 0.00, 𝑆𝑝𝑒𝑎𝑟𝑚𝑎𝑛′𝑠 𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛) Top, right: Recovery results for the inverse temperature β (𝑟 = 0.870, 𝑝 = 0.00, 𝑆𝑝𝑒𝑎𝑟𝑚𝑎𝑛′𝑠 𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛). Bottom, left: Recovery results for the lapse rate 𝜖 (𝑟 = 0.886, 𝑝 = 0.00, 𝑆𝑝𝑒𝑎𝑟𝑚𝑎𝑛′𝑠 𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛). Bottom, right: Recovery results for the side bias 𝑏 ! (𝑟 = 0.981, 𝑝 = 0.00, 𝑆𝑝𝑒𝑎𝑟𝑚𝑎𝑛′𝑠 𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛). B. Histogram of the number of trials across fitted data. As described above, this particular animal’s data, on which the recovery is done, had n = 4,335 trials. Since parameter recovery fundamentally depends on the number of trials in the data, in order to show that this is representative of the number of trials for the population, we also show the histogram of the number of trials across animals, and denote the example session we chose for the recovery. X-axis: number of trials in each fit, y-axis: histogram counts, black dotted line: mean number of trials, red dotted line: number of trials in the recovery plots in part A. Supplemental Figure 9. Q-learning model validation. Model validations for the Q-learning model. To make sure the fitted Q-learning model is able to capture the key behavioral aspects in the data, we simulated the fitted model on the task data by using the concatenated data for each mouse before and after injection, as the models were fit on the concatenated data. Then we analyzed the behavior of the model in the exact same way we analyzed the mice data. The model captured the same key behavioral effects that we reported in the main figures 2 and 3, namely, a significant decrease in the performance, a significant decrease in the win-stay, and a non-significant change in lose-shift after injection. Note that, despite the model having only one learning rate and inverse temperature terms across all trials, it can capture the decrease in win-stay, while its lose-shift change is non-significant. A. Performance of the model’s simulated behavior with fitted parameter values, before and after injection. The model shows a significant decrease in performance for the VS-PMCA group (p = 0.004), but not for the other DMS-PMCA (p = 0.92) or DLS-PMCA groups (p = 0.46). B. Win-stay values of the model’s simulated behavior with fitted parameter values, before and after injection. The model shows a significant decrease in win-stay for the VS-PMCA group (p = 0.01), but not for the other DMS-PMCA (p = 0.56) or DLS-PMCA groups (p = 0.08). C. Lose-shift values of the simulated model’s simulated behavior with fitted parameter values, before and after injection. The model does not show a significant decrease in lose - shift for VS-PMCA (p = 0.47), DMS-PMCA (p = 0.32), and DLS-PMCA (p = 0.12). Statistical tests were Wilcoxon sign rank tests run pre- vs post-injection. *= p<0.05, ** = p<0.01, ***=p<0.001. Supplemental Figure 10. Psychometric function shifts for all cohorts pre- and post-injection. Q-learning model simulated psychometric functions using behavioral data from all cohorts. Supplemental Figure 11. Q-learning model fits for all cohorts of mice. Q-learning model parameter fits pre- and post-injection for mice in VS-PMCA, DMS-PMCA, DLS-PMCA, VS-control, and VS-iβark cohorts. Gray lines represent individual animal parameters; black line indicates average across animals. Error bars are S.E.M. across animals. Statistical tests were Wilcoxon sign rank tests run pre- vs post-injection. P-values are uncorrected for multiple comparisons. DLS-PMCA mice showed changes in model side bias parameter, consistent with shifts in their overall side bias poking (Supplemental Figure 3A). We note that VS-control mice showed significantly increased learning rate after injection. However, we also note that the VS-control group had significantly lower baseline learning rates compared to the other cohorts (mean VS-control = 0.44; mean of VS-PMCA, DMS-PMCA, and DLS-PMCA pre-injection learning rates = 0.57; p = 0.00, Wilcoxon rank sum). *= p<0.05, ** = p<0.01, ***=p<0.001. Supplemental Figure 12. Performance and GLM models for behavior and ACD in head-fixed task. A. Performance (left; % pokes to the better port) and switching (right) behavior for mice in the head-fixed decision-making task, aligned to Bayesian-inferred block switches. Thick line is averaged behavior over animals (averaged over all blocks for each animal first); error bar is S.E.M. over animals. B. Mean fitted GLM beta weights of regressors representing past trial choice (Left: 1, Right: -1), outcome (Rewarded: 1, Unrewarded: -1), Choice x Outcome, and side bias (constant) on the log odds of choosing left in the current trial. The model was fitted to trials pooled from all included behavioral sessions (Methods) using a logistic link function. We included separate regressors representing those variables 1, 2, and 3 trials prior. In this task Past Choice weights were high because mice often made the same choice as in previous trials (because the better option is consistent for several trials in a row before it reverses), especially the most recent choice. Crucially, however, we also observed a key signature of reinforcement learning, indicating that animals learned by adjusting their choices based on their consequences. Specifically, Past Choice x Outcome had additional predictive value beyond just Past Choice, with larger effects of more recent outcomes. Finally, individual mice had some degree of side bias in their behavior. On average, mice modestly preferred the right option indicated by a negative weight of Constant, and this appears to be stronger following a reward, indicated by negative weights of Past Outcome (though this may be a result of the mice’s tendency to stick to the same side following a reward). Error bars are 1 SE; asterisks represent significance of each weight. *= p<0.05, ** = p<0.01, ***=p<0.001. Supplemental Figure 13. GLM model of ACD in head-fixed task. Mean fitted GLM beta weights of model fitted to VS astrocyte calcium fluorescence 1-3.5 seconds after outcome delivery. The model was fitted to trials pooled from all included behavioral sessions (Methods), including data from 1-3 trials prior, using a linear link function. The model included beta weights of regressors representing the effects of the current outcome, past outcomes from choices of the same side as the current trial, past outcomes from choices of the opposite side, and a constant term. Crucially, we observed a key signature of RPE coding. Because RPE is roughly expressed as “Actual reward - predicted reward,” RPE coding activity in most RL tasks (where animals incrementally build reward predictions about each option based on its past outcomes) should be fit by a positive weight of the current outcome (“Actual reward”) and by negative weights of past outcomes from the same choice, with larger negative weights for more recent outcomes (“Predicted reward”) ( Bayer & Glimcher, 2005 ). Indeed, this is exactly what we observed in the model’s significant positive weight for current outcome and significant negative weights for same choice x outcome (p<0.001). The model also had a significant negative effect of the most recent past trial’s outcome if the animal had chosen the opposite side on that trial. This likely occurs because this effect is based on a rare subset of trials where animals had previously built up an unusually high reward prediction attracting them to the currently chosen side (as indicated by the fact that they chose the current side even in spite of having just gotten a reward from choosing the opposite side on the previous trial), thus resulting in an unusually negative RPE signal in response to the current outcome. *= p<0.05, ** = p<0.01, ***=p<0.001. Supplemental Figure 14. Ventral striatum astrocyte calcium responses to outcome prediction error during long inter-trial intervals. A. Astrocyte calcium responses split out by RPE (conventions following Figure 4H ) but aligned to the start of the next trial. From top to bottom, showing responses to trials where the ITI ranged from short (5 seconds) to long (9 seconds). Calcium transients, particularly to strong positive RPEs, remained elevated for 5-7 after the outcome, which led to an elevated baseline calcium in trials with shorter ITIs. Supplemental Figure 15. Ventral striatum medium spiny neurons do not show differences in post-synaptic temporal dynamics after ACD attenuation. A-B. Histograms of the means obtained from rise (left) or decay (right) times of sEPSCs recorded from NAc D1-MSNs (A) or D2-MSNs (B) in brain slices from PMCA (D1: n = 11 cells, 5 mice; D2: n = 19 cells; 6 mice) or control mice (D1: n = 8 cells, 4 mice; D2: n = 18 cells; 6 mice). C-D. Histograms of the means obtained from rise (left) or decay (right) times of sIPSCs recorded from NAc D1-MSNs (C) or D2-MSNs (D) in brain slices from PMCA (D1: n = 13 cells, 5 mice; D2: n = 10 cells; 4 mice) or control mice (D1: n = 12 cells, 4 mice; D2: n = 10 cells; 4 mice). Data are mean ± S.E.M. ns, not significant. Supplemental Figure 16. Circuit model variability as a function of baseline shift and input sharing. A. Illustration of the variability in the firing rates in the presence and absence of baseline shift (blue and black shades). B. The difference in the ratio of poke to better port, win stay and lose shift between models with or without baseline shifts. C. Illustration of the variability in the RPE signal δi in the presence and absence of input-sharing (black vs red shades). D. The difference in the ratio of poke to better port, win stay and lose shift between models with or without input sharing. Acknowledgements This work was seeded by the MURI DOD grant to Fabio Pasqualetti, Samet Oymak, Bruno Sinopoli, Thomas Papouin, ShiNung Ching, and Ilya Monosov. Subsequent funding was provided by support from Washington University School of Medicine (to IEM), the National Institute of Mental Health (F31MH135646 to JP), the McDonnell Center for Systems Neuroscience, McDonnell Center for Cellular and Molecular Neuroscience, and the Taylor Family Institute at Washington University in St. Louis. We thank Drs. Bruno Averbeck and Ethan S. Bromberg-Martin for modelling advice, and Drs. Ethan S. Bromberg-Martin and Daeyeol Lee for detailed comments about the manuscript during its preparation. We thank Dr. Robert Gereau for providing necessary resources. The Animal Behavior Core at Washington University performed behavioral testing. We thank Rebecca Mellor for assistance with tissue processing and subsequent tissue imaging. Funder Information Declared Army Research Office , 78259-NS-MUR National Institute of Mental Health , F31MH135646 References ↵ Adamsky , A. , Kol , A. , Kreisel , T. , Doron , A. , Ozeri-Engelhard , N. , Melcer , T. , Refaeli , R. , Horn , H. , Regev , L. , Groysman , M. , London , M. , & Goshen , I . ( 2018 ). Astrocytic Activation Generates De Novo Neuronal Potentiation and Memory Enhancement . Cell , 174 ( 1 ), 59 – 71 .e14. doi: 10.1016/j.cell.2018.05.002 OpenUrl CrossRef PubMed ↵ Adermark , L. , Stomberg , R. , Söderpalm , B. , & Ericson , M . ( 2024 ). Astrocytic Regulation of Endocannabinoid-Dependent Synaptic Plasticity in the Dorsolateral Striatum . International Journal of Molecular Sciences , 25 ( 1 ), 581 . doi: 10.3390/ijms25010581 OpenUrl CrossRef PubMed ↵ Alexander , G. E. , Delong , Mahlon R , & Strick , Peter L . ( 1986 ). Parallel organization of functionally segregated circuits linking basal ganglia and cortex. Ann . Rev. Neurosci ., 9 , 357 – 381 . OpenUrl CrossRef ↵ Allen , N. J . ( 2014 ). Astrocyte Regulation of Synaptic Behavior . Annual Review of Cell and Developmental Biology , 30 (Volume 30, 2014), 439 – 463 . doi: 10.1146/annurev-cellbio-100913-013053 OpenUrl CrossRef PubMed ↵ Araque , A. , Parpura , V. , Sanzgiri , R. P. , & Haydon , P. G . ( 1999 ). Tripartite synapses: Glia, the unacknowledged partner . Trends in Neurosciences , 22 ( 5 ), 208 – 215 . doi: 10.1016/S0166-2236(98)01349-6 OpenUrl CrossRef PubMed Web of Science ↵ Averbeck , B. , & O’Doherty , J. P . ( 2022 ). Reinforcement-learning in fronto-striatal circuits . Neuropsychopharmacology , 47 ( 1 ), 147 – 162 . doi: 10.1038/s41386-021-01108-0 OpenUrl CrossRef PubMed ↵ Bai , Y. , Zhou , Z. , Han , B. , Xiang , X. , Huang , W. , & Yao , H . ( 2024 ). Revisiting astrocytic calcium signaling in the brain . Fundamental Research , 4 ( 6 ), 1365 – 1374 . doi: 10.1016/j.fmre.2023.11.021 OpenUrl CrossRef PubMed ↵ Balleine , B. W. , & O’Doherty , J. P . ( 2010 ). Human and Rodent Homologies in Action Control: Corticostriatal Determinants of Goal-Directed and Habitual Action . Neuropsychopharmacology , 35 ( 1 ), 48 – 69 . doi: 10.1038/npp.2009.131 OpenUrl CrossRef PubMed Web of Science ↵ Ballesta , S. , Shi , W. , Conen , K. E. , & Padoa-Schioppa , C . ( 2020 ). Values encoded in orbitofrontal cortex are causally related to economic choices . Nature , 588 ( 7838 ), 450 – 453 . doi: 10.1038/s41586-020-2880-x OpenUrl CrossRef PubMed ↵ Bartolo , R. , & Averbeck , B. B . ( 2020 ). Prefrontal Cortex Predicts State Switches during Reversal Learning . Neuron , 106 ( 6 ), 1044 – 1054 .e4. doi: 10.1016/j.neuron.2020.03.024 OpenUrl CrossRef PubMed ↵ Bayer , H. M. , & Glimcher , P. W . ( 2005 ). Midbrain dopamine neurons encode a quantitative reward prediction error signal . Neuron , 47 ( 1 ), 129 – 141 . doi: 10.1016/j.neuron.2005.05.020 OpenUrl CrossRef PubMed Web of Science ↵ Becker , S. , Nold , A. , & Tchumatchenko , T . ( 2022 ). Modulation of working memory duration by synaptic and astrocytic mechanisms . PLOS Computational Biology , 18 ( 10 ), e1010543 . doi: 10.1371/journal.pcbi.1010543 OpenUrl CrossRef ↵ Beeler , J. A. , Daw , N. , Frazier , C. R. M. , & Zhuang , X . ( 2010 ). Tonic dopamine modulates exploitation of reward learning . Frontiers in Behavioral Neuroscience , 4 . doi: 10.3389/fnbeh.2010.00170 OpenUrl CrossRef PubMed ↵ Bekar , L. K. , He , W. , & Nedergaard , M . ( 2008 ). Locus Coeruleus α-Adrenergic–Mediated Activation of Cortical Astrocytes In Vivo . Cerebral Cortex , 18 ( 12 ), 2789 – 2795 . doi: 10.1093/cercor/bhn040 OpenUrl CrossRef PubMed Web of Science ↵ Benoit , L. , Hristovska , I. , Liaudet , N. , Jouneau , P.-H. , Fertin , A. , Ceglia , R. de , Litvin , D. , Castro , M. A. D. , Jevtic , M. , Zalachoras , I. , Kikuchi , T. , Telley , L. , Bergami , M. , Usson , Y. , Hisatsune , C. , Mikoshiba , K. , Pernet-Gallay , K. , & Volterra , A. ( 2025 ). Astrocytes functionally integrate multiple synapses via specialized leaflet domains . Cell , 0(0). doi: 10.1016/j.cell.2025.08.036 OpenUrl CrossRef ↵ Berridge , M. J . ( 1998 ). Neuronal Calcium Signaling . Neuron , 21 ( 1 ), 13 – 26 . doi: 10.1016/S0896-6273(00)80510-3 OpenUrl CrossRef PubMed Web of Science ↵ Bertsekas , D. , & Tsitsiklis , J. N . ( 1996 ). Neuro-Dynamic Programming . Athena Scientific . ↵ Britten , K. H. , Newsome , W. T. , Shadlen , M. N. , Celebrini , S. , & Movshon , J. A. ( 1996 ). A relationship between behavioral choice and the visual responses of neurons in macaque MT . Visual Neuroscience , 13 ( 1 ), 87 – 100 . doi: 10.1017/S095252380000715X OpenUrl CrossRef PubMed Web of Science ↵ Bull , C. , Freitas , K. C. , Zou , S. , Poland , R. S. , Syed , W. A. , Urban , D. J. , Minter , S. C. , Shelton , K. L. , Hauser , K. F. , Negus , S. S. , Knapp , P. E. , & Bowers , M. S . ( 2014 ). Rat Nucleus Accumbens Core Astrocytes Modulate Reward and the Motivation to Self-Administer Ethanol after Abstinence . Neuropsychopharmacology , 39 ( 12 ), 2835 – 2845 . doi: 10.1038/npp.2014.135 OpenUrl CrossRef PubMed ↵ Burton , A. C. , Nakamura , K. , & Roesch , M. R . ( 2015 ). From ventral-medial to dorsal-lateral striatum: Neural correlates of reward-guided decision-making . Neurobiology of Learning and Memory , 117 , 51 – 59 . doi: 10.1016/j.nlm.2014.05.003 OpenUrl CrossRef PubMed ↵ Bushong , E. A. , Martone , M. E. , Jones , Y. Z. , & Ellisman , M. H . ( 2002 ). Protoplasmic Astrocytes in CA1 Stratum Radiatum Occupy Separate Anatomical Domains . Journal of Neuroscience , 22 ( 1 ), 183 – 192 . doi: 10.1523/JNEUROSCI.22-01-00183.2002 OpenUrl Abstract / FREE Full Text ↵ Calipari , E. S. , Bagot , R. C. , Purushothaman , I. , Davidson , T. J. , Yorgason , J. T. , Peña , C. J. , Walker , D. M. , Pirpinias , S. T. , Guise , K. G. , Ramakrishnan , C. , Deisseroth , K. , & Nestler , E. J . ( 2016 ). In vivo imaging identifies temporal signature of D1 and D2 medium spiny neurons in cocaine reward . Proceedings of the National Academy of Sciences , 113 ( 10 ), 2726 – 2731 . doi: 10.1073/pnas.1521238113 OpenUrl Abstract / FREE Full Text ↵ Castro , D. C. , & Bruchas , M. R . ( 2019 ). A Motivational and Neuropeptidergic Hub: Anatomical and Functional Diversity within Nucleus Accumbens Shell . Neuron , 102 ( 3 ), 529 – 552 . doi: 10.1016/j.neuron.2019.03.003 OpenUrl CrossRef PubMed ↵ Chai , H. , Diaz-Castro , B. , Shigetomi , E. , Monte , E. , Octeau , J. C. , Yu , X. , Cohn , W. , Rajendran , P. S. , Vondriska , T. M. , Whitelegge , J. P. , Coppola , G. , & Khakh , B. S . ( 2017 ). Neural Circuit-Specialized Astrocytes: Transcriptomic, Proteomic, Morphological, and Functional Evidence . Neuron , 95 ( 3 ), 531 – 549 .e9. doi: 10.1016/j.neuron.2017.06.029 OpenUrl CrossRef PubMed ↵ Chakroun , K. , Mathar , D. , Wiehler , A. , Ganzer , F. , & Peters , J . ( 2023 ). Dopaminergic modulation of the exploration/exploitation trade-off in human decision-making . Nature Communications , 14 ( 5369 ), e51260 . doi: 10.7554/eLife.51260 OpenUrl CrossRef ↵ Charles , A. C. , Merrill , J. E. , Dirksen , E. R. , & Sandersont , M. J . ( 1991 ). Intercellular signaling in glial cells: Calcium waves and oscillations in response to mechanical stimulation and glutamate . Neuron , 6 ( 6 ), 983 – 992 . doi: 10.1016/0896-6273(91)90238-U OpenUrl CrossRef PubMed Web of Science ↵ Chen , A. B. , Duque , M. , Rymbek , A. , Dhanasekar , M. , Wang , V. M. , Mi , X. , Tocquer , L. , Narayan , S. , Legorreta , E. M. , Eddison , M. , Yu , G. , Wyart , C. , Prober , D. A. , Engert , F. , & Ahrens , M. B . ( 2025 ). Norepinephrine changes behavioral state through astroglial purinergic signaling. Science (New York , N.Y .) , 388 ( 6748 ), 769 – 775 . doi: 10.1126/science.adq5233 OpenUrl CrossRef PubMed ↵ Chen , A. B. , Duque , M. , Wang , V. M. , Dhanasekar , M. , Mi , X. , Rymbek , A. , Tocquer , L. , Narayan , S. , Prober , D. , Yu , G. , Wyart , C. , Engert , F. , & Ahrens , M. B . ( 2024 ). Norepinephrine changes behavioral state via astroglial purinergic signaling (p. 2024.05.23.595576). bioRxiv. doi: 10.1101/2024.05.23.595576 OpenUrl Abstract / FREE Full Text ↵ Chen , J. , Lambo , M. E. , Ge , X. , Dearborn , J. T. , Liu , Y. , McCullough , K. B. , Swift , R. G. , Tabachnick , D. R. , Tian , L. , Noguchi , K. , Garbow , J. R. , Constantino , J. N. , Gabel , H. W. , Hengen , K. B. , Maloney , S. E. , & Dougherty , J. D . ( 2021 ). A MYT1L syndrome mouse model recapitulates patient phenotypes and reveals altered brain development due to disrupted neuronal maturation . Neuron , 109 ( 23 ), 3775 – 3792 .e14. doi: 10.1016/j.neuron.2021.09.009 OpenUrl CrossRef PubMed ↵ Cohen , M. X . ( 2008 ). Neurocomputational mechanisms of reinforcement-guided learning in humans: A review. Cognitive, Affective , & Behavioral Neuroscience , 8 ( 2 ), 113 – 125 . doi: 10.3758/CABN.8.2.113 OpenUrl CrossRef PubMed Web of Science ↵ Collins , A. G. E. , & Frank , M. J . ( 2014 ). Opponent actor learning (OpAL): Modeling interactive effects of striatal dopamine on reinforcement learning and choice incentive . Psychological Review , 121 ( 3 ), 337 – 366 . doi: 10.1037/a0037015 OpenUrl CrossRef PubMed ↵ Cooper , M. L. , Selles , M. C. , Cammer , M. , Gildea , H. K. , Sall , J. , Chiurri , K. E. , Saab , A. S. , Liddelow , S. A. , & Chao , M. V . ( 2025 ). Astrocytes connect specific brain regions through plastic gap junctional networks . bioRxiv: The Preprint Server for Biology , 2025.07.18.665573. doi: 10.1101/2025.07.18.665573 OpenUrl Abstract / FREE Full Text ↵ Corkrum , M. , Covelo , A. , Lines , J. , Bellocchio , L. , Pisansky , M. , Loke , K. , Quintana , R. , Rothwell , P. E. , Lujan , R. , Marsicano , G. , Martin , E. D. , Thomas , M. J. , Kofuji , P. , & Araque , A . ( 2020 ). Dopamine-Evoked Synaptic Regulation in the Nucleus Accumbens Requires Astrocyte Activity . Neuron , 105 ( 6 ), 1036 – 1047 .e5. doi: 10.1016/j.neuron.2019.12.026 OpenUrl CrossRef PubMed ↵ Corkrum , M. , Rothwell , P. E. , Thomas , M. J. , Kofuji , P. , & Araque , A . ( 2019 ). Opioid-Mediated Astrocyte–Neuron Signaling in the Nucleus Accumbens . Cells , 8 ( 6 ), Article 6. doi: 10.3390/cells8060586 OpenUrl CrossRef ↵ Cornell-Bell , A. H. , Finkbeiner , S. M. , Cooper , M. S. , & Smith , S. J . ( 1990 ). Glutamate Induces Calcium Waves in Cultured Astrocytes: Long-Range Glial Signaling . Science , 247 ( 4941 ), 470 – 473 . doi: 10.1126/science.1967852 OpenUrl Abstract / FREE Full Text ↵ Costa , V. D. , Monte , O. D. , Lucas , D. R. , Murray , E. A. , & Averbeck , B. B . ( 2016 ). Amygdala and Ventral Striatum Make Distinct Contributions to Reinforcement Learning . Neuron , 92 ( 2 ), 505 – 517 . doi: 10.1016/j.neuron.2016.09.025 OpenUrl CrossRef PubMed ↵ Costa , V. D. , Tran , V. L. , Turchi , J. , & Averbeck , B. B . ( 2015 ). Reversal learning and dopamine: A Bayesian perspective . Journal of Neuroscience , 35 ( 6 ), 2407 – 2416 . doi: 10.1523/JNEUROSCI.1989-14.2015 OpenUrl Abstract / FREE Full Text ↵ Cox , J. , & Witten , I. B . ( 2019 ). Striatal circuits for reward learning and decision-making . Nature Reviews Neuroscience , 20 ( 8 ), 482 – 494 . doi: 10.1038/s41583-019-0189-2 OpenUrl CrossRef PubMed ↵ Cui , G. , Jun , S. B. , Jin , X. , Pham , M. D. , Vogel , S. S. , Lovinger , D. M. , & Costa , R. M . ( 2013 ). Concurrent activation of striatal direct and indirect pathways during action initiation . Nature , 494 ( 7436 ), 238 – 242 . doi: 10.1038/nature11846 OpenUrl CrossRef PubMed Web of Science ↵ Dabney , W. , Kurth-Nelson , Z. , Uchida , N. , Starkweather , C. K. , Hassabis , D. , Munos , R. , & Botvinick , M . ( 2020 ). A distributional code for value in dopamine-based reinforcement learning . Nature , 577 ( 7792 ), 671 – 675 . doi: 10.1038/s41586-019-1924-6 OpenUrl CrossRef PubMed ↵ Dalton , G. L. , Phillips , A. G. , & Floresco , S. B . ( 2014 ). Preferential involvement by nucleus accumbens shell in mediating probabilistic learning and reversal shifts . Journal of Neuroscience , 34 ( 13 ), 4618 – 4626 . Scopus. doi: 10.1523/JNEUROSCI.5058-13.2014 OpenUrl Abstract / FREE Full Text ↵ D’Ascenzo , M. , Fellin , T. , Terunuma , M. , Revilla-Sanchez , R. , Meaney , D. F. , Auberson , Y. P. , Moss , S. J. , & Haydon , P. G . ( 2007 ). mGluR5 stimulates gliotransmission in the nucleus accumbens . Proceedings of the National Academy of Sciences , 104 ( 6 ), 1995 – 2000 . doi: 10.1073/pnas.0609408104 OpenUrl Abstract / FREE Full Text ↵ Daw , N. D. , O’Doherty , J. P. , Dayan , P. , Seymour , B. , & Dolan , R. J . ( 2006 ). Cortical substrates for exploratory decisions in humans . Nature , 441 ( 7095 ), 876 – 879 . doi: 10.1038/nature04766 OpenUrl CrossRef PubMed Web of Science ↵ De Pittà , M. , Brunel , N. , & Volterra , A. ( 2016 ). Astrocytes: Orchestrating synaptic plasticity? Neuroscience , 323 , 43 – 61 . doi: 10.1016/j.neuroscience.2015.04.001 OpenUrl CrossRef PubMed ↵ De Pittà , M. , Volman , V. , Berry , H. , & Ben-Jacob , E. ( 2011 ). A Tale of Two Stories: Astrocyte Regulation of Synaptic Depression and Facilitation . PLoS Computational Biology , 7 ( 12 ), e1002293 . doi: 10.1371/journal.pcbi.1002293 OpenUrl CrossRef PubMed ↵ Dermietzel , R. , Hertberg , E. L. , Kessler , J. A. , & Spray , D. C . ( 1991 ). Gap junctions between cultured astrocytes: Immunocytochemical, molecular, and electrophysiological analysis . Journal of Neuroscience , 11 ( 5 ), 1421 – 1432 . doi: 10.1523/JNEUROSCI.11-05-01421.1991 OpenUrl Abstract / FREE Full Text ↵ Di Castro , M. A. , Chuquet , J. , Liaudet , N. , Bhaukaurally , K. , Santello , M. , Bouvier , D. , Tiret , P. , & Volterra , A. ( 2011 ). Local Ca2+ detection and modulation of synaptic release by astrocytes . Nature Neuroscience , 14 ( 10 ), 1276 – 1284 . doi: 10.1038/nn.2929 OpenUrl CrossRef PubMed ↵ Di Chiara , G. ( 2002 ). Nucleus accumbens shell and core dopamine: Differential role in behavior and addiction . Behavioural Brain Research , 137 ( 1 ), 75 – 114 . doi: 10.1016/S0166-4328(02)00286-3 OpenUrl CrossRef PubMed Web of Science ↵ Doron , A. , Rubin , A. , Benmelech-Chovav , A. , Benaim , N. , Carmi , T. , Refaeli , R. , Novick , N. , Kreisel , T. , Ziv , Y. , & Goshen , I . ( 2022 ). Hippocampal astrocytes encode reward location . Nature , 609 ( 7928 ), 772 – 778 . doi: 10.1038/s41586-022-05146-6 OpenUrl CrossRef ↵ Faust , T. W. , Mohebi , A. , & Berke , J. D . ( 2025 ). Reward expectation and receipt differentially modulate the spiking of accumbens D1+ and D2+ neurons . Current Biology , 35 ( 6 ), 1285 – 1297 .e3. doi: 10.1016/j.cub.2025.02.007 OpenUrl CrossRef PubMed ↵ Favetta , G. , & Bubacco , L . ( 2025 ). Beyond neurons: How does dopamine signaling impact astrocytic functions and pathophysiology? Progress in Neurobiology , 251 , 102798 . doi: 10.1016/j.pneurobio.2025.102798 OpenUrl CrossRef PubMed ↵ Frank , M. J . ( 2025 ). Adaptive Cost-Benefit Control Fueled by Striatal Dopamine . Annual Review of Neuroscience , 48 ( 1 ), 1 – 22 . doi: 10.1146/annurev-neuro-112723-025228 OpenUrl CrossRef PubMed ↵ Frank , M. J. , Seeberger , L. C. , & O’Reilly , R. C . ( 2004 ). By Carrot or by Stick: Cognitive Reinforcement Learning in Parkinsonism . Science , 306 ( 5703 ), 1940 – 1943 . doi: 10.1126/science.1102941 OpenUrl Abstract / FREE Full Text ↵ Franklin , N. T. , & Frank , M. J . ( 2015 ). A cholinergic feedback circuit to regulate striatal population uncertainty and optimize reinforcement learning . eLife , 4 ( 12029 ). doi: 10.7554/eLife.12029 OpenUrl CrossRef ↵ Fujii , Y. , Maekawa , S. , & Morita , M . ( 2017 ). Astrocyte calcium waves propagate proximally by gap junction and distally by extracellular diffusion of ATP released from volume-regulated anion channels . Scientific Reports , 7 ( 1 ), 13115 . doi: 10.1038/s41598-017-13243-0 OpenUrl CrossRef PubMed ↵ Fukushima , K . ( 1980 ). Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position . Biological Cybernetics , 36 ( 4 ), 193 – 202 . doi: 10.1007/BF00344251 OpenUrl CrossRef PubMed Web of Science ↵ Gerasimov , E. , Erofeev , A. , Borodinova , A. , Bolshakova , A. , Balaban , P. , Bezprozvanny , I. , & Vlasova , O. L . ( 2021 ). Optogenetic Activation of Astrocytes—Effects on Neuronal Network Function . International Journal of Molecular Sciences , 22 ( 17 ), 9613 . doi: 10.3390/ijms22179613 OpenUrl CrossRef ↵ Gerfen , C. R. , & Surmeier , D. J . ( 2011 ). Modulation of striatal projection systems by dopamine . Annual Review of Neuroscience , 34 , 441 – 466 . doi: 10.1146/annurev-neuro-061010-113641 OpenUrl CrossRef PubMed Web of Science ↵ Glimcher , P. W . ( 2011 ). Understanding dopamine and reinforcement learning: The dopamine reward prediction error hypothesis . Proceedings of the National Academy of Sciences , 108 (supplement_3), 15647–15654. doi: 10.1073/pnas.1014269108 OpenUrl Abstract / FREE Full Text ↵ Godynyuk , E. , Bluitt , M. N. , Tooley , J. R. , Kravitz , A. V. , & Creed , M. C . ( 2019 ). An Open-Source, Automated Home-Cage Sipper Device for Monitoring Liquid Ingestive Behavior in Rodents . eNeuro , 6 ( 5 ), ENEURO.0292-19.2019. doi: 10.1523/ENEURO.0292-19.2019 OpenUrl Abstract / FREE Full Text ↵ Gold , J. I. , & Ding , L . ( 2013 ). How mechanisms of perceptual decision-making affect the psychometric function . Progress in Neurobiology , 103 , 98 – 114 . doi: 10.1016/j.pneurobio.2012.05.008 OpenUrl CrossRef PubMed ↵ Gong , L. , Pasqualetti , F. , Papouin , T. , & Ching , S . ( 2024 ). Astrocytes as a mechanism for contextually-guided network dynamics and function . PLOS Computational Biology , 20 ( 5 ), e1012186 . doi: 10.1371/journal.pcbi.1012186 OpenUrl CrossRef ↵ Grossman , C. D. , Bari , B. A. , & Cohen , J. Y . ( 2022 ). Serotonin neurons modulate learning rate through uncertainty . Current Biology , 32 ( 3 ), 586 – 599 .e7. doi: 10.1016/j.cub.2021.12.006 OpenUrl CrossRef PubMed ↵ Guttenplan , K. A. , Maxwell , I. , Santos , E. , Borchardt , L. A. , Manzo , E. , Abalde-Atristain , L. , Kim , R. D. , & Freeman , M. R . ( 2025 ). GPCR signaling gates astrocyte responsiveness to neurotransmitters and control of neuronal activity . Science , 388 ( 6748 ), 763 – 768 . doi: 10.1126/science.adq5729 OpenUrl CrossRef PubMed ↵ Haber , S. N . ( 2003 ). The primate basal ganglia: Parallel and integrative networks . Journal of Chemical Neuroanatomy , 26 ( 4 ), 317 – 330 . doi: 10.1016/j.jchemneu.2003.10.003 OpenUrl CrossRef PubMed Web of Science ↵ Haber , S. N. , & Knutson , B . ( 2010 ). The Reward Circuit: Linking Primate Anatomy and Human Imaging . Neuropsychopharmacology , 35 ( 1 ), 4 – 26 . doi: 10.1038/npp.2009.129 OpenUrl CrossRef PubMed Web of Science ↵ Halassa , M. M. , Fellin , T. , Takano , H. , Dong , J.-H. , & Haydon , P. G . ( 2007 ). Synaptic Islands Defined by the Territory of a Single Astrocyte . The Journal of Neuroscience , 27 ( 24 ), 6473 – 6477 . doi: 10.1523/JNEUROSCI.1419-07.2007 OpenUrl Abstract / FREE Full Text ↵ Haydon , P. G. , & Carmignoto , G . ( 2006 ). Astrocyte Control of Synaptic Transmission and Neurovascular Coupling . Physiological Reviews , 86 ( 3 ), 1009 – 1031 . doi: 10.1152/physrev.00049.2005 OpenUrl CrossRef PubMed Web of Science ↵ Heiman , M. , Schaefer , A. , Gong , S. , Peterson , J. D. , Day , M. , Ramsey , K. E. , Suárez-Fariñas , M. , Schwarz , C. , Stephan , D. A. , Surmeier , D. J. , Greengard , P. , & Heintz , N . ( 2008 ). A Translational Profiling Approach for the Molecular Characterization of CNS Cell Types . Cell , 135 ( 4 ), 738 – 748 . doi: 10.1016/j.cell.2008.10.028 OpenUrl CrossRef PubMed Web of Science ↵ Henneberger , C. , Papouin , T. , Oliet , S. H. R. , & Rusakov , D. A . ( 2010 ). Long term potentiation depends on release of D-serine from astrocytes . Nature , 463 ( 7278 ), 232 – 236 . doi: 10.1038/nature08673 OpenUrl CrossRef PubMed Web of Science ↵ Humphries , M. D. , Khamassi , M. , & Gurney , K . ( 2012 ). Dopaminergic control of the exploration-exploitation trade-off via the basal ganglia . Frontiers in Neuroscience , 6 . doi: 10.3389/fnins.2012.00009 OpenUrl CrossRef PubMed ↵ Ineichen , C. , Sigrist , H. , Spinelli , S. , Lesch , K.-P. , Sautter , E. , Seifritz , E. , & Pryce , C. R . ( 2012 ). Establishing a probabilistic reversal learning test in mice: Evidence for the processes mediating reward-stay and punishment-shift behaviour and for their modulation by serotonin . Neuropharmacology , 63 ( 6 ), 1012 – 1021 . doi: 10.1016/j.neuropharm.2012.07.025 OpenUrl CrossRef PubMed Web of Science ↵ Irizarry-Valle , Y. , & Parker , A. C . ( 2015 ). An Astrocyte Neuromorphic Circuit That Influences Neuronal Phase Synchrony . IEEE Transactions on Biomedical Circuits and Systems , 9 ( 2 ), 175 – 187 . IEEE Transactions on Biomedical Circuits and Systems. doi: 10.1109/TBCAS.2015.2417580 OpenUrl CrossRef ↵ Ito , M. , & Doya , K . ( 2009 ). Validation of Decision-Making Models and Analysis of Decision Variables in the Rat Basal Ganglia . Journal of Neuroscience , 29 ( 31 ), 9861 – 9874 . doi: 10.1523/JNEUROSCI.6157-08.2009 OpenUrl Abstract / FREE Full Text ↵ Ito , M. , & Doya , K . ( 2011 ). Multiple representations and algorithms for reinforcement learning in the cortico-basal ganglia circuit . Current Opinion in Neurobiology , 21 ( 3 ), 368 – 373 . doi: 10.1016/j.conb.2011.04.001 OpenUrl CrossRef PubMed ↵ Johnson , A. R. , Christensen , B. A. , Kelly , S. J. , & Calipari , E. S . ( 2022 ). The influence of reinforcement schedule on experience-dependent changes in motivation . Journal of the Experimental Analysis of Behavior , 117 ( 3 ), 320 – 330 . doi: 10.1002/jeab.755 OpenUrl CrossRef PubMed ↵ Kang , S. , Hong , S.-I. , Lee , J. , Peyton , L. , Baker , M. , Choi , S. , Kim , H. , Chang , S.-Y. , & Choi , D.- S . ( 2020 ). Activation of Astrocytes in the Dorsomedial Striatum Facilitates Transition From Habitual to Goal-Directed Reward-Seeking Behavior . Biological Psychiatry , 88 ( 10 ), 797 – 808 . doi: 10.1016/j.biopsych.2020.04.023 OpenUrl CrossRef PubMed ↵ Keeler , J. F. , Pretsell , D. O. , & Robbins , T. W . ( 2014 ). Functional implications of dopamine D1 vs. D2 receptors: A ‘prepare and select’ model of the striatal direct vs. indirect pathways . Neuroscience , 282 , 156 – 175 . doi: 10.1016/j.neuroscience.2014.07.021 OpenUrl CrossRef PubMed ↵ Kelley , A. E . ( 2004 ). Ventral striatal control of appetitive motivation: Role in ingestive behavior and reward-related learning . Neuroscience & Biobehavioral Reviews , 27 ( 8 ), 765 – 776 . doi: 10.1016/j.neubiorev.2003.11.015 OpenUrl CrossRef PubMed Web of Science ↵ Khakh , B. S . ( 2025 ). On astrocyte-neuron interactions: Broad insights from the striatum . Neuron , 0 ( 0 ). doi: 10.1016/j.neuron.2025.08.009 OpenUrl CrossRef ↵ Khakh , B. S. , & McCarthy , K. D . ( 2015 ). Astrocyte calcium signaling: From observations to functions and the challenges therein . Cold Spring Harbor Perspectives in Biology , 7 ( 4 ), a020404 . doi: 10.1101/cshperspect.a020404 OpenUrl Abstract / FREE Full Text ↵ Kim , H. F. , & Hikosaka , O . ( 2013 ). Distinct basal ganglia circuits controlling behaviors guided by flexible and stable values . Neuron , 79 ( 5 ), 1001 – 1010 . doi: 10.1016/j.neuron.2013.06.044 OpenUrl CrossRef PubMed Web of Science ↵ Kim , H. , Sul , J. H. , Huh , N. , Lee , D. , & Jung , M. W . ( 2009 ). Role of striatum in updating values of chosen actions . The Journal of Neuroscience: The Official Journal of the Society for Neuroscience , 29 ( 47 ), 14701 – 14712 . doi: 10.1523/JNEUROSCI.2728-09.2009 OpenUrl Abstract / FREE Full Text ↵ Kim , N. , Li , H. E. , Hughes , R. N. , Watson , G. D. R. , Gallegos , D. , West , A. E. , Kim , I. H. , & Yin , H. H . ( 2019 ). A striatal interneuron circuit for continuous target pursuit . Nature Communications , 10 ( 1 ), 2715 . doi: 10.1038/s41467-019-10716-w OpenUrl CrossRef PubMed ↵ Kipf , T. N. , & Welling , M . ( 2017 ). Semi-Supervised Classification with Graph Convolutional Networks (No. arXiv:1609.02907). arXiv. doi: 10.48550/arXiv.1609.02907 OpenUrl CrossRef ↵ Kozachkov , L. , Kastanenka , K. V. , & Krotov , D . ( 2023 ). Building transformers from neurons and astrocytes . Proceedings of the National Academy of Sciences , 120 ( 34 ), e2219150120 . doi: 10.1073/pnas.2219150120 OpenUrl CrossRef PubMed ↵ Kravitz , A. V. , Tye , L. D. , & Kreitzer , A. C . ( 2012 ). Distinct roles for direct and indirect pathway striatal neurons in reinforcement . Nature Neuroscience , 15 ( 6 ), 816 – 818 . doi: 10.1038/nn.3100 OpenUrl CrossRef PubMed ↵ Kreitzer , A. C. , & Malenka , R. C . ( 2008 ). Striatal Plasticity and Basal Ganglia Circuit Function . Neuron , 60 ( 4 ), 543 – 554 . doi: 10.1016/j.neuron.2008.11.005 OpenUrl CrossRef PubMed Web of Science ↵ Krizhevsky , A. , Sutskever , I. , & Hinton , G. E . ( 2012 ). ImageNet classification with deep convolutional neural networks . Communications of the ACM , 60 ( 6 ), 84 – 90 . doi: 10.1145/3065386 OpenUrl CrossRef ↵ Krizhevsky , A. , Sutskever , I. , & Hinton , G. E . ( 2017 ). ImageNet classification with deep convolutional neural networks . Communications of the ACM , 60 ( 6 ), 84 – 90 . doi: 10.1145/3065386 OpenUrl CrossRef ↵ Kwak , S. , & Jung , M. W . ( 2019 ). Distinct roles of striatal direct and indirect pathways in value-based decision making . eLife , 8 , e46050 . doi: 10.7554/eLife.46050 OpenUrl CrossRef PubMed ↵ Lam , N. H. , Borduqui , T. , Hallak , J. , Roque , A. , Anticevic , A. , Krystal , J. H. , Wang , X.-J. , & Murray , J. D . ( 2022 ). Effects of Altered Excitation-Inhibition Balance on Decision Making in a Cortical Circuit Model . Journal of Neuroscience , 42 ( 6 ), 1035 – 1053 . doi: 10.1523/JNEUROSCI.1371-20.2021 OpenUrl Abstract / FREE Full Text ↵ Lazaridis , I. , Ahn , G. , Hirokane , K. , Choi , W. , & Graybiel , A. M . ( 2024 ). Striatal Astrocytes Influence Dopamine Dynamics and Behavioral State Transitions (p. 2024.12.01.626240). bioRxiv. doi: 10.1101/2024.12.01.626240 OpenUrl Abstract / FREE Full Text ↵ LeCun , Y. , Boser , B. , Denker , J. S. , Henderson , D. , Howard , R. E. , Hubbard , W. , & Jackel , L. D . ( 1989 ). Backpropagation Applied to Handwritten Zip Code Recognition . Neural Computation , 1 ( 4 ), 541 – 551 . doi: 10.1162/neco.1989.1.4.541 OpenUrl CrossRef ↵ Lee , D. , Seo , H. , & Jung , M. W . ( 2012 ). Neural basis of reinforcement learning and decision making . Annual Review of Neuroscience , 35 , 287 – 308 . doi: 10.1146/annurev-neuro-062111-150512 OpenUrl CrossRef PubMed Web of Science ↵ Lefton , K. B. , Wu , Y. , Dai , Y. , Okuda , T. , Zhang , Y. , Yen , A. , Rurak , G. M. , Walsh , S. , Manno , R. , Myagmar , B.-E. , Dougherty , J. D. , Samineni , V. K. , Simpson , P. C. , & Papouin , T . ( 2025 ). Norepinephrine signals through astrocytes to modulate synapses . Science , 388 ( 6748 ), 776 – 783 . doi: 10.1126/science.adq5480 OpenUrl CrossRef PubMed ↵ Lowet , A. S. , Zheng , Q. , Meng , M. , Matias , S. , Drugowitsch , J. , & Uchida , N . ( 2025 ). An opponent striatal circuit for distributional reinforcement learning . Nature , 639 ( 8055 ), 717 – 726 . doi: 10.1038/s41586-024-08488-5 OpenUrl CrossRef PubMed ↵ Lucantonio , F. , Roeglin , J. , Li , S. , Lu , J. , Shi , A. , Czerpaniak , K. , Fiocchi , F. R. , Bontempi , L. , Shields , B. C. , Zarate , C. A. , Tadross , M. R. , & Pignatelli , M . ( 2025 ). Ketamine rescues anhedonia by cell-type- and input-specific adaptations in the nucleus accumbens . Neuron , 113 ( 9 ), 1398 – 1412 .e4. doi: 10.1016/j.neuron.2025.02.021 OpenUrl CrossRef ↵ Ma , Z. , Stork , T. , Bergles , D. E. , & Freeman , M. R . ( 2016 ). Neuromodulators signal through astrocytes to alter neural circuit activity and behaviour . Nature , 539 ( 7629 ), 428 – 432 . doi: 10.1038/nature20145 OpenUrl CrossRef PubMed ↵ Mahon , S. , Vautrelle , N. , Pezard , L. , Slaght , S. J. , Deniau , J.-M. , Chouvet , G. , & Charpier , S . ( 2006 ). Distinct Patterns of Striatal Medium Spiny Neuron Activity during the Natural Sleep–Wake Cycle . Journal of Neuroscience , 26 ( 48 ), 12587 – 12595 . doi: 10.1523/JNEUROSCI.3987-06.2006 OpenUrl Abstract / FREE Full Text ↵ Maia , T. V. , & Frank , M. J . ( 2011 ). From reinforcement learning models to psychiatric and neurological disorders . Nature Neuroscience , 14 ( 2 ), 154 – 162 . doi: 10.1038/nn.2723 OpenUrl CrossRef PubMed Web of Science ↵ Martín , R. , Bajo-Grañeras , R. , Moratalla , R. , Perea , G. , & Araque , A . ( 2015 ). Circuit-specific signaling in astrocyte-neuron networks in basal ganglia pathways . Science (New York, N.Y.) , 349 ( 6249 ), 730 – 734 . doi: 10.1126/science.aaa7945 OpenUrl Abstract / FREE Full Text ↵ Matikainen-Ankney , B. A. , Earnest , T. , Ali , M. , Casey , E. , Wang , J. G. , Sutton , A. K. , Legaria , A. A. , Barclay , K. M. , Murdaugh , L. B. , Norris , M. R. , Chang , Y.-H. , Nguyen , K. P. , Lin , E. , Reichenbach , A. , Clarke , R. E. , Stark , R. , Conway , S. M. , Carvalho , F. , Al-Hasani , R. , … Kravitz , A. V . ( 2021 ). An open-source device for measuring food intake and operant behavior in rodent home-cages . eLife , 10 , e66173 . doi: 10.7554/eLife.66173 OpenUrl CrossRef ↵ Meron Asher , S. , & Goshen , I. ( 2025 ). Chemogenetic and optogenetic tools revolutionizing the study of astrocytes in memory . Current Opinion in Neurobiology , 93 , 103057 . doi: 10.1016/j.conb.2025.103057 OpenUrl CrossRef PubMed ↵ Metha , J. A. , Brian , M. L. , Oberrauch , S. , Barnes , S. A. , Featherby , T. J. , Bossaerts , P. , Murawski , C. , Hoyer , D. , & Jacobson , L. H . ( 2020 ). Separating Probability and Reversal Learning in a Novel Probabilistic Reversal Learning Task for Mice . Frontiers in Behavioral Neuroscience , 13 . doi: 10.3389/fnbeh.2019.00270 OpenUrl CrossRef ↵ Miguel-Quesada , C. , Zaforas , M. , Herrera-Pérez , S. , Lines , J. , Fernández-López , E. , Alonso-Calviño , E. , Ardaya , M. , Soria , F. N. , Araque , A. , Aguilar , J. , & Rosa , J. M . ( 2023 ). Astrocytes adjust the dynamic range of cortical network activity to control modality-specific sensory information processing . Cell Reports , 42 ( 8 ), 112950 . doi: 10.1016/j.celrep.2023.112950 OpenUrl CrossRef PubMed ↵ Miyazaki , I. , & Asanuma , M . ( 2016 ). Serotonin 1A Receptors on Astrocytes as a Potential Target for the Treatment of Parkinson’s Disease . Current Medicinal Chemistry , 23 ( 7 ), 686 – 700 . doi: 10.2174/0929867323666160122115057 OpenUrl CrossRef PubMed ↵ Mohebi , A. , Wei , W. , Pelattini , L. , Kim , K. , & Berke , J. D . ( 2024 ). Dopamine transients follow a striatal gradient of reward time horizons . Nature Neuroscience , 27 ( 4 ), 737 – 746 . doi: 10.1038/s41593-023-01566-3 OpenUrl CrossRef ↵ Molnár , T. , Dobolyi , Á. , Nyitrai , G. , Barabás , P. , Héja , L. , Emri , Z. , Palkovits , M. , & Kardos , J . ( 2011 ). Calcium signals in the nucleus accumbens: Activation of astrocytes by ATP and succinate . BMC Neuroscience , 12 ( 1 ), 96 . doi: 10.1186/1471-2202-12-96 OpenUrl CrossRef PubMed ↵ Mongrédien , R. , Anesio , A. , Fernandes , G. J. D. , Eagle , A. L. , Maldera , S. , Pham , C. , Robert , S. , Bezerra , F. , Vilette , A. , Bianchi , P. C. , Franco , C. , Louis , F. , Gruszczynski , C. , Niépon , M.-L. , Betancur , C. , Erdozain , A. M. , Robison , A. J. , Boucard , A. A. , Cruz , F. C. , … Vialou , V . ( 2025 ). Astrocytes Control Cocaine-Induced Synaptic Plasticity and Reward Through the Matricellular Protein Hevin . Biological Psychiatry , 98 ( 8 ), 612 – 623 . doi: 10.1016/j.biopsych.2025.02.904 OpenUrl CrossRef PubMed ↵ Montague , P. , Dayan , P. , & Sejnowski , T . ( 1996 ). A framework for mesencephalic dopamine systems based on predictive Hebbian learning . The Journal of Neuroscience , 16 ( 5 ), 1936 – 1947 . doi: 10.1523/JNEUROSCI.16-05-01936.1996 OpenUrl Abstract / FREE Full Text ↵ Montalban , E. , Ansoult , A. , Herrera Moro Chao , D. , Pham , C. , Franco , C. , Contini , A. , Castel , J. , Hassouna , R. , Hardonk , M. H. , Petitbon , A. , Foppen , E. , Gangarossa , G. , Trifilieff , P. , Li , D. , Luquet , S. , & Martin , C. ( 2025 ). Striatal astrocytes modulate behavioral flexibility and whole-body metabolism in mice . Nature Communications , 16 ( 1 ), 5417 . doi: 10.1038/s41467-025-60968-y OpenUrl CrossRef PubMed ↵ Mu , Y. , Bennett , D. V. , Rubinov , M. , Narayan , S. , Yang , C.-T. , Tanimoto , M. , Mensh , B. D. , Looger , L. L. , & Ahrens , M. B . ( 2019 ). Glia Accumulate Evidence that Actions Are Futile and Suppress Unsuccessful Behavior . Cell , 178 ( 1 ), 27 – 43 .e19. doi: 10.1016/j.cell.2019.05.050 OpenUrl CrossRef PubMed ↵ Murlanova , K. , Jouroukhin , Y. , Novototskaya-Vlasova , K. , Huseynov , S. , Pletnikova , O. , Morales , M. J. , Guan , Y. , Kamiya , A. , Bergles , D. E. , Dietz , D. M. , & Pletnikov , M. V . ( 2023 ). Loss of Astrocytic µ Opioid Receptors Exacerbates Aversion Associated with Morphine Withdrawal in Mice: Role of Mitochondrial Respiration . Cells , 12 ( 10 ), 1412 . doi: 10.3390/cells12101412 OpenUrl CrossRef ↵ Murphy-Royal , C. , Ching , S. , & Papouin , T . ( 2023 ). A conceptual framework for astrocyte function . Nature Neuroscience , 26 ( 11 ), 1848 – 1856 . doi: 10.1038/s41593-023-01448-8 OpenUrl CrossRef PubMed ↵ Nagai , J. , Bellafard , A. , Qu , Z. , Yu , X. , Ollivier , M. , Gangwani , M. R. , Diaz-Castro , B. , Coppola , G. , Schumacher , S. M. , Golshani , P. , Gradinaru , V. , & Khakh , B. S . ( 2021 ). Specific and behaviorally consequential astrocyte Gq GPCR signaling attenuation in vivo with iβARK . Neuron , 109 ( 14 ), 2256 – 2274 .e9. doi: 10.1016/j.neuron.2021.05.023 OpenUrl CrossRef PubMed ↵ Nassar , M. R. , Wilson , R. C. , Heasly , B. , & Gold , J. I . ( 2010 ). An approximately Bayesian delta-rule model explains the dynamics of belief updating in a changing environment . Journal of Neuroscience , 30 ( 37 ), 12366 – 12378 . doi: 10.1523/JNEUROSCI.0822-10.2010 OpenUrl Abstract / FREE Full Text ↵ Navarrete , M. , & Araque , A . ( 2010 ). Endocannabinoids potentiate synaptic transmission through stimulation of astrocytes . Neuron , 68 ( 1 ), 113 – 126 . doi: 10.1016/j.neuron.2010.08.043 OpenUrl CrossRef PubMed Web of Science ↵ Nazari , S. , Faez , K. , Amiri , M. , & Karami , E . ( 2015 ). A digital implementation of neuron– astrocyte interaction for neuromorphic applications . Neural Networks , 66 , 79 – 90 . doi: 10.1016/j.neunet.2015.01.005 OpenUrl CrossRef PubMed ↵ Nimitvilai-Roberts , S. , Gioia , D. , Zamudio , P. A. , & Woodward , J. J . ( 2021 ). Ethanol inhibition of lateral orbitofrontal cortex neuron excitability is mediated via dopamine D1/D5 receptor-induced release of astrocytic glycine . Neuropharmacology , 192 , 108600 . doi: 10.1016/j.neuropharm.2021.108600 OpenUrl CrossRef PubMed ↵ Oberheim , N. A. , Takano , T. , Han , X. , He , W. , Lin , J. H. C. , Wang , F. , Xu , Q. , Wyatt , J. D. , Pilcher , W. , Ojemann , J. G. , Ransom , B. R. , Goldman , S. A. , & Nedergaard , M . ( 2009 ). Uniquely Hominid Features of Adult Human Astrocytes . Journal of Neuroscience , 29 ( 10 ), 3276 – 3287 . doi: 10.1523/JNEUROSCI.4707-08.2009 OpenUrl Abstract / FREE Full Text ↵ Octeau , J. C. , Chai , H. , Jiang , R. , Bonanno , S. L. , Martin , K. C. , & Khakh , B. S . ( 2018 ). An optical neuron-astrocyte proximity assay at synaptic distance scales . Neuron , 98 ( 1 ), 49 – 66 .e9. doi: 10.1016/j.neuron.2018.03.003 OpenUrl CrossRef PubMed ↵ Ogata , K. , & Kosaka , T . ( 2002 ). Structural and quantitative analysis of astrocytes in the mouse hippocampus . Neuroscience , 113 ( 1 ), 221 – 233 . doi: 10.1016/s0306-4522(02)00041-6 OpenUrl CrossRef PubMed Web of Science ↵ Oliveira , J. F. , Sardinha , V. M. , Guerra-Gomes , S. , Araque , A. , & Sousa , N . ( 2015 ). Do stars govern our actions? Astrocyte involvement in rodent behavior . Trends in Neurosciences , 38 ( 9 ), 535 – 549 . doi: 10.1016/j.tins.2015.07.006 OpenUrl CrossRef PubMed ↵ Oyama , K. , Hernádi , I. , Iijima , T. , & Tsutsui , K.-I . ( 2010 ). Reward Prediction Error Coding in Dorsal Striatal Neurons . The Journal of Neuroscience , 30 ( 34 ), 11447 – 11457 . doi: 10.1523/JNEUROSCI.1719-10.2010 OpenUrl Abstract / FREE Full Text ↵ Panatier , A. , Vallée , J. , Haber , M. , Murai , K. K. , Lacaille , J.-C. , & Robitaille , R . ( 2011 ). Astrocytes Are Endogenous Regulators of Basal Transmission at Central Synapses . Cell , 146 ( 5 ), 785 – 798 . doi: 10.1016/j.cell.2011.07.022 OpenUrl CrossRef PubMed Web of Science ↵ Papouin , T. , Dunphy , J. M. , Tolman , M. , Dineley , K. T. , & Haydon , P. G . ( 2017 ). Septal Cholinergic Neuromodulation Tunes the Astrocyte-Dependent Gating of Hippocampal NMDA Receptors to Wakefulness . Neuron , 94 ( 4 ), 840 – 854 .e7. doi: 10.1016/j.neuron.2017.04.021 OpenUrl CrossRef PubMed ↵ Parent , A. , & Hazrati , L.-N . ( 1995 ). Functional anatomy of the basal ganglia. I. The cortico-basal ganglia-thalamo-cortical loop . Brain Research Reviews , 20 ( 1 ), 91 – 127 . doi: 10.1016/0165-0173(94)00007-C OpenUrl CrossRef PubMed Web of Science ↵ Parker , N. F. , Cameron , C. M. , Taliaferro , J. P. , Lee , J. , Choi , J. Y. , Davidson , T. J. , Daw , N. D. , & Witten , I. B . ( 2016 ). Reward and choice encoding in terminals of midbrain dopamine neurons depends on striatal target . Nature Neuroscience , 19 ( 6 ), 845 – 854 . doi: 10.1038/nn.4287 OpenUrl CrossRef PubMed ↵ Paukert , M. , Agarwal , A. , Cha , J. , Doze , V. A. , Kang , J. U. , & Bergles , D. E . ( 2014 ). Norepinephrine Controls Astroglial Responsiveness to Local Circuit Activity . Neuron , 82 ( 6 ), 1263 – 1270 . doi: 10.1016/j.neuron.2014.04.038 OpenUrl CrossRef PubMed Web of Science ↵ Petzold , G. C. , & Murthy , V. N . ( 2011 ). Role of Astrocytes in Neurovascular Coupling . Neuron , 71 ( 5 ), 782 – 797 . doi: 10.1016/j.neuron.2011.08.009 OpenUrl CrossRef PubMed Web of Science ↵ Philips , R. T. , Sur , M. , & Chakravarthy , V. S . ( 2017 ). The influence of astrocytes on the width of orientation hypercolumns in visual cortex: A computational perspective . PLOS Computational Biology , 13 ( 10 ), e1005785 . doi: 10.1371/journal.pcbi.1005785 OpenUrl CrossRef PubMed ↵ Planert , H. , Berger , T. K. , & Silberberg , G . ( 2013 ). Membrane properties of striatal direct and indirect pathway neurons in mouse and rat slices and their modulation by dopamine . PloS One , 8 ( 3 ), e57054 . doi: 10.1371/journal.pone.0057054 OpenUrl CrossRef PubMed Web of Science ↵ Reimsnider , S. , Manfredsson , F. P. , Muzyczka , N. , & Mandel , R. J . ( 2007 ). Time Course of Transgene Expression After Intrastriatal Pseudotyped rAAV2/1, rAAV2/2, rAAV2/5, and rAAV2/8 Transduction in the Rat . Molecular Therapy , 15 ( 8 ), 1504 – 1511 . doi: 10.1038/sj.mt.6300227 OpenUrl CrossRef PubMed Web of Science ↵ Reitman , M. E. , Tse , V. , Mi , X. , Willoughby , D. D. , Peinado , A. , Aivazidis , A. , Myagmar , B.-E. , Simpson , P. C. , Bayraktar , O. A. , Yu , G. , & Poskanzer , K. E . ( 2023 ). Norepinephrine links astrocytic activity to regulation of cortical state . Nature Neuroscience , 26 ( 4 ), 579 – 593 . doi: 10.1038/s41593-023-01284-w OpenUrl CrossRef PubMed ↵ Richardson , N. R. , & Roberts , D. C. S . ( 1996 ). Progressive ratio schedules in drug self-administration studies in rats: A method to evaluate reinforcing efficacy . Journal of Neuroscience Methods , 66 ( 1 ), 1 – 11 . doi: 10.1016/0165-0270(95)00153-0 OpenUrl CrossRef PubMed Web of Science ↵ Robbins , H . ( 1952 ). Some aspects of the sequential design of experiments . Bulletin of the American Mathematical Society , 58 ( 5 ), 527 – 535 . OpenUrl CrossRef ↵ Roberts , B. M. , Doig , N. M. , Brimblecombe , K. R. , Lopes , E. F. , Siddorn , R. E. , Threlfell , S. , Connor-Robson , N. , Bengoa-Vergniory , N. , Pasternack , N. , Wade-Martins , R. , Magill , P. J. , & Cragg , S. J . ( 2020 ). GABA uptake transporters support dopamine release in dorsal striatum with maladaptive downregulation in a parkinsonism model . Nature Communications , 11 , 4958 . doi: 10.1038/s41467-020-18247-5 OpenUrl CrossRef PubMed ↵ Roesch , M. R. , Singh , T. , Brown , P. L. , Mullins , S. E. , & Schoenbaum , G . ( 2009 ). Ventral Striatal Neurons Encode the Value of the Chosen Action in Rats Deciding between Differently Delayed or Sized Rewards . The Journal of Neuroscience , 29 ( 42 ), 13365 – 13376 . doi: 10.1523/JNEUROSCI.2572-09.2009 OpenUrl Abstract / FREE Full Text ↵ Rubin , R. , Abbott , L. F. , & Sompolinsky , H . ( 2017 ). Balanced excitation and inhibition are required for high-capacity, noise-robust neuronal selectivity . Proceedings of the National Academy of Sciences , 114 ( 44 ), E9366 – E9375 . doi: 10.1073/pnas.1705841114 OpenUrl Abstract / FREE Full Text ↵ Samejima , K. , Ueda , Y. , Doya , K. , & Kimura , M . ( 2005 ). Representation of Action-Specific Reward Values in the Striatum . Science , 310 ( 5752 ), 1337 – 1340 . doi: 10.1126/science.1115270 OpenUrl Abstract / FREE Full Text ↵ Scemes , E. , & Giuame , C . ( 2006 ). Astrocyte Calcium Waves . Glia , 54 ( 7 ), 716 – 725 . doi: 10.1002/glia.20374 OpenUrl CrossRef PubMed Web of Science ↵ Schipke , C. G. , Haas , B. , & Kettenmann , H . ( 2008 ). Astrocytes Discriminate and Selectively Respond to the Activity of a Subpopulation of Neurons within the Barrel Cortex . Cerebral Cortex , 18 ( 10 ), 2450 – 2459 . doi: 10.1093/cercor/bhn009 OpenUrl CrossRef PubMed Web of Science ↵ Schultz , W . ( 1998 ). Predictive Reward Signal of Dopamine Neurons . Journal of Neurophysiology , 80 ( 1 ), 1 – 27 . doi: 10.1152/jn.1998.80.1.1 OpenUrl CrossRef PubMed Web of Science ↵ Segura-Aguilar , J. , Mannervik , B. , Inzunza , J. , Varshney , M. , Nalvarte , I. , & Muñoz , P . ( 2022 ). Astrocytes protect dopaminergic neurons against aminochrome neurotoxicity . Neural Regeneration Research , 17 ( 9 ), 1861 . doi: 10.4103/1673-5374.335690 OpenUrl CrossRef PubMed ↵ Seibenhener , M. L. , & Wooten , M. C . ( 2015 ). Use of the Open Field Maze to Measure Locomotor and Anxiety-like Behavior in Mice . Journal of Visualized Experiments : JoVE , 96 , 52434 . doi: 10.3791/52434 OpenUrl CrossRef PubMed ↵ Serra , I. , Esparza , J. , Delgado , L. , Martín-Monteagudo , C. , Puigròs , M. , Podlesniy , P. , Trullás , R. , & Navarrete , M . ( 2022 ). Ca2+-modulated photoactivatable imaging reveals neuron-astrocyte glutamatergic circuitries within the nucleus accumbens . Nature Communications , 13 ( 1 ), 5272 . doi: 10.1038/s41467-022-33020-6 OpenUrl CrossRef PubMed ↵ Serra , I. , Martín-Monteagudo , C. , Sánchez Romero , J. , Quintanilla , J. P. , Ganchala , D. , Arevalo , M.-A. , García-Marqués , J. , & Navarrete , M . ( 2025 ). Astrocyte ensembles manipulated with AstroLight tune cue-motivated behavior . Nature Neuroscience , 1 – 11 . doi: 10.1038/s41593-025-01870-0 OpenUrl CrossRef ↵ Shen , W. , Flajolet , M. , Greengard , P. , & Surmeier , D. J . ( 2008 ). Dichotomous dopaminergic control of striatal synaptic plasticity . Science (New York, N.Y.) , 321 ( 5890 ), 848 – 851 . doi: 10.1126/science.1160575 OpenUrl Abstract / FREE Full Text ↵ Soares-Cunha , C. , Coimbra , B. , David-Pereira , A. , Borges , S. , Pinto , L. , Costa , P. , Sousa , N. , & Rodrigues , A. J . ( 2016 ). Activation of D2 dopamine receptor-expressing neurons in the nucleus accumbens increases motivation . Nature Communications , 7 , 11829 . doi: 10.1038/ncomms11829 OpenUrl CrossRef PubMed ↵ Sofroniew , M. V. , & Vinters , H. V . ( 2010 ). Astrocytes: Biology and pathology . Acta Neuropathologica , 119 ( 1 ), 7 – 35 . doi: 10.1007/s00401-009-0619-8 OpenUrl CrossRef PubMed Web of Science ↵ Stedehouder , J. , Roberts , B. M. , Raina , S. , Bossi , S. , Liu , A. K. L. , Doig , N. M. , McGerty , K. , Magill , P. J. , Parkkinen , L. , & Cragg , S. J . ( 2024 ). Rapid modulation of striatal cholinergic interneurons and dopamine release by satellite astrocytes . Nature Communications , 15 ( 1 ), 10017 . doi: 10.1038/s41467-024-54253-7 OpenUrl CrossRef PubMed ↵ Stewart , W. J . ( 1975 ). Progressive reinforcement schedules: A review and evaluation . Australian Journal of Psychology , 27 ( 1 ), 9 – 22 . doi: 10.1080/00049537508255235 OpenUrl CrossRef ↵ Stobart , J. L. , Ferrari , K. D. , Barrett , M. J. P. , Glück , C. , Stobart , M. J. , Zuend , M. , & Weber , B . ( 2018 ). Cortical Circuit Activity Evokes Rapid Astrocyte Calcium Signals on a Similar Timescale to Neurons . Neuron , 98 ( 4 ), 726 – 735 .e4. doi: 10.1016/j.neuron.2018.03.050 OpenUrl CrossRef PubMed ↵ Sutton , R. S. , & Barto , A. G . ( 2018 ). Reinforcement Learning: An Introduction (Second). MIT Press . ↵ Takata , N. , Mishima , T. , Hisatsune , C. , Nagai , T. , Ebisui , E. , Mikoshiba , K. , & Hirase , H . ( 2011 ). Astrocyte calcium signaling transforms cholinergic modulation to cortical plasticity in vivo . Journal of Neuroscience , 31 ( 49 ), 18155 – 18165 . Scopus. doi: 10.1523/JNEUROSCI.5289-11.2011 OpenUrl Abstract / FREE Full Text ↵ Tang , G. , Polykretis , I. E. , Ivanov , V. A. , Shah , A. , & Michmizos , K. P . ( 2019 ). Introducing Astrocytes on a Neuromorphic Processor: Synchronization, Local Plasticity and Edge of Chaos . Proceedings of the 7th Annual Neuro-Inspired Computational Elements Workshop , 1 – 9 . doi: 10.1145/3320288.3320302 OpenUrl CrossRef ↵ Tang , H. , Costa , V. D. , Bartolo , R. , & Averbeck , B. B . ( 2022 ). Differential coding of goals and actions in ventral and dorsal corticostriatal circuits during goal-directed behavior . Cell Reports , 38 ( 1 ), 110198 . doi: 10.1016/j.celrep.2021.110198 OpenUrl CrossRef PubMed ↵ Ugray , Z. , Lasdon , L. , Plummer , J. , Glover , F. , Kelly , J. , & Martí , R . ( 2007 ). Scatter Search and Local NLP Solvers: A Multistart Framework for Global Optimization . INFORMS Journal on Computing , 19 ( 3 ), 328 – 340 . doi: 10.1287/ijoc.1060.0175 OpenUrl CrossRef Web of Science ↵ Wu , Y. , Dai , Y. , Lefton , K. B. , Holy , T. E. , & Papouin , T . ( 2024 ). STARDUST: A pipeline for the unbiased analysis of astrocyte regional calcium dynamics . STAR Protocols , 5 ( 3 ), 103305 . doi: 10.1016/j.xpro.2024.103305 OpenUrl CrossRef PubMed ↵ Xin , Q. , Wang , J. , Zheng , J. , Tan , Y. , Jia , X. , Ni , Z. , Feng , J. , Wu , Z. , Li , Y. , Li , X. , Ma , H. , & Hu , H . ( 2024 ). Neuron-astrocyte Coupling in Lateral Habenula Mediates Depressive-like Behaviors . doi: 10.1101/2024.11.03.621722 OpenUrl Abstract / FREE Full Text ↵ Yee , A. G. , Liao , Y. , Muntean , B. S. , & Ford , C. P . ( 2025 ). Discrete spatiotemporal encoding of striatal dopamine transmission . Science , 389 ( 6756 ), 200 – 206 . doi: 10.1126/science.adp9833 OpenUrl CrossRef PubMed ↵ Yu , X. , Taylor , A. M. W. , Nagai , J. , Golshani , P. , Evans , C. J. , Coppola , G. , & Khakh , B. S . ( 2018 ). Reducing Astrocyte Calcium Signaling In Vivo Alters Striatal Microcircuits and Causes Repetitive Behavior . Neuron , 99 ( 6 ), 1170 – 1187 .e9. doi: 10.1016/j.neuron.2018.08.015 OpenUrl CrossRef PubMed ↵ Zachry , J. E. , Kutlu , M. G. , Yoon , H. J. , Leonard , M. Z. , Chevée , M. , Patel , D. D. , Gaidici , A. , Kondev , V. , Thibeault , K. C. , Bethi , R. , Tat , J. , Melugin , P. R. , Isiktas , A. U. , Joffe , M. E. , Cai , D. J. , Conn , P. J. , Grueter , B. A. , & Calipari , E. S . ( 2024 ). D1 and D2 medium spiny neurons in the nucleus accumbens core have distinct and valence-independent roles in learning . Neuron , 112 ( 5 ), 835 – 849 .e7. doi: 10.1016/j.neuron.2023.11.023 OpenUrl CrossRef ↵ Zahm , D. S . ( 1999 ). Functional-anatomical Implications of the Nucleus Accumbens Core and Shell Subterritories . Annals of the New York Academy of Sciences , 877 ( 1 ), 113 – 128 . doi: 10.1111/j.1749-6632.1999.tb09264.x OpenUrl CrossRef PubMed Web of Science ↵ Zheng , K. , Bard , L. , Reynolds , J. P. , King , C. , Jensen , T. P. , Gourine , A. V. , & Rusakov , D. A . ( 2015 ). Time-Resolved Imaging Reveals Heterogeneous Landscapes of Nanomolar Ca(2+) in Neurons and Astroglia . Neuron , 88 ( 2 ), 277 – 288 . doi: 10.1016/j.neuron.2015.09.043 OpenUrl CrossRef PubMed View the discussion thread. Back to top Previous Next Posted October 20, 2025. Download PDF Supplementary Material Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Ventral striatal astrocytes contribute to reinforcement learning Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Ventral striatal astrocytes contribute to reinforcement learning Julia Pai , Fatih Sogukpinar , Kei Ogasawara , Garrett J Smith , Francesca R Fiocchi , Yanchao Dai , Yifan Wu , Michael J Frank , ShiNung Ching , Federica Lucantonio , Thomas Papouin , Marco Pignatelli , Naoki Hiratani , Ilya E Monosov bioRxiv 2025.10.20.683205; doi: https://doi.org/10.1101/2025.10.20.683205 Share This Article: Copy Citation Tools Ventral striatal astrocytes contribute to reinforcement learning Julia Pai , Fatih Sogukpinar , Kei Ogasawara , Garrett J Smith , Francesca R Fiocchi , Yanchao Dai , Yifan Wu , Michael J Frank , ShiNung Ching , Federica Lucantonio , Thomas Papouin , Marco Pignatelli , Naoki Hiratani , Ilya E Monosov bioRxiv 2025.10.20.683205; doi: https://doi.org/10.1101/2025.10.20.683205 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Neuroscience Subject Areas All Articles Animal Behavior and Cognition (7633) Biochemistry (17680) Bioengineering (13888) Bioinformatics (41927) Biophysics (21445) Cancer Biology (18585) Cell Biology (25491) Clinical Trials (138) Developmental Biology (13373) Ecology (19897) Epidemiology (2067) Evolutionary Biology (24308) Genetics (15606) Genomics (22494) Immunology (17736) Microbiology (40385) Molecular Biology (17175) Neuroscience (88583) Paleontology (666) Pathology (2830) Pharmacology and Toxicology (4822) Physiology (7641) Plant Biology (15149) Scientific Communication and Education (2045) Synthetic Biology (4293) Systems Biology (9822) Zoology (2271)