Anterior cingulate cortex mixes retrospective cognitive signals and ongoing movement signatures during decision-making

Anterior cingulate cortex mixes retrospective cognitive signals and ongoing movement signatures during decision-making | 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 Anterior cingulate cortex mixes retrospective cognitive signals and ongoing movement signatures during decision-making View ORCID Profile Lukas T. Oesch , Makenna C. Thomas , Davis Sandberg , João Couto , Anne K. Churchland doi: https://doi.org/10.1101/2025.04.11.648398 Lukas T. Oesch 1 Department of Neurobiology, David Geffen School of Medicine, University of California , Los Angeles, Los Angeles Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Lukas T. Oesch Makenna C. Thomas 1 Department of Neurobiology, David Geffen School of Medicine, University of California , Los Angeles, Los Angeles Find this author on Google Scholar Find this author on PubMed Search for this author on this site Davis Sandberg 1 Department of Neurobiology, David Geffen School of Medicine, University of California , Los Angeles, Los Angeles Find this author on Google Scholar Find this author on PubMed Search for this author on this site João Couto 1 Department of Neurobiology, David Geffen School of Medicine, University of California , Los Angeles, Los Angeles Find this author on Google Scholar Find this author on PubMed Search for this author on this site Anne K. Churchland 1 Department of Neurobiology, David Geffen School of Medicine, University of California , Los Angeles, Los Angeles Find this author on Google Scholar Find this author on PubMed Search for this author on this site For correspondence: AChurchland{at}mednet.ucla.edu Abstract Full Text Info/History Metrics Preview PDF Abstract In dynamic environments, animals must closely monitor the effects of their actions to inform switches in behavioral strategy. Anterior cingulate cortex (ACC) neurons track decision outcomes in these environments. Yet, it remains unclear whether ACC neurons similarly monitor behavioral history in static environments and, if so, whether these signals are distinct from movement representations. We recorded large-scale ACC activity in freely moving mice making visual evidence-accumulation decisions. Many ACC neurons exhibited nonlinear mixed selectivity for previous choices and outcomes (trial history) and were modulated by movements. Trial history could be stably decoded from population activity and accounted for a separable component of neural activity than posture and movements. Trial history encoding was conserved across different subjects and was unaffected by fluctuating behavioral biases. These findings demonstrate that trial history monitoring in ACC is implemented in a conserved population code that is independent of the volatility of subjects’ task environment. Introduction A key problem for decision-makers is to discover a set of rules or internal models that map cues in the environment to actions in a way that yields a desired outcome. To build or update internal models, animals must explore different response options and their resulting outcomes and then integrate over these past experiences to determine the best course of action. The anterior cingulate cortex (ACC) is believed to play a major role in evaluating the outcomes of ongoing behavior to inform changes in behavioral strategy 1 . Evidence supporting this idea comes from several lines of research in humans, non-human primates and rodents performing probabilistically rewarded tasks in dynamic environments where the rules governing which actions are rewarded frequently switch. In these tasks ACC neurons encode predictions about expected rewards resulting from a speciOc behavior 2 , 3 . Violations of these predictions (that are sometimes called surprise signals or unsigned prediction errors) strongly drive neural activity in the ACC 4 , 5 . One hypothesis is that these prediction error signals are then used to update subjects’ current beliefs and the exploration of alternative behaviors 6 – 13 . In support of this hypothesis, inhibition of ACC neurons prevent animals from adapting their behavior to a new rule after changes of the task environment 12 – 15 . This view of ACC function would predict that the post-outcome activity of ACC neurons should generally be low and would only be elevated after an unexpected outcome, signaling a change in the environment 7 . Surprisingly, ACC neurons not only strongly encode previous choices and outcomes after block transitions when the subjects’ behavior needs to be updated but also within blocks when subjects exploit their task knowledge 16 – 18 . This existing work suggests that the ACC might not only drive belief updating but also continuously monitor trial history. However, most studies investigating the functions of ACC have employed tasks where decision-rules switch in a block-wise manner. In these tasks, close monitoring of trial history and tracking of unexpected changes is built into the task design and is required for high task performance. A new approach is needed for two reasons. First, imposing a block structure on behavioral tasks, as is done in dynamic environments, can lead to strong correlations between previous choices and outcomes, as well as previous and upcoming reporting actions 16 . It is thus unclear if the recorded signals in ACC neurons reflect true trial history monitoring or, instead, postural or movement variables 19 . Second, it remains unknown whether trial history signals in ACC neurons exclusively emerge in dynamic, unpredictable environments or whether they may also be present in deterministic tasks with stable rules governed by observable sensory inputs. If trial history signals are present even in static environments, this would argue that decision circuits are wired with flexibility in mind, allowing decision-makers to constantly evaluate their strategy. Previous behavioral observations hint at this idea: perceptual decisions in stable environments show hallmarks of strategies for economic decisions in dynamic environments 20 . These strategies, which include uncertainty guided exploration 21 , rely on retaining information about outcomes of previous decisions. The ACC is a natural candidate to track such choice – outcome combinations, but its responses have been studied almost entirely in the context of dynamic, economic decisions. To overcome these limitations, we trained mice in a freely moving perceptual evidence accumulation task with a Oxed rule and a randomized trial structure 22 , 23 while recording neural activity from ACC neurons and carefully tracking animals’ posture and movements. In keeping with the hypothesis that ACC monitors trial history, we found that ACC neurons strongly encoded choice- and outcome history and their different interactions. Using decoding analyses, we show that these representations remain stable well into the new trial. Taking advantage of methods to separate the influence of often highly collinear task variables and movements on neural activity 24 , 25 we show that trial history signals cannot simply be explained by animals’ body position or movements. Finally, we demonstrate that trial history, but not movement encoding, is low-dimensional and that the neural dynamics encoding trial history are conserved across subjects. Results Perceptual decisions in expert performers are modulated by trial history We trained mice on a self-initiated freely moving visual decision-making task that required the integration of pulsatile sensory evidence over time 22 , 23 . After initiating a trial by poking into the center port, mice were presented with a train of visual flashes ( Figure 1A, B ). On trials where the number of flashes exceeded the category boundary of 12 events per second (12 Hz), mice were rewarded for poking into the right port and conversely, were rewarded for choosing the left port after a low-rate stimulus trial. Importantly, these contingencies remained stable and were not reversed. Furthermore, whether a trial was high- or low-rate was randomly assigned. These measures ensured that the environment remained static and that there was no trial structure that animals could exploit by using trial history information. Animals learned this task (center Oxation and stimulus discrimination) over the course of 10.72 ± 1.25 weeks and achieved high performance levels ( Figure 1C ). Download figure Open in new tab Figure 1. Perceptual decisions in expert performers are influenced by trial history. (A) Illustration of perceptual task setup. Animals initiate trials by poking into a central port. This triggers a 1 second train of visual flashes. Animals are required to wait inside the center port for one second before reporting their decisions in either the left or the right port. Animals are rewarded for choosing the left port after low-rate stimuli and for choosing the right port after high-rate stimuli. Incorrect choices result in a 16 kHz tone and 2 s timeout. (B) Trials are deFned by four distinct phases. Early ITI: the second after the presentation of the previous trial’s outcome, Late ITI: 1 s before until 0.3 s after the end of the outcome, Stimulus: 0.5 s before to 1 s after stimulus onset, Action: 0.2 s before to 0.3 s after reporting movement onset. (C) Psychometric functions for individual subjects (grey, n = 12 animals, 356 sessions) and averaged across subjects (black). (D) Psychometric functions for trials that were preceded by speciFc combinations of previous choice and outcome (trial history) for an example subject. Dashed black line represents the overall Ft. (E) Weights of a logistic regression trained to predict choice based on the stimulus rate and trial history from the subject shown in (D) Lines represent individual sessions, dots show mean ± sem weight across sessions. (F) Boxplot showing the median ± interquartile range of the model weights. Whiskers depict 1.5 x the interquartile range. White dots represent the mean and dark scatter points are individual subjects. (G) Behavioral influence of trial history quantiFed as the length of the vector in the space spanned by the four distinct combinations of previous choice and outcome. (H) The difference between history strength and stimulus rate weight is negatively correlated with task performance. Points depcit individual sessions and are color-coded by subject identity. ***P < 0.001. A series of recent studies have suggested that perceptual decisions might be influenced by previous choices and/or outcomes even in expert performers 23 , 26 – 31 and that previous correct and incorrect choices might differentially impact subsequent behavior 32 . To verify that mice in our study were similarly influenced by trial history, we partitioned subjects’ trials from all the sessions based on whether the previous choice was to the left or right and whether the mice were rewarded for their previous choice. We then Otted a psychometric function to trials with these different trial history contexts and found that the curves were indeed different. Figure 1D shows the psychometric curves for an example mouse whose choices became right biased after correct right choices but were largely unaffected by previous incorrect right choices and previous left choices regardless of their outcome ( Figure 1D , Supplementary Figure 1A - D ). To compare how strongly animals’ decisions were influenced by sensory evidence or speciOc trial history contexts on a session-by-session basis, we modeled the animals’ choices as a function of the current stimulus rate plus an individual regressor for each trial history context ( Figure 1E ). As expected, we found that for all the expert mice, the model weights for stimulus rate strongly influenced animals’ decisions ( Figure 1F ). Trial history biased animals’ choice in different ways. While animals tended to repeat previously rewarded choices (negative signs indicate favoring left choices), previous incorrect choices had a weaker and less consistent influence on upcoming decisions ( Figure 1F ). Nevertheless, the magnitudes of all the model weights were signiOcantly larger than the weight magnitudes from a shuffled control (linear mixed effects model with subject as random effect). The weaker average influence of incorrect choices was not because they had large values that could be either positive or negative. Indeed, the session-by-session variance of all the trial history weights was similar ( Supplementary Figure 1E ) and we found no evidence of clusters of sessions with similar weight patterns ( Supplementary Figure 1F ). Importantly, when we looked not one-but two trials back, we found that the choices, outcomes and their interaction had a lower impact on the mice’s decisions than the ones from the last trial and they were very close to zero ( Supplementary Figure 1G ). Given that the trial structure of our task was random and that only the stimulus rate determines the rewarded side, using trial history to inform choices will decrease performance. To verify this, we Orst summarized the overall influence of the most recent trial history on the animals’ choices (history strength) by taking the length of the vector of all trial history weights ( Figure 1G ). We then subtracted the history strength from the stimulus weight and correlated this difference to the animals’ performance. We found a strong negative correlation indicating that performance decreases as history strength grows ( Figure 1H ). These Ondings conOrm that decisions in expert mice are influenced by multiple trial history contexts even when this leads to suboptimal performance 33 . Trial history can be decoded from ACC population activity We then asked whether ACC neurons encode trial history information in expert performers. To this end, we recorded neural activity from CaMKII-expressing neurons in the ACC using miniaturized head-mounted fluorescence microscopes 34 ( Figure 2A and B). The small, lightweight size of this scope meant that we were able to measure neural activity in freely moving mice. Further, this approach allowed us to record many neurons simultaneously (322 ± 42 neurons per subject) and repeatedly sample neural activity from the same subjects over multiple sessions, enabling the use of powerful population level analyses and comparisons within- and across subjects. Download figure Open in new tab Figure 2. ACC neuron populations uniquely represent multiple trial history contexts. (A) Transgenic mice expressing GCaMP6s under the control of the CaMKII promoter were implanted with a gradient refractive index lens above their ACC; neural activity was recorded using a miniaturized fluorescence microscope (top). Histological veriFcation of GCaMP6 expression and lens position (bottom). (B) (Top) Field of view from an example recording. A single frame is overlaid with the variance projection of the entire recording (green). We identiFed individual neurons using an automatic segmentation algorithm 37 (bottom). (C) Example traces from seven simultaneously recorded neurons. Horizontal lines represent the Frst second of the ITI after respective choices and outcomes. (D) Peri-event time histograms of three neurons from (C) showing distinct responses for different trial history contexts. Lines and shading represent mean ± sem. (E) Mean activity difference for previous right vs left choices (x-axis) plotted against mean activity difference for previous correct vs. incorrect trials (y-axis) during the late ITI for individual neurons from one session per mouse. Circled points mark the example neurons shown in (D), color shading schematizes increasing selectivity for speciFc trial history contexts. (F) Mean ± sem accuracy of a linear decoder trained to predict trial history based on ACC population activity. Bars represent the average across the entire trial time (n = 9 subjects). ***P < 0.001, linear mixed-effects model with session nested in subject as random effects. (G) Mean ± sem decoder accuracy over trial time for different trial history contexts. ***P 2 s is the start of the stimulus presentation. (I) Decoding stability: trial history decoders were trained for every time point and tested on every other time point. The average over all subjects is shown. (J) Mean ± sem cross-decoding accuracy divided by trial phase. Dashed red line reflects chance level. *P < 0.05, ***P < 0.001, linear mixed-effects model with session nested in subject as random effects and post-hoc comparison between model and shuffled accuracies. The activity of many ACC neurons was modulated by speciOc trial history contexts ( Figure 2C ). For example, the activity of some neurons was driven almost exclusively by previous correct right or correct left choices during the inter-trial interval (ITI, Figure 2D , N1, N4) or by both incorrect left and right choices but at different times in the trial ( Figure 2D , N7), indicating clear non-linear mixed selectivity 35 , 36 . To summarize the selectivity of the population of all ACC neurons we plotted the activity difference between previous right and left choices against the difference of activity for previous correct and incorrect trials ( Figure 2E ). We found a large variety of different response proOles. Notably however, while the population tended to show higher selectivity for previous incorrect trials than for correct ones, neurons were more selective for previous right or left choices after correct than after incorrect trials. To quantitatively assess the strength of neural representations of trial history, we trained linear decoders on ACC population activity to predict the different trial history contexts (4 classes). The decoding accuracy over the entire trial time was high (78 % as compared to 25% at chance level) and signiOcantly higher than the shuffled control ( Figure 2F ). Depending on the structure of neural representation, optimizing decoders to classify all the non-linear combinations of previous choice and outcome might lead to losses in decoding accuracy of the underlying binary decoding of previous choice or previous outcome ( Supplementary Figure 2A ). However, we found no difference in the classiOcation accuracy for previous choice or previous outcome between the full trial history decoders and binary classiOers speciOcally optimized to decode previous choice or outcome ( Supplementary Figure 2B ). This suggests that neural representations in the ACC support independent readouts of previous choice and outcome and their combinations. Next, we asked whether all trial history contexts could be decoded equally well during all phases of the trial. To answer this question, we separately computed the decoding accuracy for each trial history context across the entire trial duration ( Figure 2G ). We found that the temporal dynamics depended on the outcome of the previous trial: the decoding accuracy of previously rewarded left and right trials was similar and remained high throughout the ITI and slowly decayed thereafter. The decoding accuracy for previous incorrect left and right choices started decaying during the late ITI and was signiOcantly lower than the accuracy for either of the previous correct choices. We further found that the decoding accuracy for previous incorrect right choices was slightly lower than for previous incorrect left choices. We then asked whether the neural representation of certain trial history contexts was more similar than others and whether the similarity persisted across trial time. The more similar the neural representation of a pair of trial history contexts is, the higher the probability that the decoders misclassify them. We therefore examined the errors made by the decoders by constructing confusion matrices over trial time. We found that the types of misclassiOcations changed with time. At the very beginning of the ITI, right after the mice had Onished reporting their decision, the decoders frequently confounded correct with incorrect choices during both left and right decisions ( Figure 2H , top ). 500 ms later into the early ITI, all the different trial history contexts were highly decodable ( Figure 2H , middle ) and Onally at the beginning of the stimulus presentation decoders were less accurate at distinguishing between previous left and right choice for incorrect trials, whereas previous correct left and right trials were still well decoded ( Figure 2H , bottom ). These Ondings suggest that neural representations evolve from initially tracking previous choices to later keeping a record of combinations of previous choices and outcomes during the ITI and the stimulus and action periods. ACC neurons exhibit long intrinsic- and task-related Oring timescales 38 – 40 , making them suitable to retain information over extended periods of time. We therefore asked whether the neural representations encoding trial history information remained stable across different trial phases 41 – 43 or whether they evolve gradually as the trial unfolds 44 . To test this, we trained decoders on all time points and tested their decoding accuracy on all the other timepoints. We found that decoders were very stable within the early and late phases of the ITI and within the stimulus and action phase, with a marked split between ITI and new trial timepoints ( Figure 2I ). When we quantiOed the decoder stability, we found that trial history decoders trained on the early ITI also performed very well on the late ITI time points but were at chance level or only slightly above for the later trial phases. Decoders trained on the late ITI predicted trial history better than chance across the entire trial. Finally, decoders trained on the stimulus and action phase performed above chance on all other phases but the early ITI. Taken together, these Ondings point to a stable trial history representation during the ITI that undergoes a major reorganization when the mice transition from the ITI to initiating a new trial. Finally, we assessed whether trial history could be decoded better in sessions where subjects more strongly relied on trial history to guide their choices. We found no correlation between behavioral trial history strength and trial history decoding accuracy ( Supplementary Figure 4A ). There was a weak positive relationship between behavioral weights for trial history and decoding accuracy only for previous incorrect right trials ( Supplementary Figure 4B ). Our Ondings thus suggest that trial history information is similarly encoded in the ACC population activity regardless of the animals’ behavioral trial history biases. Neural representations of trial history cannot solely be explained by body posture or movements Recent work in mice has demonstrated that much of the neural activity not only across motor but also sensory cortical areas can be explained by spontaneous movements 24 , 45 – 47 . Some movements are highly stereotyped and aligned to task events, task variables or animal’s past actions 48 . We therefore examined whether neural signatures of trial history were simply a reflection of the animals’ movements or whether they represented a uniquely cognitive signal. To address this question, we Otted linear encoding models which relate the activity of individual neurons to task variables, and postural and movement regressors 24 . The task variable regressors were choice, outcome, trial history and visual stimulus events; movements included the pokes into the three different ports, a lower dimensional representation of either the full behavioral video (Video SVD) or the motion energy of the video (Motion energy SVD), and neurons’ tuning to head-orientation angle and body position ( Figure 3A , left and center , Supplementary Figure 3A). To avoid overOtting the models, we regularized the regressor weights using ridge regression and ran ten-fold cross validation. On held-out trials, the models on average predicted 32.8 ± 1.9% of the neural variance across the imaged ACC populations with above 85% explained variance for some individual neurons ( Figure 3A , right, compare red and gray traces for each neuron). Download figure Open in new tab Figure 3. Trial history explains considerable unique variance on the single-cell level. (A) Linear encoding model design. Neural activity was modeled as a function of task variables, instructed- and uninstructed movements (left and center). Right panel shows measured (grey) and reconstructed (red) activity of four example neurons. The number indicates the percent explained variance (B and C) Maximal amount of neural variance (B) or variance uniquely (C) explained by different regressors. Boxes represent median of subject means ± interquartile range, white circle depicts mean and grey scatters individual subjects (n = 9 subjects). (D, E and F) Unique variance of all neurons for trial history plotted against the unique variance for Video SVD (D), chest position (E), and choice (F). Having built a model that successfully predicted single-trial neural activity, we set out to determine which variables contributed to this prediction. We Orst asked how much neural variance could be explained by each model regressor alone (maximal explained variance). To address this question, we Otted encoding models with all regressors shuffled except the regressor of interest plus a time-varying intercept. We then also Otted an intercept only model and subtracted its explained variance from the single-regressor model explained variance. This ensured that the explained variance was driven by the speciOc variable in question rather than temporal fluctuations shared across all trials. We found that uninstructed movements, particularly the video and video motion-energy regressors, explained relatively high amounts of variance on their own ( Figure 3B ). This observation complements previous reports that movements explain considerable neural activity in head-restrained mice 24 , 45 , 46 and demonstrates that movements also have a major impact on neural activity in freely moving animals 49 . Because task variables can be highly correlated with each other (or with movements) the maximal explained variance is vulnerable to over-estimating the true explained variance (Supplementary Figure 3B). To account for this, we computed the unique explained variance of the different regressors. To this end we Otted the full model with only the regressor of interest shuffled and then subtracted its explained variance from the full model variance (Supplementary Figure 3C). We found that not only Video SVD but also trial history accounted for considerable unique neural variance ( Figure 3C ). This observation is in stark contrast to the impact of other task variables in other areas where task variables typically account for minimal neural variance and are dwarfed by movement signals 24 . We then assessed the temporal proOle of the maximal and unique explained variance trial history. We observed that the maximal explained variance for trial history decreased over the different trial phases, similar to the decrease in trial history decoding accuracy (Supplementary Figure 3D). In contrast, the unique explained variance for trial history remained relatively stable (Supplementary Figure 3E). We further assessed whether the maximal or the unique explained variance for trial history of the ACC neuron populations were related to the behavioral trial history biases. Similar to the decoding results, we did not observe any correlation between behavioral trial history biases and maximal or unique neural variance explained by trial history (Supplementary Figure 4C and D), providing reassurance that neural trial history encoding signals are not simply driven by behavioral biases. However, we observed that trial history encoding and decodability was signiOcantly lower in females than in males (Supplementary Figure 5). Previous work demonstrated that neurons in prefrontal and parietal cortices of primates, rats and mice have mixed selectivity to different task variables 35 , 36 , 50 – 53 . We thus asked whether multiple regressors might explain variance within the same neurons. Because multiple collinear variables could drive the maximal explained variance, we only considered the more conservative measure of unique explained variance to assess mixed selectivity. We indeed found that in many neurons where trial history explained unique variance also video SVD and to lesser extent chest position and choice could explain sizeable amounts of unique variance ( Figure 3D - F ) pointing to a high degree and diversity of mixed selectivity in ACC neurons. Taken together these data demonstrate that ACC neurons, in contrast to neurons in many regions of the dorsal cortex, strongly encode trial history information that cannot simply be explained by movements. Neural dynamics representing trial history are low-dimensional and similar between subjects Having established a method to estimate neural activity driven by trial history rather than a set of other variables, such as chest position or head-orientation, we then sought to understand whether there is a limited set of neural population dynamics representing trial history shared across many neurons in the ACC. To address this question, we Orst applied principal component analysis (PCA) to the weights of trial history, the video SVD components or the neuron’s tuning to the chest position obtained from the full encoding models. We observed that within the top two dimensions, the four trial history contexts were well-separated ( Figure 4A ) while the Orst and third dimensions of chest point encoding showed well-deOned place tuning ( Figure 4B ). We then compared the number of dimensions required to capture at least 90% of the variance of the weights for trial history, chest position tuning and video SVD ( Figure 4C ). We found that relatively few PCA dimensions were necessary to explain the trial history encoding (12.0 ± 0.6) and chest position (17.2 ± 1.7). By contrast, many more dimensions were needed to explain the same amount of variance in the video SVD (63.0 ± 6.4, Figure 4D ). These Ondings indicate that trial history information is encoded in a small set of temporal dynamics and that there are far fewer chest tuning motifs in the ACC population than video SVD motifs. Download figure Open in new tab Figure 4. Neural dynamics representing trial history are low-dimensional and similar between subjects. (A) Trial history encoding weights projected into the two Frst principal component dimensions and separated by trial history. (B) Two dimensions of the chest point encoding weights. (C) Mean ± sem fraction of cumulative explained encoding weight variance for different variables as a function of PCA dimensions. (D) Minimum number of PCA dimensions required to explain 90% of the encoding weight variance. Grey lines represent individual subject, with circle and error bar depicts mean ± sem over subjects. ***P < 0.001, linear mixed-effects model with session nested in subject as random effects (E) Schematic representation of matrix transformations for procrustes analysis. (F) Trial history dynamics of all other sessions (grey) matched to the reference session (black) on A (top). Two dimensions of chest point encoding weights of an example session (right) matched to the reference session in B (left). (G) Dissimilarity (procrustes distance) of different regressors for different sessions within subjects and across subjects. ***P < 0.001, linear mixed-effects model with random effect of subject. Next, we hypothesized that if the neural dynamics for trial history represented a reliable monitoring process, they should be similar across different sessions from the same subject and even between subjects. To test this hypothesis, we used procrustes analysis 54 analysis. Procrustes analysis compares two geometric shapes with each other by Onding a set of matrix transformations that minimize the disparity between the two shapes ( Figure 4E ). The more similar two shapes are the smaller the procrustes distance. When we matched the trial history dynamics across different sessions, we observed very similar shapes ( Figure 4F , top ), whereas not all dimensions of chest point encoding matched well between two example sessions ( Figure 4F , bottom ). We then quantiOed dissimilarity (procrustes distance) of trial history, chest position and video SVD encoding between sessions from the same subjects and between different subjects ( Figure 4G ). We found that the within- and across subject procrustes distance of trial history encoding was signiOcantly lower than the distance of chest and video SVD tuning. Importantly, there was no difference between within- and across subject procrustes distance for trial history encoding weights. These results provide evidence for the hypothesis that trial history dynamics discovered by the linear encoding model might represent a monitoring process shared by different subjects. Discussion We set out to test whether ACC neurons encode behavioral history in fully deterministic perceptual decision-making tasks. To do this, we trained freely moving mice to make decisions about abstract features of visual stimuli and measured neural activity of excitatory ACC neurons using a miniaturized, head-mounted microscope. This approach allowed us to simultaneously record neural dynamics from hundreds of ACC neurons during unrestrained and self-initiated perceptual decision-making. We found that trial history influenced decision-making and that subjects weighed trial history differently in each session ( Figure 1E and F). We observed that the activity of individual ACC neurons was strongly modulated by trial history ( Figure 2C - E ). Using linear decoders, we showed that the neural representations of distinct trial history contexts are well separated across the entire trial ( Figure 2G ) and that these representations remained stable over several seconds ( Figure 2I and J). We showed that trial history explained large fractions of neural variance on the individual neuron level and that much of this variance could not be accounted for by other correlated task variables or movements ( Figure 3B and C). Finally, we found that trial history encoding is low-dimensional and that neural dynamics representing trial history over trial time are relatively similar across different subjects, unlike representations of posture and movement ( Figure 4F, G ). Neural signals in the anterior cingulate cortex are proposed to represent multiple cognitive signals including surprise, effortful decisions, and contextual-, task rule- or strategy representations 4 , 6 , 9 , 11 , 14 – 16 , 55 – 58 . We Ond that ACC neurons in expert mice performing a fully deterministic perceptual task encode trial history more strongly than other task variables. This is remarkable for two reasons. First, the trial structure of our perceptual task was random and lacked a block structure. This ensured that the only informative feature for the animals was the sensory evidence and any reliance on trial history hampered optimal performance. Nevertheless, individual excitatory ACC neurons and the excitatory ACC population faithfully tracked trial history. This suggests that behavioral history monitoring in ACC neurons may be independent of the volatility of the environment rather than emerging as result of having learned the statistics that govern its dynamics. These Ondings contrast with reports from an earlier study that showed that previous choice and outcome signals only emerge after introducing the Orst rule reversals in a set-shifting task 18 . However, in their study in head-restrained animals, trials were automatically triggered and animals licked to report their decisions. In contrast, mice in our task actively initiated trials and reported their choices by freely moving within an arena, adding strong volitional and spatial components to the task and making the behavior costly. It is possible that a combination of these factors may influence trial history representations in the ACC. We further note that Spellman and colleagues recorded from prelimbic rather than ACC neurons, which might explain the divergent observations. Second, a series of recent studies have shown that neural activity in many cortical regions of mice can be better explained by movement signals rather than true internal or cognitive variables 24 , 45 , 46 . Indeed, internal states seem to be highly intertwined with the stereotypy of movement patterns 48 and the two might be hard to separate 59 , 60 . In line with these reports, we Ond prominent representations of postural- and movement signals in many ACC neurons. However, we also observe strong trial history encoding that was not readily explainable by movements (in contrast to encoding of upcoming choice, for example). This encoding might thus represent an internal cognitive signal. The trial history encoding in ACC neurons displayed three main characteristics that are well-suited for monitoring behavioral history: rich encoding allowing for flexible readouts of information, temporal stability, and consistency across behavioral biases. First, many ACC neurons exhibited non-linear mixed selectivity with respect to animals’ previous choices and outcomes. Non-linear mixed selectivity supports efOcient encoding of task variable combinations in population activity while at the same time enabling flexible read-outs in down-stream regions 35 , 36 and neurons with non-linear mixed selectivity are wide-spread in frontal 50 , 52 , 53 and parietal 51 , 61 , 62 cortical regions. In agreement with this idea, we show that all combinations of previous choices and outcomes can be decoded from the ACC population activity without impairing the decoding of previous choice or outcome themselves. The ACC is thus well positioned to represent different aspects of trial history information and broadcast it to downstream targets. Our recordings also revealed that trial history representations remained stable for 1 - 2 seconds. This is reminiscent of reports suggesting that ACC neurons have long intrinsic- and task-related timescales 38 – 40 that may enable them to retain relevant task variable information into and beyond subsequent trials 16 , 18 . ACC neurons might thus integrate salient information about subjects’ actions and their outcomes over multiple trials, as has been shown in M2 63 and accumulate evidence that would suggest a change in the environment 11 . We observed that the trial history representations in ACC were highly consistent and did not change with behavioral biases or across sessions or even subjects. We took advantage of the Onding that animals performing our evidence accumulation task showed trial history biases that fluctuated between sessions to probe whether trial history representations would be influenced by these biases. We found that the encoding strength and decodability of trial history from ACC neurons was unaffected by subjects’ biases. This indicates that the ACC holds a faithful representation of trial history and might not directly drive biases associated with it. In agreement with this, recent studies in non-human primates and mice reported that ACC and prelimbic activity did not predict choice biases 17 and disruption of ACC activity had no impact on ongoing decision-making 12 , 13 , 15 , 18 . We further found that the dynamics encoding trial history were not only similar across sessions from the same subject but also consistent between subjects, in strong contrast to movement encoding that was highly variable between sessions and subjects. This result lends further support to the hypothesis that ACC neurons track information about previous actions and their outcomes in a generalized form as a cognitive signal that is distinct from highly variable and individualized movement signals. Accurate monitoring of past choices and outcomes is vital to discover regularities in one’s environment early in learning, but it also enables subjects to compare expected with obtained outcomes in dynamic environments to update these expectations if necessary. Our Ondings propose that the trial history monitoring in ACC provides a substrate for this comparison. The presence of trial history signals in expert mice performing a deterministic perceptual task with a Oxed rule may indicate that the animals continue evaluating their task knowledge even after attaining high performance and that they may be poised to explore alternative strategies at any moment 20 . We speculate that the evolutionary advantages of this readiness to switch to different strategies in light of new evidence have greatly outweighed the possible metabolic costs of constantly maintaining an active representation of behavioral history. Our observations raise questions that can be addressed in future studies. First, we do not know what aspects of the trial history representations might be exclusive to the ACC. Strong trial history signals have also been observed in other prefrontal regions, such as the prelimbic 18 or orbitofrontal cortex 64 , the M2 29 , 63 or the PPC 33 , 41 , 62 . It remains unclear whether trial history signals in these regions are similarly independent of movements or of other tracked task variables, and whether they display similar temporal stability and between subject similarity. Second, we have not directly tested the effects of ACC disruption on task performance because optogenetic intervention would only yield limited new insights. Due to the high non-linear mixed selectivity of ACC neurons for different trial history contexts and their mixing of trial history and movement signals it is difOcult to perturb neural dynamics that are speciOcally linked to trial history. Indeed, repeated strong optogenetic inhibition of ACC neurons was shown to decrease task engagement rather than influencing subjects’ decision-making 65 , 66 . Third, given that head-restrained behavioral tasks have become a staple in circuit neuroscience it will be crucial to directly compare the trial history and movement signals in the ACC of mice making perceptual decisions in head-Oxed and freely moving tasks. This is particularly important because head-restrained tasks may differ from freely moving behavior in aspects that were shown to influence ACC activity, including volitional task engagement, effort or affective components of discomfort and pain 67 . It thus remains an open question how these task-modality speciOc factors may influence trial history representations in the ACC. In summary, our study provides evidence that ACC neurons maintain an internal record of past decisions and outcomes that is separate from physical movements suggesting that ACC continuously monitors behavioral success even in highly predictable environments. Methods Animal subjects We used transgenic mice on a C57/B6 background that expressed the genetically encoded calcium indicator GCaMP6s in calmodulin-kinase II expressing cells under the control of a tetracycline response element (GCaMP6s was constitutively expressed in the absence of doxycycline). We obtained these subjects by crossing TRE-GCaMP6s mice (B6;DBA-Tg(tetO-GCaMP6s)2Niell/J, Jax stock no. 024742) 68 with CaMKII-tTA mice (B6.Cg-Tg(Camk2a-tTA)1Mmay/DboJ, Jax stock no. 007004) 69 . For our study we used at total of 14 mice of both sexes that had attained expert performance in our visual evidence accumulation task. Two of these animals were excluded because they did not meet our task performance and session metrics (see behavioral analysis below). We therefore included behavioral data from 8 males and 4 females. We obtained high quality imaging data from 6 of the 8 males and from 2 of the females and included these subjects into the imaging dataset. We further included imaging data from one male and one female performing auditory evidence accumulation that had initially been trained to discriminate visual stimuli. We did not Ond differences in trial history encoding in these mice and pooled their data with the other mice. For a complete summary please refer to Supplementary Table 1 . Mice were 8-25 weeks old when handling and water restriction started. Mice were housed in a room with controlled humidity (30 - 70%) and temperature (71 - 76 °F / 21.67 - 24.44°C). Mice were kept on a reverse dark-light cycle, with lights on at 10 am and lights off at 10 pm. Animals were provided with cotton nestlets and received an acrylic shelter and running wheel as additional enrichment. They were fed regular laboratory chow (NIH-31 ModiOed Open Formula Mouse/Rat Diet 7013) ad libitum and given additional treats (Mini Yogurt Drops, Bio-Serv) at the beginning of the day and after performing the task. Animals were water restricted during behavioral training and testing (see below). All behavioral and surgical procedures adhered to the guidelines established by the National Institutes of Health and were approved by the Institutional Animal Care and Use Committee of the University of California, Los Angeles, David Geffen School of Medicine. No statistical methods were used to pre-determine sample sizes. Sample sizes are similar to previous publications. View this table: View inline View popup Download powerpoint Supplementary Table 1: Summary of included subjects, task conditions, imaging sessions and coordinates. Habituation and water restriction A minimum of one week before the start of the behavioral training we began habituating the mice to being handled by the experimenters and let them walk and climb on the experimenters’ hands. During the handling animals were also presented with a treat (Mini yoghurt drop). At the same time, we restricted the mice’s access to water to 1 ml per subject and day and the water was given right after the mouse handling. To ensure that the subjects maintained healthy hydration levels, we measured their body weight daily during the handling or before running the behavioral task. If their weight fell beneath 80% of their body weight before water restriction they were supplemented with 0.4 ml of water after the behavioral session 70 . Perceptual evidence accumulation task We trained animals on an audio-visual evidence accumulation task 22 , 23 . In this task mice were required to integrate pulsatile sensory evidence and compare the number of experienced stimuli to an implicit category boundary (12 Hz). Trials were randomly assigned to be high- or low-rate and we did not impose any limitations as to how many times in a row certain trial types could repeat. Mice self-initiated trials by poking into a centrally located port (and thereby breaking an infrared laser beam). After a very brief random delay this triggered the playback of a train of visual (or auditory stimuli). The train was composed of a set of 15 ms long stimulus events that occurred pseudo-randomly distributed within a duration of 1 second. Stimulus events were required to be separated by at least 25 ms. We generated stimulus trains with 4, 6, 8, 10, 14, 16, 18, and 20 stimulus events, omitting 12 Hz stimulus trains, which would lie on the category boundary. After the stimulus train had Onished playing, we repeated the same stimulus train again. This made that stimulus information was still available when the mice were allowed to make their decisions after center Oxation. Mice had to remain poking in the central port for at least 1 second plus a random delay drawn from an exponential distribution, after which an auditory go-cue (7kHz) signaled to them that they could report their choice. The random delay was added to make sure that the animals could not simply time their response and to minimize movement preparation signals in the neural activity before the go-cue. If animals left the central port before the go-cue the trial was judged as an early withdrawal and the mice were presented with a white noise stimulus and 2 second timeout. Animals then reported their decision by poking into one of two ports either located on the left or the right side. Pokes in the left port after a 12 Hz stimulus were rewarded with a 4 - 6 µl drop of water, while left pokes after high-rate trials and right pokes after low-rate trials were punished with a tone (15 KHz) and 2 second timeout. If animals failed to report a choice within 30 s the trial ended and was deemed a no-choice trial. All the behavioral testing took place in sound attenuated booths that were illuminated only with infrared light (for video recordings). Animals performed the task inside a custom-made modular transparent red acrylic enclosure (20 x 20 cm, 24 high). Ports were installed along the back wall of the enclosure and separated by two 5 cm dividers made from clear acrylic. The task was implemented using custom written matlab functions and an arduino-based state machine (Bpod, Sanworks). All auditory and visual cues were delivered via a low-latency soundcard (Fenix, HT Omega) and sounds were played from two speakers located on either side of the central port (Harman-Kardon) while visual flashes were played from two LED panels attached above the center port. Water was delivered to the left and right ports through suspended syringes and silicone tubing and the volume was controlled via the opening time of a solenoid valve (Lee Company). Animal behavior was Olmed through the transparent acrylic floor from underneath the animals (Chameleon3, Teledyne-FLIR, Fujinon 3 MP Varifocal Lens). The videos were acquired at 100 fps and the camera was controlled using labcams ( https://github.com/jcouto/labcams ). Camera acquisition was synchronized to behavioral events via TTL pulses at the beginning of every trial. All the code to run the behavioral task, and assembly instructions and a parts list are publicly available ( https://github.com/churchlandlab/chipmunk ). Behavioral training procedure One day before the start of the behavioral training, water restricted mice were habituated to the task enclosure. They were presented with 1 ml water inside the left and right ports and in a weighing boat in front of the ports and allowed to drink and explore the enclosure for 5 - 10 minutes. At the beginning of the Orst full training sessions, the inside of all the three ports was baited with a drop of water. For the Orst 3 - 5 sessions the wait time and all the delays were set to 0 s, so that the mice did not need to Oxate inside the central port after triggering the stimulus. Similarly, for these sessions we removed the dividers between the ports to facilitate exploration, and we set the stimulus to keep repeating until mice made their decision or for up to 30 s (which is the maximum allowed time to choose a response port). We did this to help the mice temporally bind causal task events (center poke -> stimulus, stimulus playing -> response port poke). Once animals exhibited stable center-to-side sequences, we introduced the dividers and started increasing the required center Oxation time. For most animals we slowly increased the wait time by 0.3 ms after each successfully completed trial (i.e. no early withdrawal). It took the animals up to 3 months (10.72 ± 1.25 weeks) to reach > 1 s wait-times. We also gradually lowered the volume of the water rewards the animals earned after correct trials, starting at 8 μl at the beginning and ending at 4 - 6 μl at the expert stage. Most of the subjects were exposed to rewards and punishments from the Orst day of training. This allowed them to learn the task rule (with 4 and 20 Hz stimuli) within 2 - 3 weeks (10 - 15 sessions) and avoided the animals developing strong side-biases as has been reported before 23 . One cohort of mice (LO028 and LO032) included in this study was allowed to revise their choices and be rewarded on the correct port after initially selecting the incorrect one during the initial training process. This procedure, however, was more prone to producing Orst a switching bias that would have to be counteracted by manipulating the trial structure and present trials rewarded on one side more frequently leading in turn to a side bias. This side bias had then to be counteracted by presenting trials for the other side more frequently before returning to a balanced number of trials on both sides and removing any correlations in the trial structure. This approach therefore slowed down the animals’ learning of the stimulus rate - side association. We started introducing the more difOcult stimulus rates once the animals waited for > 1 s and performed at above 80% correct on the easy stimulus rates. Behavioral analysis ( Figure 1 and Supplementary Figure 1) We included subjects for the behavioral analyses based on a two-step selection process: First, we only included subjects that learned to wait for > 1 s (including the random delay) and that experienced all the different stimulus rates at least during one session and that performed above 80% on easiest trials at least once. Second, from these subjects we included only sessions where the animals actually waited > 1 s plus delay, experienced at least two different pairs of stimulus rates (for example 4-20 and 6 - 18 Hz), performed at least 100 trials and where the early withdrawal rate did not exceed 55%. We deliberately did not Olter individual sessions for high performance to capture the session-by-session variability in biases that might affect performance levels. Only animals performing the visual version of the task were included for the behavioral analyses. We Otted individual psychometric curves for each subject using all the completed trials from all their sessions and then averaged the recovered psychometric functions ( Figure 1C ). Psychometric functions were modeled using cumulative Gaussian functions parametrized by four parameters, perceptual bias (α), sensitivity (β), lower (γ)- and upper lapse rate (λ): where F (x; α, β) is the cumulative Gaussian function: We also Otted separate psychometric functions to trials with different preceding trial history ( Figure 1D ). To do this we only included completed trials that were also directly preceded by a completed trial (as opposed to an early withdrawal). Again, to have sufOcient amounts of data to estimate the psychometric parameters, we pooled trials from multiple sessions. We used a custom python package to Ot the four model parameters. ( https://github.com/jcouto/Ot_psychometric.git ). To assess the overall influence of sensory evidence and trial history on animals’ decisions and to analyze session-by-session variability we constructed logistic regression models that sought to predict the mice’s choice based on stimulus rate and trial history ( Figure 1E and F). Our models featured a regressor for stimulus rate plus a categorical regressor for each combination of previous choice and outcome (for example previous correct left choice) and an intercept (sensory bias): where x → Trial history is given by the vector: For the stimulus rate regressor the original stimulus rates were centered around 0 and scaled to their minimum and maximum. A 4 Hz stimulus was therefore encoded as −1 whereas a 20 Hz stimulus assumed a value of 1 with 0 representing a 12 Hz stimulus. The four trial history regressors were categorical regressors that were set to 1 whenever the trial history matched and were 0 otherwise. In these models, the intercept represents the log odds of choosing the right side during a 12 Hz stimulus in the absence of any trial history. When Otting the models we balanced the number of left- and right-choice trials by subsampling from the class with more observations. We iteratively sampled observations from the majority class 20 times, making sure to include all the observations with similar frequencies. For each of these 20 rounds of sampling we applied 10-fold cross-validation. Models were Otted with an l2 penalty set to 1. We chose not to optimize this penalty because the penalty influences the model weights. Thus, keeping the penalty the same allowed us to compare the model weights between different subjects and sessions. To capture the overall influence of trial history on animals’ choices we calculated the length of the vector of the four trial history weights (history strength, Figure 1G and H). All these analyses were performed using custom python code ( https://github.com/churchlandlab/chiCa ). We also modeled how choices and outcomes from the last and second last trials influenced animals’ decisions (Supplementary Figure 1G). We did this by including a regressor for the choice, the outcome and their interaction for the last (t-1) and second last (t-2) trial instead of the trial history regressors. These main effects were coded as 1 when the choice at trial t-1 or t-2 was to the right side or when the choice was correct and 0 otherwise. The interaction regressor was set to 1 when the choices were to the right and correct and 0 otherwise. The stimulus regressor was modeled the same way as in the trial history model. Note that when modeling the main effects of previous choices and previous outcomes, the intercept term no longer reflects a trial history free bias (at 12 Hz stimulus rate) but rather represents the condition with all dummy variables set to 0, an incorrect left choice at t-1 and t-2 with 12 Hz stimulus presented in this case. We also assessed whether sessions cluster with respect to their behavioral decoding weights from the trial history models (Supplementary Figure 1F). To this end, we used UMAP (using umap-learn ) to Ond a non-linear embedding to project the session’s 6 regressor weights into a 2-dimesional space and plotted all the sessions color-coded by subject (Supplementary Figure 1F). Surgical procedures For our neural recordings, we targeted an anterior and ventral portion of anterior cingulate cortex, which was proposed to be a homolog of the primate Brodmann Area 24 71 . To gain optical access to ACC on the right hemisphere we implanted 4 x 1 mm gradient refractive index (GRIN) lens centered at the following target coordinates (with respect to Bregma): AP: +1.3 mm, ML: 0.5 mm, VD: −1.3 mm (−1.5 mm from the skull surface). Before the surgery, animals were subcutaneously injected with meloxicam (2 mg/kg) for analgesia and enrofloxacin (5 mg/kg) as antibiotic treatment. Mice were then Orst anesthetized in 3% isoflurane before being maintained on 1 - 1.5% isoflurane throughout the surgical procedures. Body temperature was monitored and maintained close to 37°C using a feedback-controlled heating pad (PhysioSuite, Kent ScientiOc). Animals were placed in a stereotaxic frame (Kopf, Stoelting). We cut the hair on the scalp with scissors and subsequently used a hair removal cream (Nair) to remove remaining hair. After disinfecting the scalp, an incision was made and a diamond-shaped part of the skin was removed. We then carefully detached the neck muscles from the skull using forceps and scraped the skull clean with a scalpel. We next applied a layer of tissue adhesive glue (Vetbond, 3M) around the edges of the skull to seal off any exposed soft tissue. We then marked the position of the lens, drilled three very small craniotomies arranged in a triangle around the lens site, and inserted three micro screws into the skull (Fine Science Tools). We made a circular craniotomy with 1.2 mm diameter to later insert the GRIN lens. The dura underneath the craniotomy was carefully removed with sharp 30-gauge needles. We aspirated tissue using blunt 28- or 30-gauge needles using a custom vacuum suction apparatus. We were careful to aspirate a cylindrical volume rather than a conical one and to create a flat surface at the bottom of the aspirated volume. All excess blood was aspirated before lowering the lens into the craniotomy using a suction-based lens holder. After placing the lens, we cemented it to the skull and anchored it to the three screws using a three-component dental cement (C&B metabond, Parkell). We then built a support structure around the lens using black dental cement (Orthojet, Lang Dental). The tip of the lens was protected using the lid of a 1 ml eppendorf tube glued to the dental cement structure using acrylic glue (Zap-A-Gap). Mice were allowed to recover for 7 days after the surgery, and they were checked daily. On the Orst 4 days after surgery, they were injected with meloxicam and enrofloxacin. A minimum of 14 days animals were implanted with a miniscope baseplate. To this end animals were again anesthetized in 3% and maintained on 1 - 1.5% isoflurane anesthesia, and their body temperature was held constant. The eppendorf lid was detached using a drop of acetone and the lens surface was cleaned with ethanol. To provide additional contrast we painted the area around the GRIN lens with black nail polish. We then positioned the miniaturized microscope with a baseplate attached above the lens surface. At this point the anesthesia was lowered to 0.5% to decorrelated cortical activity in order to identify the best imaging focal plane. The baseplate was then attached using black dental cement. After the cement had dried, we removed the miniature microscope and covered the baseplate with a protective cap. We monitored the animals’ behavior immediately after recovery. This procedure was non-invasive and did not require administration of analgesics or antibiotics. Histology At the end of the experiments, animals were deeply anesthetized with an intraperitoneal injection of pentobarbital (Euthasol) and transcardially perfused with phosphate-buffered saline followed by 4% paraformaldehyde. Subsequently the brain was extracted and stored in 4% paraformaldehyde overnight for post-Oxation before being stored in phosphate- buffered saline. The brains were then embedded in 4% agarose and cut into 80 μm sections. Sections were mounted and nuclei were stained using DAPI inside the mounting medium (Fluoromount-G with DAPI, Invitrogen). Photomicrographs of GCaMP6s expression and DAPI were taken using a Nikon Eclipse Ti2 microscope equipped with a Yokogawa CSU-22 Spinning Disk unit to enable confocal fluorescence microscopy. Approximate coordinates of the GRIN lens center were obtained by comparing major landmarks on the sections (position of the anterior commissure, presence of corpus callosum) with a standard mouse brain atlas (Paxinos and Keith B. J. Franklin, 2007, 4 th edition). Miniature microscope imaging and image processing We imaged the neural activity of ACC neurons using a head-mounted miniaturized fluorescence microscope (Cai et al. , 2016, UCLA miniscope V4, https://github.com/Aharoni-Lab/Miniscope-v4 ). For every subject we determined the best focal plane by performing brief recordings at different imaging depths (by adjusting the electro tunable lens). We then kept this depth constant across recording sessions. We used excitation powers ranging from 22 to 37 mW and adjusted them to optimize imaging quality. Along with the images we also acquired data from gyroscopes on the miniaturized microscope and later reconstructed the animals head-orientation angles from these data. Images and gyroscope data were acquired at 30 frames per second. We synchronized the behavioral events and the imaging by logging the TTL pulses from the behavioral state machine indicating the start of a new trial and TTLs from the miniscope signaling the acquisition of a frame on the same computer using a Teensy (Teensy 3.2). To segment images and identify individual neurons we used a custom data handling- and processing pipeline ( https://github.com/jcouto/labdata-tools ) and built a plugin to run caiman-based image segmentation routines within this framework 37 ( https://github.com/flatironinstitute/CaImAn ). We Orst spatially down sampled the videos by a factor of 2 using the imresize function in Matlab (which yielded better results than comparable methods in python). We then corrected the videos for motion along the x and y axes using the normCorre implementation in caiman and subsequently ran the segmentation using constrained non-negative matrix factorization (CNMF-E). Raw traces were temporally detrended with the built-in caiman methods. We then visually inspected the identiOed putative neurons and excluded signals that were not likely originating from neurons. We used the following exclusion criteria: Location outside of the lens perimeter, aberrant shape and size, absence of activity above local background fluorescence. We then aligned the neural recording to the behavior by identifying the recording frames during which new trials started. At this stage, we also identiOed frame drops (unexpectedly long intervals between frame acquisition TTLs) and accounted for these by interpolating the neural signals. Movement and posture analysis From the gyroscope data acquired using the miniature microscope we reconstructed the animals’ head orientation at every imaging time point, which was given by the three angles: pitch, roll and yaw. To track the position of the animals in space we used deeplabcut 73 ( https://github.com/DeepLabCut/DeepLabCut ). We tracked the following 12 body parts that were visible at the underside of the mice: tail tip, tail center, rectum, genital, chest, nose tip, left foot, right foot, left hand, right hand, left ear, right ear. Although we ultimately only used the chest position for further analysis, we found that increasing the number of tracked body parts increased the tracking accuracy overall. We chose to focus on the chest point as it represents the most central tracked body part (to estimate position in space) and because it was rarely occluded. Models were trained with a total of 160 labeled frames from two different sessions, 80 frames each. We also generated low-dimensional versions of the full behavioral video. To do this we used an approximate singular value decomposition ( https://github.com/jcouto/wOeld ) on the mean-centered videos and retained the top 200 components. For every video we also generated a motion-energy video by subtracting each frame from its preceding one and then taking the absolute value of this difference. We then similarly decomposed these motion energy videos and retained the top 200 components for further analysis. Neural decoding analyses ( Figure 2 and Supplementary Figure 2) For the neural decoding analysis, we aligned the neural activity to four different phases of the inter-trial interval and the trial time. We deOned the start of the ITI as the moment when the animals had reported their choice, and we included neural activity for up to 1 s after (Early ITI). The next phase (Late ITI) was based on when the outcome ended. This moment was deOned as the animal leaving the rewarded port after a correct trial and at the end of the punishment noise and timeout (2 s after choice report) when the animals had made an incorrect choice. For the analysis we included the 1 s preceding this moment to 0.3 s after. We deOned the stimulus phase as lasting from 0.5 s before the onset of the visual stimuli to 1 s after and aligned the neural activity accordingly. Finally, we included data from 0.2 s before to 0.3 after animals left the Oxation in the center port (Action). We decoded the speciOc trial history context from the neural population activity at the different aligned time points. We Otted logistic regression models using the inferred spike rate signal convolved with a Gaussian kernel (1 frame standard deviation) to classify combinations of previous choice and outcome ( Figure 2F, G ), hence the decoders distinguished between 4 different classes using a one-versus-all approach. We balanced the number of classes in a similar way as described in behavioral analysis, subsampling from the majority classes. We then partitioned the subsampled data into a training set containing 80% of the data and a test set containing the remaining 20% while maintaining balanced labels and a test set containing the remaining 20% of the data. The activity of individual neurons in both sets were standardized by subtracting the mean of the training set and dividing it by the standard deviation of the training set. This avoided data leakage from the training to the test set. We then Orst optimized the magnitude of the L2 penalty by Onding the value that yielded best model predictions in the test fold within an exponential series from 10 - 10 to 10 12 . The best parameter was then used to Ot the logistic regression models. These procedures were repeated for the other 7 training-test splits. Finally, this procedure was repeated 20 times using different subsampled trials. The Onal model accuracies represent the averages over all sub-samplings and cross-validation folds. As a control, in parallel we Otted models where we shuffled the labels yielding chance level decoding accuracy. Decoding analyses to determine the neural decoding of previous choice or previous outcome only rather than all of their combinations were performed in the same way (Supplementary Figure 2A). When decoding previous choice or outcome alone, we made sure to equally balance the previous outcomes or choices, respectively, in the datasets. To estimate the trial history decoder performance on the binary problems of decoding only previous choice or outcome, we calculated the fraction of trials correctly classiOed with respect to the variable of interest, ignoring what the class label was for the irrelevant variable. To analyze the similarity of the neural representations between the different classes we constructed confusion matrices ( Figure 2H ). To do this, we counted how many times decoders correctly and incorrectly classiOed trials as each of the different trial history contexts for every given true trial history. Values are reported as the number of predicted trials divided by the number of true trials for each class. We also looked at how stable the neural encoding of trial history was on the population level ( Figure 2I and J). We did this by predicting the trial history contexts from neural activity at a given time point using the relationship (the logistic regression model) found at another time point. We did this for every pair of included time points. We made sure to do the predictions with the originally trained models from all folds and sub-samplings and then averaged the prediction accuracies rather than averaging the model coefOcients Orst and then performing the predictions. Linear encoding models ( Figure 3 and Supplementary Figure 3) To assess what task variables or movements drive neural variance we constructed a linear encoding model 24 . Importantly, our model was built such that it could not only capture relationships that were Oxed across trial time, such as neural tuning for speciOc locations in the task arena, but also relationships that could vary over time, such as how choice influences neural activity. This was achieved by including kernel regressors that could span parts or the entire aligned trial time ( Figure 3A ). These regressors were binary vectors that were set to 1 only at the time of an event. To for example capture how choice affected neural activity we added time-shifted versions of these vectors to the model until the time of the events spanned the entire trial time. Variables for which the kernel regressors spanned the entire trial time included choice, outcome and trial history (one for each combination of previous choice and outcome). Importantly, we also included a trial time variable, which consisted of regressors that were 1 for the same time point across trials. This trial time regressor effectively served as an intercept for the other time-shifted regressors. We also time-shifted regressors that tracked the stimulus events as kernels, so that we could detect neural responses to the stimulus events for up to 0.5 s after the stimuli had occurred. We also included pokes into all the different ports as regressors for instructed movements and allowed them to capture neural variance from 0.5 s before to 1 s after the poke happened. Head-orientation angle tuning was assessed by dividing the full range of 360° into 60 bins (6° per bin). We then built a binary vector whose value was one whenever the animal’s head-orientation angle was within that bin and zero otherwise. We did this separately for pitch, roll and yaw. In an analogous manner we constructed regressors that tracked whether a mouse’s chest was within a spatial bin or not. The bins covered the entire camera Oeld of view and covered a 1.28 x 1.28 cm area. Finally, we included the top 200 video and video motion energy components as analog regressors in the design matrix. To make sure that the video and video motion energy components only contained information about the movements rather than the visual flashes that were also visible on the video, we orthogonalized these regressors against the stimulus regressors using QR-decomposition before Otting the linear models. We then aligned the raw fluorescence traces for all the neurons to the speciOed trial events in the four different trial phases (as described in decoding analysis). To Ot the models we used ridge regression, which regularizes the model weights using an L2 penalty. We did this to handle the high degree of multi-collinearity of the different regressors (trial history and chest point, for example) and to counteract over-Otting due to the large number of included regressors. To Ot the models, we Orst split the data in 10 folds (see decoding analysis above for detailed description of the procedure) making sure that we equally sampled all trial time points. We then standardized our analog regressors (the video SVD and video motion energy SVD) in the training set and used the mean and standard deviation from the training set to standardize the regressors in the test set. We then standardized the neural activity in a similar fashion. We then Orst determined the optimal regularization strength for each neuron by Otting the models using an exponential scale of regularization strengths from 10 -3 to 10 5 and selecting the value that yielded the highest explained variance in the test set as assessed using the coefOcient of determination (R 2 ). We then reOtted the models with that value before repeating this procedure in the next fold. Besides Otting the full models we also Otted single-variable only models, where the regressors for all but one variable were shuffled, and one-removed models, where the regressors of only a single variable were shuffled. We then computed the maximal explained variance per neuron for each variable ( Figure 3B ) as the single- variable only variance of that variable minus the trial time-only model variance (representing the intercept). Analogously, we computed the unique variance ( Figure 3C ) as the difference between the full model and the one-removed model variances. To reconstruct the explained variance for speciOc variables over trial time ( Figure 3D ) we retained the predicted activity for each neuron of all the folds and then re-ordered them to reflect original trial time. We then calculated the coefOcient of determination for each time point separately. Note however, that the models were Otted and regularized to explain variance across the entire trial time and so the estimates of explained variance for individual trial times tends to be lower than the full model explained variance. Variable dimensionality and procrustes analysis ( Figure 4 ) We assessed the dimensionality of the neural representation of speciOc movements or task variables by separately applying principal component analysis to the encoding model weights (from the full model) of the different variables. We then deOned the dimensionality as the minimum number of dimensions required to explain 90% of the weight variance across the neuron population ( Figure 4C and D). We measured the similarity of the population representation of task variables within and between subjects using procrustes analysis (Horrocks, Rodrigues and Saleem, 2024, Figure 4E, F and G). Procrustes analysis seeks to minimize the (least-squares) euclidean distance between two shapes by identifying the optimal matrix translation, scaling and rotation of the target shape: Where ||X|| = {trace(X T X)}½ is the eudlidean norm, X 1 denotes the shape being matched, X 2 the reference shape, β is the scale parameter, Γ the rotation matrix and κ the location vector. To match the shapes of the matrices for procrustes analysis and to ensure that we were considering meaningful low-dimensional dynamics or behavioral tuning motifs we determined the sessions with maximal dimensionality (i.e. dimensions necessary to explain 90% of the variance) and used this number of dimensions for all the other sessions. We did this for every variable separately. The dissimilarity between shapes was expressed as the procrustes distance, where identical shapes have a distance of 0, whereas a value of 1 indicates maximal dissimilarity. Data reporting and statistical analyses Unless indicated otherwise data are presented as mean ± standard error of the mean (sem) over subjects. All boxplots show the median ± interquartile range of the data and the whiskers depict 1.5 times the interquartile range. Means are shown on boxplots as white circles. Where multiple sessions were acquired per subject the sessions were averaged Orst before plotting the averages over subjects. However, the statistics were done on the original data. Because we repeatedly sampled neural activity from the same Oeld of view, several neurons are likely present in multiple recordings. To avoid representing multiple samples from the same neurons, on Ogures showing data from individual neurons we only include the session with the most recorded neurons per mouse. All statistical comparisons were performed using linear-mixed effects models (lmerTest, R) with subject identity as a random effect. When multiple measures within a session were analyzed, we included the session identity as a random effect nested within subject. We only included random intercepts in our mixed-effects models. Post-hoc group means were compared by estimating the marginal means from the mixed-model Ots using emmeans (R). Pearson correlation coefOcients and their signiOcance were computed using scipy.stats.pearsonr. Detailed descriptions and results of all the mixed effects models are provided in the Supplementary Statistics Table 1. Data and code availability All code is publicly available. Wherever speciOc packages are used the packages are cited in the corresponding methods section. The code to run the analyses and statistical models, and to generate the Ogure panels can be found in the following repository: https://github.com/LukasOesch/Oesch_et_al_Trial_History_ACC.git . All the data used in this study will be made available upon publication. Author contributions LTO and AKC designed the project, LTO and JC adapted the behavioral task from a previous version, LTO, MCT and DS trained animals, LTO performed surgeries and collected imaging data, LTO and JC wrote analysis tools, LTO analyzed and visualized the data, LTO and AKC wrote the manuscript with inputs from all other authors. Supplementary Material Download figure Open in new tab Supplementary Figure 1. The behavioral influence of most recent trial history varies between sessions. (A, B, C, D) Estimates of the psychometric parameters, bias (A), sensitivity (B), upper lapse rate (C), and lower lapse rate (D). Scatter points on the left represent individual subjects while dots with errorbars depict mean ± sem over subjects. *P < 0.05, **P < 0.01. (E) Variance of the session-by-session weights of the behavioral choice decoding. Boxes represent median ± interquartile range and dots depict the mean. All comparisions are made agains the bias (intercept). ***P < 0.001. (F) Non-linear embedding of the behavioral choice decoding weights. Scatter points represent sessions color- coded by subject. (G) Behavioral decoding models including regressors for most recent trial history (t-1) and trial history from two trials back (t-2). Lines are individual subjects while dots with errorbars show mean ± sem over subjects. Download figure Open in new tab Supplementary Figure 2. Decoders trained on trial history decode previous choice and outcome equally well as corresponding binary classiGers. (A) Mean ± sem. decoding accuracy across trial time for decoders trained on previous choice- (orange) or previous outcome (teal) only. (B) Mean ± sem. neural decoding accuracy for previous choice or previous outcome for their corresponding binary decoders or trial history decoders (trained to decode combinations of previous choices and outcomes), linear mixed-effects model with session nested in subject as random effects. Download figure Open in new tab Supplementary Figure 3. The linear encoding model identiGes the unique contribution of the regressors to the neural activity. (A) Comparison of encoding model weights from an example neuron for model where the regressors for all except a single variable are shuffled (red) and the full models without shuffling (dark grey). Trial history weights (left), yaw tuning (center) and chest point tuning (right) are shown. (B) Correlation between maximal explained variance for different variables in all neurons from an example session. Note the high correlation between trial history and yaw and chest point. (C) Correlation of the unique explained variance between variables for the same session as in (B). (D and E) Mean ± sem reconstructed time course of maximal (D) and unique explained variance (E) for different variables. Download figure Open in new tab Supplementary Figure 4. Neural encoding of trial history in ACC is not correlated with the strength of trial history biases. (A) Neural population decoding of trial history is not correlated with the behavioral usage of trial history information. Each point is a single animal. (B) Decoding accuracy for speciFc trial history contexts is independent of the magnitude of the corresponding behavioral model weight. (C and D) Neither maximal nor unique variance of individual neurons explained by trial history are correlated with behavioral trial history biases. Download figure Open in new tab Supplementary Figure 5: Female mice show weaker trial history representations than male mice. (A) Anatomical reconstruction of GRIN lens position (its center) separated by subject sex. (B and E) Mean trial history decoding accuracy by lens position along the anterior-posterior (B) and dorsal-ventral axis (E) separated by subject sex. We found a signiFcant main effect of sex (p > 0.01) but no effects for AP or DV lens position or their interaction, linear-mixed effects model with subject as random effect. (C and F) Mean maximal explained variance for trial history as a function of GRIN lens position and color-coded by sex. We observed signiFcant effects of sex (p < 0.01), AP position (p < 0.01), DV position (p < 0.01) and the interaction between AP and DV position (p < 0.01), linear mixed effects model with random effect for subject. (D and G) Average unique explained variance for trial history by GRIN lens AP (D) and DV (G) position separated by subject sex. We found signiFcant effects of sex (p < 0.01), but no effects for AP or DV position or their interaction, linear mixed-effects model with subject as random effect. Acknowledgements We thank Letizia Ye and Marvin Vasquez for animal training and care, Marvin Vasquez for acquiring the histology photomicrographs and Anup Khanal for technical help. We thank Pinping Zhao for guidance on tissue aspiration and GRIN lens implantation procedures and Federico Sangiuliano Jimka for technical assistance with the V4 minscopes. We would also like to thank all the lab members for their feedback and discussions. This work was supported by NSF-NCS collaborative award (2219946), the NIH R01 EY022979 to AKC and the Swiss National Science Foundation Early Postdoc Mobility fellowship (P2BEP3) to LTO. Funder Information Declared National Institutes of Health, https://ror.org/01cwqze88 , NIH NEI 022979 Footnotes We replaced Supp. Fig 3 with a corected version. References 1. ↵ Monosov , I. E. , Haber , S. N. , Leuthardt , E. C. & Jezzini , A . Anterior Cingulate Cortex and the Control of Dynamic Behavior in Primates . Curr. Biol . 30 , R1442 – R1454 ( 2020 ). OpenUrl CrossRef PubMed 2. ↵ Alexander , W. H. & Brown , J. W . Medial prefrontal cortex as an action-outcome predictor . Nat. Neurosci . 14 , 1338 – 1344 ( 2011 ). OpenUrl CrossRef PubMed 3. ↵ Soltani , A. & Izquierdo , A . Adaptive learning under expected and unexpected uncertainty . Nat. Rev. Neurosci . 20 , 635 – 644 ( 2019 ). OpenUrl CrossRef PubMed 4. ↵ Hayden , B. Y. , Heilbronner , S. R. , Pearson , J. M. & Platt , M. L . Surprise signals in anterior cingulate cortex: Neuronal encoding of unsigned reward prediction errors driving adjustment in behavior . J. Neurosci . 31 , 4178 – 4187 ( 2011 ). OpenUrl Abstract / FREE Full Text 5. ↵ Hyman , J. M. , Holroyd , C. B. & Seamans , J. K . A Novel Neural Prediction Error Found in Anterior Cingulate Cortex Ensembles . Neuron 95 , 447 – 456 .e3 ( 2017 ). OpenUrl CrossRef PubMed 6. ↵ Seo , H. & Lee , D . Temporal Oltering of reward signals in the dorsal anterior cingulate cortex during a mixed-strategy game . J. Neurosci . 27 , 8366 – 8377 ( 2007 ). OpenUrl Abstract / FREE Full Text 7. ↵ Quilodran , R. , Rothé , M. & Procyk , E . Behavioral Shifts and Action Valuation in the Anterior Cingulate Cortex . Neuron 57 , 314 – 325 ( 2008 ). OpenUrl CrossRef PubMed Web of Science 8. Karlsson , M. P. , Tervo , D. G. R. & Karpova , A. Y . Network resets in medial prefrontal cortex mark the onset of behavioral uncertainty . Science (80-.) . 338 , 135 – 139 ( 2012 ). OpenUrl Abstract / FREE Full Text 9. ↵ Tervo , D. G. R. et al. Behavioral variability through stochastic choice and its gating by anterior cingulate cortex . Cell 159 , 21 – 32 ( 2014 ). OpenUrl CrossRef PubMed 10. ↵ Bissonette , G. B. , Powell , E. M. & Roesch , M. R . Neural structures underlying set- shifting: Roles of medial prefrontal cortex and anterior cingulate cortex . Behav. Brain Res . 250 , 91 – 101 ( 2013 ). OpenUrl CrossRef PubMed Web of Science 11. ↵ Sarafyazd , M. & Jazayeri , M . Hierarchical reasoning by neural circuits in the frontal cortex . Science (80-) . 364 , ( 2019 ). 12. ↵ Cole , N. , Harvey , M. , Myers-Joseph , D. , Gilra , A. & Khan , A. G . Prediction-error signals in anterior cingulate cortex drive task-switching . Nat. Commun . 15 , 7088 ( 2024 ). OpenUrl CrossRef PubMed 13. ↵ Chen , W. , Liang , J. , Wu , Q. & Han , Y . Anterior cingulate cortex provides the neural substrates for feedback-driven iteration of decision and value representation . Nat. Commun . 15 , 1 – 15 ( 2024 ). OpenUrl CrossRef PubMed 14. ↵ Tervo , D. G. R. et al. The anterior cingulate cortex directs exploration of alternative strategies . Neuron 109 , 1876 – 1887 .e6 ( 2021 ). OpenUrl CrossRef PubMed 15. ↵ Akam , T. et al. The Anterior Cingulate Cortex Predicts Future States to Mediate Model-Based Action Selection . Neuron 109 , 149 – 163 .e7 ( 2021 ). OpenUrl CrossRef PubMed 16. ↵ Sul , J. H. , Kim , H. , Huh , N. , Lee , D. & Jung , M. W . Distinct roles of rodent orbitofrontal and medial prefrontal cortex in decision making . Neuron 66 , 449 – 460 ( 2010 ). OpenUrl CrossRef PubMed Web of Science 17. ↵ Li , Y. S. , Nassar , M. R. , Kable , J. W. & Gold , J. I . Individual neurons in the cingulate cortex encode action monitoring, not selection, during adaptive decision-making . J. Neurosci . 39 , 6668 – 6683 ( 2019 ). OpenUrl Abstract / FREE Full Text 18. ↵ Spellman , T. , Svei , M. , Kaminsky , J. , Manzano-Nieves , G. & Liston , C . Prefrontal deep projection neurons enable cognitive flexibility via persistent feedback monitoring . Cell 184 , 2750 – 2766 .e17 ( 2021 ). OpenUrl CrossRef PubMed 19. ↵ Muysers , H. et al. A persistent prefrontal reference frame across time and task rules . Nat. Commun . 15 , 1 – 14 ( 2024 ). OpenUrl CrossRef PubMed 20. ↵ Pisupati , S. , Chartarifsky-Lynn , L. , Khanal , A. & Churchland , A. K . Lapses in perceptual decisions reflect exploration . Elife 10 , 1 – 27 ( 2021 ). OpenUrl CrossRef PubMed 21. ↵ Gershman , S. J. Uncertainty and exploration . Decision 6 , 277 – 286 ( 2019 ). OpenUrl CrossRef PubMed 22. ↵ Raposo , D. , Sheppard , J. P. , Schrater , P. R. & Churchland , A. K . Multisensory decision-making in rats and humans . J. Neurosci . 32 , 3726 – 3735 ( 2012 ). OpenUrl Abstract / FREE Full Text 23. ↵ Odoemene , O. , Pisupati , S. , Nguyen , H. & Churchland , A. K . Visual evidence accumulation guides decision-making in unrestrained mice . J. Neurosci . 38 , 10143 – 10155 ( 2018 ). OpenUrl Abstract / FREE Full Text 24. ↵ Musall , S. , Kaufman , M. T. , Juavinett , A. L. , Gluf , S. & Churchland , A. K . Single-trial neural dynamics are dominated by richly varied movements . Nat. Neurosci . 22 , 1677 – 1686 ( 2019 ). OpenUrl CrossRef PubMed 25. ↵ Miri , A. , Daie , K. , Burdine , R. D. , Aksay , E. & Tank , D. W . Regression-based identiOcation of behavior-encoding neurons during large-scale optical imaging of neural activity at cellular resolution . J. Neurophysiol . 105 , 964 – 980 ( 2011 ). OpenUrl CrossRef PubMed Web of Science 26. ↵ Lak , A. et al. Reinforcement biases subsequent perceptual decisions when conOdence is low: A widespread behavioral phenomenon . Elife 9 , 1 – 26 ( 2020 ). OpenUrl CrossRef PubMed 27. Gupta , D. , DePasquale , B. , Kopec , C. D. & Brody , C. D . Trial-history biases in evidence accumulation can give rise to apparent lapses in decision-making . Nat. Commun . 15 , 662 ( 2024 ). 28. Braun , A. , Urai , A. E. & Donner , T. H . Adaptive history biases result from conOdence-weighted accumulation of past choices . J. Neurosci . 38 , 2418 – 2429 ( 2018 ). OpenUrl Abstract / FREE Full Text 29. ↵ Jiang , W. , Liu , J. , Zhang , D. , Xie , T. & Yao , H . Short-Term Influence of Recent Trial History on Perceptual Choice Changes with Stimulus Strength . Neuroscience 409 , 1 – 15 ( 2019 ). OpenUrl CrossRef PubMed 30. Mochol , G. , Kiani , R. & Moreno-Bote , R . Prefrontal cortex represents heuristics that shape choice bias and its integration into future behavior . Curr. Biol . 31 , 1234 – 1244 .e6 ( 2021 ). OpenUrl CrossRef PubMed 31. ↵ Busse , L. et al. The detection of visual contrast in the behaving mouse . J. Neurosci . 31 , 11351 – 11361 ( 2011 ). OpenUrl Abstract / FREE Full Text 32. ↵ Hermoso-Mendizabal , A. et al. Response outcomes gate the impact of expectations on perceptual decisions . Nat. Commun . 11 , 1 – 13 ( 2020 ). OpenUrl CrossRef PubMed 33. ↵ Hwang , E. J. , Dahlen , J. E. , Mukundan , M. & Komiyama , T . History-based action selection bias in posterior parietal cortex . Nat. Commun . 8 , 1 – 14 ( 2017 ). OpenUrl CrossRef PubMed 34. ↵ Cai , D. J. et al. A shared neural ensemble links distinct contextual memories encoded close in time . Nature 534 , 115 – 118 ( 2016 ). OpenUrl CrossRef PubMed 35. ↵ Rigotti , M. et al. The importance of mixed selectivity in complex cognitive tasks . Nature 497 , 585 – 590 ( 2013 ). OpenUrl CrossRef PubMed Web of Science 36. ↵ Kaufman , M. T. et al. The implications of categorical and category-free mixed selectivity on representational geometries . Curr. Opin. Neurobiol . 77 , 102644 ( 2022 ). 37. ↵ Zhou , P. et al. EfOcient and accurate extraction of in vivo calcium signals from microendoscopic video data . Elife 7 , 1 – 37 ( 2018 ). OpenUrl CrossRef PubMed 38. ↵ Murray , J. D. et al. A hierarchy of intrinsic timescales across primate cortex . Nat. Neurosci . 17 , 1661 – 1663 ( 2014 ). OpenUrl CrossRef PubMed 39. Maisson , D. J. N. et al. Choice-relevant information transformation along a ventrodorsal axis in the medial prefrontal cortex . Nat. Commun . 12 , 1 – 14 ( 2021 ). OpenUrl CrossRef PubMed 40. ↵ Song , M. et al. Hierarchical gradients of multiple timescales in the mammalian forebrain . Proc. Natl. Acad. Sci . 121 , 2017 ( 2024 ). OpenUrl CrossRef 41. ↵ Akrami , A. , Kopec , C. D. , Diamond , M. E. & Brody , C. D . Posterior parietal cortex represents sensory history and mediates its effects on behaviour . Nature 554 , 368 – 372 ( 2018 ). OpenUrl CrossRef PubMed 42. Bonaiuto , J. J. , De Berker , A. & Bestmann , S . Response repetition biases in human perceptual decisions are explained by activity decay in competitive attractor models . Elife 5 , 1 – 28 ( 2016 ). OpenUrl CrossRef PubMed 43. ↵ Cavanagh , S. E. , Towers , J. P. , Wallis , J. D. , Hunt , L. T. & Kennerley , S. W . Reconciling persistent and dynamic hypotheses of working memory coding in prefrontal cortex . Nat. Commun . 9 , 3498 ( 2018 ). OpenUrl CrossRef PubMed 44. ↵ NajaO , F. et al. Excitatory and Inhibitory Subnetworks Are Equally Selective during Decision-Making and Emerge Simultaneously during Learning . Neuron 105 , 165 – 179 .e8 ( 2020 ). OpenUrl CrossRef PubMed 45. ↵ Stringer , C. et al. Spontaneous behaviors drive multidimensional, brainwide activity . Science (80-.) . 364 , ( 2019 ). 46. ↵ Salkoff , D. B. , Zagha , E. , McCarthy , E. & McCormick , D. A . Movement and Performance Explain Widespread Cortical Activity in a Visual Detection Task . Cereb. Cortex 30 , 421 – 437 ( 2020 ). OpenUrl CrossRef PubMed 47. ↵ Musall , S. et al. Pyramidal cell types drive functionally distinct cortical activity patterns during decision-making . Nat. Neurosci . 26 , 495 – 505 ( 2023 ). OpenUrl PubMed 48. ↵ Yin , C. et al. Spontaneous movements and their impact on neural activity fluctuate with latent engagement states . bioRxiv 2023.06.26.546404 ( 2024 ). 49. ↵ Tremblay , S. , Testard , C. , DiTullio , R. W. , Inchauspé , J. & Petrides , M . Neural cognitive signals during spontaneous movements in the macaque . Nat. Neurosci . 26 , 295 – 305 ( 2023 ). OpenUrl CrossRef PubMed 50. ↵ Mante , V. , Sussillo , D. , Shenoy , K. V. & Newsome , W. T . Context-dependent computation by recurrent dynamics in prefrontal cortex . Nature 503 , 78 – 84 ( 2013 ). OpenUrl CrossRef PubMed Web of Science 51. ↵ Kira , S. , Safaai , H. , Morcos , A. S. , Panzeri , S. & Harvey , C. D . A distributed and efOcient population code of mixed selectivity neurons for flexible navigation decisions . Nat. Commun . 14 , ( 2023 ). 52. ↵ Diehl , G. W. & Redish , A. D . Differential processing of decision information in subregions of rodent medial prefrontal cortex . Elife 12 , ( 2023 ). 53. ↵ Posani , L. , Wang , S. , Muscinelli , S. , Paninski , L. & Fusi , S . Rarely categorical, always high-dimensional: how the neural code changes along the cortical hierarchy . bioRxiv ( 2024 ). doi: 10.1101/2024.11.15.623878 OpenUrl Abstract / FREE Full Text 54. ↵ Horrocks , E. A. B. , Rodrigues , F. R. & Saleem , A. B . Flexible neural population dynamics govern the speed and stability of sensory encoding in mouse visual cortex . Nat. Commun . 15 , ( 2024 ). 55. ↵ Reinert , S. , Hübener , M. , Bonhoeffer , T. & Goltstein , P. M . Mouse prefrontal cortex represents learned rules for categorization . Nature 593 , 411 – 417 ( 2021 ). OpenUrl CrossRef PubMed 56. Hart , E. E. , Blair , G. J. , O’Dell , T. J. , Blair , H. T. & Izquierdo , A . Chemogenetic modulation and single-photon calcium imaging in anterior cingulate cortex reveal a mechanism for effort-based decisions . J. Neurosci . 40 , 5628 – 5643 ( 2020 ). OpenUrl Abstract / FREE Full Text 57. Stolyarova , A. et al. Contributions of anterior cingulate cortex and basolateral amygdala to decision conOdence and learning under uncertainty . Nat. Commun . 10 , ( 2019 ). 58. ↵ Holroyd , C. B. & Yeung , N . Motivation of extended behaviors by anterior cingulate cortex . Trends Cogn. Sci . 16 , 122 – 128 ( 2012 ). OpenUrl CrossRef PubMed Web of Science 59. ↵ Hasnain , M. A. et al. Separating cognitive and motor processes in the behaving mouse . Nat. Neurosci . ( 2024 ). doi: 10.1038/s41593-024-01859-1 OpenUrl CrossRef 60. ↵ Cazettes , F. et al. A reservoir of foraging decision variables in the mouse brain . Nat. Neurosci . 26 , 840 – 849 ( 2023 ). OpenUrl CrossRef PubMed 61. ↵ Raposo , D. , Kaufman , M. T. & Churchland , A. K . A category-free neural population supports evolving demands during decision-making . Nat. Neurosci . 17 , 1784 – 1792 ( 2014 ). OpenUrl CrossRef PubMed 62. ↵ Morcos , A. S. & Harvey , C. D . History-dependent variability in population dynamics during evidence accumulation in cortex . Nat. Neurosci . 19 , 1672 – 1681 ( 2016 ). OpenUrl CrossRef PubMed 63. ↵ Siniscalchi , M. J. , Wang , H. & Kwan , A. C . Enhanced Population Coding for Rewarded Choices in the Medial Frontal Cortex of the Mouse . Cereb. Cortex 29 , 4090 – 4106 ( 2019 ). OpenUrl CrossRef PubMed 64. ↵ Hocker , D. , Brody , C. D. , Savin , C. & Constantinople , C. M . Subpopulations of neurons in LOFC encode previous and current rewards at time of choice . Elife 10 , 1 – 26 ( 2021 ). OpenUrl CrossRef PubMed 65. ↵ Vázquez , D. et al. Optogenetic Inhibition of Rat Anterior Cingulate Cortex Impairs the Ability to Initiate and Stay on Task . J. Neurosci . 44 , 1 – 11 ( 2024 ). OpenUrl CrossRef PubMed 66. ↵ González , V. V et al. A common stay-on-goal mechanism in anterior cingulate cortex for information and effort choices . eneuro ENEURO .0454-24.2025 ( 2025 ). doi: 10.1523/ENEURO.0454-24.2025 OpenUrl Abstract / FREE Full Text 67. ↵ Acuña , M. A. , Kasanetz , F. , De Luna , P. , Falkowska , M. & Nevian , T . Principles of nociceptive coding in the anterior cingulate cortex . Proc. Natl. Acad. Sci . 120 , 2017 ( 2023 ). OpenUrl CrossRef 68. ↵ Wekselblatt , J. B. , Flister , E. D. , Piscopo , D. M. & Niell , C. M . Large-scale imaging of cortical dynamics during sensory perception and behavior . J. Neurophysiol . 115 , 2852 – 2866 ( 2016 ). OpenUrl CrossRef PubMed 69. ↵ Mayford , M. et al. Control of memory formation through regulated expression of a CaMKII transgene . Science (80-.) . 274 , 1678 – 1683 ( 1996 ). OpenUrl Abstract / FREE Full Text 70. ↵ Guo , Z. V. et al. Procedures for behavioral experiments in head-Oxed mice . PLoS One 9 , ( 2014 ). 71. ↵ van Heukelum , S. et al. Where is Cingulate Cortex? A Cross-Species View . Trends Neurosci . 43 , 285 – 299 ( 2020 ). OpenUrl CrossRef PubMed 72. Paxinos , G. & Keith B. J. , Franklin , M. A. The Mouse Brain in Stereotaxic Coordinates . ( Elsevier Science , 2007 ). 73. ↵ Mathis , A. et al. DeepLabCut: markerless pose estimation of user-deOned body parts with deep learning . Nat. Neurosci . 21 , 1281 – 1289 ( 2018 ). OpenUrl CrossRef PubMed View the discussion thread. Back to top Previous Next Posted July 02, 2025. Download PDF 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 Anterior cingulate cortex mixes retrospective cognitive signals and ongoing movement signatures during decision-making 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 Anterior cingulate cortex mixes retrospective cognitive signals and ongoing movement signatures during decision-making Lukas T. Oesch , Makenna C. Thomas , Davis Sandberg , João Couto , Anne K. Churchland bioRxiv 2025.04.11.648398; doi: https://doi.org/10.1101/2025.04.11.648398 Share This Article: Copy Citation Tools Anterior cingulate cortex mixes retrospective cognitive signals and ongoing movement signatures during decision-making Lukas T. Oesch , Makenna C. Thomas , Davis Sandberg , João Couto , Anne K. Churchland bioRxiv 2025.04.11.648398; doi: https://doi.org/10.1101/2025.04.11.648398 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 (7635) Biochemistry (17697) Bioengineering (13895) Bioinformatics (41951) Biophysics (21456) Cancer Biology (18594) Cell Biology (25520) Clinical Trials (138) Developmental Biology (13381) Ecology (19903) Epidemiology (2067) Evolutionary Biology (24323) Genetics (15612) Genomics (22510) Immunology (17738) Microbiology (40401) Molecular Biology (17184) Neuroscience (88622) Paleontology (667) Pathology (2833) Pharmacology and Toxicology (4825) Physiology (7644) Plant Biology (15158) Scientific Communication and Education (2046) Synthetic Biology (4296) Systems Biology (9825) Zoology (2271)