Excitatory neurons of the anterior cingulate cortex encode chosen actions and their outcomes rather than cognitive state

23 The anterior cingulate cortex (ACC) causally influences cognitive control of goal -directed 24 behaviour. However, it is unclear whether ACC directly encodes cognitive vari ables like 25 attention or impulsivity, or implements goal -directed action selection mechanisms that are 26 modulated by them. We recorded ACC activity with miniature endoscopic microscopes in mice 27 performing the 5- choice-serial-reaction time task, and applied decoding and encoding 28 analyses. ACC pyramidal cells represented specific actions before and during the behavioural 29 response, whereas the response type (e.g. correct/incorrect/premature) – indicating the state 30 of attentional and impulse control – could only be decoded during and after the response with 31 high reliability. Devaluation and extinction experiments further revealed that action encoding 32 depended on reward expectation. Our findings support a role for ACC in goal-directed action 33 selection and monitoring, that is modulated by cognitive state, rather than in tracking levels of 34 attention or impulsivity directly in individual trials. 35 36 37 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 3

38 Maintenance of high levels of sustained attention and inhibition of impulsive responding are 39 key to successful goal-directed behaviour, and impaired in a variety of psychiatric disorders 40 [1,2]. Both aspects can be measured by the 5-choice-serial-reaction-time task (5-CSRTT) [3,4] 41 in both humans and rodents. In this task, subjects can make four different types of response, 42 indicative of different cognitive states: (i) they can correctly respond to a stimulus presented 43 briefly and after a considerable waiting time to earn a reward (correct response), requiring 44 high attentional and impulse control. (ii) they can, instead, follow the impulsive urge to respond 45 before cue -presentation (premature response, indicating reduced impulse control), or ( iii) 46 respond into a non-cued hole (incorrect response, indicating reduced attentional control). (iv) 47 Alternatively, they may not respond at all (omission, indicating reduced task engagement or 48 inattention). Therefore, the measurement of neurophysiological correlates of these four 49 response options, promises to identify circuits that regulate attention, impulse control, and 50 possibly other aspects of deterministic goal-directed behaviour. 51 Several rodent studies have implicated the anterior cingulate cortex (ACC) in this regulation. 52 Manipulations of rodent ACC have been shown to produce shifts in the relative occurrence of 53 these behavioural outcomes in the 5- CSRTT which support a causal role of this brain 54 structure. For example, the activation of G i-protein signalling in excitatory pyramidal cells, 55 either in all layers or in layer 5 exclusively, may reduce premature and, partly, increase correct 56 responding [5]. In contrast, the chemogenetic inhibition of a subgroup of ACC neurons 57 projecting to the visual cortex may induce a shift from correct responding to response omission 58 [6], whereas their pre-cue stimulation at 30 Hz after such errors may have the opposite effect 59 [7]. The chemogenetic activation of ACC parvalbumin interneurons, in turn, reduces both 60 premature and incorrect responses, but not response omissions [8]. 61 Studies with physiological measurement of neural activity in rodent ACC during the 5-CSRTT 62 and related tasks, have partly supported the possibility of such a causal role. One st udy 63 revealed that excitatory and inhibitory neurons in rat ACC may change their firing rate 64 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 4 differently both before and after correct vs. incorrect choices in the 5-CSRTT [9]. Specifically, 65 ramping neural activity in the ACC and the adjacent prelimbic cortex (PrL, upper part of the 66 rodent medial prefrontal cortex, mPFC) before cue- presentation has been interpreted as 67 preparatory signal under conditions that require high sustained attention, as this activity 68 increase was smaller before incorrect (low attention) responses and lowest during omissions 69 [9,10]. 70 Several other studies using physiological measurements have, however, failed to find such an 71 indication of a causal role of ACC for modulating the occurrence of response options on a trial-72 by-trial basis in the form of distinct pre -choice activity. They rather suggest that ACC may 73 monitor ongoing behaviour, and potentially provide feedback or error signals . For example, 74 neurons in rat ACC were shown to encode behaviour-related information mostly during and 75 after a choice, in a deterministic lever-based working memory task, thereby monitoring action 76 and outcome [11]. A specific subpopulation of ACC neurons that project to visual cortex was 77 selectively excited after incorrect choices or omissions (i.e. they conduct error-monitoring), but 78 their activity did not differ between those erroneous and correct choices while they were made 79 [7]. Imaging during a head-fixed Go/No-Go paradigm even found no evidence for a selective 80 recruitment of these neurons and projections for enhanced stimulus discrimination, but rather 81 that they simply represent rewarded action and stimuli [12]. Using a Go/No-go paradigm with 82 visual cues in mice, another group confirmed that ACC neurons are generally more likely 83 activated by cues that imply reward than those that do not, but also suggested that these cells 84 fire selectively either to signals that imply action or action restraint [13]. Another study in the 85 5-CSRTT also failed to detect much increase of firing rates of pyramidal neurons in the dorsal 86 PrL/ACC region before cue- onset, but found the modulation of their firing times by gamma-87 oscillations in this period [14]. The role of the ACC may also depend on the task structure, as 88 it was shown that, in a probabilistic task, rat ACC neurons represent expected outcome first, 89 before switching to actual outcome in case of a mismatch between the two [11], which could 90 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 5 constitute a feedback signal for updating prior believes. This is in line with selective activity of 91 some ACC neurons after incorrect choices in the 5-CSRTT, constituting an error signal [7,9]. 92 In summary, no clear picture of the mechanistic role of the rodent ACC in attention and impulse 93 control has emerged yet; whereas some studies found representations several seconds 94 before the choice event, which could indicate a causal role for choices or , generally, the 95 present cognitive state, others have found representations rather around the time of choice 96 itself and most profoundly during rewarded or after incorrect responses, which is more in line 97 with action- and outcome-monitoring, at least in deterministic tasks. Likewise, in monkeys, 98 ACC activity has been linked to error -monitoring, value representation and belief -updating 99 [15,16] rather than to attention per se [12]. Therefore, we here use simultaneous monitoring 100 of dozens of excitatory neurons with miniature endoscopic microscopes (miniscopes) in the 5-101 CSRTT, in mice, in combination with time-resolved encoding and decoding analysis to reveal 102 which aspects of attentional, impulse and motor control are represented in the ACC at which 103 point in time. 104

105 Miniscope-based recording of neocortical activity in the 5-CSRTT 106 To monitor activity of individual pyramidal neurons, we transduced ACC with an AAV5-vector 107 expressing the fluorescent calcium sensor GCaMP6m under the CamKIIα-promoter [17], in 108 male C57BL/6J wildtype mice ( N = 12), and implanted a gradient refractory index lens in a 109 separate surgery at the same location (Figure 1B). For comparison, we also generated a 110 smaller, second subgroup (N = 6), where activity was monitored in the ventral mPFC (Figure 111 1B), a region that was previously shown to represent rewarded choices [18]. Mice had been 112 pre-trained in the 5-CSRTT, and their training was continued after recovery from the second 113 surgery, until they reached a stable baseline. 114 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 6 115 Figure 1. Behavioural performance with simultaneous miniscope recording . ( A) 116 Structure of an individual trial of the 5 -CSRTT (see Methods for description). (B) Selective 117 transfection of the ACC (comprised of regions Cg1 and Cg2; left, AP 1 mm) and the ventral 118 mPFC (at the border between the regions PrL and IL; right, AP 2 mm) with GCaMP6m 119 expressed in excitatory cells; black gap in the right hemispheres indicates GRIN lens location. 120 (D) Measures of task engagement and performance in the 5- CSRTT (as indicated above 121 panels) during training sessions without miniscope or dummy (baseline, BL- untethered) and 122 during the first three sessions performed with tethered miniscope (Day 1- 3). Dots indicate 123 individual animals, bars show mean ± s.e.m.. Asterisks represent Dunnett pairwise post-hoc 124 test comparing tethered days against baseline after significant effect of day in a one -way 125 ANOVA. ( E) Key performance indicators of the 5- CSRTT (as indicated above panels) 126 measuring attention (accuracy, incorrect responses), impulse control (premature responses) 127 and task engagement (omissions) for the baseline protocol and six challenge conditions during 128 which miniscope recordings were conducted. Same display of mean ± s.e.m. and statistics 129 (comparison of each challenge against baseline) as in (D). (F) Example of a 400 µm x 400 µm 130 raw image obtained from ACC of an individual mouse during the 5-CSRTT with a miniscope, 131 with exemplary identified active neurons encircled in different colours corresponding to traces 132 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 7 shown in (H). (G) Similar display as in (F), but overlay of fields of view as maximum projection 133 from 12 animals as imaged in the same 5-CSRTT protocol in ACC (circles indicate the same 134 neurons as in (F). (H) Example of z-scored calcium activity traces over 10 min measured with 135 GCaMP6m in the FOV shown in (F) from exemplary individual active neurons. 136 137 The recording of neural activity with miniscopes during operant tasks constitutes a challenge 138 due to the relatively large and protruding form factor of such microscopes and a 139 disadvantageous design of many operant box systems with deep and low recesses 140 constituting the reward receptacle and poke holes. To enable miniscope recordings during the 141 5-CSRTT, we designed a custom-made operant box system with shallow and elevated poke-142 holes that reside in a protruding inner wall-layer (Figure 1C) [19]. This allowed mice to conduct 143 the task with little disturbance by the mounted and tethered miniscope (UCLA model v3 or v4; 144 Supplementary Video 1), as was confirmed by a lack of changes of achieved trial numbers, 145 response latency, attentional accuracy (number of correct responses/(number of correct and 146 incorrect response), and omissions (number of trials with omitted responses relative to total 147 number of trials, %) beyond the first day of tethered training (Figure 1D). With repeated 148 tethered training, animals performed well over 100 trials with less than 50% omissions on 149 average, providing sufficient numbers of active responses for further analysis (Figure 1D). 150 In order to maximally engage attentional and impulse control – and to obtain suffic ient 151 numbers of incorrect and premature responses per session for later analysis - we performed 152 six behavioural challenges with simultaneous miniscope recordings; this included a further 153 shortening of the stimulus duration (SD) from 2 s at baseline to 0.8 or 1.0 s in challenge 154 conditions, and/or an extension of the waiting time (inter -trial interval, ITI) before stimulus 155 presentation from 5 s at baseline to fixed durations of 7 or 9 s or to variable lengths (7, 9, 11, 156 or 13 s randomly at equal distribution, varITI). As expected , attentional performance, as 157 indicated by accuracy, was lower with decreased stimulus duration (0.8 s SD challenge), 158 which, however, also increased omissions, making it less suitable for analysis (Figure 1E). 159 Overall, the varITI challenge appeared to produce the most suitable dataset for further 160 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 8 physiological analysis, given that the relative number of premature responses was increased 161 (P < 0.0001; Dunnet’s post-hoc test after significant main effect of challenge in mixed-effects 162 ANOVA, N = 18) and the relative number of omissions was decreased (P < 0.001) compared 163 to the baseline protocol, whereas incorrect responses were still present at a level comparable 164 to the other test conditions (Figure 1E). We obtained stable recordings over 30 min sessions, 165 yielding 10-72 cells per field of view (FOV) in the ACC and 14-72 cells/FOV in the mPFC (see 166

for details on trace extraction; Figure 1F-H). 167 Individual ACC neurons have time -locked activity peaks around correct and 168 premature responses 169 We first investigated qualitatively, if identified neurons display activity that is related to any of 170 the four behavioural response options. Therefore, the calcium signal traces of each neuron 171 were extracted from 4 s before until 7 s after each behavioural choice-poke event (note that, 172 for omissions, the end of the stimulus presentation, was used as reference time point of choice 173 for all time-locked analysis). Such individual episodic traces split into two populations of traces 174 from trials with either even or odd order number; the averages of traces from even trials were 175 then plotted in vertical order according to the peak latency of the averages of the 176 corresponding odd trials (Figure 2A -B; Supplementary Figure 1). This indicated that ACC 177 neurons often showed activity patterns that were time-locked to correct and premature 178 responses. In support of this conclusion, Pearson correlations between the averages of odd 179 and even trials indicated a high reproducibility of time-locked activity around correct and 180 premature responses, with particularly high correlations (> 0.8) around and after the time of 181 choice, in ACC and mPFC (Figure 2A-B, bottom; Supplementary Figure 1). Such correlated 182 patterns were largely absent for incorrect responses and omissions. This constitutes a first 183 indication that neural activity in both structures represents aspects of choices related to high 184 attention (correct responses) and impulsivity (premature responses). 185 Notably, correct and incorrect responses (in contrast to the other two event types) involve a 186 similar global sensory stimulation (one poke -hole illuminated at the time of responding) and 187 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 9 the same motor output (poking), suggesting that the time-locked activity seen only for correct 188 responses, does likely not reflect sensory or motor aspects. However, we cannot fully exclude 189 the possibility that deviations in the local stimulus - illuminated vs. non-illuminated hole into 190 which the mouse pokes - may partially account for such differences. 191 192 Figure 2. Event-locked activity of individual neurons in the variable ITI challenge. (A) 193 Top: Average z-scored calcium activity in individual ACC neurons time-locked to the onset of 194 the behavioural event stated above each sub-panel, shown for -4 - +7 s around the event. For 195 cross-validation, averaging was done across the even trials only and the cells sorted according 196 to the average peak latency across the odd trials (see also Supplementary Figure 1). Bottom: 197 Pearson correlations between the averaged z-scored activity of the odd and even trials at each 198 time point. No te that temporal order of peaks is maintained for correct and premature 199 responses with resulting high correlations, but not for incorrect choices and omissions. N = 12 200 animals and 443 cells. (B) Same display and analysis as in (A) but for all neurons recorded in 201 mPFC. N = 6 animals and 229 cells. (C) Same data as in (A) but clustered into four clusters 202 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 10 (I-IV, indicated on the right) according to their activity from -4 - 7 s around correct responses. 203 Equivalent to (A), clustering was done based on the average of odd trials, only, whereas the 204 plot shows the average of the corresponding even trials. Activity around incorrect, premature 205 and omitted responses for the same cells is shown in the same order as was determined 206 according to the clustering around correct responses resulting in a lack of emerging temporal 207 patterns. (D) Same analysis as in (C) but for cells in mPFC. Grey lines in (A-D) indicate cells 208 for which a response cannot be shown because the mouse made no incorrect response. ( E) 209 Z-scored single- trial calcium activity of exemplary individual cells from each cluster (I -IV, 210 indicated on the right). The timepoint of cue presentation (for correct and incorrect responses 211 only), reward receptacle entry and exit (only correct responses) are shown by the white short 212 vertical lines. The consistent white line represents the timepoint of the choice. (F) Same as in 213 (E), but for cells from the mPFC. 214 215 To further investigate the temporal relationship of neural activity, we k- means-clustered the 216 cells [20] into four clusters according to their average activity in odd correct response trials, 217 sorted neurons within each cluster according to time of peak -activity, and then displayed the 218 corresponding average of even trials (Figure 2C, D). Qualitatively, this resulted in three 219 clusters with relatively clear activity peaks either before, during, or after correct responses , 220 respectively, in addition to a fourth cluster with increased activity during reward collection only, 221 in ACC (Figure 2C). In contrast, in mPFC a cluster with a well-defined peak at the time of the 222 correct choices was lacking, and the emerging clusters showed activity either before or after 223 the response (Figure 2D). To assess the response-specificity of these temporal patterns, we 224 conducted the same temporal alignment and averaging for the other three response options 225 but sorted the cells according to their order number obtained for clustering by correct 226 responses. For all three response types, this resulted in the loss of clear temporal response 227 patterns, indicating that the temporal relationship cells displayed for correct responses, were 228 largely specific for this one response type (Figure 2C-D). Finally, when plotting the activity in 229 individual trials of one randomly selected neuron for each cluster, the trial-to-trial reliability of 230 activity as time-locked to correct responses was qualitatively confirmed (Figure 2E-F). 231 Distinct choices are represented by ACC population activity 232 While the analyses described above confirm that individual ACC and mPFC neurons are 233 modulated by ongoing attention- and impulsivity-related choices or actions, a comprehensive 234 and multi-variate encoding of behaviour is expected only at the level of populations of multiple 235 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 11 neurons. To evaluate such behavioural representations in ACC , we performed a decoding 236 analysis, training linear support vector machine ( SVM) classifiers to predict the type of 237 behavioural event performed in any given trial based on the population activity , at different 238 time points around the choice event, and during the first 4 s of the preceding ITI (starting with 239 the end of the time -out or of reward consumption). We first focused on binary discrimination 240 between correct and either omitted or premature responses, given that incorrect responses 241 occurred in insufficient numbers for this analysis, in the varITI-challenge in most mice (Figure 242 1E; Supplementary Table 1) . To estimate significant predictions, we compared the resulting 243 accuracies with accuracies obtained from classifiers that were trained on data with randomly 244 shuffled labels ( paired t-tests at each time point with Benjamini-Hochberg correction for 245 multiple comparisons). 246 No appreciable prediction of correct responses (vs. omissions or vs. premature responses) 247 was possible during the ITI, indicating that ACC activity did not reflect, if a mouse was going 248 to act in a goal-directed, attentive fashion or to avoid task engagement or to act impulsively in 249 an upcoming trial (Figure 3A, left). Although, there was a significant decoding of correct vs. 250 omitted or premature responses at low accuracy of around 60% (vs. 50% chance level), 251 already from at least 4 s before the choice poke onwards, average decoding accuracies only 252 started rising around cue- onset and reached their maximum of >90% only approx. 600 ms 253 after the choice-poke. They remained at >90% throughout the time of reward consumption 254 (Figure 3A, right). This indicates that the pre- cue and pre-choice representations were very 255 minor compared to the same representation around and after the choice. These results appear 256 inconsistent with the notion that the primary driver of variance in ACC activity are slowly 257 varying cognitive states of attention or impulsivity, but rather that ACC representations seems 258 to be tightly tied to actions and outcomes. In mPFC, decoding accuracies had a similar 259 temporal trajectory (Figure 3B). 260 261 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 12 262 Figure 3. Decoding of behavioural choice from population activity in ACC and mPFC. 263 (A-B) Cross -validated decoding accuracies derived from binary classification using linear 264 SVMs calculated from z -scored amplitude values at each 200 ms time bin and predicting 265 behavioural events from population activity in ACC (A; N = 11 mice; one mouse not analyzable 266 due to low omission rate) or mPFC ( B; N = 6 mice) in individual varITI sessions; solid lines 267 and shading represent averages across animals ± s.e.m., respectively. Decoding accuracies 268 were first averaged across 100 classifiers calculated on data from each session, and then 269 across sessions (i.e. animals). Dashed lines indicate results from the same analysis but 270 performed on control data obtained by random shuffling of event -labels relative to neural 271 activity data; dots at the bottom indicate a significant difference in the pairwise comparisons 272 between those two accuracy values at each time point ( t-test with Benjamini -Hochberg 273 correction for multiple comparisons). Binary classification was done differentiating correct 274 responses against omissions (black) or against premature responses (magenta). Chance level 275 is 50%. (C) Same analysis as in (A) but classifying correct vs. incorrect responses by using 276 sessions from across all challenge protocols, if more than 5 incorrect responses were made 277 (N = 6) . mPFC was not analysed because only 3 sessions had the sufficient number of 278 incorrect responses. See Supplementary Table 1. 279 280 In the pairwise discriminations described above, a confound by non- choice-related aspects 281 such as presence of the cue ( correct vs. premature) or the motor response ( correct vs. 282 omission) cannot be ruled out. Only correct and incorrect responses are sufficiently similar in 283 most parameters and differ mainly in the choice per se. To enable a cross-validated decoding 284 analysis involving incorrect responses, we gathered sessions from all six behavioural 285 challenge conditions given they had a sufficient number of incorrect responses (≥ 6). Using 286 such data, we found that – in contrast to premature and omitted responses - incorrect choices 287 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 13 could be distinguished significantly from correct ones only from the time of the choice- poke 288 onwards (approximately 1 s after cue-onset), indicating a lack of a consistent signature of both 289 preparatory attention and of sensory stimuli in the overall ACC activity (Figure 3C ). As seen 290 with the other decoding attempts, high and saturating average prediction accuracies (>80%) 291 could only be reached from around 600 ms after the choice- poke and were maintained 292 throughout reward consumption or its omission (Figure 3C). Overall, the temporal distribution 293 of decoding accuracies is more aligned with the notion that ACC represents ongoing action 294 and its (expected or actual) consequence rather than controlling levels of sustained attention 295 or impulsivity on a trial-by-trial basis. 296 ACC neuron populations encode spatial aspects of ongoing action 297 The observation that decoding of response types available in the 5-CSRTT was only possible 298 with high accuracy from the choice -poke onwards , suggests that ACC and mPFC may 299 represent selected actions rather than high-level cognitive states like attention and impulsivity. 300 To further scrutinize this hypothesis, we investigated whether these neurons encode a more 301 fine-grained representation of current action by analy sing the responses to each individual 302 choice poke-hole. We aligned the average activity of each neuron in even trials to the time of 303 correct choice for each hole individually and sorted the neurons first according to the hole (1-304 5) which evoked the strongest response during the poke ( ±1 s) into five groups, and then 305 sorted by peak latency of the average of odd trials within each group. A reproducible pattern 306 emerged for even trials that correlated strongly to that of odd trials from around the time of 307 choice onwards, in ACC and mPFC (Figure 4A-B , bottom), and which appeared the more 308 dissimilar between poke-holes around the time of poking, the further the holes were apart from 309 each other (Figure 4A-B, top). 310 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 14 311 Figure 4. Activity in the ACC represents spatial action selection. (A-B) Same data as in 312 Figure 2A-B (correct responses), but separated by poke-hole and arranged into five groups 313 (indicated on the right, separated by black lines) based on the hole for which the response of 314 a neuron had the maximum AUC in the period ± 1s around the poke (indicated by a black 315 rectangle around the cluster). Top: Average z-scored calcium activity in individual ACC (A) or 316 mPFC (B) neurons time-locked to the correct choice-poke, shown for -4 - +7 s around the 317 event. For cross-validation, averaging was done across the even trials only and the cells were 318 sorted according to the average peak latency across the odd trials of poke 1 . Grey lines 319 indicate sessions in which the given hole was not poked into. Bottom: Pearson correlations 320 between the averaged z-scored activity of the odd and even trials at each time point. Note that 321 the qualitative similarity to the pattern of a given poke gets reduced the further away the poke-322 hole is, especially around the time of poking. Supplementary Figure 2A -B shows the same 323 data without prior sorting into clusters. ( C) Based on the data shown in (A -B), Pearson 324 correlations between response patterns of pairs of poke holes, coded in colour according to 325 the distance between the holes ; correlation values for hole- combinations with the same 326 distance (e.g. 1-4 and 2-5 for distance 3) were averaged; see Supplementary Figure 2C for 327 the individual correlation values of each hole -pair. A repeated -measures ANOVA was 328 calculated across the population of 11 observations in the time period ±1 s around the poke 329 (grey bar; P -value for main effect of hole -distance indicated at the top); results from paired 330 Sidak post-hoc tests are indicated for adjacent hole distances in the colour -legend of each 331 sub-panel. *** P 0.5. (D) Accuracies of the decoding of the identity of the 332 poke-hole, either for all active responses (correct, incorrect, premature; black) or for correct 333 responses only (blue) aligned to the time of poking (0, vertical line); average latency to c ue 334 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 15 onset (response latency), reward-poke entry (reward latency) and start of the next ITI (reward 335 consumption time, indicted only for correct trials) are indicated by arrowheads. Decoding 336 accuracies were first averaged across 100 classifiers calculated on data from each session, 337 and then across sessions (i.e. animals). Dashed lines indicate results from the same analysis 338 but performed on control data obtained by random shuffling of event -labels relative to neural 339 activity data; dots at the top indicate significant pairwise comparisons between those two 340 accuracy values (t-test after correcting for multiple comparisons across time bins). Shaded 341 area, s.e.m. across mice. 342 343 To assess this quantitatively, we calculated Pearson correlations between such population 344 activity patterns for all pairs of poke-holes and averaged them across hole-pairs with the same 345 distance (Figure 4C; Supplementary Figure 2). Indeed, within approximately ±1 s around the 346 choice-poke, correlations were the higher the closer the holes of the pair were to each other. 347 After the choice, in contrast, correlations were consistently high, irrespective of distance. This 348 suggested that ACC activity displays a certain similarity related to spatial proximity of poke 349 holes before the poke, while being dominated by non- spatial aspects of the choice after it is 350 made (Figure 4A-C). 351 To further investigate this early spatial selectivity, we trained multi -class SVM classifiers to 352 decode the identity of the poke-hole. Surprisingly, this identity could be predicted from about 353 1 s before the poke (just after cue- onset) onwards, reaching an average peak accuracy of 354 close to 70% in ACC and close to 55% in mPFC (against a chance level of 20%) approximately 355 200 ms after the poke. In contrast to the representation of event-type (Figure 3A-C), average 356 accuracy decreased again immediately after the poke, suggesting that the representation of 357 the precise action fades after its occurrence (Figure 4D). Interestingly, the decoding accuracy 358 was virtually identical, when performing the same analysis on pokes of all three active 359 response types (correct, incorrect, premature) combined, suggesting the existence of a 360 representation of action that is independent from the representation of response- type and 361 outcome. 362 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 16 Independent encoding of action and outcome in ACC 363 The hypothesis of an independent encoding of motor action (poke hole identity) and choice or 364 outcome (event-type) entails the question to what extent each factor determines neural activity 365 in the ACC, at every time point. To answer this question, we trained linear regression models 366 predicting the activity of each neuron using predictors coding the active poking (vs. omissions), 367 rewarded choice (correct responses vs. absence of reward due to erroneous choices), and 368 poke-hole identity (spatial location; encoded by four predictors, see predictor matrix in Figure 369 5A). We ran a separate regression for each time -point in the aligned activity and calculated 370 the cross-validated coefficient of partial determination (CPD) for each predictor at each time-371 point, which quantifies the share of the variation in the neural activity that is uniquely explained 372 by that predictor. To estimate statistical significance for a given predictor at a given timepoint, 373 the distribution of CPD values across subjects for that predictor and timepoint was compared 374 to 0% [21]. 375 From the average onset of cue presentation (1 s before the poke), matching the time course 376 of the decoding analysis (Figure 4D), the spatial identity of the poke-hole started to gain ever 377 more influence over ACC- activity, and dominated it compared to the other predictors from 378 approximately 600 ms before until 400 ms after the choice poke (Figure 5B). This effect was 379 mostly driven by an encoding of the left (hole 1-2) vs. the right (hole 4-5) side of the 5-choice 380 wall, although most other tested predictors of the selected action (poke discrimination left, 381 right, and middle) and of active responding in general (vs. omission) displayed significant 382 CPDs during and after the time of poking, as well (Supplementary Fig. 3). From 600 ms after 383 the poke onwards, however, the factor of rewarded (correct vs. erroneous) response had the 384 single strongest influence on neural activity out of the tested predictors (Figure 5B). This 385 overall pattern suggests, that AAC simultaneously encodes fine-grained selected action and 386 a high-level representation of both active and correct responding from the time of choice-poke 387 onwards for almost 2 s, but action representation dominates around the time of execution 388 whereas high- level representation dominates subsequently (Figure 5B). Given that the 389 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 17 influence of outcome starts rising immediately after the choice poke - and hence more than a 390 second before actual reward consumption - this activity reflects expected (vs. omitted) reward 391 at this early stage, but it might later on reflect actual outcome, as previously shown [11]. mPFC 392 neurons, in contrast, also encoded outcome (from 200 ms after the poke) but lacked a 393 consistent encoding of motor action, especially before the poke (Figure 5C). 394 395 Figure 5. Encoding of poking action and reward in the population activity of the ACC 396 and mPFC. (A) Orthogonal predictor matrix designed to indicate the representation of the 397 poke (1: poke in either hole, 0: omission), the poke directionality (1: poke in left holes 1 or 2; -398 1: poke in right holes 4 or 5; 0: poke in middle hole 3 or omission), right poke discrimination 399 (1: poke in hole 4; -1: poke in hole 5; 0: poke in holes 1,2 and 3 and omission), left poke 400 discrimination (1: poke in hole 1; -1: poke in hole 2; 0: poke in holes 3,4 and 5 and omission), 401 middle poke discrimination (1: poke in hole 3; - 0.25: poke in holes 1,2,4,5; 0: omission) and 402 reward (1: rewarded; 0: not rewarded). The four predictors representing the spatial location of 403 the poke hole have been removed at once from the regression model in order to quantify the 404 encoding of motor action. (B-C) Coefficient of partial determination (CPD) averaged across 405 cells recorded in ACC (B, N = 11 mice) and mPFC (C; N = 6 mice). Time bins where CPDs for 406 a given event were significantly higher than zero after cross -validated linear regression are 407 indicated with a dot at the top of each panel, colour -coded for the respective predictor (one 408 sample t-test with Benjamini-Hochberg post-hoc correction). CPDs were determined for each 409 event-type by subtracting the sum squared errors of the full linear regression model 410 (incorporating every event type as predictor) from the sum squared error of the reduced 411 regression model where one predictor (corresponding to the event type for which the 412 population activity should be explained) was removed. 413 414 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 18 Devaluation and omission of reward alter representation of choices 415 The prominent representation of correct – and, hence, rewarded – responses in ACC , 416 particularly during and after the response , which is also in line with earlier studies [9,10], 417 suggests that these activities may represent the expectation of reward rather than elevated, 418 preparatory attention. This is also supported by the fact that incorrect and correct responses 419 cannot be discriminated from population activity before the choice-poke is made (Figure 3C). 420 To assess this conclusion further, we conducted four further 5-CSRTT experiments where the 421 value or expectancy of the reward was altered, using the combined 0.8s/7s ITI challenge (to 422 obtain more incorrect responses, Figure 1E), at the end of the sequence of tests: first we 423 recorded a baseline session with normal food- deprivation and reward delivery. In a second 424 session, the reward was devalued by pre-feeding (by providing 6 g of food overnight and 2 ml 425 of milk reward 1 h before session start). In the third and fourth recording sessions the food-426 deprivation (i.e. value of reward) was normal, but the delivery of reward was omitted (extinction 427 1 and 2). Normal training sessions in the baseline protocol were conducted after the first tw o 428 sessions, but not between the extinction sessions. 429 At the behavioural level, both devaluation and extinction caused a significant decrease of the 430 number of correct responses, driven by an increase in omissions, whereas accuracy remained 431 relatively stable (P < 0.05, Dunnett’s post-hoc test, performed after significant main effect of 432 condition in RM -ANOVA; N = 15; Figure 6A-C ). During devaluation, also active erroneous 433 (incorrect and premature) responses decreased, and reward latency increased, in line with a 434 reduced motivation (Figure 6D; Supplementary Figure 4); such effects appeared qualitatively 435 also during extinction sessions, but mostly without reaching significance. Response latencies 436 were not significantly altered indicating unperturbed responsiveness in any of the conditions 437 (Supplementary Figure 4). 438 439 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 19 440 Figure 6. Decoding of behavioural choice from population activity in ACC during reward 441 devaluation and extinction experiments. ( A-D) Measures of task engagement and 442 performance in the 5- CSRTT (as indicated on y -axes) during training sessions in the 0.8s -443 SD/7s-ITI combined challenge at normal reward conditions (baseline, BL), after devaluation 444 of reward (Deval), or with omission of reward (extinction; Extinct 1/2), as indicated on x -axes 445 with tethered miniscope. Dots indicate individual animals coded by colour, bars show mean ± 446 s.e.m.. Asterisks represent Dunnett pairwise post-hoc test comparing BL condition against the 447 other conditions after RM-ANOVA. * P < 0.05; ** P < 0.01; *** P < 0.001. See Supplementary 448 Figure 4 for further behavioural measures. (E-F) Cross-validated decoding accuracies derived 449 from binary classification using linear SVMs calculated at each 200 ms time bin and predicting 450 correct vs omitted (E) or vs. premature (F) responses from population activity in ACC during 451 test sessions in the 0.8s -SD/7s-ITI combined challenge with normal reward conditions 452 (baseline, black) , after devaluation of reward (blue) or with omission of reward on two 453 consecutive test sessions (e xtinction 1, red; extinction 2, orange). Whereas 10 mice 454 participated in these test sessions, actual N-numbers for the analyses (stated in figure legends 455 on the right and in Supplementary Table 1) vary mainly due to mice that did not perform 456 sufficient numbers of correct or premature responses, in rare cases also due to technical 457 failures. Solid lines and shading represent averages across animals ± s.e.m., respectively. 458 Decoding accuracies were first averaged across 100 classifiers calculated on data from each 459 session, and then across sessions (i.e. animals). Dashed lines indicate results from the same 460 analysis but performed on control data obtained by random shuffling of event -labels relative 461 to neural activity data. Dots below the traces of each panel indicate time bin and alignment 462 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 20 (coded by colour) where decoding accuracies from these two classifiers differed significantly 463 (paired t-test with Benjamini-Hochberg correction; coded by colour for each reward condition). 464 Dots at the bottom represent Dunn -Sidak post-hoc tests comparing accuracy values from 465 devaluation or extinction sessions (coded by colour) to those from the baseline condition 466 (conducted after significant effect of reward condition in the repea ted-measures ANOVA). 467 Significance level is encoded by dot size. Triangles indicate average latency to cue-468 presentation, reward receptacle entry and the start of the next ITI (the latter was often close 469 after reward receptacle entry in extinction conditions as no reward was provided). Chance 470 level is 50%. Decoding of incorrect responses could not be performed due to their small 471 numbers in most of the respective sessions. 472 473 To evaluate if the change of the value or contingency of the reward affected the representation 474 of rewarded responses, we repeated the time -resolved binary decoding of correct vs. 475 premature or omitted responses, as conducted previously for the varITI-challenge (Figure 3), 476 in each of the four conditions. Whereas, under normal reward conditions, the trajectories 477 looked like those found before, with increasing representation of the choice and its outcome 478 with the cue-onset, somewhat altered decoding accuracies were found in the other conditions 479 (Figure 6E-F). Generally, decoding accuracy was lower in the conditions with altered reward 480 value or occurrence, as indicated by comparisons to the accuracy achieved by classifiers 481 trained with shuffled control datasets . Significant decoding of correct responses was hardly 482 possible before they actually occurred, possibly reflecting a certain lack of representation of 483 preparatory attentional or impulse control required for such responses (Figure 6E -F). When 484 comparing the accuracy achieved under baseline condition with that achieved under reward 485 devaluation (RM-ANOVA with pairwise Dunne tt post-hoc tests), these two conditions rarely 486 differed, suggesting that the value of the reward has relatively little influence on the 487 discriminability of representations of rewarded vs. non -rewarded choices (Figure 6E-F). In 488 contrast, during both extinction sessions, the discrimination of correct responses vs. omissions 489 and, to a lesser extent, vs. premature responses, deteriorated from about 2 s after the choice 490 poke onwards and differed significantly from the accuracy achieved in the baseline condition, 491 in line with the lack of a reward (Figure 6E-F). This implies that beyond this time point, ACC 492 largely encodes the actual outcome, which is no longer distinct between the choice options in 493 the extinction sessions. Similarly, in the second extinction session, accuracy for decoding 494 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 21 correct responses vs. omissions differed already from 600 ms before the choice poke onwards 495 and on most subsequent time points (Figure 6E). This supports the idea that ACC partly 496 represents expected outcome, which has been learned to be identical between these two 497 choice options by the time of the second extinction session. 498 499 Figure 7. Encoding of reward and action in ACC depends on the relative value and 500 presence of reward. (A) Coefficient of partial determination (CPD) averaged across cells 501 recorded in ACC in the four different reward-related conditions encoded by colour (number of 502 animals): black, baseline (10); blue, devaluation (6); red, extinction session 1 (6); green, 503 extinction session 2 (6). As in Figure 5, CPDs were determined for each predictor, stated on 504 the left (rows) by subtracting the sum squared errors of the full linear regression model 505 (incorporating every event type as predictor) from the sum squared error of the reduced 506 regression model where one predictor was removed. To evaluate the encoding of spatial 507 location as such, all four predictors reflecting spatial location were combined by removing 508 them at once (see Figure 5A) . Time bins where CPDs for a given event were significantly 509 higher than zero after cross -validated linear regression are indicated with a dot at the top of 510 each panel, colour -coded for the respective reward condition (one sample t-test with 511 Benjamini-Hochberg post-hoc correction). Dots below represent Sidak tests comparing CPD 512 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 22 values from devaluation or extinction sessions (coded by colour) to those from the baseline 513 condition (conducted after significant effect of reward condition or time-reward interaction in 514 RM-ANOVA). Significance level is encoded by dot size. Triangles indicate average latency to 515 cue-presentation, reward receptacle entry and the start of the next ITI (the latter was often 516 close after reward receptacle entry in extinction conditions as no reward was provided). To 517 ensure, that CPDs in the baseline session are not simply higher because of a larger number 518 of trials or responses used for their calculation, the number of baseline- trials was down -519 sampled to roughly match those of the reward condition it is compared to in each panel f or 520 each poke-hole and response type. 521 522 To further elucidate the hypothesis that ACC encodes expected and, subsequently, actual 523 outcome, we performed an encoding analysis, as previously done for the varITI -challenge 524 (Figure 4), for the data obtained from these four conditions. Across time intervals, we 525 compared CPD values to the control value of 0% for all conditions (t -test with Benjamini -526 Hochberg correction) and we compared the devaluation and extinction conditions to baseline 527 (RM-ANOVA followed by Sidak post-hoc test; Figure 7A). As expected, extinction – especially 528 when repeated – led to a reduction of the representation of reward. However, also devaluation 529 ensued a strong decrease of reward representation compared to baseline, suggesting that 530 reward determines ACC activity the stronger the higher its v alue is (a result that the prior 531 decoding analysis could not reveal, see Figure 6E -F). Strikingly, both devaluation and 532 extinction also led to a virtual loss of the spatial representation of the poke- hole: in all three 533 conditions, CPD values for the combined spatial predictors were significantly lower than those 534 in the baseline condition around the time of poking and, in contrast to the baseline, rarely 535 exceeded 0% (Figure 7A). This suggests that the representation of specific actions in ACC 536 depend on their expected (reward) value. 537 538

539 Using miniscope recordings during the 5- CSRTT in mice , we here demonstrated that 540 excitatory neurons of the posterior ACC represent the choices available in this complex task, 541 both at the high level of response options that depend on cognitive state and at the low level 542 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 23 of the selected action (poke-hole identity) . We found that coding at both levels happens 543 simultaneously, but that ACC activity is dominated by the selected action (graded spatial 544 representation of poke-hole location) during the execution of the choice poke, and by outcome 545 expectation from about 400 ms after the choice, followed by encoding of actual outcome, once 546 obtained. Whereas action representation faded , the high- level representation of the chosen 547 option (or its consequence) remained stable for at least 7 s after the choice, but deteriorated 548 much faster when reward was omitted during extinction sessions. When value or occurrence 549 of reward were changed, the share of cells selectively activated by reward and of cells 550 selectively inhibited during omissions – i.e., the share of cells representing those options 551 whose relative utility changed – increased. This suggests that ACC represents action-outcome 552 contingency rather than outcome itself. Likewise, the representation of rewarded (correct) 553 responses, but also of low-level spatial action parameters decreased when reward was either 554 devalued or omitted. This underlines that motor and reward representations in ACC depend 555 on the expectation and value of a choice’s outcome. 556 Importantly, significant preparatory network activity during the ITI and before the cue which 557 could indicate a general level of high or low attention, impulsivity, or task engagement in a 558 given trial was either not found (attention-related activity discriminating correct from incorrect 559 choices) or was rather weak (discrimination between impulsive responses or omissions and 560 correct responses). In fact, all conducted analyses assessing response-type coding suggest 561 that choices are represented in ACC mainly during and after their time of occurrence, and that 562 rewarded responses are represented most reliably, as seen in other studies [12,13]. 563 How can these observations be reconciled with earlier reports of signatures of attention and 564 impulsivity in ACC [9–11,22] and, most importantly, with the well- documented ability to 565 modulate these cognitive states by manipulation of ACC neurons [5,6,8,23]? Firstly, with 566 respect to pre-choice activity, the discrimination of correct responses from either premature or 567 omitted responses was actually possible, albeit at low (~60%) accuracy , already during at 568 least 4 s before the choice poke (but not at the beginning of the ITI; Figure 3, 6). Furthermore, 569 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 24 correct and premature responses were accompanied by time-locked activity of neurons in the 570 same pre-choice time period (Figure 2). This indicates a representation of the state of impulse 571 control that exists, albeit weakly, well before the choice, though not at the beginning of the ITI 572 – in line with ramping activity before premature responses observed in rats [10]. In contrast, 573 incorrect responses could only be distinguished from correct responses at time of occurrence 574 (Figure 3) and were not accompanied by time-locked activity before (Figure 2). This suggests 575 that the trial-to-trial variation of sustained attention is not represented in excitatory cells of 576 posterior ACC. The divergence between both phenomena is in line with the fact, that 577 impulsivity, but not accuracy, was modulated by chemogenetic manipulation of such neurons 578 in ITI-challenges [5]. Since modulation of ACC parvalbumin-interneurons (which we did not 579 record in this study) could alter both aspects of executive function [8], they could constitute a 580 locus of the attentional component in this task [14]. Thirdly, chemogenetic or pharmacological 581 modulation of ACC could exert its effect anatomically and temporally more globally, rather 582 than controlling behaviour directly and in individual trials. The ACC might be acting through 583 the demonstrated strong representation of obtained reward, errors, and action- outcome 584 contingencies, to shape tendencies of attentional and impulse control in upcoming trials 585 through other circuits, as shown for a subpopulation of ACC neurons already [6,7]. Fourthly, 586 and alternatively to the previous scenario, ACC modulation may influence the occurrence of 587 distinct response-types because of its role in action selection [21,24], as we suggested based 588 on our [5,8] and other [6] previous chemogenetic and optogenetic [7] data before [8]. In this 589 scenario, elevated activity of certain ACC pyramidal neurons triggers the response into a 590 certain poke-hole (in line with the early encoding of action found in this study). This could be 591 caused by the ACC due to reward expectation leading to strong excitatory AAC activity 592 entailing a correct response. But it could also be caused erroneously by higher order inputs 593 received by the ACC without the appropriate cue leading to incorrect or premature responses, 594 accompanied by somewhat less effective, less time-locked ACC excitation. Activation of PV-595 interneurons [8] or direct partial inhibition of excitatory neurons in ACC [5] could increase the 596 threshold needed for such activation that ultimately triggers the selected response, so that the 597 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 25 strong activation triggering correct responses remains supra-threshold, whereas the weaker 598 activity which could otherwise cause erroneous responses falls sub -threshold. Further 599 physiological investigations are required to probe these mechanistic models. 600 Regarding mPFC, the number of mice we recorded in our study for comparative purposes is 601 too small to draw general conclusions regarding the coding in this region, which has also been 602 examined in other studies in the 5 -CSRTT [10,14,26,27]. Nevertheless, t he relatively minor 603 differences between representations in ACC and mPFC we observed are consistent with an 604 earlier finding of oscillatory coupling of both regions during choice events in the 5-CSRTT [22] 605 and, more generally, with a model of widely distributed encoding of behaviour across 606 neocortex [16]. Population activity in mPFC also represented choice options but less reliabl y 607 and partly shifted to later time points, i.e., after the choice, and again with a bias to encoding 608 rewarded choices more than non- rewarded ones. Also, a fine- grained encoding of spatial 609 aspects of a selected action was virtually absent. This is in line with previous 610 electrophysiological measurements in rats during this task, indicating that a majority of 611 responsive mPFC cells respond after the choice, representing trial outcome, and cells show 612 considerably stronger firing rate increases during correct than during premature responses 613 [10]. 614 In conclusion, our temporally resolved analysis suggests that ACC excitatory neurons 615 represent a chosen action as it is made as well as its expected and actual outcome. Our results 616 uncover parallel encoding of fine- grained spatial parameters of selected actions - in 617 dependence on their outcome value - and of action-outcome contingency in ACC, and suggest 618 that trial-by-trial encoding of high-level cognitive states before the choice is either minimal (for 619 task engagement and impulsivity) or absent (for attention). 620 621 622 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 26

623 Animals 624 In total, 46 male C57BL/6J wildtype mice, were used for this study. Animals were group - or 625 single-housed in Type II -Long individually ventilated cages (Greenline, Tecniplast, G), 626 enriched with sawdust, sizzle-nestTM, and cardboard houses (Datesand, UK), and maintained 627 at a 13 h light / 11 h dark cycle. Water was available ad libitum . All experiments were 628 performed in accordance to the German Animal Rights Law (Tierschutzgesetz) 2013 and were 629 approved by the Federal Ethical Review Committee (Regierungsprädsidium Tübingen) of 630 Baden-Württemberg, Germany (licence number TV1469). 631 5-CSRTT training and testing with calcium imaging 632 Mice started training in the 5-CSRTT at 3-5 mo of age and were kept under food-restriction at 633 85-95% of their prior average free- feeding weight which was measured over 3 days 634 immediately prior to the start of food restriction at the start of the behavioural training. Testing 635 was conducted in operant chambers placed individually in melamine- MDF sound-insulated 636 and ventilated outer boxes and fitted internally with an array of five nose -poke holes on one 637 wall and a reward receptacle on the opposite wall. All six apertures could be illuminated to 638 instruct the entry into them and were fitted with IR break -beams to detect entry and exist of 639 the animal’s snout. All experiments were conducted in custom -made trapezoidal chambers 640 based on the pyControl system [19,28] (https://pycontrol.readthedocs.io). 641 The 5CSRTT training protocol was similar to what we previously described [5]. In brief, after 642 initiation of food-restriction, mice were accustomed to consume the reward (strawberry milk, 643 MüllermilchTM, G) first in their home cage, and then in the operant box (2-3 exposures each). 644 Subsequently, mice were trained in 2-13 sessions (30 min, once daily) of habituation training. 645 In each trial, all holes of the 5-poke wall were illuminated for an unlimited time and the mouse 646 could poke into any one of them to earn a 40 µl milk reward subsequently disposed from the 647 illuminated receptacle. If mice attained at least 30 rewards each in two consecutive sessions 648 or (in exceptional cases) had reached the 16th session of habituation training, they were moved 649 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 27 to the 5- CSRTT training, during which mice transitioned through five stages of increasing 650 difficulty, based on reaching certain performance criteria in each stage, as described 651 previously[5]. The difficulty was determined by the length of time the stimulus was presented 652 (stimulus duration, SD) and the length of waiting time between the end of the previous trial 653 and the stimulus presentation of the next trial (inter -trial-interval, ITI). In case a reward was 654 collected on the previous trial, the ITI was initiated by the removal of the snout of the animal 655 from the reward receptacle. In all 5- CSRTT protocols only one pseudo- randomly selected 656 aperture of the 5-choice wall was lit up after the ITI, indicating that this hole needs to be poked 657 into ( correct response ) in order to earn a 20 µl milk reward (Figure 1A). Trials were not 658 rewarded but instead terminated immediately with a 5 s time-out period during which the 659 house light was turned off, if the animals either poked into any hole during the ITI (premature 660 response), poked into a non-illuminated hole (incorrect response) during the SD and limited-661 hold time (LH, until 2 s after SD), or failed to poke throughout the trial (omission). The relative 662 numbers of such response types were used as performance indicators measuring premature 663 responding [%premature = 100*(number of premature responses)/(number of trials)], 664 sustained attention [accuracy = 100*(number of correct responses)/(number of correct and 665 incorrect responses combined)], and lack of participation [%omissions = 100*(number of 666 omissions)/(number of trials)]. A trial was considered to start at the beginning of the ITI, i.e. 667 included premature responses. Additionally, the time required to poke into the indicated hole 668 after it was illuminated (response latency) and the time from the exit from the correct hole until 669 the entry into the reward receptacle (reward latency) were measured, whereby the latter is 670 usually used as a compound indicator of motivation and locomotor drive [3]. In all stages, 671 sessions lasted 30 min and were performed once daily at the same time of day and in the 672 same box for each animal. 673 After surgery (see below), animals were trained until they had reached the final baseline stage 674 (BL; 2 s SD, 5 s ITI) obtaining an accuracy >80% and an omission rate <50% in two 675 consecutive sessions. For the last 5 d before the first imaging session, mice were accustomed 676 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 28 to be gently fixed in the experimenter’s hand for the fixation of the miniscope before the 677 session and were trained with dummy ‘miniscopes’ that were equal in height and weight, 678 attached to the baseplate. Training with dummy miniscopes lasted until an accuracy >80% 679 and an omission rate <50% in two consecutive sessions was reached again. On subsequent 680 days, mice were trained for 3 d with an actual, tethered miniscope in distinct operant chambers 681 set up for simultaneous imaging, followed by the first imaging session conducted also in the 682 baseline stage. Subsequently, the imaging sessions (30 min) were repeated with different 683 challenge conditions in the same order for every mouse (see Figure 1E). Some imaging 684 sessions needed to be repeated due to technical failures. In between imaging sessions, mice 685 were trained in the same testing chambers in the baseline stage with the miniscope attached. 686 After imaging sessions with the challenge protocols were completed, some mice underwent 687 a separate set of pharmacological experiments in the 5 -CSRTT (data not shown in this 688 manuscript), after which training in the baseline protocol and then further imaging sessions in 689 the combined 0.8s -SD/7s-ITI challenge protocol with concomitant manipulation of value or 690 occurrence of reward followed: Firstly, imaging was conducted under normal conditions of 691 food-restriction and reward- delivery (baseline), secondly imaging was conducted after 692 devaluation of reward by pre-feeding (providing 6 g of food overnight and 2 ml of milk reward 693 1 h before session start), thirdly, two sessions followed under conditions of normal food-694 restriction but omission of reward (extinction). Training in the baseline protocol without imaging 695 was conducted before and after the devaluation session, but not between the extinction 696 sessions. 697 Surgical procedures 698 After the mice reached at least stage 4 (4 s SD, 5 s ITI) of the 5- CSRTT, a nimals were 699 anaesthetized using isoflurane (AbbVie, G), received subcutaneous injections of analgesics 700 (0.08 mg/kg buprenorphine, Bayer, G; 1 mg/kg meloxicam, Boehringer Ingelheim, G), and 701 local scalp anaesthesia (200 µl of 0.025 % bupivacaine, AstraZeneca, UK) before placement 702 in a stereotaxic frame (Kopf, US; manual digital frame, World Precision Instruments, US) with 703 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 29 non-rupture mouse ear bars. The body temperature was stabilized using a feedback -704 controlled heating blanket (Harvard Apparatus, US) and the anaesthesia was maintained with 705 1.5 % isoflurane. The following stereotaxic coordinates (from bregma) and volumes were used 706 for bilateral transfection of the stated areas; ACC: injection at AP +1.0, ML 0.4, DV 1.3 (300 nl) 707 and 1.7 (2 00 nl); ventral mPFC: injection at AP +2.2, ML 0.3, DV 2.2 (200 nl). An AAV5-708 CamKIIα-GCaMP6m vector suspension (6.2*1012 vg/ml; University of Zürich viral vector 709 facility; UZH-VVF, CH) was diluted down 1:1 to a final titre of 3.1*1012 vg/ml in 5 % sorbitol/PBS 710 (Sigma, G) and infused using a 10 µl precision syringe (WPI, US) at an infusion rate of 711 100 nl/min. To minimize backflow of the virus, the needle was kept in place for 5 min at each 712 site after infusion, and additionally for another 5 min 0.1 mm above the last infusion site . 713 Subsequently, the wound was sutured, the mouse was allowed to recover in a temperature-714 controlled chamber at 36°C, and provided with mesh- food, gel-food and daily post-operative 715 monitoring for 7 d, including application of meloxicam (Metacam, 1 mg/kg, Boehringer 716 Ingelheim, G) on the first 3 d. The mice were kept on ad libitum food. 717 Approximately, one week after the injection of the viral construct , a gradient refractory index 718 (GRIN) lens (Inscopix, CA, USA) was implanted. The surgery initially followed the steps 719 described above for the virus injection and was followed by two craniotomies into the occipital 720 and parietal bone where a screw (1 mm diameter, Precision Technologies, GB) was placed 721 into each hole for later implant stability. A craniotomy for the lens (1 or 0.5 mm in diameter for 722 ACC or mPFC, respectively) was made above the original infusion site. Before lowering the 723 lens into the brain tissue, the skull was dried for better glue attachment and the lens was 724 cleaned with 70% ethanol or 100% isopropanol. From the brain surface, the lens was held 725 vertically by a pipette tip with negative internal pressure created by a vacuum pump and 726 lowered by 20 µm every 30 s into the brain tissue at the original infusion site until reaching a 727 depth of 1 mm (ACC). For mPFC, a custom-made GRIN-lens injector (“GRINjector”) was used 728 placing the lens at 2 mm from the brain surface. Super-glue (Loctite 401, Henkel, DE) was 729 applied to attach the lens to the skull, followed by light-curable dental adhesive (BreezeTM, 730 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 30 Pentron, US). Dental cement was applied on the exposed skull with approx. 1 mm of the lens 731 protruding out from the skull. Kwik-Sil™ (WPI, US ) was applied on the lens to protect it from 732 later mechanical damage. Post-operative care was applied as after the first surgery. 733 Between 2-6 wks after lens implantation, the GCaMP6m expression was checked. For this, 734 the mouse was ana esthetized and fixed to the stereotaxic frame as described above. After 735 removing the Kwik -Sil™ cone, the lens surface was wiped with lens tissue soaked in 70% 736 ethanol or 100% isopropanol. The baseplate was attached to the miniscope and to the 737 stereotactic frame using a clamp. Depending on the quality and quantity of GCaMP 6m-738 expressing cells, the baseplate was persistently fixated to the implant by applying light-curable 739 dental cement (Flow-ItTM, Pentron, US) around the lens, layer by layer leaving a small (approx. 740 1 mm) gap below the baseplate, which was filled with 2- component adhesive (Loctite 3090, 741 Henkel DE) for ultimate fixation. After drying, the miniscope was de- attached and a custom-742 made protective cap was put on the baseplate and fixated by the baseplate screw. 743 Calcium imaging 744 Calcium imaging was done using UCLA miniscopes v3 or v4 [29], including their data 745 acquisition (DAQ) box ( https://open-ephys.org) and acquisition software 746 (www.miniscope.org). The temporally aligned recording of behavioural events and imaging 747 frames was achieved through pyControl [19,28], connecting the miniscope DAQ -box 748 (https://open-ephys.org) via an i nput trigger GPIO SMA connector to the pyControl 749 microcontroller board through which the start and the end of image acquisition was controlled 750 by TTL-pulses sent to the miniscope DAQ-box. Prior to the session start, mice were equipped 751 with the miniscope and the optimal focus was set by adjusting the focus slider of the miniscope 752 manually (v3) or the focus electronically (v4). A thin and flexible coaxial cable (CW2040 -753 3650SR, Cooner Wire, US) connected the miniscope to the DAQ -box for power supply, LED 754 control, and CMOS data transmission. For some recordings a custom -made moto rized 755 commutator [30] was used to eliminate the need to manually un-twist the cable. Images were 756 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 31 recorded at 20 fps, maximum gain, and with an excitation intensity that was adjusted for each 757 mouse individually. 758 Histology 759 Animals were given an over -dose of ketamine/medetomidine (≥200 mg/kg ketamine, Zoetis, 760 G; ≥2mg/kg medetomidine, Pfizer, US) and perfused with 0.01 M phosphate- buffered saline 761 (PBS) followed by 4 % PFA/PBS. The baseplate and the implant were carefully detached from 762 the scull and the brain was removed and then stored in 4 % PFA/PBS overnight before 763 placement in 20 % sucrose for dehydration before sections were cut at 100 µm thickness on 764 a vibratome (VT1000, Leica, DE). Every second section was stained with DAPI (10 -4 % w/v) 765 for 30 min, washed with PBS twice and mounted on glass slides. A Leica DM6B 766 epifluorescence microscope (Leica, DE) was used to scan the slides with a 5x objective and 767 determine virus expression offline. 768 Data analysis 769 Pre-processing of calcium traces 770 Single-photon imaging data for each session were pre-processed using MATLAB as described 771 previously[17]. Each image frame was spatially down-sampled to a 400x400 pixel frame and 772 divided by its low -pass filtered version to remove wide- field fluctuations and brightness 773 gradients over the field of view. After band- pass filtering each frame to enhance structural 774 features of the image to facilitate the alignment of different frames, the TurboReg algorithm 775 [31] was used for motion correction. Each movie was temporally smoothed and temporally 776 down-sampled from 20 Hz to 5 Hz followed by signal normalization of each image frame in 777 units of relative changes in fluorescence, ΔF(t)/F0 = (F(t) − F0)/F0, where F0 is the mean 778 image obtained by averaging the entire movie. For cell sorting, spatial filters corresponding to 779 individual neurons were identified using an automated cell sorting routine based on principal 780 and independent component analysis (PCA/ICA) [32]. Extracted spatial filters were verified as 781 neural cells upon visual inspection based on size, morphology and the activity trace. 782 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 32 Temporal alignment of calcium traces to behavioural events 783 Calcium traces of each cell were z-transformed using the Matlab function zscore to control for 784 variations between animals and sessions. Custom written Matlab scripts were used first to 785 align the behavioural and the imaging data determining the timestamps for the ITI start, cue 786 presentation, choice and outcome (referred to as epoc hs) of each trial in terms of frame 787 numbers, and for labelling the trial with the corresponding choice of the mouse, i.e. correct, 788 incorrect or premature response, or omission (referred to as response type). In each trial, the 789 calcium signals were extracted within defined time wind ows from 4 s before to 7 s after the 790 event timestamp of the onset of each epoch. The extracted traces were then averaged across 791 trials of the same response type, thereby forming population vectors that represented each 792 response type aligned to the onset of each epoch. Peri-event time histograms were created 793 by plotting heat maps of the population vectors for each response type aligned to the choice 794 onset (Figure 2). The cells were sorted based on their average peak latency. For cross -795 validation, heat maps were created based on the population vectors created from averaging 796 only across even or odd trials, which were then correlated across cells within each time bin to 797 receive a measure for the reliability of the temporal activity pattern. K-means clustering was 798 applied using the Matlab function kmeans with a preset number of four clusters and the 799 distance metric set to 'cosine'. Thereby, the cells were grouped based on the similarity of their 800 calcium signals, which were extracted within defined time windows relative to the onset of the 801 respective choice (see above) and averaged across odd trials. Subsequently, the cells were 802 sorted according to their cluster assignment and peak latency applying the sorting order to the 803 calcium signals averaged across even trials and plotting them using peri -event time 804 histograms (Figure 2C-D). For each cluster, the single-trial calcium signal of a corresponding 805 exemplary cell was plotted for each response type using peri -event time histograms (Figure 806 2E-F). 807 Decoding analysis of population activity 808 Binary linear support vector machine (SVM) classifiers (Figure 3A -C and 6E-F) were trained 809 and tested on differentiating between trials with correct responses and those with either 810 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 33 omitted, premature or incorrect responses based on the amplitude of the calcium trace of all 811 neurons in a FOV combined in 200 ms time-bins in a time- window from 4 s before until 7 s 812 after the onset of each choice poke. For training the classifier, the Matlab function fitcsvm was 813 used with the kernel function, box constraint, kernel scale and standardization set to ’linear’, 814 ’1’, ’auto’ and 'false', respectively. The dataset of every session (with a minimum of 6 events 815 for each response type) was randomly partitioned into the training (each observation was 816 labelled with the response type) and test set (lacking label assignment) in a 80/20 ratio, while 817 ensuring the test set maintained balance, resulting in an imbalanced distribution within the 818 training set for most sessions. For sessions, where the number of trials varied extensively 819 across response types, trials from the response type with a higher number of events were 820 randomly removed until achieving bal ance in the test set. For training the SVM classifier, a 821 balanced training set is essential to prevent a classification bias towards the majority class 822 (i.e. the behavioural event class with the highest number of observations in the respective data 823 set) [35]. Using the synthetic minority oversampling technique (SMOTE) [36,37] on the training 824 set, the number of observations in each event class was equalized by artificially synthesizing 825 new samples in the minority classes (i.e. the behaviou ral event classes with a lower number 826 of observations than the majority class). This algorithm randomly selects an observation from 827 the underrepresented event class and identif ies its four nearest neighbours, of which one is 828 randomly chosen. A value is randomly picked in the Euclidean distance between the 829 observation and the neighbour and is assigned to the new synthesized sample. The smote 830 approach requires the number of events in the minority class to be greater than the number of 831 set neighbours (i.e. four) and the ratio between the number of events in the majority and 832 minority class to be less than the set number of neighbours (i.e. four). In sessions where this 833 was not the case, events were up- sampled for the minority class using SMOTE with the 834 number of neighbours set to the number of events in the minority class and/or trials were 835 randomly removed from the majority class until the required conditions were met. The entire 836 procedure of random data set partitioning, SMOTE up- sampling of the training set, and the 837 subsequent training and testing of the decoder was repeated 100 -times, thereby producing 838 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 34 100 averages and 100 s.e.m. values, which were averaged to yield a grand mean and grand 839 s.e.m value representing the decoding performance based on each session. As a control for 840 the test decoder, a second binary classifier model was established as described above, but 841 labels of the training set in each fold within each session were shuffled prior to classifier 842 training, thereby creating distributions that represented chance level (null distribution) [38]. 843 Since mice mostly made an insufficient number of incorrect responses during the varITI 844 challenges, to allow this type of decoding analysis, the data from several sessions with at least 845 8 incorrect responses were grouped across challenge protocols (five combined sessions, two 846 varITI sessions, two fixed-ITI sessions) to perform a separate decoding analysis that included 847 incorrect responses (Figure 3C). 848 Multiclass decoding was performed to predict into which of the five poke holes the mouse 849 poked into during a given trial including all response types or correct responses only, using a 850 multi-class SVM classifier [39,40] with the same approach as described above (Fig ure 4D). 851 The linear SVM multi-classifier was trained using the matlab function fitcecoc with the kernel 852 function, box constraint, kernel scale and standardization set to ’linear’, ’1’, ’auto’ and ’off’, 853 respectively. Additionally, the option coding was set to ’onevsall’ ; t he one -vs-all strategy 854 performs a separate binary classification for each class in the dataset (i.e. in total four) treating 855 it as the positive class, whereas all other classes combined are treated as the negative class. 856 Testing is performed by independently applying every sample from the test data set on each 857 trained binary classifier yielding confidence values with the highest selecting the predicted 858 class for this sample. 859 Encoding analysis of the modulation of neural activity by behavioural events 860 Linear regression models were created to predict the calcium signal in 200 ms timebins in a 861 time window from 4 s before until 7 s after the onset of the choice epoch for each individual 862 neuron (Figure 5 and 7). Regularized linear regression was performed using the Matlab 863 function lasso applying L1 (lasso) regularization with 10-fold cross-validation to find the optimal 864 regularization strength λ that minimizes the loss. Binary predictors were used to code for the 865 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 35 presence of a poke (active poke vs. omission), the spatial poke identity (four one-hot predictors 866 corresponding to left-right directionality, right poke discrimination, left poke discrimination and 867 middle poke discrimination) , and correct (rewarded) responses (vs. erroneous and omitted 868 responses combined) in each trial (see Figure 5A for predictor matrix). To test how much of 869 the variance of the activity of individual neurons at every time bin could be explained by each 870 predictor, the coefficient of partial determination (CPD) was calculated, measuring how much 871 further the predictor contributed to the explanation of the full regression model [21,33,34]. 872 CPDs were determined for each predictor by subtracting the mean squared error of the full 873 linear regression model from the mean squared error of the reduced regression model, where 874 the predictor for the specif ic event type in question was removed. The CPD for predictor i is 875 defined as: 876 𝐶𝐶𝐶𝐶𝐶𝐶𝑖𝑖= 𝑀𝑀𝑀𝑀𝑀𝑀𝑋𝑋−𝑖𝑖− 𝑀𝑀𝑀𝑀𝑀𝑀𝑋𝑋 𝑀𝑀𝑀𝑀𝑀𝑀𝑋𝑋−𝑖𝑖 877 w here 𝑀𝑀𝑀𝑀𝑀𝑀𝑋𝑋−𝑖𝑖 is the mean squared error in a regression model that includes all of the relevant 878 predictor variables except i, and 𝑀𝑀𝑀𝑀𝑀𝑀𝑋𝑋 is the mean squared error in a regression model that 879 includes all of the relevant predictor variables. To compute the CPD for spatial poke identity, 880 all of the four spatial one-hot predictors were removed together. 881 For the devaluation and extinction conditions (Figure 7), which led to reduced number s of 882 correct responses, pseudo-randomly chosen events from the baseline (control) condition were 883 selected to roughly match the number of the respective experimental condition for each poke-884 hole and response type , to ensure that CPDs in the baseline session are not simply higher 885 because of a larger number of trials or responses used for thei r calculation . This down-886 sampling was repeated 100 times and the CPDs of the respective predictors averaged across 887 iterations before plotting. 888 889 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 36 Statistics 890 Behavioural data was analysed using M atlab (R2019a, The MathWorks Inc, USA) and only 891 t wo-sided tests were used. 5- CSRTT performance during calcium imaging sessions (Figure 892 1D-E, Figure 6A-D) was analysed using an ANOVA involving the task paradigm as between-893 subject independent variable and one of the behaviou ral parameters as dependent variable. 894 In case of a significant effect of task paradigm, Dunnet t’s post-hoc tests were conducted 895 between the baseline and any other challenge. Decoding accuracies (Figure 3 ) were 896 statistically compared using repeated- measures ANOVA with the time -bin and epoch type 897 variable as within-subject factors. A Dunn-Sidak-test was used for post -hoc test ing. 898 Comparisons against accuracies of control classifiers (trained with shuffled labels, performing 899 at chance level) in decoding analyses or against 0% CPD in encoding analyses have been 900 done with paired- sample or one- sample t-tests, respectively, with Benjam ini-Hochberg 901 corrections for the repeated testing in each time interval. All applied statistical tests are stated 902 in the corresponding figure legends. All bar and line graphs display mean ± s.e.m. or data from 903 individual mice, as indicated. 904 905 Data availability 906 All raw data can be obtained from the corresponding author upon reasonable request. Scripts 907 of all task files applied in custom -made operant boxes can be obtained from 908 https://github.com/KaetzelLab/Operant-Box-Code and design files for such operant boxes are 909 deposited at https://github.com/KaetzelLab/Operant-Box-Design-Files. 910 911 Code availability 912 Analysis scripts are available from GitHub at 913 https://github.com/martinjendryka/Jendryka_et_al_ACC_imaging_5CSRTT.git. 914 915 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 37

916 1. Dalley JW, Robbins TW. Fractionating impulsivity: neuropsychiatric implications. Nat Rev 917 Neurosci. 2017;18: 158–171. doi:10.1038/nrn.2017.8 918 2. Millan MJ, Agid Y, Brüne M, Bullmore ET, Carter CS, Clayton NS, et al. Cognitive 919 dysfunction in psychiatric disorders: characteristics, causes and the quest for improved 920 therapy. Nat Rev Drug Discov. 2012;11: 141–168. doi:10.1038/nrd3628 921 3. Bari A, Dalley JW, Robbins TW. The application of the 5-choice serial reaction time task 922 for the assessment of visual attentional processes and impulse control in rats. Nat Protoc. 923 2008;3: 759–767. 924 4. Bari A, Robbins TW. Inhibition and impulsivity: Behavioral and neural basis of response 925 control. Prog Neurobiol. 2013;108: 44–79. doi:10.1016/j.pneurobio.2013.06.005 926 5. van der Veen B, Kapanaiah SKT, Kilonzo K, Steele-Perkins P, Jendryka MM, Schulz S, 927 et al. Control of impulsivity by Gi- protein signalling in layer -5 pyramidal neurons of the 928 anterior cingulate cortex. Commun Biol. 2021;4: 1– 16. doi:10.1038/s42003-021-02188-929 w 930 6. Norman KJ, Koike H, McCraney SE, Garkun Y, Bateh J, Falk EN, et al. Chemogenetic 931 suppression of anterior cingulate cortical neurons projecting to the visual cortex disrupts 932 attentional behavior in mice. Neuropsychopharmacol Rep. 2021;41: 207– 214. 933 doi:10.1002/npr2.12176 934 7. Norman KJ, Riceberg JS, Koike H, Bateh J, McCraney SE, Caro K, et al. Post -error 935 recruitment of frontal sensory cortical projections promotes attention in mice. Neuron. 936 2021 [cited 25 Feb 2021]. doi:10.1016/j.neuron.2021.02.001 937 8. Jendryka MM, Lewin U, van der Veen B, Kapanaiah SKT, Prex V, Strahnen D, et al. 938 Control of sustained attention and impulsivity by Gq -protein signalling in parvalbumin 939 interneurons of the anterior cingulate cortex. Transl Psychiatry. 2023;13: 1– 12. 940 doi:10.1038/s41398-023-02541-z 941 9. Totah NKB, Kim YB, Homayoun H, Moghaddam B. Anterior Cingulate Neurons 942 Represent Errors and Preparatory Attention within the Same Behavioral Sequence. J 943 Neurosci. 2009;29: 6418–6426. doi:10.1523/JNEUROSCI.1142-09.2009 944 10. Donnelly NA, Paulsen O, Robbins TW, Dalley JW. Ramping single unit activity in the 945 medial prefrontal cortex and ventral striatum reflects the onset of waiting but not imminent 946 impulsive actions. Eur J Neurosci. 2015; n/a-n/a. doi:10.1111/ejn.12895 947 11. Hyman JM, Whitman J, Emberly E, Woodward TS, Seamans JK. Action and Outcome 948 Activity State Patterns in the Anterior Cingulate Cortex. Cereb Cortex. 2013;23: 1257–949 1268. doi:10.1093/cercor/bhs104 950 12. Broom E, Imbriotis V, Sengpiel F, Connelly WM, Ranson A. Recruitment of frontal 951 sensory circuits during visual discrimination. Cell Rep. 2022;39: 110932. 952 doi:10.1016/j.celrep.2022.110932 953 13. Wal A, Klein FJ, Born G, Busse L, Katzner S. Evaluating Visual Cues Modulates Their 954 Representation in Mouse Visual and Cingulate Cortex. J Neurosci. 2021;41: 3531–3544. 955 doi:10.1523/JNEUROSCI.1828-20.2021 956 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 38 14. Kim H, Ährlund -Richter S, Wang X, Deisseroth K, Carlén M. Prefrontal Parvalbumin 957 Neurons in Control of Attention. Cell. 2016;164: 208–218. doi:10.1016/j.cell.2015.11.038 958 15. Hunt LT, Malalasekera WMN, Berker AO de, Miranda B, Farmer SF, Behrens TEJ, et al. 959 Triple dissociation of attention and decision computations across prefrontal cortex. Nat 960 Neurosci. 2018;21: 1471–1481. doi:10.1038/s41593-018-0239-5 961 16. Hunt LT, Hayden BY. A distributed, hierarchical and recurrent framework for reward-962 based choice. Nat Rev Neurosci. 2017;18: 172–182. doi:10.1038/nrn.2017.7 963 17. Grewe BF, Gründemann J, Kitch LJ, Lecoq JA, Parker JG, Marshall JD, et al. Neural 964 ensemble dynamics underlying a long-term associative memory. Nature. 2017;543: 670–965 675. doi:10.1038/nature21682 966 18. Lui JH, Nguyen ND, Grutzner SM, Darmanis S, Peixoto D, Wagner MJ, et al. Differential 967 encoding in prefrontal cortex projection neuron classes across cognitive tasks. Cell. 968 2021;184: 489-506.e26. doi:10.1016/j.cell.2020.11.046 969 19. Kapanaiah SKT, van der Veen B, Strahnen D, Akam T, Kätzel D. A low-cost open-source 970 5-choice operant box system optimized for electrophysiology and optophysiology in mice. 971 Sci Rep. 2021;11: 22279. doi:10.1038/s41598-021-01717-1 972 20. Li Y, Mathis A, Grewe BF, Osterhout JA, Ahanonu B, Schnitzer MJ, et al. Neuronal 973 Representation of Social Information in the Medial Amygdala of Awake Behaving Mice. 974 Cell. 2017;171: 1176-1190.e17. doi:10.1016/j.cell.2017.10.015 975 21. Akam T, Rodrigues-Vaz I, Marcelo I, Zhang X, Pereira M, Oliveira RF, et al. The anterior 976 cingulate cortex predicts future states to mediate model-based action selection. Neuron. 977 2020;0. doi:10.1016/j.neuron.2020.10.013 978 22. Totah NKB, Jackson ME, Moghaddam B. Preparatory Attention Relies on Dynamic 979 Interactions between Prelimbic Cortex and Anterior Cingulate Cortex. Cereb Cortex. 980 2013;23: 729–738. doi:10.1093/cercor/bhs057 981 23. Koike H, Demars MP, Short JA, Nabel EM, Akbarian S, Baxter MG, et al. Chemogenetic 982 Inactivation of Dorsal Anterior Cingulate Cortex Neurons Disrupts Attentional Behavior 983 in Mouse. Neuropsychopharmacology. 2016;41: 1014–1023. doi:10.1038/npp.2015.229 984 24. Sul JH, Kim H, Huh N, Lee D, Jung MW. Distinct Roles of Rodent Orbitofrontal and 985 Medial Prefrontal Cortex in Decision Making. Neuron. 2010;66: 449– 460. 986 doi:10.1016/j.neuron.2010.03.033 987 25. Sohal VS, Zhang F, Yizhar O, Deisseroth K. Parvalbumin neurons and gamma rhythms 988 enhance cortical circuit performance. Nature. 2009;459: 698– 702. 989 doi:10.1038/nature07991 990 26. Terra H, Bruinsma B, de Kloet SF, van der Roest M, Pattij T, Mansvelder HD. Prefrontal 991 Cortical Projection Neurons Targeting Dorsomedial Striatum Control Behavioral 992 Inhibition. Curr Biol. 2020;30: 4188-4200.e5. doi:10.1016/j.cub.2020.08.031 993 27. de Kloet SF, Bruinsma B, Terra H, Heistek TS, Passchier EMJ, van den Berg AR, et al. 994 Bi-directional regulation of cognitive control by distinct prefrontal cortical output neurons 995 to thalamus and striatum. Nat Commun. 2021;12: 1994. doi:10.1038/s41467-021-22260-996 7 997 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 39 28. Akam T, Lustig A, Rowland JM, Kapanaiah SK, Esteve-Agraz J, Panniello M, et al. Open-998 source, Python-based, hardware and software for controlling behavioural neuroscience 999 experiments. Kemere C, Wassum KM, Kemere C, Siegle J, editors. eLife. 2022;11: 1000 e67846. doi:10.7554/eLife.67846 1001 29. Aharoni D, Hoogland TM. Circuit Investigations With Open -Source Miniaturized 1002 Microscopes: Past, Present and Future. Front Cell Neurosci. 2019;13. 1003 doi:10.3389/fncel.2019.00141 1004 30. Kapanaiah SKT, Kätzel D. Open -MAC: A low-cost open-source motorized commutator 1005 for electro - and opto-physiological recordings in freely moving rodents. HardwareX. 1006 2023;14: e00429. doi:10.1016/j.ohx.2023.e00429 1007 31. Thevenaz P, Ruttimann UE, Unser M. A pyramid approach to subpixel registration based 1008 on intensity. IEEE Trans Image Process. 1998;7: 27–41. doi:10.1109/83.650848 1009 32. Mukamel EA, Nimmerjahn A, Schnitzer MJ. Automated Analysis of Cellular Signals from 1010 Large-Scale Calcium Imaging Data. Neuron. 2009;63: 747– 760. 1011 doi:10.1016/j.neuron.2009.08.009 1012 33. Cai X, Kim S, Lee D. Heterogeneous Coding of Temporally Discounted Values in the 1013 Dorsal and Ventral Striatum during Intertemporal Choice. Neuron. 2011;69: 170– 182. 1014 doi:10.1016/j.neuron.2010.11.041 1015 34. Chiang F- K, Wallis JD. Neuronal encoding in prefrontal cortex during hierarchical 1016 reinforcement learning. J Cogn Neurosci. 2018;30: 1197– 1208. 1017 doi:10.1162/jocn_a_01272 1018 35. Krawczyk B. Learning from Imbalanced Data: Open Challenges and Future Directions. 1019 Prog Artif Intell. 2016;5: 221–232. doi:10.1007/s13748-016-0094-0 1020 36. Chawla NV, Bowyer KW, Hall L O, Kegelmeyer WP. SMOTE: Synthetic Minority Over -1021 sampling Technique. J Artif Intell Res. 2002;16: 321–357. doi:10.1613/jair.953 1022 37. Larsen BS. Synthetic Minority Over-Sampling Technique (SMOTE). 2020. 1023 38. Combrisson E, Jerbi K. Exceeding chance level by chance: The caveat of theoretical 1024 chance levels in brain signal classification and statistical assessment of decoding 1025 accuracy. J Neurosci Methods. 2015;250: 126–136. doi:10.1016/j.jneumeth.2015.01.010 1026 39. Boser BE, Guyon IM, Vapnik VN. A Training Algorithm for Optimal Margin Classifiers. 1027 Proceedings of the Fifth Annual Workshop on Computational Learning Theory. New 1028 York, NY, USA: Association for Computing Machinery; 1992. pp. 144– 152. 1029 doi:10.1145/130385.130401 1030 40. Vapnik V. The Nature of Statistical Learning Theory. Springer Science & Business Media; 1031 2013. 1032 1033 1034 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint 40 Acknowledgments 1035 We thank S tefanie Schulz (Ulm University) for assistance with histology. Funding: This work 1036 was funded by the Boehringer Ingelheim-Ulm University (BIU) Center (TPN010, to B.L., A.P., 1037 D.K.), the Else -Kroener-Fresenius/German-Scholars-Organization Programme for excellent 1038 medical scientists from abroad (GSO/EKFS 12; to D.K.), the DFG (KA 4594/2-1; to D.K.), and 1039 the Alfred-Krupp Foundation (to B.L.). 1040 1041 Author Contributions 1042 M.M.J., B.L., A.P., T.A. and D.K. designed the study. M.J. and U.L. conducted behavioural 1043 experiments. M.M.J. conducted surgeries. S.K.T.K., T.A., and D.K. developed pyOS-5 operant 1044 box hardware and software; S.K.T.K. programmed operant box task protocols and integration 1045 of miniscope recordings. B.F.G. and H.D. provided assistance with manufacturing and usage 1046 of UCLA v3 miniscopes. B.F.G., B.L. and D.K. provided essential resources. M.M.J. analysed 1047 the data with advise from B.F.G., T.A. and D.K.. M.M.J. and D.K. wrote the manuscript, which 1048 was revised by all authors. 1049 1050 Competing Interests statement 1051 The authors declare no competing interest. A.P. is an employee of Boehringer Ingelheim. 1052 1053 .CC-BY-NC-ND 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted April 15, 2024. ; https://doi.org/10.1101/2024.04.12.589244doi: bioRxiv preprint