Task-dependent representations of stimulus and choice in mouse parietal cortex — TEI-tagged view

PhopersName, GoardpersName, WoodsonpersName, CrawfordpersName, SurpersName (2018date). “Task-dependent representations of stimulus and choice in mouse parietal cortex.” Nature Communicationstitle·j.

Object of study — whether the role of posterior parietal cortex is specific to sensoryterm processing or to sensorimotor transformation

Research question — whether PPCterm signals encode the stimulusterm or the animal's impending choiceterm, tested by imaging the same neuronsterm before and after a reversed sensorimotor contingencyterm

Contribution — mouse PPCterm is strongly task-dependent and reflects choiceterm more than stimulusterm, and may play a role in the transformation of visualterm inputs into motorterm commands

Full text — tagged

The reference list is not part of the text we hold for this paper. Our copy was extracted from the publisher's page and the bibliography did not survive that extraction, so the article below ends before its references. Nothing has been removed by us.

Task-dependent representations of stimulusterm and choiceterm in mouse parietal cortex Gerald N PhopersName Gerald N PhopersName 1 Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, MA 02139 USA 2 Picower Institute for Learning and Memory, Massachusetts Institute of Technology, Cambridge, MA 02139 USA 3 Department of Organismic and Evolutionary Biology, Harvard University, Cambridge, MA 02138 USA 1, 2, 3, # , Michael J GoardpersName Michael J GoardpersName 1 Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, MA 02139 USA 2 Picower Institute for Learning and Memory, Massachusetts Institute of Technology, Cambridge, MA 02139 USA 4 Department of Molecular, Cellular, and Developmental Biology, University of California, Santa Barbara, Santa Barbara, CA 93106 USA 5 Department of Psychological & Brain Sciences, University of California, Santa Barbara, Santa Barbara, CA 93106 USA 1, 2, 4, 5, # , Jonathan WoodsonpersName Jonathan WoodsonpersName 1 Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, MA 02139 USA 2 Picower Institute for Learning and Memory, Massachusetts Institute of Technology, Cambridge, MA 02139 USA 1, 2 , Benjamin CrawfordpersName Benjamin CrawfordpersName 1 Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, MA 02139 USA 2 Picower Institute for Learning and Memory, Massachusetts Institute of Technology, Cambridge, MA 02139 USA 1, 2 , Mriganka SurpersName Mriganka SurpersName 1 Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, MA 02139 USA 2 Picower Institute for Learning and Memory, Massachusetts Institute of Technology, Cambridge, MA 02139 USA 1, 2, ✉ 1 Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, MA 02139 USA 2 Picower Institute for Learning and Memory, Massachusetts Institute of Technology, Cambridge, MA 02139 USA 3 Department of Organismic and Evolutionary Biology, Harvard University, Cambridge, MA 02138 USA 4 Department of Molecular, Cellular, and Developmental Biology, University of California, Santa Barbara, Santa Barbara, CA 93106 USA 5 Department of Psychological & Brain Sciences, University of California, Santa Barbara, Santa Barbara, CA 93106 USA ✉ Corresponding author.

# Contributed equally.

Received 2017 Sep 22; Accepted 2018 Jun 11; Collection date 2018.

© The Author(s) 2018 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/ .

PMCID: PMC6030204 PMID: 29968709 This article has been corrected. See Nat Commun. 2019 Jan 18;10:389 .

Abstract The posterior parietal cortex (PPCterm) has been implicated in perceptual decisionsterm, but whether its role is specific to sensoryterm processing or sensorimotor transformation is not well understood. Here, we trained mice to perform a go/no-go visualterm discrimination taskterm and imaged the activity of neuronsterm in primary visualterm cortex (V1term) and PPCterm during engaged behavior and passiveterm viewing. Unlike V1term neuronsterm, which respond robustly to stimuli in both conditions, most PPCterm neuronsterm respond exclusively during task engagementterm·var. To test whether signals in PPCterm primarily encoded the stimulusterm or the animal’s impending choiceterm, we image the same neuronsterm before and after re-training mice with a reversed sensorimotor contingencyterm. Unlike V1term neuronsterm, most PPCterm neuronsterm reflect the animal’s choiceterm of the new targetterm stimulusterm after re-training. Mouse PPCterm is therefore strongly task-dependent, reflects choiceterm more than stimulusterm, and may play a role in the transformation of visualterm inputs into motorterm commands.

The precise role of PPCterm in transforming sensoryterm signals to relevant actions is not yet clear. Here, the authors show that unlike V1term, which is largely driven by visualterm input, PPCterm is strongly task-dependent and exhibits a mixture of stimulusterm and choiceterm signals in a visualterm decisionterm taskterm.

Introduction Perceptual decision-making involves multiple cognitive processes, including processing of sensoryterm stimuli, accumulation of evidence, and transformation of sensoryterm information into an appropriate motorterm plan. Although many brain regions have been implicated in perceptual decisionsterm, dissociating their individual contribution to these different processes remains a challenge.

In particular, the posterior parietal cortex (PPCterm) has been hypothesized to play a key role in at least some types of decisionterm tasksterm in both primates 1 , 2 and in rodents 3 – 5 . However, the specific role of rodent parietal cortex, and whether its function is homologous to that of primates, remains unclear. While rodent PPCterm plays a minimal role in simple auditory 6 and whisker-based 7 decisionterm tasksterm (but see 8 ), multiple groups have demonstrated that it is causally necessary for make decisionsterm on the basis of visualterm stimuli 3 , 4 , 9 , 10 .

However, the specific role of PPCterm in visualterm decision-making remains unclear. Some argue that rodent PPCterm may ultimately be more homologous to extrastriate cortex in processing sensoryterm signals that are accumulated elsewhere for decision-making 10 . Indeed, both anatomical projection studies 11 , 12 , as well as functional mapping studies 13 , 14 indicate that PPCterm may overlap with or contain a group of retinotopically-organized extrastriate areas that are rostral to primary visualterm cortex (V1term). Others suggest that PPCterm reflects internal biases related to the value of past stimuli 8 or actions 9 . However, removal of internal biases cannot fully explain the decision-making deficits induced by inactivation of PPCterm during visualterm decisionterm tasksterm 3 , 4 , 10 . A third, alternative possibility is that PPCterm may play a role in the mapping of visualterm stimuli to motorterm commands. If this were the case, one may expect that activity in PPCterm would be highly task-dependent, reflecting the animal’s decisionterm depending on learned sensorimotor contingencies.

Here, we use populationterm calcium imagingterm·mth to investigate these possibilities by measuring activity in PPCterm and in V1term during a go/no-go lick-based visualterm discrimination taskterm. Having previously demonstrated the necessity of PPCterm in the performance of this taskterm 3 , we seek in this work to investigate its specific role in either sensoryterm processing or sensorimotor transformation by manipulating task engagementterm·var and learned taskterm contingencies. V1term neuronsterm exhibit robust visualterm responses during both task engagementterm·var and passiveterm viewing of stimuli that remain stable after taskterm contingencyterm reversal. By contrastterm, PPCterm responses are largely specific to taskterm performance and reflect the animal’s choiceterm before and after taskterm contingencyterm reversal. Our results are consistent with a role of the mouse posterior parietal cortex in transforming visualterm information to motorterm commands during perceptual decisionsterm.

Results Imaging V1term and PPCterm during taskterm performance and passiveterm viewing We trained mice on a head-fixed lickterm/no-lick visualterm discrimination taskterm (Fig. 1a, b ), similar to previous designs 3 , 15 , 16 . Water-restricted mice discriminated between a targetterm stimulusterm (horizontal grating drifting upwards, 0° from vertical, Stimulusterm A) which was rewarded with water, and a non-target stimulusterm of an orthogonal orientation (vertical drifting upwards rightwards, 90°, Stimulusterm B). Lickterm responses to the non-target grating were discouraged by punishment with a small, aversive drop of quinine. Quinine concentrationsterm·var were chosen to discourage licking to non-targets, but without inducing long-lasting changes in motivation on subsequent trials (Supplementary Fig. 1 ). A retractable lickterm spout was presented immediately after stimulusterm presentation (2 s) and retracted on every trial. This restricted the animal’s lickterm response to the “response” period (1.5 s) and allowed us to separately assess perception and action (Fig. 1b ). Video recording of the mice during the stimulusterm period confirmed that mice withheld licking until the spout became available during the response epoch 3 . Mice ( n = 15) achieved high levels of performance on the taskterm (Fig. 1c ; d-primeterm·def or d' , 2.23 ± 0.18; mean ± SEM), with a bias towards licking, resulting in more false alarms than misses.

Fig 1.

Open in a new tab Imaging calcium responses in V1term and PPCterm during engaged taskterm performance and passiveterm viewing. a Head-fixed mice were trained to perform a go, no-go lick-based visualterm discrimination taskterm. A retractable lickterm spout was used to restrict lickterm responses to a specific epoch of the taskterm. Licksterm following a targetterm stimulusterm (red, horizontal drifting upwards, Stimulusterm A) were rewarded with water, while licksterm to non-target stimulusterm (blue, vertical drifting rightwards, Stimulusterm B) were punished with quinine. b Trial structure for Engaged and Passiveterm conditions. After a brief auditory preparatory cue, a drifting grating was presented for 2 s. During Engaged trials, the retractable lickterm spout was presented immediately after stimulusterm offset for a minimum of 1.5 s. During Passiveterm trials, the spout was inaccessible. Engaged and Passiveterm trials were presented in blocks, which were usually interleaved. c Rate of licking on targetterm (T, red) and non-target (NT, blue) for each mouse used in imaging experiments ( n = 15). Behavioral performance was quantified as d-primeterm·def (mean across mice: 2.22). Errorterm bars in this and all subsequent figures depict mean ± SEM. d Stimulus-evoked response of two V1term neuronsterm, one target-selective (left column), and one non-target selective (right). Top, raw calcium response to multiple presentations of targetterm (red) and non-target (blue) stimuli during both Engaged and Passiveterm conditions. Middle, heatmap of trial-to-trial responses to targetterm and non-target stimuli, presented in alternating blocks of Engaged (left) and Passiveterm (right) trials, normalized to max response. Light gray shaded regions in this and subsequent figures demarcates duration of stimulusterm. Bottom, overlay of trial-averaged responses for each stimulusterm during Engaged (left) and Passiveterm (right) conditions. Line thickness indicates mean ± SEM. e Same as d but for two PPCterm neuronsterm We have previously demonstrated using a version of this taskterm with a 4 s delay period that inactivation of either V1term or PPCterm during the stimulusterm period disrupts behavioral performance, whereas inactivation during later periods has no effect 3 . Stimulus-period activity in PPCterm is therefore necessary for the taskterm, but it is unclear whether such activity is important for sensoryterm processing, attentional engagementterm, decisionterm formation, or motorterm planning. To help distinguish between these possibilities, we measured neural activity in PPCterm and V1term under two different conditions: engagementterm in the behavioral taskterm as well as passiveterm viewing of the same visualterm stimuli (Fig. 1b ). While “task engagementterm·var” may not be able to fully separate various non-sensory signals, we reasoned that this approach could allow us to identify pure sensoryterm responses that were robust during both passiveterm and engaged conditions.

We used two-photon microscopy and a volumetric imaging approach to image hundreds of neuronsterm simultaneously in either V1term or PPCterm (see “Methods” section). After behavioral training, we injected AAV2/1 syn-GCaMP6s 17 into the two areas under stereotaxic guidance. We used a resonant scanning system combined with a z-piezo to concurrently record activity from several hundreds of GCaMP6-infected cells within layer 2/3 in a volume comprising four planes (850 × 850 µm) 20 µm apart in depth, at an overall stack rate of 5 Hz. Images were corrected for X – Y movement, and fluorescence traces were extracted from semi-automatically generated ROIs based on a pixel-wise activity map (see “Methods” section).

To investigate the effects of behavioral performance on neural responses, we imaged the same neuronsterm in alternating blocks (5–10 min, 40–80 trials) of engaged behavior and passiveterm viewing (Fig. 1b ). During “engaged” behavior trials, the spout was extended on each trial during the response period. During “passiveterm” trials, the spout was not extended during the response period but remained withdrawn and inaccessible from the animal. To avoid extraneous stimulusterm confounds, no additional cue was provided to signal engaged or passiveterm blocks. Nonetheless, mice rapidly became aware after the first 1–2 trials of a block whether the spout would be available for a behavioral response, as confirmed in video recordings by the complete lack of licking during passiveterm blocks. Discrimination performance was similarly high for both the first and subsequent engaged blocks.

We imaged from an average of 606 cells (range: 257–1057) in each populationterm within either V1term or PPCterm, of which an average 135 cells (range: 21–364) exhibited significant task-related responses. Pilot experiments using transgenic mice with tdTomato expressed in PV + and SOM + interneuronsterm revealed that calcium signals from these neuronsterm were generally too weak to be measured with volume scanning, and therefore the vast majority of task-responsive cells were likely to be excitatory pyramidalterm neuronsterm 3 . We first analyzed only the neural responses measured during correct Engaged trials (ignoring errorterm trials), or during Passiveterm viewing. In V1term, many neuronsterm showed a robust and reliable response to either the targetterm or non-target stimulusterm, though some were modulated by engagementterm in the taskterm. For example, one target-preferring cell showed a reliable passiveterm response to the targetterm stimulusterm that was moderately enhanced during taskterm performance (Fig. 1d , left). A non-target preferring cell, however, exhibited a suppressed response during engagementterm in the taskterm (Fig. 1d , right). By contrastterm, PPCterm neuronsterm exhibited much stronger responses during task engagementterm·var compared to passiveterm viewing, specifically for targetterm stimuli. For example, one target-preferring cell had robust activity only during engagementterm (Fig. 1e , left), whereas another cell showed relatively weak passiveterm responses that became stronger during taskterm performance (Fig. 1e , right).

Responses are visually-driven in V1term but task-gated in PPCterm To compare the overall response properties of V1term and PPCterm, we analyzed the responses of all neuronsterm that had significant stimulus-period activity during the taskterm. We focused on stimulus-period responses given that inactivation during this period disrupts behavioral performance 3 . A total of 1915 neuronsterm (18% of all neuronsterm, 18 fields, 9 mice) in V1term (Fig. 2a, b ) and 3524 neuronsterm (26% of all neuronsterm, 22 fields, 10 mice) in PPCterm (Fig. 2c, d ) were significantly responsive during the taskterm. Two striking differences between V1term and PPCterm were immediately apparent when examining the trial-averaged responses. First, while V1term was about evenly split between target- and nontarget-preferring cells (60.6% ± 5.8% target-preferring, Fig. 2e ), there was a stronger bias in PPCterm towards the targetterm stimulusterm (87.6% ± 7.4% target-preferring; V1term vs. PPCterm, p = 4.68 × 10 −5 , Wilcoxon rank-sum testterm·stat). Secondly, while most task-responsive V1term neuronsterm also responded during passiveterm viewing (88.3% ± 6.8% with significant passiveterm response, Fig. 2f ), the majority of PPCterm neuronsterm had responses that were gated by task engagementterm·var (only 29.7% ± 10.3% with significant passiveterm response; V1term vs. PPCterm, p = 4.71 × 10 −7 , Wilcoxon rank-sum testterm·stat).

Fig 2.

Open in a new tab V1term neuronsterm respond passively, whereas PPCterm responses are gated by task engagementterm·var. a Trial-averaged responses of all task-responsive V1term neuronsterm ( n = 1915). Heatmap of all trial-averaged responses (preferred stimulusterm only) in both Engaged (left) and Passiveterm (right) conditions, normalized by peak response. Neuronsterm are separated into target-preferring (red) and non-target preferring (blue), and then into passive-responding (solid) and task-gated (hatched). Vertical dashed lines demarcate duration of stimulusterm. b Bottom, average response across V1term neuronsterm to the targetterm (red) and non-target (blue) in Engaged (left) and Passiveterm (right) conditions. Line thickness indicates mean ± SEM. c , d Same as ( a , b ), but for PPCterm neuronsterm ( n = 3524). e Percentage of neuronsterm in V1term (left) and PPCterm (right) that prefer targetterm (T) or non-target (NT) stimulusterm. Errorterm bars represent bootstrapped SEM across sessionsterm·stat. f Percentage of target-preferring neuronsterm in V1term (left) and PPCterm (right) that are task-gated, i.e. that respond only during engagementterm (E-only), or that respond during both engaged and passiveterm conditions (E + P). g Histogram of stimulusterm selectivity indexterm·def for V1term (left) and PPCterm (right) during Engaged (filled bars) and Passiveterm conditions (gray line). Positive selectivityterm indicates preference for Targetterm. Colored bars indicate neuronsterm with significant individual selectivityterm during Engaged trials. h Histogram of engagement modulation indexterm·var for V1term (left) and PPCterm (right) for target- (red) and nontarget-preferring (blue) neuronsterm The dramatic effect of task engagementterm·var on PPCterm responses could also be observed by comparing the selectivityterm of responses during Passiveterm and Engaged conditions (Fig. 2g ). We quantified selectivityterm using an ROC-based index that ranged from −1 to 1, with positive values indicating preference for the targetterm stimulusterm (see “Methods” section for details). V1term had a large proportion of significantly selective ( p < 0.05, permutation testterm·stat) neuronsterm in Passiveterm conditions (85.3% ± 3.8% of cells), with a small but significant increase in Engaged conditions (91.8% ± 2.5% of cells; Passiveterm vs. Engaged, p = 4.97 × 10 −3 , Wilcoxon signed-rank testterm·stat). By contrastterm, PPCterm neuronsterm were largely unselective during Passiveterm conditions (41.9% ± 10.8% selective), and yet became significantly selective to targetterm trials during task engagementterm·var (89.5% ± 4.7% of cells; Passiveterm vs. Engaged, p = 5.95 × 10 −5 , Wilcoxon signed-rank testterm·stat). We also quantified the change in responses for engagementterm vs. passiveterm viewing for each neuronterm using an engagement modulation indexterm·var (Fig. 2h ), which also ranged from −1 to 1, with positive values indicating increases with engagementterm. The mean modulation indexterm·def was significantly above zero for target-preferring neuronsterm in both V1term (0.227 ± 0.009, n = 1148; p < 5.0 × 10 −4 , permutation testterm·stat, 2000 permutations) and PPCterm (0.351 ± 0.004, n = 3292; p < 5.0 × 10 −4 ). Nontarget-preferring neuronsterm in V1term showed weaker modulation (modulation indexterm·def: 0.059 ± 0.010 for non-target, n = 767) than target-preferring neuronsterm ( p = 9.07 × 10 −4 , clustered Wilcoxon rank-sum testterm·stat), indicating that the effect of engagementterm was stimulus-specific.

Comparing responses during engaged behavior and passiveterm viewing therefore revealed different response properties in V1term and PPCterm. In V1term, task engagementterm·var did not merely increase overall responsiveness (as would be expected with arousal), but instead modulated firing rates to enhance the contrastterm between targetterm and non-target stimuli. By contrastterm, PPCterm responses were strongly target-selective and gated by behavior (Fig. 2 ). Furthermore, because this activity is selective to targetterm trials in which the animal licksterm, PPCterm responses likely signal some aspect of the animal’s choiceterm (decisionterm formation or motorterm planning), instead of simply reflecting overall increases in arousal or attention during task engagementterm·var. A subset (~30%) of PPCterm neuronsterm, however, do have significant passiveterm responses, and could play a role in sensoryterm processing.

Errorterm trials reveal sensitivity of PPCterm to both stimulusterm and choiceterm How should one interpret this target-selective task engagementterm·var signal in PPCterm? One possibility is that PPCterm encodes a stimulus-specific signal that is perhaps boosted by task engagementterm·var. An alternative is that PPCterm activity reflects movement or action planning, since the animal licksterm on Engaged targetterm (Hit) trials but not on Passiveterm trials or Engaged nontarget (CR) trials.

To help disambiguate these possibilities, we analyzed the responses of V1term and PPCterm neuronsterm on errorterm trials. We compared activity in different behavioral or sensoryterm conditions using single-neuron ROC analysisterm·stat to isolate stimulus-specific signals from those related to the choiceterm (Supplementary Table 1 ). Due to the difficulty in distinguishing between motorterm preparation and decision-related signals with our go/no-go paradigm, we define “choice” selectivityterm·def broadly as any premotor signal related to behavioral output independent of the stimulusterm. We focused on target-preferring neuronsterm, as these represented the vast majority of responses in PPCterm, and restricted analysis to those imaging fields and behavioral sessions in which the mouse committed at least five False Alarm (FA) trials (V1term, n = 1053 cells from 16 fields; PPCterm, n = 3034 cells from 21 fields).

For V1term, individual target-preferring neuronsterm varied in their responsiveness on nontarget trials (Fig. 3a , left), but the average strength of the nontarget response was similar on FA trials compared to CR trials (Fig. 3b , left). By contrastterm, PPCterm neuronsterm invariably showed stronger responses on FA vs. CR trials (Fig. 3a, b , right), indicating the presence of choice-related signals. This modulation was strongest during the response period, though still apparent in the stimulusterm period before any licksterm or punishment occurred. We quantified the selectivityterm of each neuronterm’s stimulus-period response using an ROC-based index to compare trial types (Fig. 3c ). Target-preferring V1term neuronsterm exhibited strong selectivityterm for Hit compared to either CR or FA trials (Hit vs. FA, 0.605 ± 0.009, p = 1.3 × 10 −4 , clustered Wilcoxon signed-rank testterm·stat), but weak modulation for FA vs. CR trials (0.014 ± 0.006, p = 0.462). On the other hand, PPCterm neuronsterm showed significant selectivityterm for FA vs. CR trials (0.109 ± 0.003, p = 0.014), as well as selectivityterm for Hit compared to FA trials (0.348 ± 0.004, p = 6.1 × 10 −5 ).

Fig 3.

Open in a new tab Errorterm trials reveal sensitivity of PPCterm to both stimulusterm and choiceterm. Only Hit-preferring neuronsterm were used in all panels of this figure. a Trial-averaged responses during False Alarm (FA) trials, for recordings with at least five FA trials, in V1term (left; 1053 neuronsterm across 16 sessions) and in PPCterm (right; 3034 neuronsterm across 21 sessions). Responses for each neuronterm are normalized by peak response across Correct Reject (CR, left, blue), FA (center, dark blue), and Hit trials (right, red). Neuronsterm are sorted by descending strength of FA response. Vertical dashed lines demarcate duration of stimulusterm. b Populationterm trial-averaged responses (top) during Hit, FA (dashed), and CR trials, for V1term (left) and PPCterm (right). Average lickterm rate (bottom). c Violin plots (rotated histograms) of selectivity indexterm·def for V1term (left) and PPCterm (right) neuronsterm, computed using stimulus-period responses. Three comparisons were made for each neuronterm in each area (Hit > CR, Hit > FA, FA > CR, see Supplementary Table 1 ). Horizontal lines indicate mean, and colored bars indicate neuronsterm with significant individual modulation ( p < 0.05). d Average selectivityterm between trial types, computed separately at each time-point. Colored bars at top indicate time points with significant non-zero selectivityterm ( p < 0.05, clustered Wilcoxon signed-rank testterm·stat). e Percentage of neuronsterm with significant ( p < 0.05) positive selectivityterm at each time-point. Shading indicates mean ± SEM across sessions. Dotted line indicates the expected percentage by chance We also performed a time-dependent ROC analysisterm·stat 3 , 18 , 19 to examine the dynamics of stimulus- and choice-related signals over the course of a trial. We quantified both the average selectivityterm (Fig. 3d ) as well as the fraction of selective neuronsterm (Fig. 3e ) for each area as a function of time. While both areas exhibited significant stimulus-related selectivityterm (Hit vs. FA) shortly after stimulusterm onset ( p < 0.05 starting at t = 0.2 s for both areas, clustered Wilcoxon signed-rank testterm·stat), only PPCterm exhibited any significant choice-related selectivityterm (FA vs. CR) in the stimulusterm period. The onset of choiceterm selectivityterm was delayed ( p < 0.05 starting at t = 0.8 s) compared to the stimulusterm signals and peaked during the response period. By contrastterm, choice-related selectivityterm in V1term did not become significant until the response period ( p < 0.05 starting at 2.4 s).

These analyses suggest that PPCterm, unlike V1term, is sensitive to both the stimulusterm and the impending choiceterm of the animal. However, the comparison of Hit and FA trials to compute stimulusterm selectivityterm may be confounded by overt differences in motorterm output (see lickterm rate in Fig. 3b ) or by covert differences in motorterm preparation. Could PPCterm activity be entirely explained by motorterm or choice-related signals?

While we cannot rule out covert preparatory effects, in a subset of sessions we eliminated differences in motorterm output by selecting Hit and FA trials with the same number of licksterm·var (Supplementary Fig. 2a, b ). PPCterm responses in these sessions ( n = 846 cells across 3 sessions) still exhibited stimulusterm selectivityterm (0.229 ± 0.004), as well as some motor-related sensitivity (0.130 ± 0.009) when comparing FA trials with different numbers of licksterm (Supplementary Fig. 2 c-e). We also analyzed the responses of neuronsterm on Miss trials, in which there is no motorterm output and presumably no motorterm preparation, using the subset of sessions with sufficient number of trials ( n = 1165 cells across 8 sessions for PPCterm). Interestingly, we found that a large fraction of PPCterm neuronsterm exhibited strong responses on Miss trials that were never apparent on CR trials (Supplementary Fig. 3 ; Miss vs. CR: 0.305 ± 0.010) or even on passively-presented Targetterm trials. Together, these data argue that PPCterm responses cannot be parsimoniously explained as encoding purely motor- related signals.

GLMterm·mth reveals multiplexed sensoryterm and motorterm signals in PPCterm We further used a generalized linear model (GLMterm·mth) 20 – 22 to disambiguate the contributions of stimulusterm, task engagementterm·var, and motorterm action on single neuronterm calcium responses across both Passiveterm and Engaged conditions. Responses of each V1term and PPCterm cell were modeled (see “Methods” section) as a linear combination of components that were time-locked to stimulusterm presentation or licking onset (Fig. 4 ). We quantified the performance of the GLMterm·mth as the proportion of variance explained (R 2 ) for a separate test dataset not used for fitting (Fig. 4e ). Using all significantly task-responsive neuronsterm, the average model performance was higher for V1term (0.334 ± 0.005, n = 1915 cells) than for PPCterm (0.190 ± 0.003, n = 3524; V1term vs. PPCterm, p = 4.01 × 10 −5 , clustered Wilcoxon rank-sum testterm·stat).

Fig 4.

Open in a new tab Generalized linear model (GLMterm·mth) of V1term and PPCterm responses. A generalized linear model was fit to training data to predict the calcium responses of individual V1term and PPCterm neuronsterm using lagged predictors for stimulusterm, engagementterm and motorterm action (licking). a GLMterm·mth model performance for an example V1term neuronterm. GLMterm·mth fits on Targetterm (red) and Non-target (blue) trials were compared with trial-averaged test data (black) across Passiveterm (PassNT, PassT), Engaged No-Lick (Miss, FA), and Engaged Lickterm (FA, Hit) conditions. GLMterm·mth components related to Stimulusterm (purple), Engagementterm (gray), and Motorterm (green) were summed to generate overall GLMterm·mth fit. Shaded region represents time of stimulusterm presentation. b Same as a , but for a different, nontarget-preferring V1term neuronterm. c , d GLMterm·mth model performance for two example PPCterm neuronsterm. One neuronterm has a strong Engagementterm component ( c ), while the other has a stronger Motorterm component ( d ). e Cumulative histogram of model prediction performance on test trials, quantified for each cell as fraction of variance explained ( R 2 ), for both V1term (blue, n = 1915 cells) and PPCterm (red, n = 3524). f Population-averaged GLMterm·mth model components related to Stimulusterm (left), Engagementterm (middle), and Motorterm (right), for both V1term (blue) and PPCterm (red). For Stimulusterm and Engagementterm components, the preferred stimulusterm component for each cell was used for averaging. Calcium responses were z -scored before model fitting. Light shaded gray region demarcates duration of stimulusterm We then assessed the relative contribution of stimulusterm, engagementterm, and motorterm model components for cells in each area, and illustrate the performance of the model on a few examples (Fig. 4a–d ). Most V1term cells exhibited a strong stimulusterm component (0.494 ± 0.013, component of z -scored calcium response, see “Methods” section), which reflected sensoryterm drive on Passiveterm trials, whether the neuronterm preferred the targetterm (Fig. 4a ) or the nontarget (Fig. 4b ) stimulusterm. By contrastterm, PPCterm neuronsterm had much weaker stimulusterm components (Fig. 4f , left; 0.117 ± 0.013; V1term vs. PPCterm, p = 6.84 × 10 −5 , clustered Wilcoxon rank-sum testterm·stat). The engagementterm component of the GLM-reflected stimulus-specific signals that occurred exclusively on Engaged trials. This could be disambiguated from the stimulus-independent motorterm component, which would be present on both Hit and False Alarm (FA) trials. Individual PPCterm neuronsterm varied in the relative contribution of engagementterm and motorterm components (Fig. 4c, d ), but on average PPCterm cells exhibited a slightly (but not significantly) stronger engagementterm component (Fig. 4f , middle; V1term, 0.183 ± 0.009; PPCterm, 0.233 ± 0.006; V1term vs. PPCterm, p = 0.144, clustered Wilcoxon rank-sum testterm·stat) and a much stronger motorterm component (Fig. 4f , right; V1term, 0.155 ± 0.005; PPCterm, 0.371 ± 0.005; V1term vs. PPCterm, p = 0.015, clustered Wilcoxon rank-sum testterm·stat) as compared to V1term cells.

Linear models that consider the stimulusterm, taskterm context, and licking together can explain some of the variance in calcium responses. But could even simpler models be sufficient? For each cell, we trained three additional partial models: a stimulus-only model, a motor-only model, and a stimulusterm + engagementterm model that was insensitive to the motorterm response (Supplementary Fig. 4 ). For V1term cells, the stimulus-only model performed nearly as well as the full model (median relative R 2 : 0.905), but the motor-only model performed quite poorly (median relative R 2 : 0.156). For the majority of V1term neuronsterm, the stimulusterm + engagementterm model (which excluded licking terms) performed just as well as the full model (median relative R 2 : 0.995; only 15.9% ± 5.3% of cells with worse fit). By contrastterm, while PPCterm neuronsterm were better explained by the motor-only model (median relative R 2 : 0.809) than by the stimulus-only model (median relative R 2 : 0.353), both models performed significantly worse than the full model for the majority of neuronsterm (81.9% ± 6.2% of cells for stim-only, 55.9% ± 11.3% for motor-only). PPCterm neuronsterm were reasonably well-explained by a stimulusterm + engagementterm model (median relative R 2 : 0.910) but a large proportion of cells still exhibited worse fits compared to the full model (34.1 ± 10.7% of cells with worse fit). These results add further evidence that PPCterm encodes a combination of sensoryterm and motorterm signals.

PPCterm reflects both stimulus contrastterm·var and behavioral state Previous work has shown that in both primates 1 , 2 and rats 23 , neuronsterm in PPCterm encode not only the choiceterm of the animal but also the amount of sensoryterm evidence for that decisionterm. To test whether neuronsterm in mouse PPCterm similarly reflected the decisionterm process, we varied the amount of sensoryterm evidence from trial-to-trial by manipulating stimulus contrastterm·var. We also compared responses during engaged and passiveterm conditions to examine how sensoryterm and motorterm signals may be combined in PPCterm responses.

A subset of the mice ( n = 6) were trained to perform a variant of the discrimination taskterm, in which the contrastterm of the grating stimulusterm varied randomly from trial-to-trial (Fig. 5a ). Mice performed well above chance, even at very low contraststerm ( d' at 2% contrastterm, 1.05 ± 0.25; p = 0.031, Wilcoxon signed-rank testterm·stat), although performance degraded as contrastterm was lowered ( p = 1.10 × 10 −4 , Friedman test) reflecting the decrease in the strength of sensoryterm evidence (Fig. 5b ). We imaged from neuronsterm in V1term ( n = 8 fields, 3 mice) and PPCterm ( n = 11 fields, 6 mice) and computed contrast-response functions for both passiveterm and engaged conditions. We again focused our analysis on target-preferring neuronsterm (V1term, n = 250; PPCterm, n = 611), which constituted the vast majority of task-responsive PPCterm neuronsterm (95% in this dataset). Neuronsterm were included for further analysis if they demonstrated significant Hit responses at multiple contraststerm that could be well fit with a hyperbolic ratio functionterm·def 24 (see “Methods” section).

Fig 5.

Open in a new tab PPCterm reflects both stimulus contrastterm·var and behavioral state. a Mice discriminated between orthogonally oriented Targetterm and Non-target stimuli with contrastterm varying from 2 to 64%. b Behavioral performance (d-primeterm·def) on the variable contrastterm discrimination taskterm, averaged across 19 sessions from 6 mice. Errorterm bars indicate SEM. c Populationterm trial-averaged response of all target-preferring V1term neuronsterm ( n = 250) across contraststerm during correct Engaged (left) and Passiveterm (right) targetterm (red) and nontarget (blue) trials. Responses are averaged across low (2 or 4%, light shade), medium (8 or 16%, medium shade), and high (32 or 64%, dark shade) contrastterm. Light gray shaded regions demarcate duration of stimulusterm. d Same as c but for PPCterm ( n = 611 target-preferring neuronsterm). e , f Histograms comparing engagementterm modulation in V1term ( e ) and PPCterm ( f ), computed by comparing responses on Engaged vs. Passiveterm high contrastterm trials. Colored bars indicate neuronsterm with significant modulation. g , h Histograms comparing contrastterm modulation in V1term ( g ) and PPCterm ( h ), computed by comparing responses on high vs. low contrastterm Engaged trials. Colored bars indicate neuronsterm with significant modulation As in the single-contrast taskterm, the populationterm response in V1term during the variable contrastterm taskterm was robust in both Engaged and Passiveterm conditions (Fig. 5c ), with most neuronsterm showing a significant response to at least one contrastterm during both conditions (68.9% ± 11.8% of cells). Conversely, PPCterm populationterm activity was more robust in Engaged vs. Passiveterm conditions (Fig. 5d ), with only a subset of PPCterm neuronsterm showing significant passiveterm responses (27.8% ± 14.9%; V1term vs. PPCterm, p = 7.59 × 10 −3 , Wilcoxon rank-sum testterm·stat). The modulation of PPCterm activity by contrastterm cannot be explained by changes in motorterm behavior, as lickterm rate was unchanged as a function of contrastterm (Supplementary Fig. 5a ).

We quantified the strength of modulation by contrastterm and engagementterm for each neuronterm using ROC-based indices that ranged from −1 to 1 (Fig. 5e–h ). Engagement modulation indexterm·var was similarly high in V1term and PPCterm (Fig. 5e, f ; V1term, 0.149 ± 0.019; PPCterm, 0.222 ± 0.008; V1term vs. PPCterm, p = 0.209, clustered Wilcoxon rank-sum testterm·stat), but modulation by contrastterm was stronger on average in V1term compared to PPCterm (Fig. 5g, h ; V1term, 0.487 ± 0.027; PPCterm, 0.233 ± 0.009; V1term vs. PPCterm, p = 0.021, clustered Wilcoxon rank-sum testterm·stat). Nonetheless, the mean contrast modulation indexterm·var in PPCterm was significantly greater than zero ( p = 4.86 × 10 −3 , clustered Wilcoxon signed-rank testterm·stat).

We also analyzed the errorterm trials to see whether stimulusterm and choiceterm signals could be separately extracted from responses in PPCterm, and whether these signals depended on contrastterm (Supplementary Fig. 5 ). We compared False Alarm trials with Hit and Correct Reject trials, using imaging fields and sessions with at least five False Alarm trials at each contrastterm (7 of 11 fields, n = 391 neuronsterm). We did not make comparisons with Miss trials given the low number of trials. The populationterm response on FA trials was weak, but distinguishable from the response on CR trials (Supplementary Fig. 5a ), especially during the response period (Supplementary Fig. 5b ; FA greater than CR for all contraststerm, p < 0.05, clustered Wilcoxon signed-rank testterm·stat). Using an ROC-based approach, we again found that PPCterm encoded both stimulusterm and choiceterm with differing time courses (Supplementary Fig. 5e ; stimulusterm selectivityterm significant from 1.0 to 6.6 s for all contraststerm, choiceterm selectivityterm significant from 2.2 to 5.6 s for all contraststerm, clustered Wilcoxon signed-rank testterm·stat). Interestingly, some PPCterm cells encoded the stimulusterm in a contrast-dependent manner, but also encoded the choiceterm in a contrast-independent manner (Supplementary Fig. 5c ). We quantified the contrast-dependence of the selectivity indexterm·def for each neuronterm by measuring its slope as a function of contrastterm (Supplementary Fig. 5d ). A larger proportion of neuronsterm exhibited significant contrast-dependence in stimulusterm selectivityterm (Supplementary Fig. 5 f, 23.6% ± 19.9% of cells) compared to the proportion with significant contrast-dependence in choiceterm selectivityterm (2.4% ± 1.6%). PPCterm may therefore simultaneously encode both contrast-dependent sensoryterm signals and contrast-independent motorterm signals in the same populationterm of neuronsterm.

Heterogeneous PPCterm responses to contrastterm and engagementterm PPCterm is on average modulated by both engagementterm and stimulus contrastterm·var, but closer examination of the individual PPCterm contrast-response functions revealed a great deal of heterogeneity. We therefore divided neuronsterm into groups based on whether they exhibited significant modulation by contrastterm and/or engagementterm (Fig. 6a, b ). A subset of PPCterm neuronsterm (21.2% ± 13.5% of cells) showed strong modulation by contrastterm, but very weak modulation by engagementterm (Fig. 6c, d , left column). Such neuronsterm therefore faithfully represented the sensoryterm stimulusterm regardless of behavioral state. Conversely, a larger group of PPCterm neuronsterm (28.1% ± 12.3%) were gated by task engagementterm·var but showed little to no modulation with contrastterm (Fig. 6c, d , middle column). These neuronsterm reflected the behavioral state and impending action of the animal irrespective of sensoryterm drive. Lastly, a third subset of PPCterm neuronsterm (28.2% ± 10.6%) were significantly modulated by both contrastterm and engagementterm (Fig. 6c, d , right column). PPCterm therefore appears to contain both contrast-modulated “sensoryterm” neuronsterm as well as engagement-modulated “motorterm” neuronsterm. This differs qualitatively from V1term (Fig. 6a , left; p = 7.10 × 10 −17 , χ 2 test of independence, see also Supplementary Fig. 6 ), where most neuronsterm (74.8% ± 7.0%) are modulated by contrastterm, and much fewer by engagementterm alone (9.2% ± 7.1%).

Fig 6.

Open in a new tab PPCterm neuronsterm are heterogeneous in their sensitivity to contrastterm and engagementterm. a Venn diagrams indicating proportions of target-preferring neuronsterm in V1term and PPCterm with significantly positive modulation by contrastterm alone (cyan), engagementterm alone (yellow), or both contrastterm and engagementterm (green). b Scatter plot and histograms of contrastterm modulation vs. engagementterm modulation for all target-preferring PPCterm neuronsterm ( n = 611) imaged during the variable contrastterm taskterm. Colored bars in histograms (same as Fig. 5f, h ) indicate neuronsterm with significantly positive modulation by contrastterm (cyan) or engagementterm (yellow). Colored dots on scatter plot demarcate neuronsterm with significantly positive modulation by contrastterm alone (cyan), engagementterm alone (yellow), or both contrastterm and engagementterm (green). Individual examples in c are marked with the corresponding number. c Trial-averaged responses (top row) and contrast-response functions (middle and bottom rows) of example PPCterm neuronsterm that were significantly modulated by contrastterm (left column), engagementterm (middle column), or both (right column). Contrast-response functions were computed from responses during the stimulus-period (gray shaded region in top row). Modulation indexterm·def values for each example can be found by referring to b . Errorterm bars indicate SEM across trials. d Group-averaged contrast-response functions. Percentages indicates proportion of PPCterm neuronsterm within each group. Errorterm bars indicate SEM across cells. e Left, approximate location of imaging fields within PPCterm (blue). Schematic modified from the Allen Mouse Brain Atlas ( http://mouse.brain-map.org/static/brainexplorer ). Right, relative spatial location of neuronsterm with modulation by contrastterm, engagementterm, or both, in two example imaging sessions. Spatial locations of cells from 4 planes are collapsed into one. Scale bar, 100 µm. f Average within-group and across-group distances for each cell, colored by group. No significant difference was found for any individual group or for the whole populationterm ( p = 0.102, clustered Wilcoxon signed-rank testterm·stat). g Difference in engagementterm (yellow) and contrastterm (cyan) modulation indices as a function of distance between cells. Errorterm bars indicate SEM across cells. Distances beyond 400 µm are not shown. No significant correlation between modulation indexterm·def and distance was found We then considered whether there was any anatomical organization of these functional properties within PPCterm. For each imaged PPCterm volume, functional subpopulations of contrast-modulated and engagement-modulated cells appeared to be intermingled across space (Fig. 6e ). We computed the pairwise distance between all target-preferring neuronsterm, and found no significant difference between within-group and across-group distances (Fig. 6f ; within-group, 268 ± 5 µm; across-group 273 ± 5 µm; p = 0.102, clustered Wilcoxon signed-rank testterm·stat). We also compared the functional properties of pairs of neuronsterm as a function of distance, and found that the difference in contrastterm modulation or engagementterm modulation did not depend on distance (Fig. 6g ; Pearson’s correlation with distance, difference in contrast modulation indexterm·var, r = 0.003, p = 0.660; difference in engagement modulation indexterm·var, r = 0.001, p = 0.960). Therefore, PPCterm contains neuronsterm with diverse visuomotor response properties that are spatially intermingled.

PPCterm reflects choiceterm independent of stimulus-reward contingencyterm Most PPCterm neuronsterm respond exclusively during targetterm trials, while a smaller subset of neuronsterm encode the sensoryterm stimulusterm in a contrast-dependent manner. Errorterm and GLMterm·mth analyses suggest that these PPCterm populationsterm encode choiceterm and stimulusterm, respectively. However, because errorsterm can reflect other factors such as impulsivity or inattention and thus be difficult to interpret, a more conclusive demonstration would require a clear dissociation of stimulusterm and choiceterm. We therefore manipulated the stimulus-reward structure of the taskterm and re-trained mice on a reversed sensorimotor contingencyterm. By measuring activity from the same cells before and after reversal, we could test whether individual V1term and PPCterm neuronsterm were more sensitive to stimulusterm identity or to the animal’s choiceterm.

After imaging the responses of neuronsterm in V1term and PPCterm in the original go/no-go taskterm, we reversed the rewardterm contingencies of the stimuli (Fig. 7a ). Licking in response to the original non-target stimulusterm (Stimulusterm B) was now rewarded with water, whereas licking to Stimulusterm A was punished with quinine. Three mice successfully learned the taskterm after 7–11 days of training, although performance was slightly worse than before (Fig. 7b ; d-primeterm·def, original, 2.96 ± 0.50; reversed, 1.51 ± 0.22). We then measured responses from the same populationsterm of neuronsterm in V1term and in PPCterm under the reversed reward contingencyterm·var, and identified the same neuronsterm across imaging sessions using a semi-automated procedure 25 (see "Methods" section for details).

Fig 7.

Open in a new tab PPCterm reflects choiceterm independent of stimulus-reward contingencyterm. a Mice were trained on a go/no-go discrimination taskterm with a reversed reward contingencyterm·var. b Behavioral performance before and after reversal. Left, response rate for Stimulusterm A (red) and Stimulusterm B (blue). Right, d-primeterm·def assuming Stimulusterm A as targetterm. Performance after reversal was significantly different from zero ( p < 0.001, Wilcoxon signed-rank testterm·stat). c Neural response of two example V1term neuronsterm before and after reversal of reward contingencyterm·var. Colors indicate stimulusterm (red, A; blue, B) and line styles indicate whether the stimulusterm is targetterm (solid) or non-target (dashed). Top row, response of a neuronterm selective to Stimulusterm A both before (left) and after (right) reversal. Bottom row, response of a neuronterm selective to Stimulusterm B. Insets indicate the average projection image before and after reversal, and ROI (dashed red line) that was used to extract the neural response. d Same as c but for two PPCterm neuronsterm. These neuronsterm show a switch in selectivityterm with reversal of reward contingencyterm·var. e Histogram of selectivity indexterm·def before and after reversal for V1term neuronsterm with significant responses both before and after reversal ( n = 488). Positive values indicate selectivityterm to the original Targetterm, Stimulusterm A (red). Colored bars indicated neuronsterm with significant selectivityterm. f Same as e but for PPCterm neuronsterm ( n = 509). g Scatter plot comparing selectivity indexterm·def in V1term before and after reversal. Colored points indicate neuronsterm with significant selectivityterm both before and after reversal. Neuronsterm in the first and third quadrants are stimulus-selective, as they prefer Stimulusterm A (red) or Stimulusterm B (blue) both before and after reversal. Neuronsterm in the second and third quadrants are choice-selective, as they prefer either the Go stimulusterm (green) or the No-Go stimulusterm (yellow) in both conditions. h Same as g , but for PPCterm We analyzed the trial selectivityterm of individual neuronsterm that had significant responses both before and after reversal (V1term, n = 488 cells, 8 fields in 3 mice; PPCterm, n = 509 cells, 8 fields in 3 mice). Many neuronsterm in V1term that were selective to a particular stimulusterm remained selective to the same stimulusterm after reversal, whether Stimulusterm A or Stimulusterm B (Fig. 7c ). By contrastterm, many PPCterm neuronsterm did not exhibit stable stimulusterm selectivityterm, but instead appeared to track the animal’s choiceterm. These neuronsterm preferred targetterm stimulusterm A before reversal while preferring new targetterm stimulusterm B after reversal (Fig. 7d ). This strongly suggests that PPCterm neuronsterm encode choice-related signals related to decisionterm formation or motorterm planning.

Looking at the whole populationterm, we measured the selectivityterm of neuronsterm to Stimulusterm A trials before and after reversal. Stable selectivityterm would indicate sensoryterm signals, whereas changes in this measurement would reflect signals related to choiceterm. We found that V1term selectivityterm did not change significantly after reversal (Fig. 7e , before, 0.161 ± 0.034; after, 0.095 ± 0.030; p = 0.256, clustered Wilcoxon signed-rank testterm·stat), with a relatively symmetrical distribution of neuronsterm selective to either stimulusterm A or B. However, in PPCterm, this selectivityterm measure was dramatically altered with reversal of reward contingencyterm·var (before, 0.513 ± 0.013; after, −0.373 ± 0.013; p = 4.15 × 10 −3 , clustered Wilcoxon signed-rank testterm·stat), with the majority of responsive neuronsterm preferring the new targetterm stimulusterm B after reversal.

We also compared the selectivityterm of individual neuronsterm before and after reversal using a scatter plot (Fig. 7 g, h). Purely stimulus-selective neuronsterm will remain close to the unity line and in the first (bottom-left) and third (top-right) quadrants, whereas choice-selective neuronsterm will lie in either the second quadrant (top-left; for no-go selective cells) or the fourth quadrant (bottom-right; for go-selective cells). V1term neuronsterm were strongly stimulus-selective, with the majority (74.3% ± 5.2% of cells, Fig. 8a , top) of neuronsterm lying within the first and third quadrants (Fig. 7g ). By contrastterm, the majority (66.8% ± 13.0% of cells, Fig. 8a , bottom) of PPCterm neuronsterm were found in the fourth quadrant, indicating a preference for the rewarded targetterm stimulusterm, regardless of its actual identity (Fig. 7h ). These results indicate that the majority of PPCterm neuronsterm reflected the choiceterm of the animal.

Fig 8.

Open in a new tab PPCterm cells with reversed selectivityterm exhibit stronger modulation by engagementterm and by errorterm trials. a Proportion of significantly responsive V1term (top) and PPCterm (bottom) neuronsterm with stable stimulusterm selectivityterm (Stim A, 0°; or Stim B, 90°) or reversed selectivityterm (Go or No-Go). Cells with non-significant selectivityterm either before or after reversal are categorized as non-selective (NS). b Boxplots (with center line as median, box as interquartile range, and whiskers as 1.5 × IQR) of engagement modulation indexterm·var for each response group, computed as the mean before and after reversal. “Go”-selective neuronsterm (green) show stronger engagementterm modulation, i.e., weaker passiveterm responses, than stimulus-selective neuronsterm (red or blue). c Boxplots of errorterm modulation indexterm·def for each response group, computed by comparing False Alarm trials with Correct Reject trials both before and after reversal. “Go”-selective PPCterm neuronsterm show stronger errorterm modulation (green) than stimulus-selective neuronsterm (red or blue). d Responses of two PPCterm neuronsterm with stable selectivityterm to Stimulusterm A both before and after reversal of reward contingencyterm·var. Colors indicate stimulusterm (red, A; blue, B) and dark, dashed lines indicate passiveterm trials. These neuronsterm show a robust passiveterm response in both conditions, and have a low engagement modulation indexterm·var. e Same as d but for two PPCterm neuronsterm with reversed selectivityterm after reversal of reward contingencyterm·var. These neuronsterm show virtually no passiveterm response in both conditions, and have a high engagement modulation indexterm·var Taskterm contingencyterm reversal reveals distinct PPCterm subpopulations A small subset of PPCterm neuronsterm (6.9% ± 3.7%) did show stable stimulusterm selectivityterm before and after reversal. If PPCterm includes distinct populationsterm of “stimulusterm” neuronsterm and “choiceterm” neuronsterm, we hypothesized that they could be distinguished based on their responses on passiveterm and errorterm trials. Stimulusterm selectivityterm after reversal should be predictive of modulation strength with engagementterm and choiceterm: with “stimulusterm” neuronsterm showing robust passiveterm responses and weak modulation by engagementterm and errorsterm, and “choiceterm” neuronsterm expressing strong engagementterm and errorterm modulation.

To test this hypothesis, we measured the strength of engagementterm modulation in both stimulus-selective (to either Stimulusterm A or B) neuronsterm and go-selective neuronsterm, as defined by their selectivityterm before and after reversal (Fig. 8a ). We found that go-selective PPCterm neuronsterm had higher (near significant) modulation by engagementterm than stimulus-selective neuronsterm (go-selective, 0.393 ± 0.007; stimulus-selective, 0.084 ± 0.040; p = 0.065, clustered Wilcoxon rank-sum testterm·stat; Fig. 8b , bottom). Indeed, many PPCterm neuronsterm that maintained selectivityterm to Stimulusterm A after reversal also had strong passiveterm responses (Fig. 8d ), and many neuronsterm with reversed selectivityterm were strongly gated by engagementterm (Fig. 8e ). We also tested to see whether a PPCterm neuronterm’s responses on errorterm trials related to its selectivityterm after reversal. Go-selective PPCterm neuronsterm had significantly higher errorterm modulation (computed by comparing False Alarm trials to Correct Reject trials) compared to stimulus-selective neuronsterm (go-selective, 0.273 ± 0.009; stimulus-selective, 0.019 ± 0.019; p = 0.040, clustered Wilcoxon rank-sum testterm·stat; Fig. 8c ). Interestingly, the small subset of go-selective neuronsterm in V1term (5.9% ± 2.1% of cells) also differed from stimulus-selective V1term neuronsterm in their degree of engagementterm (Fig. 8b , top, p = 0.016) and errorterm (Fig. 8c , top, p = 0.021) modulation, suggesting that these different properties are strongly related and can be used to distinguish functional cell types across different areas.

These findings demonstrate distinct subsets of neuronsterm within PPCterm. One subset of neuronsterm faithfully reflects the sensoryterm input, both in passiveterm conditions and after learning a new reward contingencyterm·var. The larger proportion of PPCterm neuronsterm, however, track the choiceterm before and after re-training, and are strongly modulated by task engagementterm·var. The flexibility and heterogeneity of PPCterm responses suggests a role for PPCterm in the mapping of sensoryterm inputs onto appropriate motorterm actions.

Discussion We developed a simple head-fixed visualterm decisionterm taskterm for mice with separate stimulusterm and response epochs, and used populationterm imaging of single neuronterm responses to investigate the role of PPCterm in perceptual decisionsterm. Our key findings are that PPCterm encodes both sensoryterm and motorterm signals across a heterogeneous pool of neuronsterm, and that its activity reflects taskterm performance and demands. Together these results suggest that mouse PPCterm is responsible for neither pure sensoryterm processing, nor for the control of motorterm output, but rather is important for the decisionterm process itself – the process of mapping sensation to action.

The small size of the mouse brain has made it difficult to identify precise borders between different functional areas. PPCterm in the rodent, as classically defined by its thalamic inputs 26 , is located in the region between anterior to V1term and posterior to somatosensory cortex. However, this location is essentially where both anatomical 11 , 12 and functional 13 , 14 mapping studies have identified the retinotopically-organized areas RL, A, and AM. The degree to which PPCterm overlaps with these secondary visualterm areas is a matter of debate. Some have argued that rodent PPCterm may have more in common with primate extrastriate cortex, given that inactivation specifically disrupts sensitivity on visualterm decisionsterm 10 . Indeed, the stereotaxic coordinates used by us 3 and others 4 , 27 to targetterm mouse PPCterm most directly overlap with area AM, which exhibits directionally-tuned responses even under anesthesia 14 . But is PPCterm simply a sensoryterm visualterm area?

We have presented multiple pieces of evidence that point to a role for PPCterm beyond mere sensoryterm processing. First, activity in PPCterm is strongly dependent on behavioral state, with only a minority of task-responsive PPCterm neuronsterm exhibiting significant responses during passiveterm viewing of stimuli (Fig. 2f ). Secondly, the selectivityterm of PPCterm neuronsterm is strongly biased toward targetterm stimuli (Fig. 2e ). This bias is likely due to the asymmetry of the Go/No-go paradigm, and may reflect a learned association of stimulusterm and rewardterm 28 . Third, PPCterm responses are modulated during errorterm trials (Fig. 3 , Supplementary Fig. 5 ). Information about the eventual choiceterm of the animal can be decoded from the activity of PPCterm, as previously shown in both mice 3 , 4 and rats 5 . Finally, and most conclusively, the biased selectivityterm of most PPCterm neuronsterm towards targetterm stimuli is dramatically reversed when the animal is retrained on a different reward contingencyterm·var (Fig. 7 ). Together, these results demonstrate that the stimulus-period responses in PPCterm are task-dependent and not purely sensoryterm.

Alternatively, one may argue that the activity patterns we observed in PPCterm can be most parsimoniously explained as movement or motor-planning related signals. After all, movement-related activity would be present only during engagementterm, it would exhibit choiceterm selectivityterm, and it would change with reward contingencyterm·var. It is unlikely that PPCterm is directly involved in executing motorterm plans, as we have previously shown that optogenetic inactivation of PPCterm during the response period, or even during the delay between stimulusterm and response, has no effect on behavior 3 . However, the possibility remains that the PPCterm responses recorded in our taskterm reflect planning- or movement-related signals that originate elsewhere. Although some PPCterm neuronsterm (~30%) do show activity that appears motor-related, due to their contrast-independent nature (Fig. 6 ), we provide evidence that PPCterm is not a purely motorterm (or motor-planning) area either.

First, passiveterm visualterm stimulation does induce a response in some PPCterm neuronsterm (~20%), as previously shown in parietal area AM of anesthetized mice 14 . This subset of neuronsterm tends to stably reflect the stimulusterm even with changes in reward contingencyterm·var (Fig. 8 ). Second, we find a subset of PPCterm neuronsterm that reflect the targetterm stimulusterm on Miss trials when the animal fails to lickterm, even though such neuronsterm are inactive during Correct Reject trials (Supplementary Fig. 3 ). Finally, the responses of many PPCterm neuronsterm (~40%) shows modulation with stimulus contrastterm·var, even for the same decisionterm and motorterm output (Fig. 6 ). This is reminiscent of previous primate 2 and rodent 23 PPCterm studies, where it has been shown that responses vary with the strength of incoming sensoryterm evidence. These findings cohere with prior work in a delayed-response version of our taskterm 3 , in which choice-related selectivityterm was highest shortly after stimulusterm presentation and became weaker over the course of the trial. This argues against a motor-planning explanation which would instead predict increased choiceterm coding with time.

In every taskterm condition reported here, PPCterm responses were heterogeneous. A subset of PPCterm neuronsterm have significant passiveterm visualterm responses (~30%, Fig. 2f , E + P) and are modulated by contrastterm in both engaged and passiveterm conditions (~20%, Fig. 6a , contrastterm). These neuronsterm maintain their stimulusterm selectivityterm even after reversal of reward contingencyterm·var (~10%, Fig. 8a , Stim). These “stimulusterm” neuronsterm are spatially intermingled with the larger proportion of “choiceterm” neuronsterm that have task-gated responses (~70%, Fig. 2f , E only), weak contrastterm modulation (~30%, Fig. 6a , engagementterm), and choice-selective responses after contingencyterm reversal (~70%, Fig. 8a , Go). Furthermore, a third group of neuronsterm have multiplexed selectivityterm to stimulusterm, engagementterm, and motorterm signals, exhibiting complex responses to combinations of sensoryterm or motorterm signals in engaged conditions. These cells exhibit modulation to both contrastterm and engagementterm (~30%, Fig. 6a , both) and their responses cannot be explained by partial GLMterm·mth models that only use a subset of the taskterm variables (~35%, Supplementary Fig. 4f ).

Heterogeneous response properties have been previously reported in both primate 29 , 30 and rodent 5 , 31 PPCterm during decisionterm tasksterm. Our work adds to this literature, and additionally provides evidence that such heterogeneous responses exist in a spatially intermingled fashion within PPCterm of the mouse 32 , 33 . Do these response types form bona-fide cell classes or does PPCterm represent a category-free populationterm, as others have proposed 5 ? Although we do find that various properties (such as modulation by engagementterm, contrastterm, and reversal) are correlated with one another, more work needs to be done to determine whether functional neuronal subgroups are truly separable.

Recent studies in rodents have suggested that some of the heterogeneity observed in PPCterm responses may be due to encoding of information from recent experience, including information about past sensoryterm stimuli 8 , previous choicesterm 9 , 27 , and the presence or absence of rewardterm. Such information about history could bias performance on sensorimotor tasksterm by providing animals with prior expectations about the value of particular stimuli 8 or actions 9 . While our findings do not allow us to directly address history-dependent biases in action selection, they are consistent with a role for PPCterm in selecting future motorterm actions.

Our results build on previous work in the field in several ways: First, we show that PPCterm neuronsterm are actively gated by engagementterm in a sensorimotor taskterm, with a substantial subpopulation of neuronsterm exhibiting sensoryterm responses only during taskterm performance. Second, we used analysis of variable contrastterm stimuli and a GLMterm·mth model to delineate the relative contributions of stimulusterm, motorterm preparation, and engagementterm to the responses of both V1term and PPCterm. These analyses showed that visualterm and motorterm signals are multiplexed within individual neuronsterm. Finally, we showed that reversing rewardterm contingencies causes the response selectivityterm to swap for the vast majority of PPCterm cells, indicating that the multiplexing of stimulusterm and response can be flexibly mapped depending on taskterm contingencies. Taken together, these results bolster the evidence that PPCterm neuronsterm are capable of mediating visuomotor transformations, although further evidence is necessary to establish this possibility conclusively.

The mouse posterior parietal cortex encodes behaviorally-relevant variables in a highly task-dependent manner, in analogy to prior work in primates. Our understanding of how decisionsterm are computed and visuomotor transformations are made will be greatly aided by future circuit-level analyses of PPCterm function in this powerful model system 34 .

Methods Mice and surgery All experiments were carried out in mice of either sex using protocols approved by Massachusetts Institute of Technology’s Animal Care and Use Committee and conformed to National Institutes of Health guidelines. Data were collected from adult (3–5 months old) wild-type (C57BL/6; n = 15) mice of either sex. The animals were housed on a 12-hour light/dark cycle in cages of up to 5 animals before the implants, and individually after the implants. All surgeries were conducted under isoflurane anesthesia (3.5% induction, 1.5–2.5% maintenance). Meloxicam (1 mg kg −1 , subcutaneous) was administered pre-operatively and every 24 h for 3 days to reduce inflammation. Once anesthetized, the scalp overlying the dorsal skull was sanitized and removed. The periosteum was removed with a scalpel and the skull was abraded with a drill burr to improve adhesion of dental acrylic. Stereotaxic coordinates for future viral injections were marked with a non-toxic ink and covered with a layer of silicon elastomer (Kwik-Sil, World Precision Instruments) to prevent acrylic bonding. The entire skull surface was then covered with dental acrylic (C&B-Metabond, Parkell) mixed with black ink to reduce light transmission. A custom-designed stainless-steel head plate (eMachineShop.com) was then affixed using dental acrylic. After head plate implantation, mice recovered for at least 5 days before beginning water restriction.

After behavioral training was complete, animals were taken off water restriction for 5 days before undergoing a second surgery to implant the imaging window. Procedures for anesthetic administration and post-operative care were identical to the first surgery. The dental acrylic and silicon elastomer covering the targeted region were removed using a drill burr. The skull surface was then cleaned and a craniotomy was performed over left V1term/PPCterm, leaving the dura intact. Neuronsterm were labeled with a genetically-encoded calcium indicator by microinjection (Stoelting) of 50 nl AAV2/1.Syn.GCaMP6s.WPRE.SV40 (University of Pennsylvania Vector Core; diluted to a titer of 10 12 genomes ml −1 ) 300 µm below the pial surface. Between two and five injections were made in each exposed region, centered at V1term (4.2 mm posterior, 2.5 mm lateral to Bregma) and PPCterm (2 mm posterior, 1.7 mm lateral to Bregma). Since the viral expression spreads laterally from the injection site, exact stereotaxic locations were photographed through the surgical microscope for determining imaging areas. Finally, a cranial window was implanted over the craniotomy and sealed first with silicon elastomer then with dental acrylic. The cranial windows were made of two rounded pieces of coverglass (Warner Instruments) bonded with optical glue (NOA 61, Norland). The bottom piece was a circular coverglass (4 mm diameter) that fit snugly in the craniotomy. The top piece was a larger circular coverglass (3–5 mm, depending on size of bottom piece) and was bonded to the skull using dental acrylic. Mice recovered for 5 days before commencing water restriction.

Behavioral tasksterm We trained mice to perform a head-fixed go/no-go visualterm discrimination taskterm, similar to previous designs 3 . Stimuli consisted of full-contrast sine wave gratings (spatial frequency: 0.05 cycles deg −1 ; temporal frequency: 2 Hz) drifting at either 0° (upwards, targetterm, Stimulusterm A) or 90° (rightwards, non-target, Stimulusterm B) away from vertical. Stimuli were presented to the right eye alone by placing the screen at an oblique angle to the animal. Behavioral training and testing was implemented with custom software written in Matlab (Mathworks) using Psychtoolbox-3 35 and Data Acquisition toolbox. Spout position was controlled by mounting the spout apparatus on a pneumatically-driven sliding linear actuator (Festo) controlled by two solenoids. Licksterm were detected using an infrared emitter/receiver pair (Digikey) mounted on either side of the retractable lickterm spout. Mice were water-restricted and earned most of their daily ration (1 mL) during training.

An auditory cue tone (5 kHz, 0.5 s, 65 dB SPL) indicated the beginning of each trial. After a 1 s delay, a visualterm stimulusterm was presented for 2 s. At the end of the stimulusterm epoch, the spout was rapidly moved within reach of the tongue, and remained within reach for 1.5 s. Correct licksterm during this period were rewarded with 5–8 µl water and a brief rewardterm tone (10 kHz, 0.1 s). Licksterm to the non-target stimulusterm were punished with a white noise auditory stimulusterm alone (early training) or white noise plus 1–3 µl of 5 mM quinine hydrochloride in water (late training). This concentration was chosen to deter licking to non-targets without causing mice to lose motivation altogether (Supplementary Fig. 1 ). At the end of the response epoch, the spout was then rapidly retracted and remained out of reach until the next trial (3 s inter-trial interval).

Mice were trained in successive stages, as previously described 3 . Briefly, mice were first trained to lickterm a stationary lickterm spout during presentation of the targetterm stimulusterm, and then non-target stimuli were gradually introduced. Spout withdrawal was introduced once mice showed good discrimination performance ( d' > 1 and R HIT − R FA > 30% for consecutive sessions), with the spout initially being extended before stimulusterm onset, but gradually delayed to extend after stimulusterm offset. Once mice reached high levels of performance at the final stage of the taskterm ( d' > 1.5 and R HIT − R FA > 50%), they were removed from water restriction for window implantation. Mice reached criterion performance after an average of 92 ± 11 sessions. After recovery from window implantation surgery, they were re-trained to a level of high performance (2–7 days) before beginning experimental sessions. Any sessions with poor performance were discarded (minimum performance criterion: d' > 1 and R HIT − R FA > 30%).

During imaging experiments, blocks of engaged behavior trials were alternated with blocks of passiveterm viewing. Blocks were 5–10 min in duration (40–80 trials per block). During passiveterm blocks, the spout was out of reach for the duration of the block. A few extra passiveterm trials were given (without imaging) before each passiveterm block to ensure that mice did not expect spout presentation during all imaged passiveterm trials. The sequence of targetterm and non-target stimuli presented for a given passiveterm block was matched to the sequence of stimuliterm·var used for the preceding engaged block. In some cases, instead of alternating between engaged and passiveterm blocks, the engaged blocks were all grouped together at the beginning of the session, followed by an equal number of consecutive passiveterm blocks. No difference in results was observed for alternating vs. grouped blocks.

Some mice ( n = 6) were trained on a variable contrastterm version of the taskterm. On each trial, the stimulusterm was randomly set to one of six contraststerm (2, 4, 8, 16, 32, or 64%), regardless of whether the stimulusterm was a targetterm or non-target. The mouse therefore could not predict the contrastterm of the stimulusterm from trial to trial.

Some mice ( n = 3) were re-trained after initial imaging experiments on a reversed reward contingencyterm·var. Reward contingencyterm·var was switched abruptly, with rewardterm given for licksterm to Stimulusterm B and no rewardterm for licksterm to Stimulusterm A. Because mice were quickly discouraged by the reversal, no quinine punishment was initially given. Additionally, the rewardterm tone (10 kHz, 0.1 s) was paired with the onset of the new targetterm stimulusterm (Stimulusterm B) early in re-training, in order to encourage licking. Three of the five mice trained on this reversed contingencyterm achieved criterion performance after re-training for 10 ± 2 days; the other two were removed from the study.

For all tasksterm, behavioral d-primeterm·var ( d' ) was computed by norminv(Hit rate) – norminv(False alarm rate), where norminv() is the inverse of the cumulative normal function 34 , 36 . Values of Hit and False alarm rate were truncated between 0.01 and 0.99, setting the maximum d' to 4.65. For illustration purposes, d-primeterm·def before and after reversal was computed using Stimulusterm A as the targetterm in both conditions (Fig. 7b ).

Two-photon imaging GCaMP6s fluorescence was imaged 14–35 days after virus injection using Prairie Ultima IV 2-photon microscopy system with a resonant galvo scanning module (Bruker). For fluorescence excitation, we used a Ti-Sapphire laser (Mai-Tai eHP, Newport) with dispersion compensation (Deep See, Newport) tuned to λ = 910 nm. For collection, we used GaAsP photomultiplier tubes (Hamamatsu). To achieve a wide field of view, we used a 16× / 0.8 NA microscope objective (Nikon), which was mounted on a z-piezo (Bruker) for volume scanning. An optical zoom of 2× was used in most cases to improve spatial resolution. Resonant scanning (15.9 kHz line rate, bidirectional) was synchronized to z-piezo steps in the acquisition software for volume scanning. For volume scanning, four 441 × 512 pixel imaging planes separated by 20 or 25 μm were imaged sequentially at a stack rate of 5 Hz in 5–10 min imaging blocks. There was very little redundant sampling of neuronsterm between imaging planes (<1%) as assayed by correlation coefficient of spontaneous activity. Laser power ranged from 40–75 mW at the sample depending on GCaMP6s expression levels. Photobleaching was minimal (<1% min −1 ) for all laser powers used. A custom stainless steel plate (eMachineShop.com) attached to a black curtain was mounted to the head plate before imaging to prevent light from the visualterm stimulusterm monitor from reaching the photomultiplier tubes. During imaging experiments, the polypropylene tube supporting the mouse was suspended from the behavior platform with high tension springs (Small Parts) to dampen movement.

Image preprocessing and cell selection Calcium imagingterm·mth data were acquired using PrairieView acquisition software and sorted into multi-page TIF files. All analyses were performed using custom scripts written either in ImageJ or MATLAB (Mathworks).

Images were first corrected for X–Y movement by registration to a reference image (the pixel-wise mean of all frames) using 2-dimensional cross correlation. To identify responsive neural somata, a pixel-wise activity map was calculated as previously described 37 . Neuronterm cell bodies were identified using local adaptive threshold and iterative segmentation. Automatically-defined ROIs were then manually checked for proper segmentation in a MATLAB-based graphical user interface (allowing comparison to raw fluorescence and activity map images). To subtract the influence of local neuropil on somatic signals, the fluorescence in the somata was estimated as F corrected_soma ( t ) = F raw_soma ( t ) − 0.7 × F neuropil ( t ), where F neuropil was the defined as the fluorescence in the region 0–15 mm from the ROI border (excluding other ROIs) 17 . Δ F/F for each neuronterm was calculated as Δ F / F t = ( F t − F 0 ) / F 0 , with F 0 defined as the mode of the raw fluorescence density distribution.

To align ROIs between different imaging sessions across days (Figs. 7 and 8 ), we used a semi-automated method similar to prior work 25 . First, for each plane, anchor points were manually defined by visualterm comparison of the two average projection images. These anchor points helped to define a predicted displacement vector field that would be used to map coordinates from one session to the other. For each coordinate, the predicted vector was defined by the average (weighted inversely by distance) of the vectors for all defined anchor points.

Next, for each ROI, a square region (~4 × the size of the ROI) around the ROI was selected. To determine the displacement across sessions, we computed the normalized cross-correlation of this square with the average projection of the other session. This was multiplied point-by-point with a mask that decayed gradually with distance from the predicted displacement vector, and then smoothed with a 2-D Gaussian filter. The peak of the resulting image was taken to be the actual displacement vector of the ROI. This process biases the displacement of each ROI towards the vector predicted from the manually defined anchor points. Finally, any ROIs with a computed displacement vector that differed by greater than 5 pixels from the predicted vector were flagged for manual inspection, and then either redrawn or removed.

After image preprocessing and Δ F/F extraction, traces were sorted by trial type (Hit, Miss, Correct Reject, False Alarm) and condition (Engaged, Passiveterm). The baseline response (1 s before stimulusterm onset) was subtracted from each trial. A neuronterm was considered taskterm responsive if its mean Δ F/F during the last 1.6 s (8 frames) of the stimulusterm period was significantly ( p < 0.01, t -test) greater than the pre-stimulus baseline (1 s), for either hit or correct reject trials. Neuronsterm also had to meet a signal-to-noise criterion, needing a trial-averaged response that exceeded a threshold of at least two standard deviations above baseline during either the stimulusterm or response period. All further analyses are based on responses during the last 1.6 s of the stimulusterm period, unless noted otherwise. Cell selection criteria for errorterm analyses, variable contrastterm analyses, and reversal analyses are described in the appropriate sections below.

Statistics The data were obtained from 15 mice, 5 with V1term only, 6 with PPCterm only, and 4 with both V1term and PPCterm. For most mice, multiple fields of view were sampled within V1term or within PPCterm. For mice with both V1term and PPCterm windows, fields from each area were sampled on interleaved sessions. Each field was imaged for a single session, consisting of multiple Engaged and Passiveterm blocks and yielding on average 94 trials (minimum 47) per stimulusterm per condition. For variable contrastterm tasksterm, an average of 25 trials (minimum 8) were acquired per contrastterm. Significantly task-responsive cells from different fields were pooled by area. No tests were conducted to determine sample size. For the full-contrast taskterm, the data came from 1915 V1term cells (18 fields, 9 mice) and 3524 PPCterm cells (22 fields, 10 mice). For the variable-contrast taskterm, the data came from 250 V1term cells (8 fields, 3 mice) and 611 PPCterm cells (11 fields, 6 mice). For comparisons before and after contingencyterm reversal, the data came from 488 V1term cells (8 fields, 3 mice) and 509 PPCterm cells (8 fields, 3 mice).

All statistical analysis was performed using custom-written scripts in MATLAB or R. In all cases, data was not assumed to be normal, and nonparametric and/or permutation teststerm·stat (2000 permutations) were used to assess statistical significance of results. All tests were two-tailed, and a significance level of p < 0.05 was considered significant. Unless otherwise noted, all measures are reported as meanterm·stat ± SEM. When estimating the percentage of selective or modulated neuronsterm, bootstrapping across imaging fieldsterm·stat was used to generate confidence intervals on the percentages for each area. When testing the statistical significance of differences between V1term and PPCterm neuronsterm that were pooled across imaging fields, clustered nonparametric teststerm·stat 38 , 39 were used to account for intra-cluster correlations 31 , 40 using the clusrank package in R.

Selectivityterm and errorterm trial analyses Neuronsterm were marked as target- or non-target-preferring (Fig. 2e ) based on their mean response during Engaged trials. Neuronsterm were marked as task-gated if they did not exhibit a significant response to their preferred stimulusterm during Passiveterm trials (Fig. 2f ).

All comparative indices (engagement modulation indexterm·var, errorterm modulation indexterm·def, contrast modulation indexterm·var, selectivity indexterm·def) were computed using a receiver operating characteristic (ROC) analysisterm·stat, which quantifies the ability of an ideal observer to discriminate between trial types based on single trial responses 18 , 36 . Each index was derived from the area under the ROC curve (auROC), and defined as 2×( auROC –0.5); this value ranged from −1 to 1 5 . Unless otherwise noted, comparative indices were computed by comparing the stimulusterm period response (averaged over the last 1.6 s). To determine whether comparative indices were significant for individual neuronsterm, we used a permutation testterm·stat. We shuffled the labels for each trial and recomputed the index 2000 times to create a distribution of indices that could have arisen by chance. Indices outside the center 95% interval of this distribution were considered significant ( p < 0.05).

For errorterm trial analyses, only behavioral sessions and imaging fields with at least five trials per condition were used. Consecutive miss trials occurring at the end of a session were excluded, as these trials are confounded by lack of motivation. Analysis was additionally limited to Hit-preferring neuronsterm. For analysis of False Alarm trials (Fig. 3 ), we analyzed 1053 V1term cells from 16 fields, and 3034 PPCterm cells from 21 fields. For analysis of Miss trials (Supplementary Fig. 3 ), we analyzed 202 V1term cells from 3 sessions, and 1165 PPCterm cells from 8 fields.

We assessed “stimulus-related” selectivityterm by comparing Hit to FA trials or Miss to CR trials. We define “choice-related” selectivityterm as any premotor- or decision-related signal, assessed by comparing FA to CR trials or Hit to Miss trials (Supplementary Table 1 ). For most comparisons, any selectivityterm measured after the stimulusterm period may include signals related to rewardterm, punishment, or motorterm output. We therefore restricted analyses of selectivityterm to the stimulusterm period.

Selectivityterm was computed across time using a ROC-based index evaluated independently at each time bin (200 ms) 19 , 29 . A clustered Wilcoxon signed-rank testterm·stat 39 was used at each time bin to test whether average selectivityterm for each area was significantly different from zero ( p < 0.05, no correction for multiple comparisons).

To assess the dependence of PPCterm responses on motorterm output (Supplementary Fig. 2 ), we noticed that a few animals exhibited natural variability in licking behavior across False Alarm (FA) trials. We analyzed neuronsterm from the sessions (846 PPCterm cells from 3 fields) that had at least five FA trials with 1–3 licksterm and five FA trials with 5 or more licksterm. We additionally randomly selected Hit trials with the same number of licksterm·var (5–9) to match the average number of licksterm·var in the FA (5–9) condition. Selectivityterm between these conditions was evaluated using a ROC-based index as above.

Generalized linear models We used generalized linear models (GLMterm·mth) to regress recorded calcium signals against a time series of taskterm events 20 – 22 . Calcium responses for each cell were z -scored and modeled as the linear combination of various taskterm events, each convolved with a filter: ŷ t = β 0 + ∑ i = 0 20 β i S tar x t - i S tar + ∑ i = 0 20 β i S nt x t - i S nt + β 0 E x E + ∑ i = 0 20 β i E tar x t - i E tar + ∑ i = 0 20 β i E nt x t - i E nt + ∑ i = - 10 15 β i L x t - i L 1 The response of a neuronterm at frame t is modelled ( ŷ t ) by the sum of a bias term ( β 0 ) and the weighted ( β i ) sum of various additional binary predictors at different lags ( i ). Binary predictors for the targetterm stimulusterm ( x t S tar ) and non-target stimulusterm ( x t S nt ) indicated the onset of stimulusterm presentation in either engaged or passiveterm trials. Binary predictors for engagementterm included a constant offset ( x E ) that was 1 during engaged trials and 0 otherwise, as well as stimulusterm predictors ( x t E tar , x t E nt ) that indicated the duration of stimulusterm presentation during engaged trials only. Binary predictors ( x t L ) for licking indicated the duration of lickterm bouts, which were defined as groups of licksterm with an inter-lick interval <1 s. The number of lags were chosen to capture both the duration of the event (2 s for stimulusterm, 1 s for licking) as well as the slow offset dynamics of the calcium response 17 . Lags were chosen to be strictly positive (causal) for stimulusterm and engagementterm predictors, but both positive and negative (anti-causal) for licking predictors. The final model had 112 coefficients including a constant bias term. Models were fit for all task-responsive neuronsterm, where the first 5 s from each trial (following auditory cue onset) was extracted and concatenated for analysis.

Models were fit using ridge regressionterm·stat using procedures similar to that of previous studies 20 , 22 , 41 . We first set aside 20% of trials from each condition (Hit, Correct Reject, Miss, False Alarm, Passiveterm Targetterm, Passiveterm Nontarget) for testing. A regularization parameter λ was estimated for each cell from among a range of λ values (10 −2 to 10 4 ) using cross-validationterm·stat. Model performance was measured by computing the proportion of explained variance, or the coefficient of determinationterm·stat ( R 2 ). Five-fold cross-validationterm·stat was used to choose the λ that maximized the performance on the remaining 80% of training trials. The final model for each cell was fit using the best λ , and then performance was evaluated by measuring R 2 of predictions for the holdout test set.

The strength of each model component (stimulusterm, engagementterm, motorterm/licking) was evaluated for each V1term and PPCterm cell (Fig. 4f ) by setting all other model coefficients to zero and finding the model predictions using just one component, averaging across both training and test trials. Model components were therefore in units of z -scored calcium response and can be directly compared with calcium response waveforms. For stimulusterm and engagementterm components, the preferred trial type (Hit or Correct Reject) was used for each cell, whereas for the motorterm component, Hit trials were used for all cells. To evaluate the relative strength of the various components across V1term and PPCterm cells, the component strength for each cell was averaged over a window of 0 to 4 s after stimulusterm onset, and then a Wilcoxon rank-sum testterm·stat was applied.

We also trained GLMterm·mth models with a subset of the taskterm predictors (Partial GLMterm·mth models, Supplementary Fig. 4 ) using the same procedures. Three additional models were trained for each cell: a stimulus-only model which included the stimulusterm and constant bias terms (43 coefficients), a motor-only model which included only the licking and bias terms (27 coefficients), and a stimulusterm + engagementterm model which used all terms except licking (86 coefficients). A separate cross-validationterm·stat process and regularization parameter was computed for each model, but the same training and test trials were using to evaluate performance.

Performance on the partial models was evaluated only for cells with significant fits ( p < 0.05) to the full model. The test trial dataset was bootstrapped 2000 times, and a p value was computed for each cell as the fraction of iterations with R 2 ≤ 0. ( R 2 can be negative if the model performs worse than a horizontal line on the held-out test data). The same bootstrapped datasets were also used to determine whether each partial model performed significantly above chance ( p < 0.05), or significantly worse than the full model fit ( p < 0.05).

Contrastterm taskterm analysis For data acquired during the variable contrastterm taskterm, neuronsterm were considered significantly responsive if the mean Δ F/F during the stimulusterm period was significantly above threshold for at least two of the six contraststerm of the same stimulusterm. We focused our analyses on target-preferring neuronsterm, which were included if their mean response across contraststerm was greater for Hit (targetterm) trials compared to Correct Reject trials.

Single neuronterm contrast-response functions were fit to the hyperbolic ratio functionterm·def, also known as the Naka-Rushton function 24 : R C = R max C n C n + C 50 n + R 0 2 where R(C) is the neural response as a function of contrastterm, R max is the saturation point, C 50 is the contrastterm at the half-saturation point, R 0 is the baseline response, and n is an exponent that determines the steepness of the curve. The responses for both Engaged and Passiveterm conditions were fit simultaneously, with n constrained to be constant across conditions, by minimizing the sum (across data points) of the squared errorterm between the model and the data, divided by the variance of that data point.

We evaluated the goodness of fit for each neuronterm using a bootstrap hypothesis testterm·stat, as detailed by others 42 . Briefly, we tested the null hypothesis that the mean of the probability distribution underlying the neural responses was identical to the predictions of the model. We measured the observed prediction errorterm ( e obs ) and computed the probability of observing an errorterm at least as large if the null hypothesis were true. To sample from a distribution that conformed to the null hypothesis, we shifted the data such that the mean responses equaled the model predictions, and drew 1000 bootstrapterm·stat samples from this dataset, computing a prediction errorterm ( e i ) for each. The proportion of samples for which the prediction errorterm was larger than e obs is the achieved significance level. For neuronsterm with an achieved significance level below 10% ( p < 0.1), there was sufficiently strong evidence against the model, and therefore these neuronsterm were excluded from further analysis.

Contrast modulation indexterm·var was computed to compare Engaged trial responses on high (64%, 32%) vs. low (2%, 4%) contraststerm, using a renormalized ROC index which ranged from −1 to 1, as described above. Engagement modulation indexterm·var was computed by comparing Engaged vs. Passiveterm trials, using high contrastterm trials only. A permutation testterm·stat was used to assess significance ( p < 0.05) of these indices by shuffling trial labels 2000 times and comparing the measured index to the shuffled distribution of indices.

To assess anatomical spatial organization, pairwise Euclidean distance was measured between all task-responsive, target-preferring neuronsterm in the same imaged volume (Fig. 6e–g ). Cells were grouped into four groups based on whether they were contrast-modulated, engagement-modulated only, both, or neither, and then the average within-group and across-group distance was computed for each cell (Fig. 6f ). To compute differences in modulation indexterm·def as function of distance (Fig. 6g ), pairs of cells were binned by distance in 40 µm bins.

Selectivityterm on errorterm trials (Supplementary Fig. 5 c-f) was computed by comparing FA trials to Hit and CR trials (with matched contrastterm) using a frame-by-frame ROC analysisterm·stat, as described above. The contrast-dependence of the auROC-based index was estimated by regressing the index against log contrastterm, and then finding the slope. The significance ( p < 0.05) of the slope was estimated for each cell by shuffling the trial labels 2000 times for each contrastterm and then computing the auROC on the shuffled data to generate a distribution of 2000 slopes.

Reverse contingencyterm taskterm analysis We imaged from the same field of neuronsterm before and after reversal of reward contingencyterm·var. The two sessions were separated by an average time interval of 16 ± 1 days. A semi-automated method was used to align ROIs between the two sessions (see "Image preprocessing and cell selection" section). Neuronsterm were included for analysis only if a significant response ( p < 0.01) to either stimulusterm was observed both before and after reversal. Selectivity indexterm·def was computed separately before and after reversal training (Fig. 7e–h ) by comparing responses to the Stimulusterm A (original targetterm) and Stimulusterm B (original non-target), with positive values indicating preference for the Stimulusterm A. Separate permutation teststerm·stat (2000 iterations) were used to assess significance of the selectivityterm before and after reversal. Neuronsterm with significant selectivityterm ( p < 0.05) both before and after reversal were categorized (Fig. 8a ) based on the sign of the selectivityterm, and whether it was stable or altered with reversal. Engagement modulation indexterm·var (Fig. 8b ) for each neuronterm was computed separately before and after reversal, comparing Engaged and Passiveterm responses using the neuronterm’s preferred stimulusterm, and then taking the mean of the two values. Errorterm modulation indexterm·def (Fig. 8c ) for each neuronterm was also computed separately before and after reversal, comparing FA and CR responses, and then taking the mean of the two values.

Data availability The data and custom MATLAB analysis code that support the findings of this study are available from the corresponding authors upon request.

Electronic supplementary material Supplementary Information (1.3MB, pdf) Peer Review File (244.2KB, pdf) Acknowledgments This work was supported by the NIH (M.J.G., F32-EY023523 and K99-MH104259; M.S., R01-EY007023 and U01-NS090473); the NSF (G.N.P., Graduate Research Fellowship; M.S., EF1451125); the Simons Center for the Social Brain (M.S.), and the Picower Institute Innovation Fund (M.J.G.; M.S.). We thank J. Sharma, A. Boesch, V. Li, C. Le, J. Krizan, V. Breton-Provencher, and T. Emery for technical assistance; R. Huda, M. Hu, H. Sugihara, and C. D. Harvey for helpful discussions and/or comments on the manuscript; L. L. Looger, J. Akerboom, D. S. Kim, and the Genetically-Encoded Neuronal Indicator and Effector (GENIE) Project at Janelia Farm Research Campus Howard Hughes Medical Institute for generating and characterizing GCaMP6 variants.

Author contributions G.N.P. and M.J.G. designed experiments with input from M.S. M.J.G. performed the surgeries. M.J.G. and G.N.P. performed the imaging experiments. M.J.G., G.N.P., J.W., and B.C. performed the behavioral training. G.N.P. developed the behavior software. G.N.P. analyzed the data with comments from M.J.G. and M.S.. G.N.P., M.J.G., and M.S. wrote the manuscript.

Competing interests The authors declare no competing interests.

Footnotes Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

These authors contributed equally: Gerald N. PhopersName, Michael J. GoardpersName.

Change history 1/18/2019 In the original version of this Article, the Acknowledgements section was inadvertently omitted. This has now been corrected in both the PDF and HTML versions of the Article.

Electronic supplementary material Supplementary Information accompanies this paper at 10.1038/s41467-018-05012-y.