A novel discrimination-avoidance task.

A, Top: Schematic of the task. Mice are trained to discriminate between two auditory cues (lasting 10 s) and shuttle between two distinct rooms of a shuttle box to avoid footshocks. Specifically, sounds A and B signal shocks in rooms A and B, respectively. Bottom: Two behavioral response scenarios. S1: If the mouse makes the correct response, either by staying in or shuttling to the correct room before the sound ends, no shock is administered. S2: If the mouse makes an incorrect response, either by staying in or shuttling to the incorrect room before the sound ends, up to 10 mild shocks (0.5 mA, 0.1 s; 2 s apart) are administered until the mouse shuttles to the correct room. Each training session comprises 50 trials, 60 s apart; sounds A and B are presented in a pseudorandom order. B, Learning curve showing the success rate in avoiding shocks across training sessions (mean ± s.e.m.; n = 11 mice). F9, 90 = 14.78; P < 0.001; One-way ANOVA. C, Correct and incorrect trial counts for shuttle vs. stay trials across training sessions for the same mice shown in B. Correct shuttle: F9, 90 = 14.62, P < 0.001; Incorrect shuttle: F9, 90 = 1.94, P = 0.057; Correct stay: F9, 90 = 1.55, P = 0.142; Incorrect stay: F9, 9Ũ = 18.73, P < 0.001; One-way ANOVA. D, The mean shuttle crossing latency, averaged over the last two sessions for individual mice, is significantly shorter in correct trials than that in incorrect trials (t10 = 2.62, P = 0.026; Effect Size: Cohen’s d = 0.789; Power = 0.656; paired t test). The black line indicates the mean; grey lines indicate individual mice. Shuttle crossing is defined as the body center crossing the midline opening of the shuttle box.

ACC neurons primarily encode post-action variables.

A, A representative brain section showing electrode tracks (left) and the presumed implantation sites (red dashed lines; right), located largely within the ACC. Scale bars, 0.5 mm. B, A representative shuttle response and corresponding Y-position of the animal’s body center. C&D, Peri-event rasters and histograms of a representative ACC neuron during correct-shuttle (C; trials sorted by shuttle latency) and correct-stay trials (D). In this session, there were 21 correct shuttles, 15 correct stays, 5 incorrect shuttles, and 9 incorrect stays. E&F, Heatmaps showing the activity of all recorded ACC neurons (n = 376) during correct-shuttle (E) and correct-stay trials (F). Neurons in E and F are arranged in the same order.

ACC neurons primarily respond during “shuttle” but not “stay” trials.

A, Heatmaps showing the activity of simultaneously recorded ACC neurons (n = 29) during correct-shuttle (top) and correct-stay trials (bottom) from the same recording session as shown in Fig. 2C. B, Principal component analysis (PCA) of ACC neuronal population activity as shown in A. PC1, PC2, and PC3 are the first three principal components; the numbers are the percentages of total variance explained by the corresponding PCs. Each circle indicates a time lapse of 0.1 s. Note that there is a robust neural state change in the 3-D PC space surrounding the shuttle response (top), but not the stay response (bottom). C, Maximum trajectory distance between the center of the baseline period and any post–Time 0 data point, as shown in B, across five animals (top), and the corresponding scree plots showing variance explained by the first eight PCs during shuttle trials (bottom).

Characterizing ACC neuronal activity in relation to action initiations.

A, Top, schematic illustrating the definition of shuttle initiation, crossing, and termination (pause). Bottom, distributions of half-(left) and full-shuttle durations (right). B, Z-scored activity of all ACC neurons (n = 376) during correct shuttle trials. C, Principal-component analysis (PCA) classifies ACC neuronal activity (as shown in B) into seven categories. PC1, PC2, and PC3 represent the first three principal components color coded from low (dark) to high scores (white). D, Mean activity (± s.e.m.) of the seven categories of ACC neurons. E, Fractions of individual categories of ACC neurons.

ACC neurons exhibit limited modulation by speed.

A, Speed-tuning curves of three simultaneously recorded neurons during free exploration in shuttle boxes, showing positive (red), negative (blue), or no correlation (gray) between firing rate and locomotion speed. “r” indicates Pearson’s correlation coefficient. B, Distributions of observed vs. shuffled correlation values across all recorded ACC neurons (n = 376). Dashed lines mark the 99th percentile of shuffled distribution. Overall, 7.7% of recorded neurons were positively modulated by speed, and 6.9% were negatively modulated. C, Proportions of speed-modulation for each category of ACC neurons (see Fig. 4D).

ACC neurons monitor actions independent of outcomes.

A, Peri-event rasters (trials) and histograms of a representative ACC neuron during correct (left), incorrect (middle), and post-shock shuttles (right) within a session. Red triangles indicate shock administrations. Note that incorrect shuttles are followed by a second shuttle after animals receive footshocks, and approximately half of the post­shock shuttles are preceded by incorrect shuttles. B, Heatmaps showing the activity of individual ACC neurons (n = 348) during correct (left), incorrect (middle), and post-shock shuttles (right). Neurons are arranged in the same order across the three heatmaps. Color bar indicates z-scored activity. Note that the number of incorrect shuttles in a session is often <7, leading to greater variability in mean activity. Only sessions with >4 incorrect shuttles are included in the analysis. C, Activity indexes of individual ACC neurons between correct and incorrect shuttles (left), and between correct and post-shock shuttles (right). Activity index is defined as: Activity Index = Meanpost-huttle – Meanpre-huttle, where Meanpre-huttle and Meanpost-huttle are the mean z scores calculated between -5-0 and 0-5 s, respectively, as shown in B. D, Correlation coefficients of the activity (−5-5 s) between correct and incorrect shuttles (x axis) and between correct and post-shock shuttles (y axis). Only the top and bottom quartiles of ACC neurons (as shown in B) are used for the analysis.

ACC neurons monitor action state and action content.

A&B, Peri-event rasters (trials) & histograms of four simultaneously recorded ACC neurons during two sets of shuttles: rooms A B shuttles (A) vs. rooms B A shuttles (B). Notably, neurons 1&2 exhibit indiscriminate responses, either increasing or decreasing their activity after shuttles, thereby monitoring action state changes. In contrast, neurons 3&4 are selectively activated in one set of the shuttles, thereby monitoring action content (i.e., rooms A B vs. B A shuttles). C, Z-scored activity of all ACC neurons (n = 376) during A B shuttles (left) and B A shuttles (right). The color bar indicates z score. Neurons are categorized and sorted according to coefficient β or Δβ values (see Methods): Category 1 (n = 98), β1 > β2, sorted by β1; Category 2 (n = 107), β1 < β2, sorted by β2; Categories 3&4 (n = 29/43), both β1 and β2 are significantly positive or negative, sorted by the combined magnitude of β1 and β2. D, Coefficient values β1 β2, and Δβ (β1–β2) values are shown in red, blue, and gray, respectively, for individual ACC neurons ordered in the same sequence as shown in C. E, Mean activity for the five major neuronal categories (as discussed in C) during A B (red) and B A (blue) shuttles.

Post-shuttle ACC neuronal population activity decodes action content.

C, Schematic diagram of support vector machine (SVM) decoding. ACC neuronal population activity from pre-shuttle period (−5–0), post-shuttle period (0–5), or shuffled spikes is used to train the decoder and subsequently distinguish between action content (rooms A B vs. B A shuttles). B, Mean decoding accuracy (blue line) and individual decoding accuracies for 15 sessions (grey lines). Friedman test (P < 0.001) and post-hoc Wilcoxon signed-rank test with Bonferroni correction (***P < 0.001). C, SVM decoding accuracy across all 15 sessions as a function of the fraction of neurons removed (20% per step), applied separately to action-content neurons (blue) or the remaining neurons (grey) within each session (P < 0.001 for each comparison between the two removals; Wilcoxon signed-rank test). Shaded areas denote ± s.e.m.

Post-action ACC activity influences future performance within a task session.

A, Schematic illustration. B, Mean activity (± s.e.m.) of post-shuttle activated neurons (orange lines; top 1/3), inhibited neurons (blue lines; bottom 1/3), and remaining ACC neurons (black lines; middle 1/3), which preceded either correct trials (solid lines) or incorrect trials (dashed lines). C, Further comparison of the activation strength for post-shuttle activated neurons (left) and inhibited neurons (right) between the two conditions. Each pair of dots indicates an ACC neuron. Two-way ANOVA, Interaction: F2, 328 = 5.69; P = 0.004; Simple effect for Top1/3: ***P < 0.001; Effect Size: Cohen’s d = 0.33. D–F, Similar to A–C, except that the comparison is based on the status of the preceding trials. Mean z scores in C and F were calculated between 0–5 s after shuttle crossings. Two-way ANOVA revealed no-significant difference. n.s., non-significant.

Proposed model of cue-action-outcome associative learning.

Top: Four distinct phases of information coding: pre-shuttle, shuttle, post-shuttle, and outcome phases. Bottom: Integration of cue, action, and outcome information occurs when all three components are concurrently available (indicated by the oval). Notably, without the ACC, action-relevant information could be lost after action execution, thereby disrupting cue–action–outcome associative learning.

Experimental setup.

A, The shuttle box used for behavioral training. Room A is configured with two black walls (left and right), one white wall (back), and a transparent front wall for video recording purposes. Room B is configured with two white walls (left and back), one metal wall (right), and a transparent front wall. B, Schematic diagram of the control setup utilizing MATLAB functions, with numbers indicating the sequence of control flow. C, A comprehensive flowchart illustrating the control setup as shown in B. D, Left, sounds A and B signal shocks in the bottom and top rooms of the shuttle box, respectively. Right, the Y position of a well-trained mouse in a ∼40-min session.

Extended training recording sites.

A, Correct and incorrect trial counts for shuttle vs. stay trials across 15 training sessions. In shuttle trials, mice must move to the adjacent room before the sound ends to avoid shocks, whereas in stay trials, mice must remain in their current room to avoid shocks. B, reconstructed recording sites for individual mice. C, Heatmaps showing the activity of all recorded ACC neurons during correct-shuttle trials across individual mice.

ACC neurons primarily respond during “shuttle” but not “stay” trials.

A&B, Dimension reduction analysis (i.e., PCA) indicates robust changes in ACC neuronal population activity during correct shuttles (A) but not correct stays (B) from four representative recording sessions. Time “0” indicates shuttle crossings (A) or sound onsets (B), respectively. For more details, see Fig. 3B.

ACC shows limited response to cue, stay trials, and footshocks.

Left, Heatmap showing the activity of individual ACC neurons (n = 348) in relation to auditory cue onset. Middle, Heatmap showing the activity of individual ACC neurons (n = 348) during stay trials. C, Heatmap showing the activity of individual ACC neurons (n = 336) during footshock. Note, footshock trials with a shuttle response within 1 s shock onset were excluded to avoid shuttle response confound. As a result, some behavioral sessions were excluded due to insufficient trial numbers.

ACC pyramidal neurons and interneurons both monitor action content.

A, Spike waveforms (mean ± s.d.) of two representative ACC neurons: one putative pyramidal neuron and one interneuron. B, Peri-event rasters (trials) & histograms of the same two ACC neurons surrounding shuttle responses. Both neurons exhibit differential activity changes that discriminate between rooms A B (top panels) vs. B A shuttles (bottom panels). C, Cross-correlation histograms between the putative interneuron (Neuron #2; the same as shown in A) and three other pyramidal neurons (Neurons #1, #3, and #4). Neuron #4 appears to excite Neuron #2, which in turn inhibits Neuron #3, as indicated by short-latency (∼2 ms) excitatory or inhibitory interactions. The four ACC neurons were recorded simultaneously.

ACC neurons do not display place cell characteristics.

A, A representative navigation path during inter-trial intervals of a discrimination-avoidance task session. B, Place field activity of two representative ACC neurons. Both neurons show spiking activity across the chamber without place preference. The color bar indicates normalized firing rate. C, Spatial information coding distribution of all recorded ACC neurons (n = 376). D–F, Same as in A–C, but for neurons recorded from hippocampal dorsal CA1 (n = 161). Notably, more than half of CA1 neurons exhibit high information coding (>2 bits/spike; F), whereas very few ACC neurons show comparable levels of information coding (C).

SVM decoding of multiple behaviors from ACC neuronal population activity.

A, Mean (blue line) and individual session decoding accuracies (grey lines; 15 sessions) for decoding A B vs. B A shuttles using short (left, 2.5 s), medium (middle, 7.5 s), or long decoding windows (right, 10 s) surrounding shuttle crossings. **P < 0.01, ***P < 0.001, Wilcoxon signed-rank test. B, Mean (blue line) and individual session decoding accuracies (grey lines) for decoding correct vs. incorrect stays (left) or Room-A vs. Room-B stays (right), using a 10-s window after sound onset. Notably, in several sessions, lower-than-chance decoding accuracy for Corr/Incorr stays (colored dots; left) corresponded to high decoding accuracy for Room A/B stays (shown in the same colors; right), suggesting a potential confound between these variables.