Extended Data Fig. 3. Further analyses of single neuron and neuronal population activity.
a, Unsupervised agglomerative clustering of task- or satiety-modulated neuronal dynamics surrounding Go trials. Clusters are manually color-coded approximating the predominant information in the cluster: orange, food choice; blue, water choice; gray, satiety states; purple, odor; green, food-related action activity; red, water-related action activity. Clusters are shown separated by horizontal lines. Neurons are individually trial-averaged by condition (Go-odor food choices, Go-odor water choices, Go-odor non-responsive/sated) and Z-scored across condition. Vertical dashed line, Go-odor onset. s.d., standard deviation. Leftmost column, brain region assignments for each neuron, color coded according to the Allen Brain Atlas colormap (see Extended Data Table 1). b, Single-cell cluster distribution by brain region, with clusters (color coded as in a) plotted as a normalized fraction of cells in a given region (parenthesis: total cell numbers per region). Regions are ordered by hierarchical clustering of the regional cluster distributions. c, Schematic of regression analysis for functional properties of individual neurons related to task or satiety. d, Quantification of mixed selectivity per neuron with significant information about upcoming choice in baseline activity. The histogram shows the distribution of counts of additionally significant regressors (described in c) per cell. A cell is counted as significant for a given regressor if the firing rate variance explained by the regressor is greater than that of a per-cell circularly permuted null distribution, with threshold P = 0.05; per-cell P-values are not adjusted for multiple comparisons. e, Co-occurrence of regressors for all task- or satiety-modulated cells is shown as the pairwise correlation coefficient of variance explained by each regressor, sorted for visualization by unsupervised hierarchical clustering. The three clusters group regressors for cross-session satiety (regressors 4, 2, 3), Go vs. No-Go trial type (regressors 5, 8), and reward choice (regressors 7, 1, 6). f–j, Per-cell visualization of pairwise association between goal-information (x axes) and other regressors (y axes). Cells are color-coded according to their assigned brain region, following the Allen Brain Atlas colormap. Information at baseline about upcoming choice tended to co-occur with all regressors except Go vs. No-Go regressors. Data in a–j pooled from n = 7 mice, 7 sessions. k–n, Video decoding analyses. Decoding analysis from video motion data of upcoming choice. Videos of the animal’s face or body were collected at 100 Hz and the principal components (PCs) of the video motion energy were used for prediction of behavioral choice (Methods). k, Predictiveness (AUC) of choice by video data of the body or face (x-axis) during the 1 s pre-odor (blue bars) or the response period (1–2 s post-odor onset, green bars). Mean ± 95% confidence interval. Dashed lines, session permuted null distribution. Test trials are taken from trials flanking behavioral switches (4 trials before the penultimate trial prior to each switch, and 4 trials following the 1st trial after a switch) to remove any spurious contribution of slow timescale motion covariates with choice. l, Face video predictiveness (AUC, y-axis) of upcoming choice during the 1 s pre-odor epoch, with variable numbers of motion principal components (PCs) used for prediction. Mean ± 95% confidence interval. k, l, n = 7 mice, 7 sessions. m, Visualization of the most predictive PC for food and water choices, superimposed on the average image of the animal. Pixels with high positive weights in the PC are colored red and pixels with high negative weights are colored blue. The pattern of PC weights does not indicate any obvious distinguishing motion feature of the animal, suggesting that the predictiveness of the video motion comes from more subtle “tells” of the animal’s facial motion. n, Predictiveness (AUC, y-axis) of the face video motion PCs (gray) or neural data (green) as a function of time, with decoders trained on each 10-ms time bin across a trial. Mean ± 95% confidence interval (n = 7 mice, 7 sessions). Dashed horizontal lines, circular permutation null distributions for bins in the 1 s pre-odor period. Vertical dashed line, odor onset. In summary, we can predict behavioral choice from pre-odor videos of the animal (indeed, we expect that neural activity should be reflected by animal behavior). However, predictiveness does not come from one or a few dominant motions, as the decoder requires dozens of PCs at least. Moreover, pre-odor prediction from videos of the animal’s pose is not as good as prediction from neural population activity. o, Analysis of redundancy of information for upcoming choice across simultaneously recorded cells. A linear decoder is trained to predict upcoming choice from the 1 s of simultaneously recorded population neural activity preceding food or water choices. The size of the population used for decoding is randomly subsampled to examine the effect on decoder performance of increasing numbers of simultaneously recorded cells, agnostic to brain region. Decoding performance (receiver operator characteristic area under the curve, AUC) is assessed for test trials flanking behavioral transitions, such that test trials with different classes occur close to each other in time and the contribution of spurious long-timescale correlations are largely removed. Mean ± 95% confidence interval (n = 7 mice, 7 sessions). Dashed lines, circularly permuted null distribution for each decoder. p, Percentage simultaneously recorded population variance explained by coding dimensions across task period. Baseline period, 1 s pre-odor activity preceding food or water choices (hit trials). Choice period, 1 s to 2 s post-odor during hit trials. Response period, 1 s to 2 s post-odor for all trials (including No-Go trials). Coding dimensions are calculated as the variance-normalized average firing rate difference between periods in trials corresponding to a given regressor (Methods). Goal regressor, 1 s pre-odor period preceding water vs. food choices. Choice regressor, 1 s to 2 s post-odor for water vs. food choices. Response regressor, 1 s to 2 s post-odor for Go vs. No-Go trials. Mean ± 95% confidence interval across recording sessions. Variance explained by each coding dimension is consistent across mice/sessions (n = 7) and largely distinct across regressors and their corresponding task periods. The amount of baseline period population variance explained by the goal regressor is comparable to the response period variance explained by the choice and response regressors. q, r, Cells significant for goal information have significantly higher noise covariance (covariance between firing rates in the 1 s pre-odor, with the average activity before food or water trials removed from the trial-by-trial firing rates, and evaluated at a 10-ms temporal resolution) than non-goal significant cells, suggesting that goal cells may share a common source of fluctuations or tend to influence each other’s activity more frequently than non-goal cells. Comparison for cells within the same region, q, and across regions, r. ****, two-sided t-test, P ≤ 1 × 10−6 (q, t = 11.52, P = 1.11 × 10−30, n = 7 mice, 21,496 pairwise correlations; r, t = 52.08, P = 0.0, n = 7 mice, 2,720,269 pairwise correlations). Y-axis is truncated. Dashed lines, means.