Experimental paradigm.

A: Schematic of the behavioral task. The blue rectangle indicates the temporal window over which all analyses were conducted, from 1 s before the first central lick, T, to 1 s after the first lateral lick, D. B: Psychometric curve averaged across sessions and subjects. Circles and error bars represent mean ± s.e.m. (N = 23 sessions across 13 subjects) for the probability of a choice in the sucrose-associated direction for each stimulus value; the continuous curve is a sigmoidal curve fit to the means. C: Schematic of acute probe insertion. D: Example histological section indicating accurate probe placement in GC (blue: Hoechst; yellow: DiI applied to probe). E: Example spike raster plot for neurons simultaneously recorded within a single session from GC of behaving mice.

Population activity and information encoding during taste mixture-based decision-making.

A: Population PSTH (N = 626). Vertical dashed lines indicate the first central (T) and lateral licks (D), respectively. The trace represents mean firing rate; shading represents s.e.m. B: Population heatmaps for single unit differential activity between correct predominantly-sucrose and correct predominantly-NaCl trials. White dots indicate each unit’s time of peak differential activity. Traces are ordered by peak time and separated by whether peak differential activity is in favor of NaCl (“NaCl-preferring”) or sucrose (“Sucrose-preferring”). White trace is the mean auROC across neurons. C-D: Decoding of task-relevant variables. For each session (grey trace), accuracy is plotted over time with colored shaded traces representing the mean ± s.e.m. over sessions. Trial labels to be decoded were mixtures (C) and choice (D). Horizontal solid line represents theoretical chance level. Horizontal dashed line represents theoretical significant decoding threshold (α = 0.01). Dashed rectangles mark the “sampling” and “delay” analysis windows.

Low-dimensional population activity trajectories.

A: Euclidean distances between pairs of trial-averaged pseudo-population activity trajectories. B: t-SNE of trial-averaged pseudo-population trajectories for all stimuli (%Sucrose/%NaCl) based on pairwise Euclidean distances between activities. C: One-dimensional linear projections of trial-averaged pseudo-population trajectories onto the demixed principal component explaining maximum stimulus-specific variance. Solid lines are correct trial averages; dotted lines are incorrect trial averages. D: Same as C, but for the demixed principal component explaining maximum choice-specific variance. T: time of first central lick; D: time of first lateral lick.

Classification of single unit coding types.

A: Representative single unit PSTH (top) and response profiles (bottom) exemplifying the different coding types within a time window (grey bar, top): linear (left), step-perception (middle), and step-choice (right). Step-perception (middle) and step-choice (right) types were disentangled by comparing correct trials to error trials (dashed lines in bottom plots). Color scale corresponds to different mixture stimuli (%Sucrose/%NaCl). B: Visualization of each neuron’s coding type label (vertical axis) between two time windows (horizontal axis). Each neuron is a point in both windows, with lines connecting the same neurons. T: time of first central lick; D: time of first lateral lick. C: Distribution of coding types across all neurons (pooled over all sessions) over time. For each time point (a window ~200 ms wide), the coding type classification analysis depicted in A was applied to each neuron.

Recurrent neural network design and behavior.

A: Model architecture. N neurons are modeled as dynamic units with internal activity h that is influenced by the external stimulus input (m(x); the time course of an example x for mixture stimulus 75/25 is shown), recurrent input (via Wrec), and noise input (not shown). A decision unit z measures the network’s choice by taking a weighted sum of activities via wz. The loss function L is minimized during training based on choice (z) and the activity of the constrained units (grey dots). T: time of stimulus onset; D: decision time. B: Psychometric curve fit to across-model means for the probability of the sucrose choice as a function of the stimulus. Circles and error bars represent mean and s.e.m. C: Example of experimentally observed PSTH (left) and the corresponding firing rate activity for the unit in the network trained to match it (right). D: Example firing rate activity for a unit in the network not explicitly trained to match any experimentally observed PSTH. Color scale corresponds to different mixture stimuli (%Sucrose/%NaCl).

Modeled population activity and single unit coding properties.

A: Trial-averaged pseudo-population activity trajectories projected onto demixed principal component of maximal stimulus-specific variance. Solid lines are correct trial averages; dotted lines are incorrect trial averages. Color scale corresponds to different mixture stimuli (%Sucrose/%NaCl). T: time of stimulus onset; D: decision time. B: Same as A but for the demixed principal component of maximal choice-specific variance. C: Left: Venn diagram showing percentages of neurons with all possible combinations of coding types over time. Right: Distribution of coding types across all units (pooled over all models) over time.

Effect of selective ablations on model dynamics and behavior.

A: Model dynamics after selectively ablating linear coding units, step-perception coding units, step-choice coding units, or “other” units. Post-ablation pseudo-population activity is projected onto the stimulus-(left column) and choice-coding (right column) components identified in the control condition (i.e., the same ones in Figure 6A-B). Color scale corresponds to different mixture stimuli (%Sucrose/%NaCl); solid and dashed lines correspond to correct and error trials. T: time of stimulus onset; D: decision time. B: Pairwise overlaps between stimulus-(top) and choice-coding (bottom) components for control (-) and each ablation condition (o: “other,” l: linear, p: step-perception, c: step-choice). C: Behavioral performance of the model after selectively ablating categories of coding units. Left: across-model distributions of task accuracy vs ablation condition. Bars represent means. * indicates significant difference vs control condition (post-hoc paired t-test Bonferroni-adjusted p < 0.01). Right: psychometric functions fit to across-model mean probability of sucrose choice for different ablation conditions. Circles and error bars represent mean and s.e.m.

Session-by-session responsive and selective neuron counts for experimental data.

Responsivity indicates a difference in firing rate distributions between baseline and a window of interest (from the first central lick to 500 ms after it for taste; from 500 ms before the first lateral lick to the first lateral lick for delay). Selectivity indicates a difference in firing rate distributions between categories within the window of interest (predominantly-sucrose vs predominantly-NaCl for taste; left vs right for delay).

Session-by-session neuron coding type counts for experimental data.

Neurons are assigned coding type labels if they exhibit the response profile pattern in any time bin (as per analysis in Figure 4C) and, thus, the labels are not mutually exclusive.

Session-by-session responsive and selective unit counts for model data.

Responsivity indicates a difference in firing rate distributions between baseline and a window of interest (from stimulus onset to 500 ms after it for taste; from 500 ms before the decision to the decision for delay). Selectivity indicates a difference in firing rate distributions between categories within the window of interest (predominantly-sucrose vs predominantly-NaCl for taste; left vs right for delay).

Session-by-session unit coding type counts for model data.

Units are assigned coding type labels if they exhibit the response profile pattern in any time bin (as per analysis in Figure 6C) and, thus, the labels are not mutually exclusive. Con.: constrained units; Unc.: unconstrained units.

Neuropixels probe trajectory reconstruction.

A: 3D reconstruction of the 23 probe trajectories from the experimental dataset. B: 2D reconstruction of the same 23 probe trajectories, overlaid on the Allen Brain Atlas at varying anteroposterior (AP) distances (relative to Bregma in mm) around GC. At these coordinates, both GU (gustatory areas) and AI (anterior insular areas) account for GC. Reconstructions performed with open-source Allen CCF Tools (Shamash et al., 2018; github.com/cortex-lab/allenCCF).

Model constrained unit activity.

A: Three examples (columns) of experimentally-observed PSTHs (top) and corresponding model unit firing rate activities trained to match them (bottom). Color scale corresponds to different mixture stimuli (%Sucrose/%NaCl). B: Comparison of firing rate activities (stimulus-averaged PSTHs) between all experimental neurons and their corresponding model constrained units. Vertical whitespace separates individual sessions/models. Firing rates are normalized to the maximum within each session. T: time of first central lick/stimulus onset; D: time of first lateral lick/decision time.

RNN unit responsiveness.

A: Activity of all RNN units grouped by responsiveness during the sampling period. If the unit’s firing rate distribution during the sampling period (T to T + 0.5 s for T the stimulus onset time) was significantly different from its baseline (T – 0.5 s to T) firing rate distribution, it was sampling responsive and grouped by whether its mean firing rate increased (left) or decreased (middle); otherwise it was non-responsive (right). B: Activity of all RNN units grouped by responsiveness during the delay period. Same as A except the firing rate distribution of interest is calculated over D – 0.5 s to D for D the decision time. C and D: Same as A and B, respectively, except that the only units considered are those labeled “other” by the response profile analysis of Figure 6C. Firing rates are expressed relative to baseline, and traces are population mean ± s.e.m.

RNN unit activity patterns.

A: Heatmaps of firing rate activities for units that responded significantly during the sampling and/or delay periods, broken down into coding units (linear, step-perception, and/or step-choice) (top) and “other” units (not linear, not step-perception, and not step-choice) (bottom). Firing rates are expressed relative to baseline and normalized to the maximum absolute value. T: time of stimulus onset; D: decision time. B: Two example “other” unit responses. Both respond significantly during the sampling period, but neither response pattern matches the linear or step templates. Color scale corresponds to different mixture stimuli (%Sucrose/%NaCl).

Modeled population activity and single unit coding properties: constrained units only.

Compare with Figures 6 and S6. A: Trial-averaged constrained pseudo-population activity projected onto demixed principal component of maximal stimulus-specific variance. Solid lines are correct trial averages; dotted lines are incorrect trial averages. Color scale corresponds to different mixture stimuli (%Sucrose/%NaCl). T: time of stimulus onset; D: decision time. B: Same as A but for the demixed principal component of maximal choice-specific variance. C: Left: Venn diagram showing percentages of constrained units (pooled over all models) with all possible combinations of coding types over time. Right: Distribution of coding types across constrained units over time.

Modeled population activity and single unit coding properties: unconstrained units only.

Compare with Figures 6 and S5. A: Trial-averaged unconstrained pseudo-population activity projected onto demixed principal component of maximal stimulus-specific variance. Solid lines are correct trial averages; dotted lines are incorrect trial averages. Color scale corresponds to different mixture stimuli (%Sucrose/%NaCl). T: time of stimulus onset; D: decision time. B: Same as A but for the demixed principal component of maximal choice-specific variance. C: Left: Venn diagram showing percentages of unconstrained units (pooled over all models) with all possible combinations of coding types over time. Right: Distribution of coding types across unconstrained units over time.

Effect of selective ablations on model dynamics and behavior: constrained vs unconstrained.

A-B: Model dynamics after selectively ablating linear coding units, step-perception coding units, step-choice coding units, or “other” units in the constrained (left columns) or unconstrained (right columns) populations. Post-ablation pseudo-population activity is projected onto the stimulus (A) and choice-coding (B) components identified in the control condition (top). Color scale corresponds to different mixture stimuli (%Sucrose/%NaCl); solid and dashed lines correspond to correct and error trials. * indicates significant difference in mean absolute projections vs corresponding control condition (Dunnett’s test p < 0.01). T: time of stimulus onset; D: decision time. C: Behavioral performance of all models after selectively ablating categories of coding units. Bars represent means. * indicates significant difference in task accuracy vs control condition (Dunnett’s test p < 0.01).

Effect of temporally restricted selective ablations on model dynamics and behavior: beginning vs end.

A-B: Model dynamics after selectively ablating linear coding units, step-perception coding units, step-choice coding units, or “other” units at the beginning of the trial (left columns) or the end of the trial (right columns). Post-ablation pseudo-population activity is projected onto the stimulus (A) and choice-coding (B) components identified in the control condition (top). Color scale corresponds to different mixture stimuli (%Sucrose/%NaCl); solid and dashed lines correspond to correct and error trials. * indicates significant difference in mean absolute projections vs corresponding control condition (Dunnett’s test p < 0.01). T: time of stimulus onset; D: decision time. The beginning is [T, T + 1.2 s]; the end is [D – 1.2 s, D]. C: Behavioral performance of all models after selectively ablating categories of coding units in the beginning or end of the trial. Bars represent means. * indicates significant difference in task accuracy vs control condition (Dunnett’s test p < 0.01).