Excitatory and Inhibitory Subnetworks Are Equally Selective during Decision-Making and Emerge Simultaneously during Learning

Abstract

Inhibitory neurons, which play a critical role in decision-making models, are often simplified as a single pool of non-selective neurons lacking connection specificity. This assumption is supported by observations in the primary visual cortex: inhibitory neurons are broadly tuned in vivo and show non-specific connectivity in slice. The selectivity of excitatory and inhibitory neurons within decision circuits and, hence, the validity of decision-making models are unknown. We simultaneously measured excitatory and inhibitory neurons in the posterior parietal cortex of mice judging multisensory stimuli. Surprisingly, excitatory and inhibitory neurons were equally selective for the animal's choice, both at the single-cell and population level. Further, both cell types exhibited similar changes in selectivity and temporal dynamics during learning, paralleling behavioral improvements. These observations, combined with modeling, argue against circuit architectures assuming non-selective inhibitory neurons. Instead, they argue for selective subnetworks of inhibitory and excitatory neurons that are shaped by experience to support expert decision-making.

Figure 1.. Simultaneous imaging of inhibitory and excitatory populations during decision-making.

A. Behavioral apparatus. Multisensory stimuli are presented via a visual display and a speaker. To initiate trials, mice lick the middle waterspout. To report decisions about stimulus rate, mice lick left/right spouts. Objective: 2-photon microscope used to image neural activity through an implanted window. B. Psychometric function showing the fraction of trials in which the mouse chose “high” as a function of stimulus rate. Dots: mean (10 mice). Line: Logit regression model (glmfit.m); mean across mice. Shaded area: standard deviation of the fit across mice. Dashed vertical line: category boundary (16Hz). C, Average image (10,000 frames). Left: green channel, GCaMP6f. Middle: red channel, tdTomato. Right: merge of left and middle. Cyan circles: GCaMP6f-expressing neurons identified as inhibitory. D, Example neurons identified by the Constrained Nonnegative Matrix Factorization algorithm (Methods) (arrow: inhibitory neuron). Left: raw ΔF/F traces. Middle: de-noised traces. Right: inferred spiking activity. Imaging was not performed during inter-trial intervals; traces from 13 consecutive trials are concatenated; dashed lines: trial onset. E, Example session; 568 neurons. Rows: trial-averaged inferred spiking activity of a neuron (frame resolution: 32.4ms). Neurons are sorted based on timing of peak activity. To ensure peaks were not driven by noisy fluctuations, we first computed trial-averaged activity using 50% of trials for each neuron. We then identified the peak-activity time for the trial-averaged response. Finally, these peak times determined the plotting order for the trial-averaged activity for the remaining 50% of trials. This cross-validated approach ensured that the tiling appearance of peak activities was not due to the combination of sorting and false-color-plotting. Red ticks at right: inhibitory neurons (n=45). Red vertical lines: trial events; Duration between events varied across trials; to make trial-averaged traces, traces were separately aligned to each trial event, and then averaged across trials. Next, averaged traces (each aligned to a different trial event) were concatenated. Vertical blue lines: border between the concatenated traces. F, Trial-averaged inferred spiking activity of 4 excitatory (top) and 4 inhibitory (bottom) neurons, for ipsi- (black) and contralateral (green) choices (mean±standard error (SEM); ~250 trials per session). G, Inferred spiking activity for excitatory (blue) and inhibitory (red) neurons. Example mouse; mean±SEM across days (n=46). Each point corresponds to an average over trials and neurons. Inferred spiking activity was downsampled by averaging over three adjacent frames (Methods). H, Distribution of inferred spiking activity 0-97ms before choice (averaged over three frames) for all mice/sessions (41,723 excitatory; 5,142 inhibitory). I, Inferred spiking activity 0-97ms before the choice (averaged over three frames) for individual mice (mean±SEM across days).

Figure 2.. Single-cell and pairwise analyses argue for non-random connections between excitatory and inhibitory neurons.

All panels: blue/red indicate excitatory/inhibitory neurons, respectively. A, Distribution of AUC values (area under the curve) of an ROC analysis for distinguishing choice from the activity of single neurons in an example session. Data correspond to the 97 ms window preceding choice (285 excitatory and 29 inhibitory neurons). Values larger than 0.5 indicate preference for ipsilateral choice; values smaller than 0.5 indicate preference for contralateral choice. Shaded areas: significant AUC values (compared to a shuffle distribution). Distributions were smoothed (moving average, span=5). 5 inhibitory and 24 excitatory neurons (17% and 8%, respectively) were significantly choice selective. B, ROC analysis on 97 ms non-overlapping time windows. Vertical axis: fraction of excitatory or inhibitory neurons with significant choice selectivity; example mouse; mean±SEM across days (n = 45 days). C, Fraction of excitatory and inhibitory neurons that are significantly choice-selective (0-97 ms before the choice) summarized for each mouse; mean±SEM across days (n = 45, 48, 7, 35 sessions per mouse). Star indicates significant difference (t-test; p<0.05); see also Figure S3D. Fraction selective neurons at 0-97ms before choice (median across mice): excitatory: 13%; inhibitory: 16%, (~6 inhibitory and 43 excitatory neurons with significant choice selectivity per session). D, ROC analysis for 97 ms non-overlapping time windows. Time course of normalized choice selectivity (defined as twice the absolute deviation of AUC from chance, given explicitly by 2*∣AUC-0.5∣) for excitatory and inhibitory neurons in an example mouse; mean±SEM across days, n = 45 sessions. E, Average of normalized choice selectivity for excitatory and inhibitory neurons (0-97 ms before choice) summarized for each mouse; mean±SEM across days. “Shuffled” denotes AUC was computed using shuffled trial labels.

Figure 3.. Linear classifiers can predict the animal’s choice with equal accuracy from the activity of excitatory or inhibitory populations.

A, Schematic of decoding choice from all neurons (left), only excitatory (middle), subsampled to the same number as inhibitory neurons, and only inhibitory (right). A linear SVM assigns weights of different magnitude (indicated by lines of different thickness) to each neuron in the population. B, Top: classification accuracy of decoders trained on all neurons (black), subsampled excitatory neurons (blue), and inhibitory neurons (red) (cross-validated; decoders trained on every 97ms time bin; example session; mean±SEM across 50 cross-validated samples). Data are aligned to the animal’s choice (black dotted line). Unsaturated lines : performance on shuffled trials. Bottom: distribution of stimulus onset, offset, go tone, and reward occurrence for the example session above. C, Classification accuracy (0-97 ms before the choice, mean±SEM across days) for real (saturated) and shuffled (unsaturated) data. D, Absolute value of weights for excitatory and inhibitory neurons in decoders trained on all neurons; mean±SEM across days. E, Distribution of classifier weights (decoder training time: 0-97 ms before the choice) for excitatory and inhibitory neurons. Neurons from all mice pooled (42,019 excitatory and 5,172 inhibitory neurons). Shading: standard error. F, Absolute value of weights in the classifier (0-97 ms before choice) for excitatory vs. inhibitory neurons. Mean±SEM across days. Star: p<0.05, t-test. G, Schematic of decoding choice from a population of subsampled excitatory neurons (top) vs. a population with half inhibitory and half excitatory neurons (bottom). H, Classifier accuracy of populations including only excitatory (blue) or half inhibitory, half excitatory neurons (magenta); example session. Classifier trained at each moment in the trial. Traces show mean±SEM (50 cross-validated samples). I, Summary of each mouse (mean±SEM across days) for the decoders (0-97 ms before the choice).

Figure 4.. Classifiers, whether trained on excitatory or inhibitory neurons, show comparable stability during decision formation.

Cross-temporal generalization of choice decoders. A, Classification accuracy of decoders for each pair of training/testing time points, for all neurons (left), subsampled excitatory neurons (middle), or inhibitory neurons (right). Diagonal: same training, testing time (as in Figure 3). Example mouse, mean across 45 sessions. B, Example classification accuracy traces showing how classifiers trained at 0-97 ms before choice generalize across time. Same mouse as in (A), mean±SEM across days C, Decoders are stable in a short window outside their training time. Red indicates that classification accuracy of a decoder tested at the time on the horizontal axis is ≤2 standard deviations of the decoder tested at the training time. Example mouse; mean across days. D, Summary of stability duration for decoders trained from 0-97 ms before choice, using inhibitory neurons (red) or subsampled excitatory neurons (blue). Mean±SEM across days, per mouse.

Figure 5.. Modeling decision circuits with different architectures.

A, Top: Non-selective decision-making model. E₁ and E₂: pools of excitatory neurons, each favoring a different choice, that excite a single pool of non-selective inhibitory neurons (I). Bottom: Classification accuracy of excitatory (blue) and inhibitory (red) neurons as a function of the relative strength of excitatory-to-inhibitory vs. inhibitory-to-excitatory connections. Arrow in this and subsequent panels: parameter value in line with experimental data. B, Top: Signal-selective model. I₁ and I₂: pools of inhibitory neurons connected more strongly to E₁ and E₂, respectively, than to E₂ and E₁; cross-pool connections are weaker than within-pool connections. Bottom: Decoding accuracy of inhibitory and excitatory neurons match at the biologically plausible regime (arrow). Cross-pool connectivity was 25% smaller than within-pool connectivity. C, Top: SNR-selective model: inhibitory neurons connect more strongly to excitatory neurons with high signal to noise ratios. Bottom: Decoding accuracy of inhibitory and excitatory neurons match near the biologically plausible regime (arrow). All plots reflect 50 excitatory and 50 inhibitory neurons out of a population containing 4000 excitatory/1000 inhibitory neurons.

Figure 6.. Pairwise noise correlations are stronger between neurons with the same choice selectivity.

A, Left: Noise correlations (Pearson’s coefficient) for pairs of excitatory-inhibitory neurons with the same (dark green) or opposite (light green) choice selectivity. Middle, right: same as left, but for excitatory-excitatory, and inhibitory-inhibitory pairs, respectively. “Shuffled” denotes quantities were computed using shuffled trial labels. Mean±SEM across days; 0-97 ms before the choice. Same vs. opposite is significant in all cases, except for mouse 3 in EE and II pairs (t-test, p<0.05). B, Example mouse: distribution of noise correlations (Pearson’s correlation coefficients, 0-97 ms before the choice) for excitatory (blue; n=11867) and inhibitory (red; n=1583) neurons. Shaded areas: significance compared to a shuffled control in which trial orders were shuffled for each neuron to remove noise correlations. C, Summary of noise correlation coefficients; mean±SEM across days.

Figure 7.. Noise correlations reduce classification accuracy.

A, Classification accuracy for an example session (0-97 ms before the choice) on neural ensembles of increasingly larger size. Mean±SEM (50 cross-validated samples). Gray: classification accuracy for pseudo-populations. Black: real populations. Both cell types were included (“All neurons”). B, Summary for each mouse; points show mean±SEM across days. Values were computed for the largest neuronal ensemble (the max value on the horizontal axis in A).

Figure 8.. Learning leads to improved choice decoding, increased fraction of choice-selective neurons, and reduced noise correlations in both populations.

A, Decoder accuracy for each training session, for all neurons (left), subsampled excitatory (middle), and inhibitory neurons (right). White vertical line: choice. Rows: average across cross-validated samples; example mouse. Colorbar applies to both plots. B, Scatter plot of classifier accuracy (0-97 ms before choice) vs. behavioral performance (fraction correct on easy trials) for all training days. r: Pearson correlation coefficient (p<0.001 in all plots); same example mouse as in (A). Correlations for behavior vs. classification accuracy for all neurons, excitatory and inhibitory: 0.55, 0.35, 0.32 in mouse 2; 0.57, 0.63, 0.32 in mouse 3. Correlations for behavior vs. choice-signal onset for all neurons, excitatory and inhibitory: −0.60, −0.34, −0.38, in mouse 2; −0.60, −0.27, −0.28 in mouse 3. All values: p<0.05 C, Same as (B), except showing the onset of choice signal (the first moment in the trial that classifier accuracy was above chance, relative to choice onset) vs. behavioral performance. D, Summary of classification accuracy averaged across early (dim colors) vs. late (dark colors) training days. E, Same as (D), but showing choice signal onset (ms). F, Same as (D), but showing pairwise noise correlation coefficients. EI:excitatory-inhibitory; EE:excitatory-excitatory; II:inhibitory-inhibitory. G, Fraction of choice-selective neurons increases over training; average across early (dim colors) and late (dark colors) training days (0-97 ms before the choice). Early days: first few training days in which the animal’s performance was lower than the 20th percentile of performance across all days. Late days: training days in which performance was above the 80th percentile of performance across all days.