Claims 38
From the paper Excitatory and Inhibitory Subnetworks Are Equally Selective during Decision-Making and Emerge Simultaneously during Learning · claims file · checked 2026-10-21: ready
- data-composition
The imaging dataset has 135 sessions from 4 mice (45, 48, 7 and 35 sessions), with on average 480 neurons per session, about 40 inhibitory, 330 excitatory and 100 unclassified; inhibitory neurons are 11% of the population.
Our simultaneous imaging and decision-making dataset includes 135 sessions from 4 mice (45, 48, 7, and 35 sessions per mouse). Median number of trials per session is 213, 253, 264, and 222, for each mouse. On average, 480 neurons were imaged per session, out of which ~40 neurons were inhibitory and ~330 were excitatory
STAR Methods, Quantification and statistical analysis; Results, paragraph 3How it will be judged
Mice are matched to the paper's mouse 1 to 4 by their session counts (45, 48, 7, 35). Neurons are the ROIs kept as good quality; counts per session from the ROI type labels. The session counts per mouse are compared as well; trials per session are counted (a) as all trials in the trial table and (b) as trials with a correct or incorrect choice, the paper not saying which. The inhibitory fraction is read (a) as inhibitory over all neurons and (b) as inhibitory over excitatory plus inhibitory. Reproduced when every value is within tolerance under one reading of each.
- fig1b-psychometric
Choices depend strongly on stimulus rate, at chance at the 16 Hz boundary and more accurate for rates further from it.
Decisions depended strongly on stimulus rate: performance was at chance when the stimulus rate was at the category boundary, and was better at rates further from the category boundary
Results, paragraph 1; Figure 1BHow it will be judged
Figure 1B uses 10 mice; only the 4 imaged mice can be tested here, which is recorded as a deviation. Per mouse, trials of all sessions with a left or right choice; the fraction of "high" choices per stimulus rate, and a logistic regression of the high choice on rate (as the legend's glmfit). The paper gives no test; the fit's slope is the measure, as a positive slope is choice accuracy rising with distance from the boundary. Chance at 16 Hz holds by design (the Methods reward 16 Hz randomly) and the fraction of high choices at 16 Hz is reported. Reproduced when, in every mouse, the slope is positive at p < 0.05.
- figs1b-trial-history
Choice depends on current stimulus strength, on the previous trial's outcome and on the time since the previous trial; the effect of a previous success is stronger after a short ITI, and ITI has no significant effect after a failure.
Top left: strength of the sensory evidence affects the animal's choices: the stronger the evidence, the higher the impact. Top middle: Success of a previous trial also affects animal's decision; the effect is stronger when the previous success occurs after a short ITI (<7sec). Top Right: Same but for previous incorrect trials; the effect of ITI after a failure was not significant
Figure S1B; Results, paragraph 1How it will be judged
The model of STAR Methods eq. 1 per session (glmfit, binomial, logit), predicting a left choice. Stimulus strength as six bins of |rate - 16 Hz| of width 2 Hz (0-2, 2-4, ..., 10-12 Hz; code, behavior/trialHist_logisRegress_setVar.m with vec_ratesdiff = 0:2:12), signed by the side the rate favours; previous success and previous failure each split at an ITI of 7 s (code, vec_iti); trials with ITI over 30 s set aside (code). The ITI is the time from the previous choice to the current stimulus onset (Methods). The figure plots coefficients averaged over animals (10 mice; here 4) with SEM. The paper names no test; the two-tailed t-test of the Methods is used across mice, and across sessions per mouse as a second reading. Reproduced when the stimulus coefficients grow with strength (the largest bin above the smallest), the short-ITI success coefficient exceeds the long-ITI one significantly, and the two failure coefficients do not differ significantly (absence judged on that t-test).
- fig1gi-inferred-spiking-differs
Inferred spiking activity differs significantly between inhibitory and excitatory neurons throughout the trial, and in the bin just before the choice in every mouse.
The magnitude of inferred spiking activity was significantly different for inhibitory vs. excitatory neurons throughout the trial ( Figure 1G ; t-test, p<0.001). Just before the choice (97 ms, average of 3 frames), this difference was clear ( Figure 1H ) and significant for all mice ( Figure 1I , t-test; p<0.001)
Results, paragraph 4; Figure 1G, H and IHow it will be judged
Correct trials (code, excInh_FRs.py). The bin 0-97 ms before the choice; activity averaged over trials per neuron. The paper does not say over what unit its t-test runs; the code (excInh_FRs.py) tests both, so the readings are (a) an unpaired two-tailed t-test across days of the day means (legend of 1I, mean +/- SEM across days) and (b) the same test on neurons pooled over days. The text does not say which cell type is higher; the direction is reported. Reproduced when p < 0.001 in every mouse, and for the example-mouse time course in every choice-aligned bin of at least one mouse under reading (a).
- fig2a-example-session
In an example session (285 excitatory and 29 inhibitory neurons), 24 excitatory (8%) and 5 inhibitory (17%) neurons were significantly choice selective just before the choice.
Data correspond to the 97 ms window preceding choice (285 excitatory and 29 inhibitory neurons) ... 5 inhibitory and 24 excitatory neurons (17% and 8%, respectively) were significantly choice selective
Figure 2A legendHow it will be judged
The session is not named; it is the session with 285 excitatory and 29 inhibitory neurons, if exactly one has these counts. ROC as in the opening comment. The shuffle makes counts vary from run to run, like a random seed; as a standard choice the 50-shuffle procedure is repeated 20 times, and the claim is reproduced when the median count is within tolerance. Not testable if no session has these counts.
- fig2bc-fraction-selective-similar
Excitatory and inhibitory neurons have a similar fraction of significantly choice-selective neurons just before the choice (median across mice 13% and 16%, about 43 excitatory and 6 inhibitory neurons per session); the difference is significant in mouse 4.
Importantly, excitatory and inhibitory neurons were similar in terms of the fraction of choice-selective neurons ( Figure 2B , C ; Figure S3B , C )
Results, Individual excitatory and inhibitory neurons are similarly choice-selective, paragraph 1; Figure 2CHow it will be judged
ROC as in the opening comment, bin 0-97 ms before the choice (Methods). Per session the fraction of each cell type that is selective; per mouse the mean over sessions, then the median over the 4 mice. A claim of absence, judged on the paper's test, an unpaired two-tailed t-test across sessions per mouse at p < 0.05 (Methods and code, ttest2). The paper reports a significant difference in mouse 4 only. Reproduced when the two medians are within tolerance and no significant difference is found in mice 1, 2 and 3; partially reproduced when only the t-tests hold. The per-session counts are read as the median across mice of the mean count per session. The normal-fit criterion is the code's; an empirical p from the 50 shuffles (two-sided, fraction of shuffles at least as far from 0.5) is shown as a variant.
- fig2bd-selectivity-time-course
The fraction of choice-selective neurons and the magnitude of choice selectivity increase gradually over the trial and peak just after the choice.
The fraction of choice-selective neurons ( Figure 2B ) and the magnitude of choice selectivity ( Figure 2D ) gradually increased over the trial, peaking just after the animal reported its choice
Results, Individual excitatory and inhibitory neurons are similarly choice-selective, paragraph 1; Figure 2B and 2DHow it will be judged
ROC per 97 ms bin of the choice-aligned traces, as in the opening comment; per mouse the mean over sessions for each bin and cell type (the figure shows an example mouse, mean over 45 days). "Just after" is not fixed; readings are (a) the peak falls in the first bin after the choice (0-97 ms) and (b) it falls within 300 ms after the choice, the window the Methods give for the confirmation lick. Reproduced when, in at least the example mouse (the one with 45 sessions) and for both measures and both cell types, the earliest choice-aligned bin is below the bin before the choice and the peak lies in the window.
- fig2de-selectivity-magnitude-similar
The magnitude and time course of choice selectivity are similar for excitatory and inhibitory neurons.
as well as the magnitude and time course ( Figure 2D , E ) of choice selectivity
Results, Individual excitatory and inhibitory neurons are similarly choice-selective, paragraph 1; Figure 2D and 2EHow it will be judged
Selectivity 2*|AUC-0.5| from correct trials (code), averaged over the neurons of each cell type per session, bin 0-97 ms before choice. A claim of absence, judged on the paper's test, an unpaired two-tailed t-test across sessions per mouse at p < 0.05 (Methods; code choicePref_ROC_exc_inh_plotsAllMice.m, ttest2). Reproduced when no mouse shows a significant difference; partially reproduced when one mouse does; the mean difference per mouse is reported either way. For the time course the same test is shown per bin.
- figs3d-same-spiking-selectivity
Restricted to excitatory and inhibitory neurons with the same spiking activity, choice selectivity is still similar, and the difference in mouse 4 of Figure 2C disappears.
D, ROC analysis restricted to those excitatory and inhibitory neurons that had the same spiking activity. Choice selectivity is still similar between the two cell types. Note that the significant difference observed for mouse 4 in Figure 2C is absent after controlling for the difference in spiking activity of inhibitory and excitatory neurons
Results, Individual excitatory and inhibitory neurons are similarly choice-selective, paragraph 1; Figure S3DHow it will be judged
Matching as the Methods give for Figure S10D, E and as the code does (choicePref_ROC_exc_inh_plotsAllMice_sameFR.m) per session, the median activity over trials in the bin 0-97 ms before choice for each neuron, the range of medians cut into 50 bins, and only the neurons of both types in the bin that holds the most inhibitory neurons kept. A claim of absence, judged on the unpaired two-tailed t-test across sessions per mouse at p < 0.05 (code, ttest2) on the session-mean selectivity and the session fraction selective. Reproduced when no mouse, mouse 4 included, shows a significant difference.
- figs3c-wider-window
ROC on the window from 200 ms before to 300 ms after the choice gives a larger fraction of selective neurons, 21% of excitatory and 27% of inhibitory (median across mice), about 69 excitatory and 11 inhibitory neurons per session.
Fraction selective neurons at this time window (median across mice): excitatory: 21%; inhibitory: 27%, resulting in approximately 11 inhibitory and 69 excitatory neurons with significant choice selectivity per session
Figure S3CHow it will be judged
As fig2bc-fraction-selective-similar, with activity averaged over the window before the ROC. The code (choicePref_ROC_exc_inh.m, aveAllTimes = -1) averages the bins from 2 before to 3 after the choice bin, 194 ms before to 291 ms after; readings are (a) that and (b) the frames from 200 ms before to 300 ms after the choice. Reproduced when all four values are within tolerance.
- figs3e-correct-error-correlation
Choice selectivity on correct and on error trials is positively correlated across neurons, so most neurons reflect the choice more than the stimulus, and within each mouse this correlation is very similar for excitatory and inhibitory neurons; it is weaker in mouse 4.
Importantly, however, within each subject, this correlation was very similar for excitatory vs. inhibitory neurons ( Figure S3E ), suggesting that the tendency for neurons to be modulated by the choice vs. the stimulus was similar in excitatory and inhibitory neurons
Results, Individual excitatory and inhibitory neurons are similarly choice-selective, paragraph 2; Figure S3EHow it will be judged
Per session and neuron, AUC on correct trials and AUC on error trials (same left against right labels, bin 0-97 ms before choice); Pearson r across neurons per cell type; per mouse the mean r across sessions. Sessions need at least 10 error trials of each choice (the code's ROC threshold, applied to error trials; a reading). Reproduced when the mean r is positive for both types in every mouse, the excitatory and inhibitory means do not differ in any mouse by the paper's unpaired two-tailed t-test across sessions at p < 0.05, and mouse 4 has the lowest mean r for at least one type.
- fig3bc-decoding-similar
Choice can be decoded from population activity well above chance, and inhibitory populations decode it as accurately as excitatory populations of the same size, in every mouse.
Classification accuracy of inhibitory neurons was well above chance and was very similar to that of excitatory neurons ( Figure 3B , red and blue traces overlap; Figure S4 ... Similar classification accuracy for excitatory and inhibitory populations was observed in all subjects ( Figure 3C )
Results, Choice is decoded with equal accuracy from both excitatory and inhibitory populations, paragraph 2; Figure 3B and 3CHow it will be judged
Decoders as in the opening comment, correct trials, bin 0-97 ms before the choice, per session the accuracy averaged over the 50 repetitions (and over the 50 excitatory subsamples). Per mouse, accuracy is above chance when it exceeds the shuffled control by an unpaired two-tailed t-test across sessions at p < 0.05. Equality is a claim of absence, judged on the paper's unpaired two-tailed t-test across sessions per mouse (code, svm_excInh_trainDecoder_eachFrame_plots_sumAllMice.py, ttest_ind) at p < 0.05. Reproduced when in every mouse both cell types are above chance and do not differ; partially reproduced when they differ in one mouse. Mean accuracies are reported.
- fig3b-decoding-time-course
Accuracy of the all-neuron choice decoder grows after stimulus onset and peaks at the time of the choice.
Classification accuracy gradually grew after stimulus onset and peaked at the time of the choice ( Figure 3B , black)
Results, Choice is decoded with equal accuracy from both excitatory and inhibitory populations, paragraph 1; Figure 3BHow it will be judged
All-neuron decoder trained and tested on each 97 ms bin of choice-aligned traces, correct trials; per mouse the mean over sessions. Figure 3B is one session; tested on the mouse means. "At the time of the choice" is read as (a) the peak in one of the two bins around the choice (97 ms before to 97 ms after) and (b) the peak within 300 ms of the choice. Reproduced when in every mouse the earliest choice-aligned bin is lower than the bin before the choice and the peak lies in the window.
- fig3df-weights-similar
In the decoder trained on all neurons, excitatory and inhibitory neurons receive weights of matched magnitude throughout the trial and with very similar distributions.
The weight magnitudes of excitatory and inhibitory neurons were matched for the entire trial ( Figure 3D ), and the distributions of weights was very similar ( Figures 3E , F )
Results, Choice is decoded with equal accuracy from both excitatory and inhibitory populations, paragraph 3; Figure 3D, E and FHow it will be judged
All-neuron decoder at the best C, weights averaged over the 50 repetitions, each session's weight vector normalised to unit length (Methods); mean |w| per cell type per session at 0-97 ms before choice and per bin. Figure 3F marks a significant difference (star) in some mice that the text does not name. Judged on the paper's unpaired two-tailed t-test across sessions per mouse at p < 0.05. Reproduced when no mouse differs, partially reproduced when one or two mice do, not reproduced otherwise; the pooled counts are compared with the expected values, and the mean |w| per mouse is reported.
- fig3hi-mixed-no-advantage
A population of half excitatory and half inhibitory neurons decodes choice with accuracy similar to that of an excitatory population of the same size, so a mixed population offers no major advantage.
Classification accuracy was similar for both decoders ( Figure 3H , I ), arguing that a mixed population offers no major advantage to decoding
Results, Choice is decoded with equal accuracy from both excitatory and inhibitory populations, paragraph 4; Figure 3H and 3IHow it will be judged
As the code (svm_excInh_trainDecoder_eachFrame.py, excInhHalf and allExc2inhSize) the mixed population is all n inhibitory neurons plus a random subsample of n excitatory neurons, and the control is 2n excitatory neurons, both over 50 subsamples; correct trials, bin 0-97 ms before choice. A claim of absence, judged on the paper's unpaired two-tailed t-test across sessions per mouse at p < 0.05. Reproduced when no mouse shows a significant difference between the two decoders in either direction; the difference per mouse is reported.
- figs5-fluorescence-decoding
Excitatory and inhibitory populations are equally choice selective when decoders are trained on fluorescence traces.
Excitatory and inhibitory populations were equally choice selective even when the analysis was performed on raw calcium traces ( Figure S5 )
Results, Choice is decoded with equal accuracy from both excitatory and inhibitory populations, paragraph 2; Figure S5How it will be judged
As fig3bc-decoding-similar, with the CNMF denoised fluorescence traces (Figure S5A uses those of Figure 1D middle; code option do_Ctrace) in place of inferred spikes. The legend calls them fluorescence traces and the text raw calcium traces; readings are (a) the denoised traces and (b) Delta F/F traces. Same test and outcome rule.
- figs6ab-previous-choice-stimulus
Population activity is weakly informative about the previous trial's choice and about the stimulus category (the stimulus category weakly except in mouse 4), and excitatory and inhibitory populations are similarly selective for both.
Excitatory and inhibitory populations were similarly selective for previous choice ( Figure S6A ). The population activity was also somewhat informative about whether the stimulus was above or below the category boundary ( Figure S6B ). Again, excitatory and inhibitory populations were similarly selective
Results, Choice is decoded with equal accuracy from both excitatory and inhibitory populations, paragraph 5; Figure S6A and S6BHow it will be judged
Decoders as in the opening comment. Stimulus category uses stimulus-aligned trials with equal numbers of left and right choices in each category (Methods). For previous choice the paper gives no alignment or window; readings are (a) choice-aligned traces, bin 0-97 ms before the choice, and (b) the first bins of the trial, before stimulus onset (the code's trialHistAnalysis option, which uses the ITI period). Above chance by the unpaired two-tailed t-test across sessions against the shuffle; equality judged on the same test between cell types. Reproduced when, for both decoders, the all-neuron decoder is above chance in most mice and no mouse shows a significant excitatory to inhibitory difference; partially reproduced when one mouse does. The Figure S6B legend excepts mouse 4 from the weak stimulus-category signal, not from the excitatory to inhibitory comparison; the all-neuron accuracy per mouse is reported. The bin for stimulus category is (a) 0-97 ms before stimulus offset and (b) the bin of peak all-neuron accuracy; the paper fixes neither.
- figs6c-outcome-decoding
Population activity is strongly selective for trial outcome, reaching about 80% accuracy (median across mice) about 400 ms after the confirmation lick, with inhibitory neurons slightly more accurate than excitatory neurons after reward.
Classification accuracy gradually increases and reaches 80% (median across mice) approximately 400ms after the animal confirms his choice (Figure S1A). Inhibitory neurons showed slightly higher selectivity for the outcome
Results, Choice is decoded with equal accuracy from both excitatory and inhibitory populations, paragraph 5; Figure S6CHow it will be judged
Outcome-aligned trials (aligned to the confirmation lick, the reward time), with equal numbers of left and right choice trials for each outcome (Methods); decoders otherwise as in the opening comment. The all-neuron accuracy at the bin containing 400 ms after the confirmation lick, mean over sessions per mouse, median over mice, within tolerance. Inhibitory above subsampled excitatory in every mouse (the text's "small but consistent difference"), in the mean over bins after the confirmation lick (the paper's "after reward delivery"); the end of the window is open, read as (a) 500 ms after the confirmation lick and (b) the end of the outcome-aligned traces. The paper gives no test for this difference; the unpaired two-tailed t-test across sessions per mouse is reported. Reproduced when both hold.
- figs6de-incorrect-trials
Decoders trained on correct trials predict the choice on incorrect trials with high accuracy, slightly lower than on correct trials, and the reduction is similar for excitatory and inhibitory populations in every mouse.
Importantly, for all mice, the change in classification accuracy was quite similar for excitatory and inhibitory neurons (right), indicating that both populations reflect choice vs. stimulus to a comparable degree
Figure S6D and S6EHow it will be judged
Decoders trained on correct trials at each bin and tested on held-out correct trials and on incorrect trials (code, testIncorr); bin 0-97 ms before choice. Accuracy on incorrect trials is high when above the shuffle by the unpaired two-tailed t-test across sessions per mouse. The reduction (correct minus incorrect) of inhibitory and subsampled excitatory decoders is compared by the same test; a claim of absence. Reproduced when incorrect-trial accuracy is above the shuffle in every mouse and the reductions do not differ significantly in any mouse; mouse 4 should have the largest reduction for the all-neuron decoder.
- fig4bd-stability-similar
Choice decoders lose accuracy at a similar rate away from their training time for excitatory and inhibitory populations, and their stability duration is indistinguishable.
This drop in accuracy occurred at a similar rate for excitatory and inhibitory populations ( Figure 4B ). To quantify this, we determined the time window over which classifier accuracy remained within 2 standard deviations of the accuracy at the training time ( Figure 4C ). This was indistinguishable for excitatory and inhibitory neurons ( Figure 4D ; Figure S7A )
Results, Choice is decoded with equal accuracy from both excitatory and inhibitory populations, paragraph 6; Figure 4B, C and D; Figure S7AHow it will be judged
Decoders trained at each bin and tested at every bin on held-out trials (Methods), correct trials. For the decoder trained at 0-97 ms before choice, a test bin counts as stable when its accuracy is at least the accuracy at the training bin minus 2 standard deviations across the 50 repetitions (Methods and Figure 4C legend). The code (svm_excInh_trainDecoder_eachFrame_stabTestTimes_setVars_post.py) uses 2 standard errors (SD over the square root of the number of repetitions); since the Methods say standard deviations, the code's choice is shown as a variant. Stability duration is the number of stable bins times 97 ms (code, bins not required to be contiguous; contiguous bins around the training bin shown as a variant). Excitatory durations are averaged over the 50 subsamples. Sessions where the training-bin accuracy does not exceed the shuffle are left out (code, one-tailed t-test across repetitions; the code's alpha 0.001 and the Methods' general 0.05 are both readings). A claim of absence, judged on the code's test, an unpaired two-tailed t-test of the four mouse means (svm_excInh_trainDecoder_eachFrame_stabTestTimes_plots.py, ttest_ind across mice); the per-mouse test across sessions is shown as a variant. Reproduced when that test finds no difference.
- figs7c-angle-stability
The angle between decoders trained at different times shows a similar time course of stability for excitatory and inhibitory populations.
An alternate method for assessing stability, computing the angle between the weights of pairs of classifiers trained at different time windows, likewise suggested that excitatory and inhibitory populations are similarly stable
Results, Choice is decoded with equal accuracy from both excitatory and inhibitory populations, paragraph 6; Figure S7CHow it will be judged
Angle in degrees between the weight vectors (averaged over repetitions) of decoders trained at each pair of bins, per session and cell type (excitatory over the 50 subsamples). The paper gives no summary or test; readings are (a) the mean angle over all pairs of bins and (b) the mean angle between the decoder at 0-97 ms before choice and each other bin. A claim of absence, judged on the paper's unpaired two-tailed t-test across sessions per mouse at p < 0.05. Reproduced when no mouse shows a significant difference.
- fig5a-nonselective-model
In a model with non-selective inhibition, inhibitory neurons decode the choice less accurately than excitatory neurons for every ratio of excitatory-to-inhibitory to inhibitory-to-excitatory connection strength.
For this value (and indeed for all other values), inhibitory neurons had lower classification accuracy than excitatory neurons ( Figure 5A , bottom; Figure S8 , left), inconsistent with our experimental results
Results, Modeling rules out decision circuits with non-selective inhibition, paragraph 2; Figure 5A; Figure S8How it will be judged
A simulation, not a test on recorded data. The linearised rate model of the STAR Methods with the printed parameters (sigma 1.25, delta 0.75, wEE 0.25, wII -2, wIE(0) 0.87, wEI(0) -0.87, c 0.1, NE 4000, NI 1000, n 50, h0 0.1), decoding accuracy Phi(sqrt(S/N)/2) from the input-averaged S/N, one network per point, the product wIE*wEI held at the reference while the ratio varies. The range of the ratio is not given; the range is read from Figure 5 if the PDF is supplied, otherwise 0.1 to 10 on a log scale (a standard choice). Reproduced when inhibitory accuracy is below excitatory at every value.
- fig5bc-selective-models-match
In signal-selective and SNR-selective models, inhibitory and excitatory neurons have matched decoding accuracy near the biologically plausible regime of equal excitatory-to-inhibitory and inhibitory-to-excitatory strengths; in all three models, accuracy rises faster for inhibitory than for excitatory neurons as the excitatory-to-inhibitory strength increases.
In this circuit architecture, inhibitory and excitatory neurons had matched classification accuracy when the connection strength from excitatory to inhibitory neurons was about the same as the strength from inhibitory to excitatory ... Thus, in all panels of Figure 5 , classification accuracy increases faster for inhibitory neurons than excitatory ones as the excitatory to inhibitory connection strength increases
Results, Modeling rules out decision circuits with non-selective inhibition, paragraphs 3 and 4; Figure 5B and 5CHow it will be judged
A simulation, not a test on recorded data. As fig5a-nonselective-model; signal-selective with rho 0.75 and the input split by population, averaged over 25 networks and 100 random picks of 50 neurons; SNR-selective with W_IE scaled by (r_j/NE)^4 and rescaled to the same mean. Reproduced when, for each model, the excitatory and inhibitory accuracy curves cross within a factor of 2 of a ratio of 1 (a reading of "about the same" and "near"), and when inhibitory accuracy rises faster than excitatory with the ratio in all three models.
- figs9-signal-selective-best
The signal-selective model has the highest decoding accuracy of the three models at every connection-strength ratio and level of differential correlations, in both cell types.
For all values of this parameter and in both cell types, signal-selective model has the highest classification accuracy for decoding the animal's choice
Results, Modeling rules out decision circuits with non-selective inhibition, paragraph 6; Figure S9How it will be judged
A simulation, not a test on recorded data. As fig5bc-selective-models-match, with differential correlations added as Sigma to Sigma + epsilon Delta h Delta h^T for epsilon 0, 0.25 and 0.5 (Figure S8 legend; which Sigma is garbled in the text, read as both Sigma_EE and Sigma_II, and as Sigma_EE only, two readings). The paper says "significantly" without a test; the claim is judged on the ordering. Reproduced when the signal-selective accuracy is highest at every ratio, epsilon and cell type.
- fig6a-same-selectivity-correlations
Noise correlations are stronger between neurons with the same choice selectivity than between neurons with opposite selectivity, for excitatory-inhibitory, excitatory-excitatory and inhibitory-inhibitory pairs, significant in every mouse except mouse 3 for EE and II pairs; the same holds in the spontaneous pre-trial period.
Indeed, neurons with the same choice selectivity had stronger correlations ( Figure 6A )
Results, Correlations are stronger between similarly tuned neurons, paragraph 1; Figure 6A; Figure S10AHow it will be judged
As the Methods and code (imaging/corr_excInh_plots.m and corr_excInh_setVars.m) neurons are "ipsi" or "contra" preferring by the sign of AUC - 0.5 at 0-97 ms before choice (AUC from correct trials; significance not required); noise correlations from the same bin on trials with a choice, after subtracting each choice's mean response from its trials; Pearson r per pair; each neuron's r averaged over its partners of the given type and preference. Spontaneous correlations use the bin 0-97 ms before the initiation tone, without subtraction. The code tests same against opposite with an unpaired t-test on neurons pooled over sessions (ttest2); the across-session test (the legend's mean +/- SEM across days) is shown as a variant. Reproduced when same exceeds opposite significantly (p < 0.05) for EI pairs in all four mice and for EE and II pairs in mice 1, 2 and 4, and spontaneous correlations show the same pattern by the same test (the text's "the same correlation effects"; Figure S10A gives no test of its own); partially reproduced when this holds for at least two of the three pair types.
- fig6bc-inhibitory-correlations-higher
Inhibitory pairs have significantly higher noise correlations than excitatory pairs, with intermediate values for excitatory-inhibitory pairs, also for spontaneous correlations and for neurons matched in spiking activity.
Inhibitory pairs had significantly higher noise correlations compared to excitatory pairs ... This was true even when we restricted the analysis to inhibitory and excitatory neurons with the same inferred spiking activity ( Figure S10D , E ). Finally, similar to previous reports ... we found intermediate correlations for pairs consisting of one inhibitory neuron and one excitatory neuron ( Figure S10B , C )
Results, Correlations are stronger between similarly tuned neurons, paragraph 2; Figure 6B and 6C; Figure S10B to S10EHow it will be judged
Per-neuron noise correlation as in fig6a-same-selectivity-correlations, averaged over partners of the same type (EE, II) or the other type (EI), bin 0-97 ms before choice; spontaneous as there; matched activity by the 50-bin "max bin" rule of the Methods. Readings for the test, an unpaired two-tailed t-test (Methods), are (a) across sessions per mouse on session means (Figure 6C, mean +/- SEM across days) and (b) on neurons pooled over sessions (code, utils/plotHist.m). Reproduced when II exceeds EE significantly in every mouse for noise, spontaneous and matched-activity correlations, with EI between them in the mean; the expected counts are matched by one mouse (the example mouse is not named).
- fig7-noise-correlations-limit-decoding
Removing noise correlations by building pseudo populations increases choice decoding accuracy of the full population in every mouse.
In all mice, removing noise correlations resulted in a consistent increase in classification accuracy ( Figure 7B ; filled vs. open circles), establishing that noise correlations limit population accuracy in PPC
Results, Noise correlations limit decoding accuracy, paragraph 2; Figure 7A and 7BHow it will be judged
All neurons, bin 0-97 ms before choice; neurons ranked by |AUC - 0.5| and added one at a time from the most selective (Methods); the AUC used for ranking is from all trials with a choice in the code (svm_excInh_trainDecoder_eachFrame.py), and from correct trials as a variant. Pseudo populations shuffle trial order independently for each neuron within left-choice and within right-choice trials (Methods). The change is pseudo minus real accuracy at the largest ensemble (Methods; code changeCA_shflTrsEachN). Trials are correct trials (as the other Figure 3 decoders); all trials with a choice is a second reading. Reproduced when the mean change across sessions is positive in every mouse and significant by the two-tailed t-test across sessions at p < 0.05, read as (a) a one-sample test of the per-session changes and (b) an unpaired test of pseudo against real accuracy, in every mouse.
- fig8ab-learning-accuracy
During learning, choice decoding accuracy increases and correlates with behavioural performance across sessions, for all, excitatory and inhibitory populations (r = 0.55, 0.35, 0.32 in mouse 2 and 0.57, 0.63, 0.32 in mouse 3).
Classification accuracy of the choice decoder increased consistently as animals became experts in decision-making ( Figure 8A , left; Figure 8D , black), leading to a strong correlation between classifier accuracy and mouse performance over training ( Figure 8B , left)
Results, Selectivity increases in parallel in inhibitory and excitatory populations during learning, paragraph 1; Figure 8A, 8B and 8DHow it will be judged
The 3 mice imaged through learning (by Figure S1C, mice 1, 2 and 4; the code's fni16, fni17, fni19). A decoder per session, bin 0-97 ms before choice. Trials are all outcomes in the code's learning script (svm_excInh_trainDecoder_eachFrame_beh_sumAllMice_plots.py, corrTrained = 0); readings are (a) all trials with a choice, balanced only by choice, and (b) equal numbers of correct and incorrect trials within each choice (the code's decodeChoiceEqCorrIncorr option). Performance is the fraction correct on easy trials (Figure 8B legend); easy as the code (svm_excInh_trainDecoder_eachFrame_stabAngles_setVars_behCA.py, thQStimStrength = [3]), the outermost of four equal bins of |rate - 16 Hz| on each side. The legend's "mouse 2" and "mouse 3" can be read as (a) the second and third of the three learning mice in the paper's numbering (mice 2 and 4) and (b) the paper's mice 2 and 3; the example mouse is the remaining one. Reproduced when in each learning mouse r is positive with p < 0.05 for all three populations and the six reported values are within tolerance under one reading of the mouse names; partially reproduced when only the signs and p-values hold.
- fig8ce-choice-signal-onset
The choice signal emerges progressively earlier in the trial as mice become experts, so the onset of above-chance decoding correlates negatively with performance (r = -0.60, -0.34, -0.38 in mouse 2 and -0.60, -0.27, -0.28 in mouse 3 for all, excitatory and inhibitory).
This resulted in a negative correlation between mouse performance and the onset of super-threshold decoding accuracy ( Figure 8C , left; Figure 8E , black)
Results, Selectivity increases in parallel in inhibitory and excitatory populations during learning, paragraph 1; Figure 8C and 8EHow it will be judged
Decoders and mice as fig8ab-learning-accuracy, at every bin of choice-aligned traces. Per session and bin, accuracy against the shuffle by a t-test across the 50 repetitions; the onset is the start of the first run of significant bins lasting at least 500 ms (5 bins, code th1), in ms from the choice; sessions without such a run are left out (Methods). The Methods give p < 0.05; the code (svm_excInh_trainDecoder_eachFrame_plots.py) uses 0.001, shown as a variant. The code tests excitatory data against the inhibitory shuffle at one line (line 600); the excitatory shuffle is used, as the Methods describe. Reproduced when the correlation is negative with p < 0.05 for all three populations in each learning mouse and the six values are within tolerance under one reading of the mouse names; partially reproduced when only the signs and p-values hold.
- fig8de-learning-parallel
The choice signal emerges at the same time in excitatory and inhibitory populations, and its magnitude and timing are matched for the two cell types throughout learning.
Importantly, the choice signal emerged at the same time in both populations, and its magnitude and timing was matched for the two cell types throughout learning ( Figure 8A - C , middle, right; Figure 8D - E , blue, red)
Results, Selectivity increases in parallel in inhibitory and excitatory populations during learning, paragraph 1; Figure 8A to 8EHow it will be judged
Early and late days as the Methods give, by behavioural performance below the 20th or above the 80th percentile of all days; readings are (a) every day beyond the percentile (code, perc_thb = [20, 80]) and (b) only the run of days at the start (early) or end (late) of training beyond it (the Methods' "initial" and "last" days). Accuracy (bin 0-97 ms before choice) and onset of inhibitory and subsampled excitatory decoders in early and in late days. A claim of absence, judged on the paper's unpaired two-tailed t-test across days per mouse at p < 0.05. Reproduced when, in every learning mouse, neither accuracy nor onset differs between the cell types in early or late days, and late accuracy exceeds early accuracy for both types (Figure 8D) by the same test.
- figs13c-equal-trials
The learning-related increase in decoding accuracy remains when the number of trials is equal in every session.
This was not due to the presence of more correct trials in later sessions: an improvement in classification accuracy was clear even when the number of correct trials was matched for each session ( Figure S13C )
Results, Selectivity increases in parallel in inhibitory and excitatory populations during learning, paragraph 1; Figure S13CHow it will be judged
Decoders trained on correct trials only (legend), with the same number of trials in every session of a mouse, the minimum over its sessions (code, same_HRLR_acrossDays). Late accuracy above early accuracy for all neurons by the unpaired two-tailed t-test across days at p < 0.05 in every learning mouse; early and late days as fig8de-learning-parallel.
- figs12-licking-running
Licking and running do not explain the learning-related changes; centre-spout licks before left and right choices are similar early and late in training, and decoding remains accurate in sessions where running did not differ between choices.
Nonetheless, when we restricted our analysis to days in which running distance was indistinguishable for the two choices (0-97 ms before the choice, t-test, P>0.05), classifiers could still accurately predict the choice ( Figure S12D )
Results, Selectivity increases in parallel in inhibitory and excitatory populations during learning, paragraph 2; Figure S12How it will be judged
Licks to the centre spout in the 250 ms before the choice, as a lick-rate trace per choice (the bin width is not given; 10 ms is a standard choice that cannot change the sign of a correlation much, shown with 1 ms as a variant); per session the Pearson r of the left-choice and right-choice traces; early against late days (as fig8de-learning-parallel) by the unpaired two-tailed t-test, a claim of absence. Running distance from the wheel encoder in the bin 0-97 ms before choice, left against right by a t-test per session; on sessions with P > 0.05, all-neuron accuracy above the shuffle. Reproduced when lick correlations do not differ between early and late days in any learning mouse and accuracy is above the shuffle in the selected sessions of every mouse.
- fig8f-correlations-decrease
Pairwise noise correlations are higher early in training than late, for all combinations of neuron types.
Pairwise correlations in neural activity were higher early in training, when mice were novices, compared to late in training, when mice approached expert behavior ( Figure 8F , unsaturated colors above saturated colors). This was observed for all combinations of neural pairs ( Figure 8F )
Results, Selectivity increases in parallel in inhibitory and excitatory populations during learning, paragraph 3; Figure 8FHow it will be judged
Noise correlations as fig6bc-inhibitory-correlations-higher (EE, II, EI), early and late days as fig8de-learning-parallel. The paper gives no test; the Methods' two-tailed t-test (unpaired, across days per mouse) at p < 0.05 is used. Reproduced when early exceeds late significantly for all three pair types in every learning mouse; partially reproduced when early exceeds late in the mean for every type and mouse but not significantly in all.
- figs13ab-pseudo-learning
Decoding accuracy of pseudo populations, which lack noise correlations, also increases significantly with training, so reduced noise correlations do not alone explain the improved accuracy.
Interestingly, a significant increase in the classification accuracy of pseudo populations was present ( Figure S13A , B )
Results, Selectivity increases in parallel in inhibitory and excitatory populations during learning, paragraph 3; Figure S13A and S13BHow it will be judged
Pseudo populations as fig7-noise-correlations-limit-decoding, for all, subsampled excitatory and inhibitory decoders, bin 0-97 ms before choice, trials as fig8ab; early and late days as fig8de-learning-parallel. Reproduced when late exceeds early by the unpaired two-tailed t-test across days at p < 0.05 in every learning mouse for the all-neuron decoder, and in the same direction for both cell types.
- fig8g-fraction-selective-triples
The fraction of choice-selective neurons increases about threefold from early to late training, in both excitatory and inhibitory neurons.
Indeed, the fraction of choice-selective neurons increased threefold during training, in both excitatory and inhibitory neurons ( Figure 8G )
Results, Selectivity increases in parallel in inhibitory and excitatory populations during learning, paragraph 3; Figure 8GHow it will be judged
ROC as in the opening comment, bin 0-97 ms before choice; early and late days as fig8de-learning-parallel; per mouse the mean fraction over early and over late days. The ratio can be read (a) per mouse and cell type, then the mean over mice, and (b) on fractions pooled over the three mice. Reproduced when, for both cell types, the ratio is within tolerance under one reading and late exceeds early by the unpaired two-tailed t-test across days in every learning mouse (code choicePref_ROC_exc_inh_plotsAllMice.m, ttest2); partially reproduced when the increase is significant but the ratio outside tolerance.
- figs2-temporal-dynamics-similar
Excitatory and inhibitory neurons have similar temporal dynamics, with similar distributions of peak-activity times across the trial.
Diverse temporal dynamics were evident in both cell types ( Figure 1E , F ) and did not appreciably differ between the two ( Figure S2 )
Results, paragraph 4; Figure S2How it will be judged
Per session, the fraction of neurons of each type whose trial-averaged activity peaks in each 100 ms window, traces aligned to each trial event and concatenated (Figure S2 and 1E legends); the alignment windows around each event are not given, and the windows used by the dataset's trial-segmented responses are used. The paper gives no test; the Methods' unpaired two-tailed t-test across sessions per mouse in each window, a claim of absence. With many windows, readings are (a) uncorrected p < 0.05 and (b) Bonferroni over windows. Reproduced when, in every mouse, the number of windows that differ is no more than expected by chance under (a) (binomial test at 0.05) and none differs under (b).
- figs14-neuropil
CNMF removes neuropil contamination as well as annulus subtraction; for an example component the two traces correlate with r = 0.96.
The comparison showed a very high degree of similarity between the two traces ( Figure S14 ; example component; r=0.96)
STAR Methods, Neuropil Contamination removal; Figure S14How it will be judged
Needs the raw movies; footprints binarised at 20% of their maximum, annulus at 3 pixels from the mask and 7 pixels wide, background subtracted, compared with the CNMF Delta F/F trace (Methods). The example component is not named; reproduced when the median r over components with annuli free of other sources is within tolerance (r cannot exceed 1, so the tolerance is the room to 1).
- figs15-labeling-specificity
In GAD-Cre;Ai14 mice 98.2% of tdTomato neurons in PPC express GABA, and all GABA neurons express tdTomato.
98.2 % of tdTomato neurons expressed GABA, and all of the GABAergic neurons expressed tdTomato
STAR Methods, Immunofluorescence staining for TdTomato and GABA; Figure S15How it will be judged
Needs the immunostained coronal sections (5 sections including PPC). Counts of tdTomato, GABA and double-labelled cells; tolerance is the room to 100%. Reproduced when the fraction is within tolerance and no GABA cell lacks tdTomato.
Pipeline
- Paper
- done (1 of 1)
- Claims
- done
- Claims check
- done (ready)
- Attempt
- todo
- Questions
- none
- Verified
- none