Coexistence of state, choice, and sensory integration coding in barrel cortex LII/III

During perceptually guided decisions, correlates of choice are found as upstream as in the primary sensory areas. However, how well these choice signals align with early sensory representations, a prerequisite for their interpretation as feedforward substrates of perception, remains an open question. We designed a two alternative forced choice task (2AFC) in which male mice compared stimulation frequencies applied to two adjacent vibrissae. The optogenetic silencing of individual columns in the primary somatosensory cortex (wS1) resulted in predicted shifts of psychometric functions, demonstrating that perception depends on focal, early sensory representations. Functional imaging of layer II/III single neurons revealed mixed coding of stimuli, choices and engagement in the task. Neurons with multi-whisker suppression display improved sensory discrimination and had their activity increased during engagement in the task, enhancing selectively representation of the signals relevant to solving the task. From trial to trial, representation of stimuli and choice varied substantially, but mostly orthogonally to each other, suggesting that perceptual variability does not originate from wS1 fluctuations but rather from downstream areas. Together, our results highlight the role of primary sensory areas in forming a reliable sensory substrate that could be used for flexible downstream decision processes.

C, Heading direction predicts the correct target side for easy (red), middle (magenta) and difficult trials (blue).D, Body movement extracted from video analysis (see methods).Left: trial sorting according to nose reaction time.Right, top row: body movement during impulsive trials.Right, bottom row: body movements during late response trials.p-value indicates difference of movement between the three behavioral categories (Friedman test, n = 7 animals).Error shades represent a 95% confidence interval following a normal distribution on a trial-by-trial basis (n= 30300 trials from 7 animals).In the initial step, linear regression is independently applied to each neuron (Rn representing the activity of neuron n) to ascertain the sensory features it represents.This regression process identifies the weights ßF1,n, ßF2,n, ßF1xF2,n, which each of the three sensory features (F1, F2 and F1xF2) drive the activity of neuron n. ß0,n represents the baseline activity in the model, and Ɛ represents the non-fitted residuals.In the subsequent step, the activities of multiple neurons are pooled using the same sensory weights that drive them.These weighted pools embody the latent representations of stimulus features encoded by the populations of neurons: Rw1 represents the population response to the stimulation frequency of whisker 1 (F1), Rw2 corresponds to the population response to the stimulation frequency of whisker 2 (F2), and Rw1xw2 carries the population response to the cross-whisker interaction.These latent representations are then leveraged for various analyses, including decoding, exemplified here as neurometric analysis, or for the trial-by-trial representational geometry of sensory versus choice signals, as depicted in Figures 6  and 7. Additional regression variables, namely choices (C1 and C2) or engagement (Eng), are incorporated explicitly in the models presented in Figure 6 and 7 respectively, to generate additional latent representations of choices and engagement ( not represented here, described in the methods).

Figure S6: Multi-whisker (MW) suppression improves decoding performance at the single neuron level, and has little net effect on the whole population average.
A, Discrimination of F1>F2 over the stimulus space depends on selectivity and MW suppression across neurons.Discrimination is measured as AUC of F1>F2 over the entire stimulus space, as described for Fig 4E .The pixel colors represent the average AUC value for neurons in the bin.The neuronal population from all animals is split into bins defined in two dimensions on their MW suppression (BF1xF2; y-axis) and their absolute selectivity index (SI; x-axis).Note the increase of AUC along both x and y axis, suggesting the MW suppression improves discriminability independent of its relationship with whisker selectivity.B, Decoding performance depends on selectivity and MW suppression.Decoding performance is quantified as the fraction of correctly decoded trials.Decoding was performed with a GLM using logit as the link function and assuming binomial distribution of F1>F2.Binning performed as in A. Please note the increase of AUC along both the x and y axis.C-D, Neural population pools Rw1 and Rw2 respond mostly to whisker 1 and 2 but also slightly to the adjacent whisker.Rw1 and Rw2 show sublinear integration of the two whiskers stimulated simultaneously.Rw1 and Rw2 were computed from two independent pools of neurons based on the sign of Bf1 -Bf2.Error shades represent s.e.m. across FOV.Firing rate and fitted firing rate in the middle and right plots are averaged across FOV.Note that the model with interaction does not fully account for the population firing rate which might be due to the diverse tuning within the population, including MW enhancement (Fig. 4b).

Figure S7: Impact of multi-whisker (MW) interaction term on (sub)-population decoding performance.
A, Schematic procedure for decoding with or without interaction.When decoding without interaction, the regression variable F1xF2 is shuffled from trial to trial in the model without interaction.Each neurons activity is fitted trial per trial as described in Fig. 4. Activity of all neurons (or subpopulations) is combined into Rw1, Rw2 and Rw1xw2, (respectively latent representations of F1, F2 and F1xF2).The target side (defined by the sign of F1-F2) is decoded separately by three combinations of latent variables.Decoding is performed using a GLM with logistic regression.B, Comparison of decoding with or without interactions in different subpopulations of neurons (rows) with decoding of different combinations of latent variables (columns).Subpopulations contain 10 % of neurons with most MW enhancement (top), 10 % null MW interaction (middle) or 10 % most MW suppression (bottom).The interaction term significantly improves decoding performance for the subpopulation of neurons with MW suppression.P-value is computed with a linear mixed effect model across FOV using mice identity as grouping variable.B, Reversed somatotopy for AUROC choice as compared to AUROC target.A LME model is used to test dependence of AUROC on position along the inter-barrel-axis (i.e. the diagonal between W1 and W2 barrel center).For graphical display, we split all neurons in ten bins of equal size.P values (p) and estimates of effect size (b) are computed with a LME model, using n = 3118 neurons from 5 animals (grouping variable).C, AUROC choice as a function of AUROC target.Statistical significance is tested independently for each neuron as being <0.025 or >0.025 from a distribution of AUROC with shuffled trial identity.As described in the text, we find a bias with more neurons preferring choice 1 (AUROC <0.5).This illustrate that most neurons code for either choice (magenta) or sensory target (blue), with a fraction of neurons having significant sensory and choice selectivity (red).D, Three scheme of possible sensory/choice coding across the neuronal population.Red areas represent significant selectivity for both sensory and choice variables.Our data indicates that coding of the two variables is orthogonal with mixed selectivity.Neural activity during engaged and disengaged states, in the sensory and engagement dimensions.Left: engagement versus W1 neural representation (Rw1).Right: engagement versus W2 neural representation (Rw2).Same method and description as in Fig. 7e.but both F1>F2 and F2>F1 trials are included in the same panel.Bars represent transition from engaged and disengaged trials.Note that all representational angles are tilted to the right, showing an increase in response of Rw1 and Rw2 during engagement across stimulation conditions (i.e.independent of the whisker stimulated at the highest frequency).

Figure S13. Example FOVs showing healthy jRGECO expression between 5 and 48 weeks following injections.
Animals RRDED 1, RRED2 and RRED8, different Fields of view.Most of our imaging was carried out before 14 weeks (with the exception of 1 FOV).Note the apparition of fluorescent aggregate in the 48 weeks example.

Figure S2 ,
Figure S2, Reaction times and facial movements A, Nose heading direction as a function of time.Trial # in x, time in y, blue indicates left heading nose; red indicates right heading nose; trials sorted according to the first detected nose movement (see methods).B, Distribution over time of first detected nose movement.

Figure S3 .
Figure S3.Highly responsive neurons are more whisker selective and clustered toward the barrel center.

Figure S4 .
Figure S4.Spatial spread of optogenetic excitation/ inhibition A, Change in fluorescence in response to optogenetic light (450 µm disk) for all cells in the FOV (n= 5034 neurons, 3 animals, 5 FOV ).23.2% of cells showed activation by optogenetic light (putative inhibitory neurons) and 23.8% cells showed inhibition (putative excitatory neurons) compared to spontaneous activity.B, Change in activity as a function of distance to center of inhibition, quantified as absolute change in activity |ΔZscore|, for the three optogenetic conditions with the smallest light amplitude (e = 1.4 mW.mm² and e= 3 mW.mm²for the 450 µm and 105 µm disks respectively).C, Inhibition and excitation as a function of distance to illumination center during selective barrel illumination.D, Visualization of the optogenetic light spread in the axial dimension of the microscope.The light spread is measured as bleaching induced on GCaMP6s expressed across cortical layers.Bleaching was induced by prolonged exposure to the optogenetic light pattern (12 minutes of continuous illumination, 105 µm disk FWHM, at ~ 44 mW/mm 2 ).Pixel intensity is computed as fluorescence before exposition minus fluorescence after exposition, thus bleaching appearing as a brighter column.Fluorescence was measured before and after bleaching with the same two photon imaging parameters in a stack with 5µm steps.Each plane is averaged over 1 second.Volume rotation and visualization was performed with ImageJ.

Figure S5 .
Figure S5.The square root linear model with interaction yields similar results to the full categorical model for predicting single neuron's activity.A, Comparison of Akaike criterion (fit performance given complexity of the model).Friedman's test with multiple comparison, shows that no other model tested outperform the model used in the rest of the study (sqrt linear with interaction; described in Equation (1); see methods).B, Variance explained for different linear, categorical, and non-linear models of activity.The higher variance explained by the full categorical model could indicate that some neurons have a modal tuning to specific frequencies.C, Graphical summary of population analysis used in Fig 4-7.In the initial step, linear regression is independently applied to each neuron (Rn representing the activity of neuron n) to ascertain the sensory features it represents.This regression process identifies the weights ßF1,n, ßF2,n, ßF1xF2,n, which each of the three sensory features (F1, F2 and F1xF2) drive the activity of neuron n. ß0,n represents the baseline activity in the model, and Ɛ represents the non-fitted residuals.In the subsequent step, the activities of multiple neurons are pooled using the same sensory weights that drive them.These weighted pools embody the latent representations of stimulus features encoded by the populations of neurons: Rw1 represents the population response to the stimulation frequency of whisker 1 (F1), Rw2 corresponds to the population response to the stimulation frequency of whisker 2 (F2), and Rw1xw2 carries the population response to the cross-whisker interaction.These latent representations are then leveraged for various analyses, including decoding, exemplified here as neurometric analysis, or for the trial-by-trial representational geometry of sensory versus choice signals, as depicted in Figures6 and 7. Additional regression variables, namely choices (C1 and C2) or engagement (Eng), are incorporated explicitly in the models presented in Figure6and 7 respectively, to generate additional latent representations of choices and engagement ( not represented here, described in the methods).

Figure S8 :
Figure S8: Neurometric functions of Rw1 and Rw2 A, B, Neurometric functions in trials with left versus right response from the animal.Decoding of Rw1 (A) or RW2 (B).C, D, Neurometric functions in trials with or without responses.Comparison of slope and bias compared with a Wilcoxon sign-rank test reveals no significant difference (p>0.05,n = 11 FOVs).Decoding of Rw1 (C) or RW2 (D).

Figure S9 :
Figure S9: Somatotopy of choice; mixed and orthogonal sensory/choice coding A, All neurons from 11 FOVs aligned to the mid-points between barrels.Left: AUROC for sensory discrimination (F1>F2 versus F1<F2); Right: AUROC for choice discrimination (C1 versus C2).Size of the dot represent deviation from chance level AUROC (0.5); color represent preference for whisker 1 (>0.5) or whisker 2 (<0.5).Note that size and color are normalized differently on the left and right plots with AUROC target having stronger deviation from chance.

Figure
Figure S10 choice information in sub-populations RC1, RC2 and RC1 -RC2.Performance of choice discriminability as a function of the number of neurons included in the pooled response.From left to right: discriminability of RC1, RC2, and RC1 -RC2.Each line represents data from one FOV (n = 11 FOV).Matched number of Choice 1/Choice 2 trials in each stimulus conditions.Decoding is 10-fold cross-validated.

Figure S11 :
Figure S11: Analysis of sensory choice coding in sub-populations of neurons with strong selectivity.We compared sensory/choice coding of the entire population (A) to 4 subpopulations that code best for sensory selectivity (B), choice selectivity (C), multi-whisker suppression (D), or intersection of sensory features and choice in correct trials (E).For the (E) population only, sensory and choice become non-orthogonal.Selection criteria are indicated on top of the column.20% with highest criterion were included in analysis (B) to (E).(E) is exemplary of collinear coding of sensory and choice variable that leads to better decoding of the target in correct trials.From top row to bottom: (row 1) Single neurons' selectivity for choice as a function of selectivity for whisker, quantified as ßC1-ßC2 and ßF1 -ßF2 respectively.(row 2) Sensory decoding performance (F1>F2) from Rw1, Rw2 and Rw1-Rw2 compared between correct versus error trials in matching stimulus condition (with |ΔF|<30).LME model analysis * indicates p<0.05; ** indicates p<0.01 and *** indicates p<0.001.(row 3) neurometric performance in left and right choice trials.(row 4) trial to trial representation of different trial categories (color coded) spanning the four possible combinations of target and choice side.Note the collinearity of representation and the representational angle between choices being ~45 degree for (E) only.