Neural activity profiles reveal overlapping, intermingled subpopulations spanning area borders in mouse sensorimotor cortex
- Committee on Computational Neuroscience, The University of Chicago, United States
- Department of Organismal Biology and Anatomy, The University of Chicago, United States
- Neuroscience Institute, The University of Chicago, United States
- NSF-Simons National Institute for Theory and Mathematics in Biology, United States
Figures
Figure 1 with 1 supplement
A dataset of densely sampled activity across mouse sensorimotor cortex.
(A) Top, dorsal view of the Allen CCF, with the area investigated here indicated by a box. Bottom, summary of several previously-identified subareas of mouse sensorimotor cortex (see Methods for descriptions, alignment, and sources). Orientation legend: A, anterior, P, posterior, L, lateral, M, medial. (B) Widefield calcium imaging in a 5-mm cranial window superimposed on the Allen CCF to show location. Activation elicited by paw vibration (Methods). (C) Counts of neurons imaged for each area, with identities colored in inset. (D) Field of view locations of the 98 sessions of two-photon imaging collected in layer 2/3 for six mice (colors). (E) Locations of the 39,398 neurons extracted from all sessions. Mouse identity colored as in D.
Figure 1—figure supplement 1
Widefield window images and vibration maps aligned to the Allen CCF for all mice.
Shown as in Figure 1B.
Figure 2 with 1 supplement
Mice performed delayed reach-to-grasp-to-drink movements to 15 distinct targets.
(A) Task timeline. ITI, inter-trial interval. (B) Infrared image of the mouse during the inter-trial interval from one of two high-speed cameras. The 15 possible target locations illustrated with colored dots. (C) Finger centroid trajectories locked to lift onset from a frontal (left panel) and dorsal (right panel) view, illustrating kinematic separation by target. Target locations and average nose and mouth locations superimposed over centroid traces for clearer visualization. Coordinate axes show 3 mm scale bars and correspond to the axes in B. (D) Box plots of lift reaction times (RTs) for each mouse. Boxes show interquartile range (IQR) with a line for median; whiskers show 1.5 times the IQR; dots show outliers. Negative RTs indicate reaches before the Go cue and were excluded from further analysis. (E) Time between lift and spout contact across mice. Three consecutive sessions per animal shown for both D and E.
Figure 2—figure supplement 1
Reach-to-grasp-to-drink centroid trajectories show moderate target-specificity.
(A) Single-trial finger centroid trajectories locked to lift onset as in Figure 2C, shown from a posterior to anterior perspective, one session per mouse. Target positions not shown. Single-trials highlighted for one target for mouse 3, the same mouse shown in Figure 2C. (B) Each trace quantifies how different reaches were on average. For each possible pair of trials, we computed the Euclidean distance between finger centroid XYZ positions at each time point. Red traces show the average distance across trials to different targets, and black traces show the average distance across trials to the same target. Shading shows SEMs across pairwise distances. In many cases, the shading is occluded by the mean trace. Horizontal bars above the traces show time points with significantly different distributions of within vs. between distances at p<0.05 from a two-sample Kolmogorov-Smirnov test.
Figure 3 with 1 supplement
Neurons in mouse sensorimotor cortex exhibited heterogeneous tuning profiles.
(A) Extracted neuron locations superimposed on the Allen CCF. Red indicates a neuron that was modulated for any of the three locking events according to our statistical test (Methods), gray indicates a neuron that was not. (B) Binned and Gaussian-smoothed (s.d. 150 µm) map derived from A. (C) Venn diagram depicting the number of modulated cells for each event alignment. (D–H) Peri-event time histograms (PETHs) of five example neurons in each area. Successful trials to each target were averaged and smoothed. Colors as in Figure 2B. Shaded regions, SEMs; scale bars, 10 events/second. Neurons were chosen to be clear and representative examples of response profiles observed in each area.
Figure 3—figure supplement 1
Modulation maps for all task events.
(A) Binned and smoothed modulation map as in Figure 3B, but for cells modulated by Cue (left), Lift (center), or First-contact (right). Cue alignment used neural data from 200 ms before to 400 ms after cue. Lift alignment used neural data from 200 ms before to 400 ms after lift. First-contact alignment used neural data from 300 ms before to 300 ms after first spout contact. (B) Binary matrix indicating which area-area comparisons had significantly different proportions of cue-modulated cells. Black squares indicate rejection of the null hypothesis at p<0.005 using a two-proportion z-test.
Figure 4
Onset of neural activity across sensorimotor cortex followed both area borders and a somatotopic organization.
(A) Smoothed histogram of time of half-max activation for neurons by anatomical area. (B) Time of half-max activation map. The metric was computed on individual neurons, neurons were binned into pixels, then the map was Gaussian smoothed (s.d. 150 µm). Colorbar ticks indicate the range of values in the binned and smoothed map.
Figure 5 with 1 supplement
PETH features were organized into distinct spatial patterns in sensorimotor cortex.
(A) Top left, schematic for the response duration metric, which measures the autocorrelation width of the trial-averaged trace for each neuron and target. Bottom left, histogram of response duration values for all modulated neurons grouped based on anatomical region. Right, metric map of the response duration values where bright regions correspond to higher response duration values and darker regions to lower. (B–F) Analogous to A for other metrics. (B) Peak-time variation, computed as the SD of peak times across targets. (C) Tuning sharpness, computed as one minus the ratio of the average peak response (excluding the strongest response condition) normalized by the strongest peak response. (D) Target tuning linearity, the coefficient of determination of the peak response with ordinal distance of each target from the target that evoked the strongest response. (E) Tuning persistence, corresponding to how consistent the tuning was across targets at all lags from the peak (Methods). Only statistically modulated targets were included for each neuron. (F) 'Local dimensionality’, the participation ratio of the 30 nearest neighbors to each neuron. Star indicates that this metric was not used for analyses below, because it is not independent for each neuron.
Figure 5—figure supplement 1
Principal component scores after VARIMAX rotation for each neuron.
Plotted as feature maps using the same method as Figure 5.
Figure 6
Derivatives of PETH feature maps produced response-property boundaries aligned with anatomy and somatotopy.
(A) Side-by-side comparison of a PETH feature map and its gradient magnitude map. Right, bright green values represent large gradient magnitudes and dark values represent small gradients. (B–E) Same as A right, for the other PETH features. (F) Quadratic mean over the five PETH features of the gradient magnitude maps. (G) Top, 10,000 1 mm random boundaries plotted in gray within the sampled region. Bottom, for each random boundary, neurons that were within 500 µm on either side of the boundary were included; neurons within 50 µm of the boundary were excluded. The target tuning linearity feature map is shown underneath for illustration. An SVM was trained to predict which side of the boundary a neuron was on using its PETH feature vector. (H) The top-performing 2.5% of random boundaries (assessed via cross-validation) plotted in yellow on top of the average gradient magnitude map.
Figure 7
Areas had distinct and multimodal distributions of PETH features.
(A) t-SNE of the five-dimensional PETH features for all neurons. Each gray point corresponds to a neuron. (B) Same embedding as in A, with neurons (points) colored by anatomical area. (C) Contour plots of the embedding from B, using two-dimensional Kernel Density Estimation to produce a smooth density estimate. Three contour levels are shown for each of the five anatomical areas of interest.
Figure 8 with 1 supplement
Neural response profiles were shared in patches that spanned anatomically defined regions.
(A) Schematic of analysis with hypothesis that neurons from different areas are well-separated in feature space. A Gaussian Mixture Model (GMM) was fit to the distribution of feature values for neurons belonging to a single anatomically defined area, then a likelihood was computed for all neurons from all areas using the fit. (B) The resulting likelihood map for M2, where bright yellow pixels correspond to high likelihood regions and dark blue pixels to low likelihood regions. Colorbar scale is logarithmic. The ends of the colormap were set to the maximum and minimum likelihood values for each map. (C–F) Same as B, where the seed regions were M1, S1-fl, S1-hl, and S1-tr, respectively.
Figure 8—figure supplement 1
Feature value histograms for all GMM components and areas.
Figure 9 with 5 supplements
Four subpopulations with distinct response profiles form an overlapping patchwork across sensorimotor cortex.
(A) Schematic of GMM modeling with the hypothesis that components from different areas correspond. Only two GMM components in each area are illustrated for simplicity, with numbers corresponding to those in B. (B) Bhattacharyya distance calculated between all pairs of GMM components across all anatomical regions. The matrix is organized based on cluster identity from hierarchical clustering of the pairwise distances. Each cluster (colored box outlines) is referred to as a subpopulation. Colormap capped to better show any within-cluster detail. Cue-locked data, modulated neurons only. (C) Anterior subpopulation likelihood map. Plotting as in the other likelihood maps with colormap as in Figure 8. (D–F) Same as C but for Forelimb motor (D), Forelimb somatomotor (E), and Hindlimb somatomotor (F) subpopulations. (G) Distributions of the five feature values (subpanels) for each of the four subpopulations (traces).
Figure 9—figure supplement 1
Highest likelihood PETHs from each subpopulation.
Two PETHs are shown from each subpopulation, for a total of eight per area. PETHs were chosen in a principled way as the highest likelihood neurons for each subpopulation within each area, with a max firing rate of at least 33 events/s. This approach identified the only two Anterior cells in S1-hl and zero cells in S1-tr; thus the S1-tr entries were left blank.
Figure 9—figure supplement 2
Separation of subpopulation distributions in feature space.
(A) Feature vectors for all neurons within a subpopulation were projected onto the axis that best linearly separated the distributions, for each pairwise comparison between subpopulations. The axis was obtained from a logistic regression and the cross-validated performance of the model is shown for each comparison. Note that subpopulation membership was determined by GMM classification, which is nonlinear, while logistic regression is linear. Because of both this model mismatch and cross-validation of the logistic regression, classification performance on the projection axis is less than 100%. (B) t-SNE plot of feature vectors as in Figure 7 but colored by subpopulation membership. (C) Time to half-max distributions for each subpopulation, plotted as in Figure 4A.
Figure 9—figure supplement 3
Additional subpopulation validation analyses.
(A) t-SNE as in Figure 7C but using the top 20 PCs of the PETHs as inputs instead of feature vectors. Multimodality was again strongly present. (B) Left, subsampled population of neurons chosen from an overlap zone in M1, to analyze for discrete subpopulations while avoiding spatial somatotopy. Center, t-SNE embedding of this M1 population using the PETH feature space. Right, t-SNE embedding of the same M1 population using the PCs. Both methods yielded clear multimodality. (C–G) Same as Figure 8B–F but using top 20 PCs instead of PETH features.
Figure 9—figure supplement 4
GMM fit to all PETH feature vectors together, agnostic to anatomical areas.
One GMM was fit to the feature vectors of cells from all five main areas. Each map plots the likelihood for all cells to each of the three components of this area-agnostic GMM.
Figure 9—figure supplement 5
Gradients of subpopulation likelihood maps.
(A–D) Gradients as in Figure 6 but calculated using the subpopulation likelihood maps depicted in Figure 9C–F instead of the PETH feature maps. As in Figure 6, each subpopulation had a distinct gradient with boundaries approximately aligned with borders between anatomical regions, and other boundaries approximately separating somatotopic representations. (E) The quadratic mean of the four subpopulation gradient maps. This pooled gradient map closely resembles the pooled map in Figure 6F.
Figure 10
Subpopulations overlapped with their member neurons spatially intermingled.
(A) Neurons were classified into their most likely subpopulation, color coded and plotted on the CCF. (B) Prevalence of subpopulation by area. (C) Schematic for distinguishing two hypotheses, if subpopulations were identified separately using data from each of two areas and the identified high-density zones of neurons were plotted as contours of the likelihood maps. Left, expected result if subpopulations varied smoothly with cortical location ; right, if subpopulations were homogeneous in their properties over cortical location. (D) Maps of prevalence of each subpopulation. (E) Likelihood map contours at 60% level, for each subpopulation identified separately using data from each area containing sufficient subpopulation members.
Author response image 1
GMM fit to spectrally transformed PETH PCs, agnostic to anatomical areas.
One GMM was fit to the spectrally-embedded PC feature vectors of cells from all 5 main areas. Each component of a 10 component model is shown.
Additional files
Download links
A two-part list of links to download the article, or parts of the article, in various formats.
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Neural activity profiles reveal overlapping, intermingled subpopulations spanning area borders in mouse sensorimotor cortex
eLife 14:RP109240.
https://doi.org/10.7554/eLife.109240.3
