Measuring Stimulus-Evoked Neurophysiological Differentiation in Distinct Populations of Neurons in Mouse Visual Cortex

Abstract

Despite significant progress in understanding neural coding, it remains unclear how the coordinated activity of large populations of neurons relates to what an observer actually perceives. Since neurophysiological differences must underlie differences among percepts, differentiation analysis-quantifying distinct patterns of neurophysiological activity-has been proposed as an "inside-out" approach that addresses this question. This methodology contrasts with "outside-in" approaches such as feature tuning and decoding analyses, which are defined in terms of extrinsic experimental variables. Here, we used two-photon calcium imaging in mice of both sexes to systematically survey stimulus-evoked neurophysiological differentiation (ND) in excitatory neuronal populations in layers (L)2/3, L4, and L5 across five visual cortical areas (primary, lateromedial, anterolateral, posteromedial, and anteromedial) in response to naturalistic and phase-scrambled movie stimuli. We find that unscrambled stimuli evoke greater ND than scrambled stimuli specifically in L2/3 of the anterolateral and anteromedial areas, and that this effect is modulated by arousal state and locomotion. By contrast, decoding performance was far above chance and did not vary substantially across areas and layers. Differentiation also differed within the unscrambled stimulus set, suggesting that differentiation analysis may be used to probe the ethological relevance of individual stimuli.

Keywords: calcium imaging; differentiation analysis; perception; population coding.

Figure 1.

Experimental design. A, Data were acquired using a standardized two-photon calcium imaging pipeline based on that described in de Vries et al. (2020) and Groblewski et al. (2020; Materials and Methods). Briefly, a custom headframe was implanted; ISI was performed to delineate retinotopically mapped visual areas; the mouse was habituated to the passive viewing paradigm over the course of approximately two weeks; and two-photon calcium imaging was performed in the left visual cortex while animals viewed stimuli presented to the contralateral eye in several experimental sessions. B, Example of an ISI map. C, Schematic of the two-photon imaging rig (reproduced with permission from Fig. 1D in de Vries et al. 2020). During the imaging sessions, head-fixed mice were free to run on a rotating disk. Locomotion velocity was recorded and pupil diameter was extracted from video of the animal’s right eye. D, Example frame from a two-photon movie. Imaging data were processed as described in de Vries et al. (2020) to obtain ΔF/F₀ traces. E, Schematic of the five visual areas targeted in this study. F, Ten randomized blocks of 12 30-s movie stimuli were presented; 4 s of mean-luminance gray was presented between stimuli. The first 60-s period was mean-luminance gray (spontaneous activity); the second 60-s period was a high-contrast sparse noise stimulus (not analyzed in this work). G, Still frames from the eight naturalistic (left) and four artificial (right) movie stimuli (see Materials and Methods, Stimuli). Two of the naturalistic stimuli, “mouse montage 1” and “mousecam,” were phase-scrambled to destroy high-level image features while closely matching low-order statistics (see Materials and Methods, Phase scrambling; Extended Data Fig. 1-1). H, Representative calcium imaging and behavioral data. A heatmap of ΔF/F₀ values is shown for 228 neurons simultaneously imaged in L2/3 of AL during presentation of four stimuli, with locomotion velocity and normalized pupil diameter plotted below. Numbers of cells recorded from each layer and area are listed in Extended Data Figure 1-2. Calcium indicator kinetics did not differ across cell populations (Extended Data Fig. 1-3).

Figure 2.

Spectral differentiation analysis. ND was computed as follows. A, For each cell, the ΔF/F₀ trace during stimulus presentation was divided into 1-s windows. B, The power spectrum of each window was estimated. C, The “neurophysiological state” during each 1-s window was defined as a vector in the high-dimensional space of cells and frequencies (i.e., the concatenation of the power spectra in that window for each cell). D, The ND of the response to a given stimulus was calculated as the median of the pairwise Euclidean distances between every state that occurred during the stimulus presentation. An illustration of how the measure behaves is shown in Extended Data Figure 2-1.

Figure 3.

ND elicited by unscrambled versus scrambled stimuli is higher in L2/3 of areas AL and AM. The difference in ND of responses to unscrambled versus scrambled stimuli is plotted for each session by layer (A), area (B), and layer-area pair (C). Each point represents the difference between the mean ND of responses to the two unscrambled and the three scrambled stimuli during a single experimental session. Similar results were found contrasting naturalistic versus artificial stimuli across the entire stimulus set (Extended Data Fig. 3-1). To demonstrate the robustness of this effect, we conducted several further analyses. Sensitivity analyses showed similar findings for various choices of analysis parameters (Extended Data Figs. 3-2, 3-3, 3-4) and when pupil diameter and locomotion were included as covariates in the LME models (Extended Data Figs. 3-5, 3-6, 3-7). We found similar results when we performed the same analysis on discrete calcium events detected from the ΔF/F₀ traces with an L₀-regularized algorithm (see Materials and Methods, Event detection), indicating that the effect is driven by differences in the large-timescale patterns of responses rather than small-timescale spectral differences within windows (Extended Data Fig. 3-8). Finally, we also found similar results when we removed event-triggered transients from the ΔF/F₀ traces, indicating that the effect is not driven solely by initial transients in the calcium response (Extended Data Fig. 3-9). A, B, Asterisks indicate significant post hoc one-sided z-tests in the layer (A) and area (B) interaction LME models (*p < 0.05; ***p < 0.001). Boxes indicate quartiles; whiskers indicate the minimum and maximum of data lying within 1.5 times the interquartile range of the 25% or 75% quartiles; diamonds indicate observations outside this range. C, Mean values are indicated by bars.

Figure 4.

Effect sizes in L2/3 of AL and AM are larger in sessions with more locomotion and larger pupil diameter. Cohen’s d is plotted against the fraction of locomotion activity (left column) and mean normalized pupil diameter (right column) during the session, with linear fit in gray. Top row: only sessions recorded from L2/3 and areas AL or AM. Bottom row: all sessions (note different scales). Top left: Pearson’s r = 0.896 (two-sided t test; t₍₄₎ = 4.030, p = 0.0157, 95% CI [0.308, 1.00]^l). Top right: r = 0.716 (t₍₄₎ = 2.054, p = 0.109, 95% CI [–0.227, 1.00]^m). Running velocity >2.5 cm/s was considered locomotion activity (see Materials and Methods, Locomotion). Normalized pupil diameter was obtained by dividing by the maximum diameter that occurred during the session (see Materials and Methods, Pupillometry).

Figure 5.

Multivariate differentiation analysis. The mean difference in the mean centroid distance of responses to unscrambled versus scrambled stimuli is plotted for each session by layer (A), area (B), and layer-area pair (C). ND elicited by unscrambled versus scrambled stimuli is higher in L2/3 and areas AL and AM, consistent with the spectral differentiation analysis. We found similar results when we analyzed discrete L₀ calcium events detected from the ΔF/F₀ traces (see Materials and Methods, Event detection; Extended Data Fig. 5-1). A, B, Asterisks indicate significant post hoc one-sided z-tests in the layer (A) and area (B) interaction LME models (**p < 0.01; ***p < 0.001). Boxes indicate quartiles; whiskers indicate the minimum and maximum of data lying within 1.5 times the interquartile range of the 25% or 75% quartiles; diamonds indicate observations outside this range. C, Mean values are indicated by bars.

Figure 6.

Stimulus category (unscrambled or scrambled) can be accurately decoded from most layers and areas. Each point represents the mean fivefold cross-validated balanced accuracy score of linear discriminant analysis performed on a single session (see Materials and Methods, Decoding analyses). Chance performance is 0.5. We found similar results when decoding stimulus identity across all 12 stimuli (Extended Data Fig. 6-1).

Figure 7.

Pairwise differences in ND among unscrambled, continuous stimuli. Post hoc pairwise comparisons using data from all neuronal populations are plotted against their p values (adjusted for multiple comparisons). Boxes show mean ND for each stimulus. ND of the “snake (predator)” stimulus is significantly greater than that of “crickets” and “man writing” at a threshold of α = 0.01, and greater than “conspecifics” at α = 0.05. ND of the “mousecam” stimulus is greater than that of “man writing” at α = 0.01. Mediation analysis showed a mixture of direct and arousal-mediated effects, indicating that changes in arousal cannot fully account for these differences (Extended Data Fig. 7-1). Pairwise differences in ND and decoding performance stratified by layer and area are shown in Extended Data Figure 7-2.

Figure 8.

SD does not explain ND. Mean ND elicited by each stimulus in L2/3 of AL and AM, plotted against SD. SD was computed by treating each pixel of the movie as a “cell” and applying the spectral differentiation measure to traces of pixel intensities over time after blurring the movie with a Gaussian filter to account for the coarseness of mouse vision. Across all stimuli, mean ND is positively correlated with SD (Pearson’s r = 0.746; one-sided t test; t₍₁₀₎ = 3.542, p = 0.00267, 95% CI [0.393, 1.00]^hh). However, here the noise stimulus is a highly influential observation (Cook’s D = 2.318, an order of magnitude larger than the next most influential observation). With the noise stimulus excluded, the correlation is weaker (r = 0.258; one-sided t test; t₍₉₎ = 0.801, p = 0.222, 95% CI [–0.307, 1.00]ⁱⁱ). Moreover, there was no evidence of a relationship with ND when considering only the scrambled stimuli and their unscrambled counterparts (r = –0.378; two-sided t test; t₍₃₎ = –0.708, p = 0.523, 95% CI [–0.945, 0.756]^jj). ND was also not explained by variation in stimulus luminance, contrast, or spectral energy (Extended Data Fig. 8-1).