Differences in the expression of cortex-wide neural dynamics are related to behavioral phenotype

Summary

Behavior differs across individuals, ranging from typical to atypical phenotypes¹. Understanding how differences in behavior relate to differences in neural activity is critical for developing treatments of neuropsychiatric and neurodevelopmental disorders. One hypothesis is differences in behavior reflect individual differences in the dynamics of how information flows through the brain. In support of this, the correlation of neural activity between brain areas, termed ‘functional connectivity’, varies across individuals² and is disrupted in autism³, schizophrenia⁴, and depression⁵. However, the changes in neural activity that underlie altered behavior and functional connectivity remain unclear. Here, we show that individual differences in the expression of different patterns of cortical neural dynamics explain observed variability in both functional connectivity and behavior. Using mesoscale imaging, we recorded neural activity across the dorsal cortex of behaviorally ‘typical’ and ‘atypical’ mice. All mice shared the same recurring cortex-wide spatiotemporal motifs of neural activity, and these motifs explained the large majority of variance in cortical activity (>75%). However, individuals differed in how frequently different motifs were expressed. These differences in motif expression explained differences in functional connectivity and behavior across both ‘typical’ and ‘atypical’ mice. Our results suggest that differences in behavior and functional connectivity are due to changes in the processes that select which pattern of neural activity is expressed at each moment in time.

eToc

MacDowell et al., use cortex-wide imaging to quantify neural dynamics of behaviorally ‘typical’ and ‘atypical’ mice. While all mice share the same set of cortex-wide patterns of neural activity, individual variability in how frequently each pattern is expressed explains individual differences in behavior and functional connectivity.

Results

To understand how differences in behavior relate to changes in cortex-wide neural dynamics, we combined behavioral assays and cortex-wide calcium imaging in mice⁶. Since we were interested in studying a heterogenous population of mice with a spectrum of individual differences, recordings were done in ‘typical’ (wildtype) mice and in mice exposed to valproic acid (VPA) in utero (Figure 1A; see Methods). VPA mice exhibit altered social and repetitive motor behaviors⁷ and are widely used to study how changes in neural structure and function relate to behavior^8-11 (often as a model of cognitive dysfunction, including autism^12,13). In particular, VPA mice show hyperconnectivity between brain regions and altered inhibitory/excitatory balance within medial prefrontal (mPFC) cortex^10,14,15. Importantly, VPA exposure provides an etiologically valid way of inducing a spectrum of behaviors in an otherwise genetically identical cohort of mice (n=52 mice; 27 VPA, 25 saline controls).

Figure 1: Cortical functional connectivity correlates with behavioral phenotype in mice.

(A) Schematic of the in utero valproic acid (VPA) model used to investigate a diversity of behavioral phenotypes. (B) Schematic of example behavioral assays. From left to right: weaning weight and developmental milestones, pup ultrasonic vocalizations (USVs) in response to maternal separation, three-chamber social approach and social novelty, self-grooming behavior, and digging behavior in a marble burying assay. See Methods for complete list of assays and measures. (C) Example results comparing behavior of VPA and saline-exposed (SAL) control animals. From left to right: animal weight at weaning, number of pup USVs in response to maternal separation, animals sociability score measured from social novelty assay, number of self-grooming epochs, percent of marbles buried in marble burying assay. Data points reflect individual mice and errorbars show mean +/− SEM. See methods for details and Figure S1A-C for complete results. (D) Contribution of individual behavioral measures to defining the hyperplane that maximally separates VPA and saline animals, quantified as the beta-weight in the linear support vector machine (SVM). (E) Animals’ positions along a ‘Phenotype Axis’ defined as the vector normal to the hyperplane that maximally separated groups, defined by the SVM fit. (F) Schematic outlining mesoscale imaging of mouse cortical neural activity at rest. (G) (top) Grid parcellation for computing functional connectivity across cortical regions of interest (ROIs). (bottom) Number of pair-wise connections between ROIs binned by distance. (H) Strength of functional connectivity as a function of distance between groups. Mean fisher-z-transformed Pearson’s Rho (${\rho }_{\mathrm{z}}$): for 3-4.5mm: Saline=0.91, CI: 0.87-0.95; VPA=0.82, CI: 0.77-0.88; p=0.033; for 4.5-6mm: Saline=0.81, CI: 0.78-0.85; VPA=0.70, CI: 0.65-0.75; p=0.003; both permutation tests. (I) Strength of long-range functional connectivity as a function of behavioral phenotype (as in panel d). Data points show individual mice. Solid and dotted lines show least squares fit and 95% confidence bounds, respectively. See also Figure S1.

Animal behavior was assessed with a battery of 9 tests targeting 12 measures of general developmental, sensory-memory, and motor behavioral domains (Figure 1B; see Methods). Consistent with previous literature^7,12, individual tests showed significant differences between the saline-exposed controls and VPA mice (Figure 1C and Figure S1A-C). Consistent with a spectrum of individual behavioral phenotypes¹, there was considerable variability across individuals, both between and within groups. To quantify this spectrum, we projected the multivariate behavioral data onto a single ‘phenotype axis’. The phenotype axis was defined as orthogonal to the linear hyperplane that maximally separated the behavior of VPA and saline mice (across all measures; Figure 1D-E; see Methods). Projecting each animal’s behavior onto this axis captured that individual’s position along the behavioral spectrum. For example, some VPA mice demonstrated more typical behaviors than others, while some saline animals showed more atypical behavioral features. Similar degrees of behavioral diversity were observed in each behavioral domain (Figure S1D-E).

To study cortex-wide neural dynamics, we used mesoscale calcium imaging to record activity of excitatory neurons across the dorsal cortex of a subset of VPA (n=11) and saline (n=9) mice (Figure 1F; all Thy1-GCaMP6f¹⁶ mice, awake and head-fixed, see Methods). These animals spanned the behavioral spectrum of the larger population (Figure 1E, yellow dots). Motivated by previous work^11,17, we tested whether functional connectivity varied across animals.

Functional connectivity between cortical regions correlates with behavioral phenotype

Functional connectivity was measured as the zero-lag pairwise correlation in neural activity between a grid of sites across the dorsal cortex (Figure 1G). In all animals, more distant cortical sites had weaker functional connectivity. However, this effect was more pronounced in VPA animals: functional connectivity between local cortical regions (<3mm) was similar between groups, but long-range (>3mm) connectivity was significantly decreased in VPA animals relative to saline controls (Figure 1H). Reflecting its relationship with behavior, the difference in long-range connectivity was significantly negatively correlated with phenotype across animals (Figure 1I; r=−0.46 p=0.033 permutation test, n=20 mice). These results are consistent with the observation that the spectrum of behavior is related to functional connectivity in humans^2-5, highlighting the face validity of the VPA model.

Individuals share the same spatiotemporal motifs in cortical neural activity

As detailed above, these differences in functional connectivity are thought to reflect changes in the spatiotemporal interactions between cortical regions. To test this, we quantified the spatiotemporal dynamics of cortical activity by decomposing cortex-wide neural activity into a set of 16 ‘motifs’ using convolutional factorization and unsupervised clustering (schematized in Figure 2A; see Methods and Figure S2A-G). Each motif captured a unique cortex-wide spatiotemporal pattern of neural activity that lasted for ~1 second and is thought to reflect a specific cognitive process¹⁸. For example, Motif 12 is a cortex-wide wave of neural activity originating in somatosensory and motor areas and travelling posteriorly to retrosplenial and visual cortices, which is thought to coordinate neural activity across areas^19,20. In contrast, Motif 5 is a spatially static, bilateral, burst of activity in the visual cortex, associated with the processing of visual stimuli^18,21. By tiling the 16 unique motifs over time, we could accurately reconstruct cortical activity in all mice (Figure 2B; see Methods, and Figure S2G and Table S1 for all 16 motifs).

Figure 2: Individuals share the same spatiotemporal motifs in cortical neural activity.

(A) Schematic outlining quantification of spatiotemporal motifs in cortical activity. (left) Convolutional matrix factorization decomposed neural activity into (right, top) recurring motifs and (right, bottom) their temporal weightings. (B) Percent variance in cortical neural activity explained by the 16 spatiotemporal motifs that were shared across all animals: mean saline: 79.59% CI: 70.88%-83.80%; VPA: 77.12%, CI:72.80%-80.38%; no difference between groups, p=0.20; Mann-Whitney U-Test and no correlation with behavioral phenotype; r=0.031, p=0.54, permutation test. (C) Example motifs. Green lines outline anatomical parcels as in Figure 1F. Red dot denotes bregma. (D) Relative percent variance in neural activity explained by each motif per mouse. This ‘barcode’ captured the pattern of expression of motifs in each animal. Data points reflect individual mice. All errorbars show mean +/− SEM. See also Figure S2 and Table S1.

Given that motifs reflect spatiotemporal patterns in neural activity across brain areas, we can use them to quantify how dynamics in neural activity specifically change between individuals. If changes in functional connectivity are due to altered patterns in the flow of neural activity across the cortex, then the spatiotemporal patterns captured by motifs will be different between individuals. In contrast, if changes in functional connectivity are due to variability in which motifs are expressed, then all mice should share the same motifs, but how often a motif occurs will differ across animals. Consistent with a change in expression, the same 16 motifs captured over 75% of the variance in cortical neural activity in both VPA and saline animals (Figure 2B; no difference between groups).

To verify that this result was not due to our analytical approach, we trained linear decoders to differentiate motifs from VPA and saline animals (see Methods). Classifiers failed to discriminate the groups, consistent with the conclusion that the same set of motifs were independently identified within each group (mean decoding accuracy across motifs=51.46%, CI 48.37%-54.81%, p=0.34, Wilcoxon Signed-Rank Test; see Methods). This was not due to an inability of linear decoders to discriminate motifs; they easily discriminated between the 16 motifs (accuracy=99.08%, CI 98.98%-99.18%, p<0.0001, Wilcoxon Signed-Rank Test). Furthermore, our approach was able to recover ground truth in simulated data (see Methods).

Differences in cortical motif expression are correlated with behavior and can explain differences in functional connectivity

The similarity of motifs across groups suggests disruptions in functional connectivity are not due to changes in the way neural activity flows through the brain. An alternative hypothesis is that individual changes in functional connectivity (and behavior) are related to differences in how frequently a motif is expressed. Consistent with this alternative hypothesis, we found that while all motifs were expressed in all mice, different motifs were active more/less often in each mouse, and so contributed to more/less of the variance in neural activity (see Methods). Motif activity was heterogenous across animals, and no single motif clearly separated VPA and saline groups (Figure 2D; no differences significant at p=0.05, Mann-Whitney U-Test). However, by measuring the relative variance captured by each motif, we could create a ‘barcode’ of neural dynamics that captured the relative expression of each motif for each animal. Using these barcodes, a cross-validated linear classifier could accurately classify VPA and saline animals (accuracy=86.61%, CI 80.00%-95.00%; see Methods). The beta weights of the classifier were distributed across all 16 motifs, supporting the conclusion that discriminability between groups relied on changes in expression of all motifs, rather than disruption of any single motif (Figure 3A). Similar to behavior, we could estimate each animal’s ‘neurotype’ by projecting its individual barcode onto the axis orthogonal to the hyperplane that best separated the two groups (Figure 3A; see Methods). Despite being defined independently and measured weeks apart, individuals’ neurotype and phenotype were significantly positively correlated (Figure 3A; r=0.76, p=0.038; statistics were tested by permuting individuals within each group in order to control for any group-level differences). In other words, animals with more atypical expression of motifs also had more atypical behavior.

Figure 3: Spatiotemporal motifs in neural activity correlate with behavioral phenotype and explain the relationship between functional connectivity and behavior.

(A) Correlation between neurotype axis and phenotype axis. To ensure correlation was not due to group-wise differences, we generated a null distribution of correlation by permuting within each group. (top inset) Red line shows observed correlation, which was significantly different from the null (p=0.038). (bottom inset) Beta weights of motif contributions to the neurotype axis. (B) Strength of animals’ long-range functional connectivity was negatively correlated with their neurotype. (C) Relative long-range vs. short-range functional connectivity within each motif was negatively correlated with the motif’s beta-weight in the neurotype axis (from panel A). (D) As in Figure 1H, shows strength of functional connectivity as a function of distance, but calculated on the timecourse of cortex-wide neural activity reconstructed from motifs. As in original data, long-range functional connectivity differed between groups: for 3-4.5mm ${\rho }_{\mathrm{z}}$ saline=1.24 CI: 1.19-1.26; VPA= 1.16 CI: 1.11-1.23 p=0.042; for 4.5-6mm saline=1.11 CI: 1.07-1.15; VPA=1.00 CI: 0.94-1.08 p=0.008; permutation test. Data points reflect individual mice, except for panel C where they reflect motifs. All errorbars show mean +/− SEM. Solid and dotted lines in panels A,B,C show least squares fit and 95% confidence bounds, respectively. See also Figures S2 and S3 and Table S1.

This effect also existed within each group of saline of VPA animals. The first principal component (PC) of motif expression within the saline group was significantly correlated with the animals’ neurotype, as defined across groups (Figure S3A; Kendall’s Tau, τ_b = 0.64, p = 0.013, permutation test). Importantly, the first ‘Neuro PC’ within the saline group was also correlated with the animals’ phenotype (Figure S3B; Kendall’s Tau, T_b = 0.55, p = 0.046, permutation test). Similar results were seen within VPA-exposed animals (Figure S3CD). Together, these results suggest the neurotype and phenotype axes are capturing the heterogeneity of neural activity and behavior across individuals, not simply between groups.

Animals’ neurotypes were broadly related to behavioral tests. Using multiple linear regression, we tested the association between the 12 behavioral measures and each animal’s neurotype. Neurotype was most strongly related to tests of motor phenotype including motor exploration, marble-burying, and self-grooming duration (Figure S4A-D). This is consistent with previous work showing the dominance of motor representations in cortical activity^21,22. Importantly, however, control analyses showed no difference in motor activity between groups during the neural recordings themselves (Figure S2H). Deeper analyses ruled out potential confounds, such as differences in brain size, GCaMP6f expression, or neural variance, confirmed that motif activity was stable across days, and that Thy1-GCaMP6f mice did not exhibit epileptiform activity²³ (Figure S2I-M).

Differences in the distribution of motif activity explained differences in functional connectivity across animals. An animals’ neurotype was significantly correlated with long-range functional connectivity (Figure 3B; r =−0.39 p=0.043, permutation test). Furthermore, removing neurotype contributions to functional connectivity eliminated the correlation between connectivity and phenotype, supporting the hypothesis that differences in motif expression drove this relationship (r =−0.17, partial regression, p=0.24, permutation test). In addition, the relative strength of long-range and short-range functional connectivity within each motif was negatively correlated with that motifs’ neurotype beta weights (Figure 3C; r =−0.47 p=0.022, permutation test, n=16 motifs). In other words, motifs capturing patterns of neural activity with stronger long-range connectivity were relatively more active in animals with more ‘typical’ behavior (compared to mice with ‘atypical’ behavior). Conversely, motifs with weaker long-range connectivity were relatively more active in ‘atypical’ mice, explaining the observed changes in functional connectivity. Finally, to test whether changes in motif activity were sufficient to explain changes in functional connectivity, we reconstructed cortical activity of all animals using the shared set of 16 motifs. These reconstructed neural dynamics showed a similar decrease in long-range functional connectivity as the original data (Figure 3D). As the same motifs were used in all animals, these changes in functional connectivity can only arise from changes in the expression of motifs, and not differences in the spatiotemporal structure of motifs across animals.

Individuals differ in adapting motif expression to a new behavioral context

Our results above suggest that individual differences in behavior (and functional connectivity) are due to individual variability in the expression of motifs. Given that motifs capture patterns of neural activity associated with specific cortical processes¹⁸, one hypothesis is that a set of control mechanisms may express motifs to engage in context-appropriate cognitive processes. If true, then motif activity should adapt to different behavioral contexts. To test this, we compared motif activity between two contexts: when mice were alone (as above) and when mice were interacting with a second animal in a stimulus-enriched environment (‘paired’ context; Figure 4A; see Methods). Consistent with the hypothesis that cortex-wide dynamics adapted to the situation, the expression of motifs was significantly different between contexts in all mice (all p-values <0.001; see Methods).

Figure 4: Individual variability in motif sampling across behavioral contexts relates to behavioral phenotype.

(A) (top) Schematic of the ‘alone’ and ‘paired’ contexts. (bottom) In each context, motif activity was measured as the barcodes in the percent explained variance (PEV) of cortical neural activity captured by each motif (as in Figure 2D). Shown barcodes are prototypical examples from a single animal. (B) Magnitude of change in motif activity between contexts was negatively correlated with behavioral phenotype. Change in motif activity was measured as the sum squared error between motif barcodes in each context. Data points show individual mice. Solid and dotted lines show fit, and 95% confidence bounds, respectively. See also Figure S4.

However, the magnitude of change in motif activity between contexts was negatively correlated with behavioral phenotype across mice (Figure 4B; r=−0.55 p=0.009, permutation test). In other words, more atypical mice showed less of a change in motif activity to the new, stimulus-rich paired context. This suggests mice with atypical behavior had a disrupted ability to adapt motif expression to match new contexts. Several more results were consistent with this hypothesis. First, how much an individual’s motif activity changed between contexts was most strongly related to their performance on behavioral tests involving novel contexts and/or social stimuli (Figure S4E-H). Second, the sequence of motifs over time was more regular in behaviorally atypical mice, suggesting they were more rigid in the ordering of motif expression (Figure S4I-J). Third, the distribution of motifs was more uniform in VPA than saline mice when they were alone (Figure S4K-N). Together, these results support the hypothesis that individual differences in behavior are, at least partially, due to variability in the control or selection of motif expression.

Discussion

Individual differences in behavior, both in health and disease, are related to changes in the functional connectivity between brain regions^2-5,17. Changes in functional connectivity have been associated with differences in structural connectivity^24-28 and deficits in synchronizing cortical regions²⁹. Our results suggest an additional mechanism – we found inter-individual differences in functional connectivity could be explained by differences in how often certain spatiotemporal motifs of cortical activity occurred.

Individual differences were not due to changes in the motifs themselves; a common set of motifs explained the majority of cortex-wide neural activity in all individuals³⁰, across the behavioral spectrum. Instead, we found individuals differed in how frequently different motifs were expressed. Motif expression was correlated with behavioral phenotype, despite the fact that motifs and behavior were measured weeks apart, suggesting this relationship reflects chronic differences between individuals. Future work is needed to investigate how biases in the expression of neural population dynamics may manifest to drive cognitive changes in health and disease.

One important caveat is that there may be differences in neural dynamics at spatial and temporal frequencies beyond what can be observed with cortex-wide calcium imaging. While motifs captured the majority of variance (>75%) in the observed neural activity, it is possible that VPA and saline-control animals differed in the details of neural activity. Future work with higher spatial and temporal resolution may provide deeper insights into individual differences in local cortical circuits and/or subcortical activity. Indeed, one interesting hypothesis is that there is a hierarchy of neural dynamics – motifs organize the general flow of neural activity across the cortex to support a behavior, carrying along the detailed representations related to a specific stimulus, decision, or action³¹.

Relatedly, our findings are limited by the two experimental groups studied and the tasks used. Future work is needed to test how well our results generalize to other etiologies of altered behavior and to other behavioral tasks.

One interesting hypothesis is that changes in expression of motifs are due to differences in cognitive control. In support of this, we found that behavioral phenotype was correlated with an individual’s ability to adapt their motif expression to a change in the environment. This suggests individual’s may differ in their ability to learn which motif to express in a given situation and is consistent with recent work suggesting that altered sampling of interactions between brain regions is associated with behavioral inflexibility³². In addition, recent work has shown that relative expression of different spatiotemporal activity patterns across brain region varies with behavioral state^33,34. Together, these findings support a ‘top-down’ model of motif expression, in which broad patterns of neural activity are selected at a given moment in time. Consistent with this, previous work has shown medial prefrontal cortex is disrupted in mice exposed to VPA¹⁴. Alternatively, differences in expression of motifs may reflect a more ‘bottom-up’ processing in which motifs spontaneously emerge from competing activity amongst local circuits. Future studies are needed to directly probe these hypothesis and investigate how the myriad of mechanisms thought to control cortex-wide dynamics – such as top-down cognitive control, neuromodulators, and internal state (e.g., homeostatic mechanisms)³¹ – can change across individuals and alter functional connectivity and behavior.

STAR Methods:

Resource Availability

Lead contact:

Further information and reasonable requests for resources should be directed to lead contact, Tim Buschman (tbuschma@princeton.edu)

Materials Availability

This study did not generate new unique reagents or materials.

Data and Code Availability

Widefield and behavioral data are publicly available (DOI: 10.5281/zenodo.10600896 and 10.5281/zenodo.10600809). Analysis code is publicly available at https://github.com/buschman-lab/CortexDynamicBehavioralPhenotype.

Experimental Model and Subject Details

All experiments and procedures were carried out in accordance with the standards of the Animal Care and Use Committee (IACUC) of Princeton University and the National Institutes of Health. Behavioral testing took place post-natal days (PND) 5-45, surgery and habituation PND 45-55, solo imaging PND 60-70, and paired imaging PND 80-100. Mice were group housed with same sex littermates prior to surgery and single housed post-surgery on a reverse 12-hr light cycle. All experiments were performed during the dark period, typically between 12:00 and 18:00. Animals received standard rodent diets and water ad libitum. Both male (n=23) and female (n=29) mice were used. All mice were offspring of 8 liters resulting from crosses between C57BL/6J and C57BL/6J-Tg(Thy1-GCaMP6f)GP5.3Dkim/J mice. 5 litters were exposed to valproic acid (VPA) in utero and 3 with saline controls. Offspring (VPA=27; saline=25) were used for behavioral experiments. Thy1-GCaMP6f heterozygous animals (VPA=11; saline=9) were also used for neural recordings. The first day of imaging in each animal was screened for potential epileptiform events associated with GCaMP6f lines, and none were observed, consistent with previous work expressing GCaMP under the Thy1 promoter²³ (Figure S2M). Histology on 8 VPA and 5 saline animals confirmed similar levels of GCaMP6f expression between groups (Figure S2J). Data from the 9 saline animals used for imaging has been reported in a previous study¹⁸.

Methods Details

In Utero Valproic Acid Model

The in utero VPA mouse model followed the procedures used in previous work³⁶. Pregnant dams were administered a single 600mg/kg subcutaneous injection of valproic acid sodium salt (Sigma Aldrich) dissolved in sterile water on embryonic day 12.5. Saline (control) litters received equivalent volume of saline.

Behavior and Development Testing

A battery of 9 behavioral tests was used to quantify animal behavior and general development. Tests, ordered by their post-natal day (PND), are reported below. The experimenter was blind to animal groups throughout behavioral testing and quantification of the behaviors. Experimental setups were cleaned with ethanol and allowed to dry between animals to remove olfactory cues. See Behavioral Measures subheading below for quantification details.

Maternal Separation Ultrasonic Vocalizations:

PND 5. Pups were assessed for isolation-induced ultrasonic vocalizations (USVs)³⁷. Tests were performed in a dark, quiet testing room after the litter had been habituated to the room for 10 minutes. The cage was kept on a warming pad throughout testing. Pups were removed from the litter and placed in a plastic dish within a soundattenuating chamber 5cm below an ultrasonic detector (Dodotronic Ultramic200K, Dodotronic; Part # UM200K). USVs were recorded for 3 minutes. Data was acquired using Audacity software on a PC (Windows 10). USV number was quantified using MUPET³⁸. Pup axillary temp was recorded prior to being replaced in the home cage. No difference in temperature was noted between groups (difference between means of each group = 0.57°F, p=0.89; Mann-Whitney U-Test).

Self-righting reflex:

PND 5-14. Pups were placed on their backs and given 10 seconds to self-right. This was repeated three times. Reflex was considered developed if the pup righted all times.

Eye Opening:

PND 12-18. The number of eyes open was recorded. Partially open eyes were considered open.

Three Chamber Maternal Scent Preference:

PND 14. A 11” x 8” plastic box was split into thirds. The lateral thirds were filled ~5mm deep with bedding from the pup’s cage (familiar) or the cage of a stranger litter. A pup was placed in the center third and the time spent in either side third was manually recorded for a duration of 60 seconds. An animal was considered in a given third if its nose was in that region. This was repeated 3 times. The cage was rotated 180 degrees between each repeat to control for side preference.

Three Chamber Social Approach and Social Novelty:

PND 40-45. Experiments were performed in a dimly lit room. Data was recorded using a video camera (Logitech Brio). Habituation: The test animal was habituated to a 23” x 14.75” custom three chamber polycarbonate box for 5 minutes (lateral chambers = 7.75” x 14.74”, Middle chamber 9” x 7.5”). Social Approach: After 5 minutes, the animal was secured in the center chamber and a metal cage was introduced to each lateral chamber. One cage contained a stranger animal, the other was empty. The test animal was released into the middle chamber, and activity recorded for 10 minutes. Social Novelty: After 10 minutes the test animal was secured in the center chamber and a new stranger mouse was added to the previously empty cage. The test animal was released, and activity recorded for 10 minutes. Stranger animals were age and sex matched and were randomly sampled from a pool of ~4-8 animals for each test mouse (with replacement). Sides were rotated between test mice. Animal activity was automatically quantified using custom MATLAB scripts that tracked the animals’ centroid (e.g., abdomen) over time.

Weaning Weight:

PND 20. Animals were weighed.

Marble Burying Assay:

PND 40-45. Performed in a dimly lit room. The test animal was habituated to a clean home cage with 1.5-inch-deep fresh bedding for 15 minutes. The animal was then removed from the cage and 14 marbles were placed on top of the leveled bedding in a grid like pattern. Animals were returned to the cage for 15 minutes after which the number of marbles >75% covered by bedding was recorded.

Repetitive Grooming Assay:

PND 40-45. Experiments were performed in a dimly lit room. The test animal was habituated to a clean home cage with no bedding for 15 minutes. Grooming behavior was then manually scored (duration and number of separate epochs) for 10 minutes.

Surgical Procedures

Surgical procedures followed previous work¹⁸. Mice were anesthetized (isoflurane: 1.5%) and administered analgesics (Buprenorphine, 0.1 mg/kg; Meloxicam, 1 mg/kg) and provided subcutaneous fluids (sterile saline, 0.01 mL/g). The dorsal scalp was shaved and disinfected with betadine and 70% isopropanol. The dorsal cranium was exposed, and the periosteum removed. The dorsal cranium was rendered optically accessible by applying a thin layer of clear dental acrylic (C&B Metabond Quick Cement System) polished with a rubber rotary tool tip (Shofu, part #0321; Dremel, Series 7700) and coated with clear nail polish (Electron Microscopy Sciences, part #72180). A custom titanium headplate with a 11mm trapezoidal window was cemented to the skull to permit head fixation. Mice recovered in a clean home cage and received post-op analgesia (Meloxicam, 1 mg/kg 24 hours post-surgery).

Widefield Imaging

Widefield imaging followed previous work¹⁸. Mice were head-fixed in a 1.5 x 4-inch polycarbonate tube under a custom-built fluorescence macroscope consisting of back-to-back 50 mm objective lens (Leica, 0.63x and 1x magnification), separated by a 495nm dichroic mirror (Semrock Inc, FF495-Di03-50x70). Excitation light (470nm, 0.4mW/mm²) was delivered through the objective lens from an LED (Luxeon, 470nm Rebel LED, part #SP-03-B4) with a 470/22 clean-up bandpass filter (Semrock, FF01-470/22-25). Fluorescence was captured in 75ms exposures (FPS = 13.3Hz) by an Optimos CMOS Camera (Photometrics). The macroscope was focused below superficial blood vessels, approximately 500um below the skull surface.

Mice were habituated to head fixation for ~5 minutes and then imaged for 12 minutes. Images were acquired as three, 4-minutes stacks of TIFF images at 980x540 resolution (~34um/pixel) using Micro-Manager software (version 1.4) on a PC. Mice were imaged for 4-6 consecutive days for a total of 105 recordings.

For paired imaging, animals were positioned in individual polycarbonate tubes. The animals faced one another, approximately eye to eye along the anterior-posterior axis and equidistant from the macroscope objective along the z axis. Animal snouts were separated by 5-7mm; close enough to permit whisking and social contact but prevent adverse physical interactions. A 1mm plexiglass divider at snout level ensured no paw/limb contact. To permit a wider (~30x20mm) field of view, the macroscope objectives were replaced with 0.63x and 1.6x magnification back-to-back objectives (lens order: mouse, 0.63x, 1.6x, CMOS camera). Images were acquired at 1960 x 1080 resolution (~34um/pixel) for 12 consecutive minutes. Each pair of animals was recorded for 3 consecutive days.

Mice pairs were age-matched, primarily non-littermate, and included same sex, opposite sex, VPA-VPA, VPA-saline, and saline-saline pairs. Each animal participated in 1-3 pairs over the course of 2-3 weeks with at least 2 days between new pairings, for a total of 165 sessions across all animals. Manual inspection identified 28 sessions in which whiskers of one animal entered the imaging field of view of the paired animal; these sessions were excluded from further analysis, yielding 137 sessions in total. Throughout imaging, animal pairs were provided with auditory stimuli consisting of naturalistic adult ultrasonic vocalizations, synthetic vocalizations, or ‘background’ noise; the details of which are supplied in our previous manuscript¹⁸.

Histology

Mice (n=13; saline=5, VPA=8) were transcardially perfused with 4.0% paraformaldehyde (PFA), brains were dissected and postfixed. Histochemistry was carried out on 50μm-thick free floating coronal brain sections and incubated in primary antibodies against green fluorescent protein (GFP) and counterstained with DAPI (Figure S2J). Sections were imaged with a Zeiss confocal microscope LSM 700 (solid-state lasers: UV 405; argon 458/488; Oberkochen, Germany). For cell density analysis sections were equally sampled to allow for cross comparisons and imaged at 20x, taking z-stacks at 2.0μm intervals. All slides were coded prior to analysis. All images were analyzed using FIJI.

Quantification and Statistical Analysis

Statistical Analysis

All analyses were performed in MATLAB (Mathworks). No statistical methods were used to predetermine sample sizes. The number of mice used in the study was based on previously published widefield imaging studies^39,40. All statistical tests, significance values, n values, and associated statistics are denoted in the main text or figure legends. Unless otherwise noted, hypothesis tests were one-tailed. Permutation tests were run using 1000 permutations. 95% confidence intervals were computed with 1000 bootstrapped samples. Correlation values were Fisher r-to-z transformed prior to computing secondary statistics (i.e., mean and confidence intervals). Unless otherwise noted, behavioral measures were taken from distinct samples (i.e., mice) while neural measures, such as motifs, were taken as the average of repeat measures in the same mouse (i.e., per recording epoch).

Behavioral Measures

Behavioral measures were quantified as follows:

Sensory-Memory:

Maternal Separation Ultrasonic Vocalizations:

Measured as the total number of USVs emitted.

Maternal Scent Preference:

Measured as the $\frac{{Time}_{familiar}}{{Time}_{familiar}+{Time}_{stranger}}$ across the three runs.

Social Approach Index.

Same as maternal scent preference but comparing the time in the compartment of the stranger mouse versus the compartment with the empty cage. Social approach index was bias-corrected for side preference by subtracting the magnitude of preference of the stranger animal compartment during habituation (relative to no bias: i.e., habituation bias - 0.5)

Social Novelty Index:

Same as maternal scent preference but comparing the time in the compartment of the ‘familiar mouse’ (e.g., stranger in approach) versus the compartment with the newly added stranger mouse.

Motor:

Marbles Buried:

The percentage of marbles (n=14) >=75% covered by bedding.

Grooming Duration:

The total duration (seconds) an animal spent self-grooming.

Number of Grooming Epochs:

The total number of separate grooming epochs. A grooming epoch was considered if it was >1 second in duration and >1 second separated from the previous epoch.

Exploration Velocity:

Computed as the average velocity (pixels/frame) of an animals’ centroid during the habituation period of the adult three chamber assay experiments.

Chamber transitions:

Computed as the total number of times an animals’ centroid crossed from one of the three chambers to another during the habituation, social approach, and social novelty experiments (collectively).

Developmental Milestones:

Self-righting reflex:

The first day the reflex was apparent.

Eye Opening:

The first day both eyes were open.

Weaning Weight:

Weight at PND 20.

Phenotype and Neurotype Axes

Phenotype and neurotype axes were defined as the vector normal to the hyperplane that maximally separated VPA and saline animals. To prevent overfitting, hyperplanes were defined by the average beta weights of 5 cross-validated hyperplanes fit on a random 50% folds of the animals. As only a subset of animals was used in imaging experiments, larger folds (60%) of the data were used to estimate the hyperplane. Hyperplanes were fit using MATLAB fitsvm function. Behavioral and motif data was scaled (i.e. z-scored across animals) prior to fitting.

An animal’s position along each axis was given by:

$$d=x^{\mathrm{T}}\beta +B$$ equation 1

Where $x^{\mathrm{T}}$ is the transpose of that animal’s 1x12 vector of the behavioral measures or percent explained variance captured by each motif (i.e. motif activity), $\beta$ is a 12 x 1 vector of hyperplane beta weights, and $B$ is the shift of the hyperplane relative to the origin (i.e. bias).

Widefield Imaging Preprocessing

Image stacks were cropped to a 540x540 pixel outline of the cortical window and aligned across sessions using user-drawn fiducials (sagittal sinus and bregma). Bregma was used to align the anatomical reference parcels overlaid in Figure 2. These overlays were created by manually tracing a 2D projection of the Allen Brain Atlas, version CCFv3⁴¹ (tracing performed with Inkscape Vector Graphics Software). To mitigate potential confounds in our widefield signal due to hemodynamic fluctuations^42-44, we used both automated and manual masking to conservatively remove non-neural pixels¹⁸. These masked pixels were ignored during subsequent spatial binning to 68x68 pixels (~136μm²/pixel). For each pixel, neural activity was computed as change in fluorescence, e.g., $\Delta \mathsf{F}∕\mathsf{F}$, over time:$\frac{F-F_0}{F_0}$ Baseline fluorescence, ${\mathsf{F}}_0$, was computed as the rolling mean of a 9750ms window (130 timepoints) across the entire recording duration. No difference in total variance of neural activity or number of imaged pixels were observed between groups (Figure S2I and K).

Next, we performed pixelwise deconvolution of the $\Delta \mathsf{F}∕\mathsf{F}$ signal into an estimate of neural activity using lucy-richardson deconvolution (lucid function from⁴⁵; kernel parameters: gamma=0.95, smt=1, p_num=30). This better estimates the underlying neural activity⁴⁶. Previously, we isolated significant activity using filtering and thresholding¹⁸; deconvolution performs the same function while minimizing data loss. After deconvolution, pixel values were normalized between zero and one to enable CNMF.

Paired-imaging recordings followed the same preprocessing steps.

Computing Functional Connectivity

Functional connectivity was computed as the fisher r-to-z transformed pairwise correlations between neural activity ($\Delta \mathsf{F}∕\mathsf{F}$) of a grid of regions of interest (ROI) within each cortical hemisphere (Figure 1F-G). For each ROI, neural activity was taken as the mean deconvolved fluorescence over time of a 9 pixels square centered on that ROI.

Identifying Spatiotemporal Motifs

We used a custom version of the seqNMF algorithm to discover spatiotemporal motifs in widefield imaging data (adapted from seqNMF MATLAB toolbox⁴⁷; changed to remove unused features and increase fitting speed). This method employs convolutional non-negative matrix factorization (CNMF) with penalty terms to facilitate discovery of repeating sequences. The general process for motif discovery and clustering across animals is detailed in our previous manuscript¹⁸. For clarity, we summarize the approach below and detail a few changes in parameter identification.

CNMF approximates the full (pixel by time) image sequence $X_{pt}$ as a sum of serial convolutions between a spatiotemporal tensor $W$ and its temporal weighting matrix $H$:

$$X_{pt}\approx {\tilde{X}}_{pt}=\sum _{k=1}^K\sum _{\ell =0}^{L-1}{W}_{pk\ell }{H}_{\mathrm{k}(\mathrm{t}-\ell )}\equiv {(W⊛H)}_{pt}$$ equation 2

where $K$ and $L$ are the maximum number of motifs and the maximum length of each motif, respectively and $⊛$ indicates the convolutional operator. The sets of motifs in $W$ and their temporal weightings in $H$ were found with an iterative multiplicative update algorithm.

The value of L was set to 13 frames (975ms). Motifs shorter than this duration were zero padded. This is above the duration of GCaMP6f event kinetics and matches the duration of spontaneous neural events found in previous work^18,48. The value of K was set to 30 motifs. This was higher than the maximum number of motifs discovered in any fit (Figure S2E).

The optimal spatiotemporal penalty term $\lambda$ of seqNMF was taken as the value that optimized the tradeoff between quality of fit, motif overlap, and number of discovered motifs. This was automatically determined for each fit by sweeping $\lambda$ across 6 orders of magnitude and identifying the cross-over point between quality of fit and motif overlap (Figure S2C). Additional orthogonality and sparsity parameters were held constant for all fits and are reported in Table S2.

The multiplicative update algorithm used for CNMF performs poorly on large datasets⁴⁹. Therefore, we fit motifs to six consecutive two-minute epochs of data from each twelve-minute session. Random initialization leads to variability in identified motifs across fits, so each fit was repeated 10 times. The best fit was taken as the fit with the lowest value of a relative AIC-like measure based on the magnitude of variance in the residuals (which were normally distributed):

$${AIC}_{relative}=n\ast log({\sigma }^2)+2\mathrm{k}$$ equation 3

Where $\mathsf{n}$ was equal to the number of data points (e.g. pixels x time) and k was the number of discovered motifs. This measure was used instead of other measures (e.g. explained variance) to avoid selecting fits that captured additional noise.

In total, 105 recordings from 20 animals yielded 4594 single motifs.

To identify motifs across animals, we clustered the single motifs using an unsupervised graph-based approach (Phenograph^50,51; see referenced paper for extensive exploration and validation of the algorithm). Phenograph generates a directed graph, where each node (here, a single motif) is connected to its $k$ (10) nearest neighbors. Distance between motifs was computed as the peak in their temporal cross correlation. Louvain community detection is then performed on this graph to group nodes into clusters. Prior to clustering, single motifs were smoothed with a 3D gaussian kernel (MATLAB imgaussfilt3; $\sigma =[1,1,0.1]$). Manual inspection revealed 8 of the identified clusters captured imaging artifacts (e.g., hair in imaging field, Figure S2F, inset) and were excluded from subsequent analyses. Collectively these noise clusters included only 182 (4%) of the identified motifs.

The overall (i.e., ‘shared’) motifs fit to all animals were estimated as the mean of the core community of single motifs in each cluster. The core community was defined as the motifs in each cluster that had the most within-cluster nearest neighbors. For a given cluster, the size of this fraction was swept from 1% to 100% of the cluster (5% increments). The optimal fraction was chosen as the value that produced an average motif with the strongest average correlation to all the underlying single motifs in that cluster. In this way, taking the average of the core community prevents averaging-out of the true spatiotemporal dynamics of each cluster. This is necessary because the convolutional nature of seqNMF can lead to motifs in the same cluster having slight temporal jitter and warping.

Prior to averaging, the motifs in the core community were aligned to a ‘template’ motif. The template motif shared the most zero-lag peak temporal cross-correlations with all other motifs within that cluster. If there were multiple templates with equivalent cross correlations, then one was chosen at random. All motifs were zero-padded to a length of $3L$ (39 timepoints) to allow for some temporal jitter and aligned by maximizing their cross-correlation lag to the template motif from their cluster. The average overall motif for the cluster was then estimated by averaging the activity of the temporally aligned core community. Finally, we aligned the motifs to one another by shifting the center of mass of activity to the middle timepoint. Timepoints with no variance across all overall motifs were removed. This removed frames without any neural activity in any motif and resulted in motifs with a duration of 27 timepoints (~2s).

The resulting 16 shared motifs identified across all animals were then refit to the each 2-minute epoch of data using the same seqNMF algorithm. Here, $W$ (e.g., the tensor of motifs) was fixed to prevent changes in the motifs while allowing their temporal weights ($H$) to be optimized. Prior to refitting, the original data were spatially smoothed with a 2D gaussian filter. The sigma of this filter was autofit for each epoch to best match the average spatial variance of the motifs (bounded between $\sigma =[0,0]$ and [1.5, 1.5]; median autofit value=[0.75, 0.75]). This same refitting procedure was used to refit the 16 motifs to the paired-imaging data.

While similar to our original approach for estimating motifs¹⁸, the approach outlined here introduces improvements in preprocessing (e.g., deconvolution), new automated approaches for parameter selection, and the addition of 11 new (VPA) animals to our analysis. Despite these changes, the number and structure of identified motifs was highly similar to our previous work (e.g. ~16 motifs shared across animals explaining ~75% of the variance in neural activity; see Table S1 for comparison of motifs between studies). The changes in preprocessing did result in fewer motifs discovered per 2-minute epoch (Figure S2E), likely due to less spatial noise in deconvolved data (verse the previously thresholded data) and the AIC-like criterion selecting more parsimonious individual fits.

The analysis pipeline was intentionally designed to avoid bias in the discovery of motifs and to prevent over clustering of motifs across animals. As noted above, experimenters were blinded to group identity throughout preprocessing, identifying, and clustering motifs. Automated heuristics were used to select CNMF parameters. Both CNMF and the Phenograph clustering algorithm are unsupervised, so the number and nature of motifs was not manually chosen. Decoder analyses (see main text and methods described below) validated that the clustering algorithm was performing as intended, not over clustering, and that within a cluster, there were no systematic differences between motifs of VPA-exposed and Saline animals. Furthermore, our previous³⁰ and concurrent⁵² work has shown that the number and structure of motifs is robust to major changes in the preprocessing, factorization, and clustering parameters.

Nonetheless, one concern might be that our approach tends to over-cluster motifs, leading to the observed similarity between groups. To validate our analysis pipeline, we created a simulated dataset with known differences in motifs. To this end, we split our 16 motifs into two groups (e.g. even motifs and odd motifs) and, for each group, created realistic simulated datasets by randomly tiling the motifs together over time (i.e., mimicking the ‘epochs’ of our original data). We then reran our motif identification pipeline de novo on these simulated datasets. As in the original pipeline, motif discovery was randomly initialized, leading to some variability in the exact structure of individual motifs. Despite this, clustering of the identified motifs re-revealed our 16 clusters. Furthermore, these clusters were accurately split by group: on average, 99% of the motifs in each cluster came from the group in which they were generated (range: 93%-100%).

Linear Classification to Compare Motifs

Linear classifiers with leave-one-out cross-validation were used to test whether the spatiotemporal structure of motifs differed between groups. For each of the 16 shared motifs, the single motifs from one animal that contributed to that shared motif were withheld, and a classifier was trained to differentiate between saline and VPA animals using the contributing motifs from all other animals (balanced between groups). The classifier was then tested to determine the group identity of the motifs from the withheld animal. This was repeated for all animals for each of the 16 overall motifs. These classifiers were unable to accurately assign group labels above chance (see Main text for statistics). Multiple classifier types were tested, including linear discriminant, support vector machine (svm) with standard linear kernels, svm with radial basis function kernels, svm with Bayesian-optimized hyperparameters, and svm with ANOVA-based feature selection. All failed to accurately distinguish motifs between groups. SVM classifiers with standard linear kernels were used for main text statistics.

To confirm classifier efficacy, a similar classification strategy was used to compare between motifs. For all 120 n-choose-two combinations of the 16 shared motifs, the contributing motifs from one animal were withheld and a classifier trained to identify the motif label of the motifs from the remaining animals. Classifiers were then tested to determine the label of the motifs in the withheld animal, which they did with high accuracy (see Main text for statistics).

Percent Explained Variance Calculations

Reconstructions of the original data were computed by convolving motifs with their temporal weightings (equation 2). Percent variance in neural activity captured by these reconstructions were computed as

$$PEV=100\ast \left(1-\frac{({\sigma }_X^2-{\sigma }_{\tilde{X}}^2)}{{\sigma }_X^2}\right)$$ equation 4

Where ${\sigma }_X^2$ and ${\sigma }_{\tilde{X}}^2$ denote the spatiotemporal variance of the original data and reconstructed data respectively. Relative PEVs, referred to as the ‘activity’ of single motifs, were calculated by dividing the PEV of each motif by the total PEV across all motifs for that epoch. These values were averaged across epochs and recordings for individual mice to yield a single value per motif per mouse.

Motif activity in paired context (Figure 4) was calculated the same way and the change in motif activity from alone to paired contexts was measured as the sum squared error between motif activity in each context. To determine if change in motif activity was greater than chance, a null distribution was estimated for each animal by randomly permuting the variance captured by motifs between the two contexts.

Key Resources Table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Experimental Models: Organisms/Strains
Mouse: Thy1-GCaMP6f: C57BL/6J-Tg(Thy1-GCaMP6f)GP5.3Dkim/J	The Jackson Laboratory	RRID:IMSR_JAX:028280
Deposited data
Widefield calcium imaging and behavioral data	This manuscript	DOI: 10.5281/zenodo.10600896 10.5281/zenodo.10600809
Software and Algorithms
MATLAB 2017a-2022b	Mathworks	N/A
Python version 3.6.4	Python Software Foundation	N/A
DeepLabCut Python Library	35	N/A
Signal Processing and Analysis Code	This manuscript	https://github.com/buschman-lab/CortexDynamicBehavioralPhenotype

Highlights.

Imaged cortex-wide neural dynamics in behaviorally ‘typical’ and ‘atypical’ mice.

All mice expressed the same set of 16 spatiotemporal patterns of cortical activity.

Individuals differed in the rate at which each of these patterns was expressed.

This predicted individual variability in behavior and functional connectivity.