Behavioral decomposition reveals rich encoding structure employed across neocortex in rats

Our understanding of sensory and motor cortical processing is often limited by experimental designs that restrict animal behavior to pre-defined actions in response to experimenter-controlled stimuli. However, observations in head-fixed animals reveal that self-generated movements profoundly affect cortical activity patterns, even in the absence of explicit cognitive tasks. While recent research has linked self-generated movements to gain modulation and predictive processing in sensory cortices, most studies use head-fixed preparations and rarely examine sensory systems in tandem. This study aims to bridge this knowledge gap by investigating how movement-associated signals in freely behaving rats reflect their natural behavioral repertoire and whether the encoded features vary across different cortical regions to support region-specific processing. Advances in quantitative pose estimation and unsupervised machine learning techniques now allow for detailed characterization of animal behavior during neural recordings, enabling researchers to analyze naturalistic actions and their neural correlates. This study utilizes these advancements to explore the representation of behavior across four major cortical areas, determining the extent to which momentary behavior is represented and whether coding is uniform or varies by region.

Existing research on sensory and motor cortical processing predominantly relies on controlled laboratory settings, limiting our understanding of how these systems respond to the broader range of actions animals exhibit in natural contexts. Studies using head-fixed preparations have demonstrated the profound influence of self-generated movements on cortical activity, including primary visual cortex, even in the absence of overt cognitive demands. While significant progress has been made in linking self-generated movements to gain modulation and predictive processing, most studies have focused on single sensory systems in head-fixed preparations, leaving a gap in our knowledge regarding the representation of natural behavior across interconnected cortical areas. Recent technological advances in motion capture, pose estimation, and unsupervised machine learning have opened up possibilities for studying neural correlates of naturalistic behavior in freely moving animals, providing insights into how subcortical circuits generate behavior, encode action space, and respond to different pharmacological substances. These techniques also allow for the inference of emotional states from facial videos and the control of virtual rodent behavior, demonstrating their potential for detailed quantification of animal actions.

Seven male Long-Evans rats were implanted with Neuropixels probes targeting primary sensory and motor cortices (four rats) or visual and auditory cortices (three rats). Three-dimensional motion capture tracked head and back markers (seven markers total, including three on the trunk and four on a rigid head mount) at 120 frames per second, providing data for pose and movement analysis. An additional head-mounted accelerometer provided independent angular velocity data. Electrophysiological data were recorded simultaneously using Neuropixels probes, sampling large neuronal ensembles. Custom software was used to extract postural variables (head and back angles in allocentric and egocentric coordinates, neck elevation) and movement variables (derivatives of angles, running speed, body direction, self-motion). A behavioral clustering pipeline, adapting previous methods, decomposed the behavioral data into 44 discrete actions based on underlying structure in the tracking data. Spike sorting using Kilosort 2.0 and manual curation in Phy 2.0 identified single units, which were further classified as fast-spiking (FS) or regular-spiking (RS) based on their spiking profiles. Neural encoding of actions was assessed by comparing neuronal firing rates to action occurrence, and decoding accuracy was evaluated using a naive Bayes classifier. Generalized linear models (GLMs) with forward selection and 10-fold cross-validation were used to identify the behavioral covariates (posture, movement features) that best explained the neuronal activity. Anatomical topography of behavioral tuning was assessed across cortical regions. Sensory modulation (light and sound) was manipulated and quantified via modulation indices to study the overlap between sensory and behavioral responses. Putative synaptic connections were identified using spike cross-correlograms, and their relationship to behavioral tuning was analyzed.

Naturalistic actions were encoded ubiquitously across all four cortical regions. Decoding analyses revealed that almost all actions could be predicted from neuronal activity in each area, although decoder accuracy varied across regions. However, finer-grained features of pose and movement showed region-specific organization. Visual and auditory cortices primarily encoded head movement in allocentric (world-referenced) coordinates, with a graded increase in allocentric head posture coding progressing laterally across visual and auditory areas. Somatosensory and motor cortices predominantly encoded trunk and head movements in egocentric (body-referenced) coordinates. GLM analyses showed that visual cortical neurons were most strongly tuned to allocentric horizontal head movement and egocentric head posture, while auditory neurons were principally tuned to gravity-relative head orientation (allocentric head roll and pitch). Motor cortex showed the largest fraction of cells encoding low-level features, primarily planar body motion, back movement, and egocentric head posture. Somatosensory cortex primarily responded to planar body motion, back movement, and posture. Adding a weight to the animals' head had minimal effect on behavior or neural tuning. The integration of sensory and behavioral signals in visual and auditory cortices was substantial, with similar overlaps between sensory and behavioral tuning in both regions. Analysis of putative synaptic connections revealed distinct patterns in different regions. In visual cortex, feedforward excitation was observed between similarly tuned cells, along with inhibitory interactions between units encoding opposing movements. Also observed was excitation between posture-encoding and luminance-modulated units, suggesting a role in distinguishing self-generated from externally generated optic flow. In auditory cortex, movement-encoding fast-spiking cells inhibited sound-modulated units, possibly reflecting a mechanism for suppressing self-generated sounds. In motor and somatosensory cortices, the most abundant connection types involved feedforward excitation between similarly tuned movement-modulated neurons. There was no correlation between synaptic strength and functional distance in any cortical area.

This study provides strong evidence that ongoing behavior is encoded at multiple levels throughout the dorsal cortex, with region-specific specialization in encoding of elementary features of pose and movement. The widespread encoding of momentary actions suggests that ongoing behavior continuously modulates cortical computations, providing context for environmental inputs and informing sensory predictions generated by internal models. The region-specific encoding of posture and movement primitives reflects optimization for different types of sensory processing, with allocentric head-centered encoding in visual and auditory cortices potentially supporting the stabilization and prediction of visual and auditory fields during active movement, respectively. The analysis of synaptic connectivity revealed different roles for these signals in different areas, potentially supporting visual self-motion subtraction and sound localization. The findings highlight the importance of considering the interplay between behavior and sensory processing to achieve a comprehensive understanding of cortical function.

This study demonstrates the ubiquitous encoding of momentary actions and region-specific encoding of elementary pose and movement features across dorsal sensorimotor cortices in freely behaving rats. These findings support the notion that ongoing behavior significantly modulates cortical computations, informing sensory processing and predictive mechanisms. Future research could explore the interaction between internal states, hierarchical behavior, and cortical encoding across longer timescales. Furthermore, high-resolution techniques could be employed to investigate the precise circuit mechanisms underlying the integration of behavioral and sensory information.

The study focused on a specific set of actions and behavioral features in a familiar environment. The generalizability of findings to other behavioral contexts or species remains to be determined. The analysis of synaptic connections relied on correlations, which may not perfectly reflect direct functional relationships. The relatively small sample size of animals in each brain region could limit the statistical power of certain analyses. The sampling of cortical layers varied across regions.