Advancing Computational Frameworks for Animal Behavioral Quantification and Understanding

Abstract

Behavior is the ultimate output of an organism's brain. The ability to precisely measure and analyze the diversity and variability manifested in animal behavior is thus required for establishing a mechanistic understanding of neural systems. Computational advances, particularly the development of automated markerless kinematic tracking methods, have enabled such measurements, significantly transforming neuroscience and neuroethology into data-driven quantitative sciences. In this dissertation, we review the recent successes of existing behavioral quantification and analysis techniques and present several novel computational systems based on machine vision and deep learning, respectively addressing the limiting factors present in the current behavioral quantification and analysis pipelines. Towards high-resolution tracking of social behavior, we developed a framework for tracking highly resolved 3D postural dynamics in freely interacting moving laboratory rodents, named social-DANNCE (s-DANNCE), together with a suite of analytical tools yielding multi-scale behavioral representations in dyadic interactions. We applied the pipeline to collect a large database of different rat models of autism. Our ASD model phenotyping surpassed conventional standards and revealed a spectrum of changes in rodent models of autism not resolved by conventional measurements. Through performing a longitudinal monitoring of a litter of juvenile rats during postnatal development, we additionally approached the challenging social tracking problem using a modular system for tracking large cohorts of laboratory rodents in 3D. Our framework was able to resolve some of the most occlusive behaviors during complex interactions and reveal the richness in behavioral repertoire exhibited by the juveniles during early development.Towards more flexible and versatile behavioral analysis, we introduced an end-to-end behavioral analysis framework named VQ-MAP capable of harmonizing heterogeneous datasets for collective analysis, resulting in fast, interpretable delineation of the shared behavioral repertoire. We deployed VQ-MAP to align rodent behavioral datasets with variable cardinality and arrangement, supporting objective comparative studies across multiple species and rodent models of autism. In longitudinal recordings of rat behavioral maturation, VQ-MAP revealed developmental trajectories connecting specific behaviors across periods of substantial bodily growth and change. Collectively, these new computational paradigms advance the quantification strategies used for ethology and neuroscience studies, expanding the boundaries of inquiry and opening new avenues for novel phenotyping assays of drug and disease.