An open question in brain disorder research is whether behavioral phenotypes in animal models are conserved across species. This question is especially pressing for autism research, where in rodent models it has been challenging to detect social phenotypes and other behavioral hallmarks of the disorder, leaving the utility of these models in doubt. Pinpointing conserved phenotypic elements would provide comparative insights generating new mechanistic hypotheses and supporting new tests of animal model validity. However, the field lacks a framework for rigorously and quantitatively comparing behavioral phenotypes across species.

A recent SFARI-funded three-dimensional (3D) pose tracking method enabled deep behavioral phenotyping in both interacting rats and mice, but technical challenges have made it difficult to map these phenotypes from one species to the other. The issue is that for existing analysis approaches, species-specific differences in body proportions and 3D pose formats, including the number and anatomical placement of tracked keypoints, make these datasets incompatible.

The goal of this proposal is to develop a new computational method permitting flexible identification of behavioral motifs across heterogeneous pose formats, thus enabling joint analysis of mouse and rat phenotypes. To achieve species and pose flexibility, the method will use transformer neural network modules inspired by the latest advances in large AI foundation models. These modules will accommodate arbitrary data structures and identify behavioral motifs from inferred true body kinematics rather than from observed keypoint coordinates directly.

The approach will be built on and used to analyze the SFARI-funded large-scale s-DANNCE dataset to draw quantitative parallels between the rat and mouse models of autism. This public dataset comprises over 140 million 3D full-body poses measuring the lone and dyadic social behavior of 7 different rat monogenic models of autism (Arid1b, Fmr1, Grin2b, Scn2a, Chd8, Arid1b, and Cntnap1 knockouts) and 2 mouse strains, including BALB/c, a mouse model of autism. Behaviors found to be modulated similarly across species will suggest potentially conserved pathways, which can be tested in future experiments. The method will also have a broad impact on neuroscience and biomedicine beyond autism.

By unifying behavioral data analysis across labs, experiments, and species, the tool can be used to assemble large-scale community data repositories that promote collective neurobehavioral knowledge building. This project will also establish a foundation for comparative behavioral phenotyping beyond just rodents, building towards a future where more rigorous links between animal models and human brain disorders transform translation from bench to bedside.