From pose to behavior: SABER integrates identity-resolved multi-animal pose tracking with language-model-based behavioral factor discovery

SABER is a framework that enhances social behavior quantification by accurately assigning posture and actions to individual animals. Developed for single overhead video streams, it overcomes challenges posed by close contact, occlusion, and identity switches. By integrating multianimal pose estimation with identity-preserving tracking, SABER provides interpretable behavioral factor mining and multiscale temporal behavior prediction.

In tests involving spontaneous social interactions among two to four mice, SABER demonstrated superior pose estimation accuracy and tracking continuity compared to existing pipelines. The behavioral factor-mining procedure identified distinct kinematic, postural, and social descriptors, while temporal integration improved classification of behavioral categories. SABER features an intuitive, open-source interface for easy use.

Applied to social defeat stress mice, SABER successfully detected reduced approach behavior and a multivariate behavioral profile that differentiated depression-susceptible from control animals. Additionally, SABER's outputs can be synchronized with miniscope calcium recordings, enabling joint analysis of behavioral states and neuronal population activity. Overall, SABER offers an accessible, identity-resolved route from single view social interaction videos to behavioral phenotyping and brain behavior analysis.