Abstract
Major depressive disorder (MDD) is an etiologically diverse psychiatric disease with heterogeneous manifestations, making it difficult to diagnose with conventional assessment standards. In addition, the obvious incompatibility of the standard survey-based tests for human MDD and the behavioral assays for depressive-like phenotypes in mice makes clear the requirement for a non-invasive method for quantifying the expression of depressive-like state in naturalistic contexts. Here, we introduce a self-supervised machine learning platform, CLOSER (Contrastive Learning-based Observer-free analysis of Spontaneous behavior for Ethogram Representation), to monitor the spontaneous behavior in a depressive disease model with enhanced precision, reliability, and efficiency. This framework incorporates 3D pose skeleton data and kinematic features in a unique data augmentation strategy to characterize semantic behavioral syllables with a high-quality feature space. Using CLOSER, we uncovered distinct motion profiles in chronically stressed mice across both sexes and different disease stages. Furthermore, we quantified the drug-specific recovery of psychomotor symptoms, highlighting CLOSER's discriminative power for identifying drug efficacy. In offering an artificial intelligence (AI)-driven decoding of exploratory behaviors, CLOSER proposes the standardization of depressive-like phenotype assessment in mouse models, thereby bridging preclinical and clinical diagnostics for psychiatric drug discovery.