GESTURE: unsupervised genotype-specific behavioral phenotyping in rodents via graph-based representation learning

Abstract

Objective quantification of complex animal behavior is fundamental to neuroscience and pharmacology, yet extracting biologically meaningful insights from high-dimensional pose data remains challenging. This is especially true for subtle phenotypic differences caused by genetic mutations, which frequently evade conventional evaluation metrics. Here, we introduce GESTURE, an unsupervised, graph-based deep generative framework that autonomously discovers and quantifies behavioral structure from raw pose dynamics. Analyzing mice with motor dysfunction alongside wild-type controls, GESTURE identifies a shared vocabulary of behavioral motifs. We show that genotype-specific differences arise from the differential usage of these motifs, yielding distinct "behavioral fingerprints" that reliably separate genotypes without supervision. By modeling behavior as a sequence rather than a static partition, GESTURE measures temporal organization directly: affected animals held motifs longer and transitioned more predictably, revealing a slowing and stereotyping of behavioral sequences rather than a simple reduction in activity. Importantly, GESTUREs graph-based representation enables training across multiple recordings and embedding behaviors into a shared latent space, supporting robust cross-animal comparisons and future cross-experiment alignment. Furthermore, automatically derived metrics of behavioral divergence track the temporal dynamics of expert-annotated disability scores, reaching agreement comparable to independent human raters. Finally, node- and edge-level explainability analyses indicate that the models latent representations are shaped by a biologically plausible focus on the animals core motor scaffold. Together, these results position GESTURE as an interpretable and scalable framework for automated behavioral phenotyping, linking genetic perturbation to quantitative behavioral phenotypes. Author summaryHow can we measure, objectively, the subtle changes in behavior caused by a disease-causing mutation? Such changes are hard for human observers to judge consistently, and scoring them by hand is slow and does not scale. We studied mice carrying a mutation in a calcium-channel gene. In people, mutations in the same gene cause episodic ataxia and related movement disorders. We developed GESTURE, which learns recurring patterns of movement directly from video of a freely moving mouse, without being told what to look for. It treats the body as a set of connected landmarks and follows how their arrangement changes over time, discovering a shared vocabulary of movements. Affected and healthy mice drew on the same vocabulary but used it differently. Each animals pattern of use formed a distinctive "behavioral fingerprint", and these fingerprints alone identified which animals carried the mutation --in every case. They also revealed something a simple activity measure would miss: affected mice hold each movement longer, and string movements together more predictably. A severity score read automatically from them matched expert judgement as closely as two trained raters matched each other, and returned the same answer every time.