Autobehaver: An AI-Based Pipeline for Animal Behavior Analysis

Abstract Behavior arises from the complex interplay between an organism’s nervous system, its genetic makeup, and the environment. High-resolution, high-throughput behavioral quantification is essential for dissecting biological function and the effects of genetic perturbation, but automated analysis remains challenging. Here, we present Autobehaver, an automated behavioral analysis pipeline based on a low-cost, high-throughput recording platform that captures videos of individual Drosophila . From each video, we extracted keypoints and used a custom Transformer to assign frame-wise behavior and orientation labels. We then converted these predictions into high-dimensional per-animal feature vectors and trained XGBoost ensembles to classify animals and identify the features that separated groups. By applying SHAP analysis to the classifier ensemble, we identified the behavioral features most informative for distinguishing groups of flies. We demonstrated the approach in several ways. First, we recovered known behavioral changes associated with heat-activated dTrpA1 activity in specific neural circuits. Second, we detected age-associated behavioral changes consistent with gradual impairment of locomotor and climbing ability. Finally, we used Autobehaver’s classifier ensemble to place animals with intermediate phenotypes along a behavioral axis and used feature-importance analysis to reveal the behavioral features underlying those intermediate states. Together, Autobehaver provides an interpretable framework for quantitative behavioral phenotyping and comparative analysis of complex genotypes.