Application of deep learning for multi-scale behavioral analysis in SNCA E46K Parkinson's disease drosophila.

Abstract

Drosophila melanogaster is widely used as a model organism in Parkinson's disease research. However, due to the complexity of motion capture and the challenges of quantitatively assessing spontaneous behavior in Drosophila melanogaster, it remains technically difficult to identify symptoms of Parkinson's disease within Drosophila based on objective spontaneous behavioral characteristics. Here, we present an automated multi-scale behavioral phenotyping pipeline that classifies phenotypes related to Parkinson's disease using motion features extracted from pose estimation data of wild-type and Synuclein Alpha E46K mutant Drosophila melanogaster. Locomotor activity was recorded in a custom-designed 3D-printed behavioral trap, and body kinematics were analyzed using a markerless pose estimation tool to extract numerical features such as movement speed, tremor-like oscillations, and limb motion patterns. Beyond kinematic analysis, we applied unsupervised clustering to the pose-derived trajectories to extract recurrent movement subtypes that characterize spontaneous behavioral sequences. We found that kinematic features alone were insufficient to distinguish mutant flies from normal individuals, whereas behavioral sequence patterns captured through unsupervised clustering enabled robust group separation. Combining both feature types further enhanced classification accuracy, with the best model achieving 85%. This system provides an objective and scalable approach for analyzing behavior related to Parkinson's disease in Drosophila melanogaster, with potential applications in monitoring disease progression and screening pharmaceutical compounds.