Modeling Withdrawal States in Opioid-Dependent Mice with Machine Learning

Abstract

Understanding opioid withdrawal behaviors in preclinical models is critical to improving therapeutic approaches for opioid use disorder (OUD). However, quantifying these withdrawal behaviors remains a difficult process for researchers, given the subtlety of behaviors and variation across individuals. To overcome these difficulties, we developed a scalable behavioral analysis pipeline using LUPE (Light aUtomated Pain Evaluator), an open-source framework integrating video acquisition, pose estimation, supervised and unsupervised classification, and expert-guided behavior discovery. Mice undergoing naloxone-precipitated opioid withdrawal were recorded and analyzed using DeepLabCut for markerless pose estimation. We hand-annotated withdrawal-specific behaviors, including jumping, genital licking, grooming, and paw tremors, and normal behaviors, including walking, rearing, and being still, using Behavioral Observation Research Interactive Software (BORIS) to generate frame-by-frame ethograms. The annotations and pose data were then imported into Active learning Segmentation of Open field in DeepLabCut (A-SOiD), an active learning platform for behavior classification. A-SOiD successfully detected some behaviors (e.g., grooming and rearing) which were of a longer duration, though other rapid behaviors (e.g., jumping and paw tremors) were inconsistently captured. While no novel behavioral motifs have been discovered yet, ongoing work aims to refine model performance. This LUPE-based pipeline sets the groundwork for standardized, high-resolution behavior quantification and is being applied to additional datasets to investigate whether new components of the withdrawal phenotype emerge across experimental conditions.