Gene–Environment Interactions in Predicting Self-Harm: A Machine Learning Approach Using Explainable Artificial Intelligence

Background: Self-harm is a critical public health concern; nevertheless, the complex interplay between genetic predispositions and environmental factors in self-harm remains poorly understood. Methods: This study employed data from 156,873 participants in the UK Biobank to investigate how polygenic risk scores (PRSs) for 15 psychiatric disorders/traits interact with environmental risk factors in predicting lifetime self-harm. Automated machine learning identified the optimal predictive model, while explainable artificial intelligence techniques were applied to assess feature importance and interactions. Results: Environmental factors, particularly interpersonal trauma, such as partner belittlement and sexual assault, demonstrated stronger predictive value than genetic factors. However, gene–environment interactions accounted for approximately 12% of the variance in predicted self-harm risk, with major depression, cannabis use disorder, and anorexia nervosa PRSs exhibiting the strongest interaction effects. Conclusions: This study’s findings suggest that individuals with genetic vulnerabilities may be particularly susceptible to interpersonal trauma, highlighting the need for personalized prevention strategies addressing combined genetic and environmental risks.