Nature or Nurture Depends on the Individual: Person Specific Artificial Intelligence Suicidality Risk Models

Refining predictive models of suicidality could improve early intervention efforts. “Heterogeneity”, which suggests there are unique individual-specific pathways through which different people come to experience suicidal thoughts and behaviors, is a pervasive problem that hinders the clinical utility of polygenic risk scores and environmental predictors. Recent strides in artificial intelligence (AI) have demonstrated that models can be trained to predict in the presence of heterogeneity. This proposal aims to parse the heterogeneity of clearly defined suicidality related phenotypes through the development of AI models using the latest data sources from large scale genetic studies of inherited and single cell transcriptomic risk and multivariate models of environmental (aka “exposome”). Specifically, a multi-PGS library will be built from summary statistics available in a GWAS catalog. Then, the model will be expanded through generation of functional annotations from a single cell RNA sequencing dataset. Using patterns across the AI model, we can infer population trends in contributions to heterogeneity. In a parallel analysis, we will build an exposome risk score (ERS) from unique environmental measures of environmental exposures and train the model to determine if using genetic predictor jointly with exposome predictors improves our ability to predict and distinguish between distinct suicidal behaviors. This work will lay the foundation for personalized prediction models with better clinical utility to improve intervention efforts.