An Actor-Critic Reinforcement Learning Framework for Variant Evidence Interpretation

We present a reinforcement learning (RL) framework that uses established genomic metrics – such as the GuRu score, variant/gene risk priors, and population frequency – to estimate the probability of observing a given genetic variant in disease. Importantly, our approach does not directly predict variant pathogenicity; instead, it quantifies the cumulative evidence supporting a variant’s clinical observability within a Bayesian context. Using simulated genetic data with a range of variability and label noise, we systematically evaluated the actor-critic algorithm performance across multiple scenarios, employing metrics including receiver operating characteristic (ROC) curves, area under the curve (AUC), calibration plots, and learning dynamics. Results indicate predictive accuracy and effective learning, demonstrating RL’s potential as a practical tool for genomic variant interpretation, setting the stage for integration into a broader Bayesian classification framework.