Toward a Self-Learning AI Agent for Drug Repurposing: Building Human-Scale Representations for Virtual Patients

Virtual patients need a useful coordinate system before they can model patient-specific dynamics. Most models in AI for science operate at molecular, cellular, or organ-specific scales. However, treatment decisions are made across a whole person and across different forms of intervention. Here, we use drug repurposing as a human-scale testbed for a central virtual-patient question: can a structured representation of biological directions organize intervention-relevant knowledge well enough to prioritize known drug-disease relationships?We introduce SteeraMed Bench, a framework for evaluating module panels built from a 332-module atlas. The atlas combines extended aging hallmarks, traditional Chinese medicine syndrome proxies, nutraceutical targets, and food-as-medicine targets. Rather than assuming that all modules should form one universal model, the framework compares panels as alternative representations for each disease task. Across 1,916 DrugBank small molecules, five chronic disease tasks, and an exploratory extension to 23 disease categories, the panels carried useful within-benchmark ranking signal. The nutraceutical and nutraceutical-extension panel (NUT+NUTX, 117 modules) achieved a mean recall@20 of 0.494 across five diseases, close to 0.524 for the full atlas. The full atlas was the strictly highest observed configuration in only 9 of 23 disease categories. Different panels were most useful for different tasks, including extended aging hallmarks for type 2 diabetes and osteoporosis, food-as-medicine for depression, and nutraceutical modules for the atherosclerosis/hyperlipidemia task.The framework also evaluates newly proposed gene sets for incremental value and redundancy. In an exploratory LLM-assisted workflow, two refined candidates showed nominal positive increments, but neither remained significant after correction for multiple testing. Performance decreased under target-family-separated evaluation and approached chance in leave-one-disease-out evaluation. SteeraMed Bench lays the coordinate and evaluation foundation on which future patient-specific dynamic models and causal intervention simulations can be built.