Foundational Model-aided Automatic High-throughput Drug Screening Using Self-controlled Cohort Study

The process of developing new drugs, from initial discovery to obtaining regulatory approval, has historically been neither cost-efficient, expeditious, nor free from risk. The growing availability of large-scale observational healthcare databases, combined with the rise of foundational models, offer an unparalleled opportunity to enable automatic high-throughput drug screening for both repurposing and pharmacovigilance. In this work, we present a general workflow for automatic high-throughput drug screening which estimates the association between various drug exposures and disease outcomes. We provide frameworks for parsing the accurate exposure length for each prescription from clinical texts and removing confounding relationships between drugs and diseases using bioinformatic mapping and foundational models. Using a self-controlled cohort study design, we tested the intention-to-treat association between 3,444 medications and 276 diseases across 6.6 million UK patients from the Clinical Practice Research Datalink (CPRD). Our analysis revealed 16,901 drug-disease pairs with significant risk reduction, indicating candidates for repurposing, as well as 11,089 pairs with significant risk increase which raise drug safety concerns. Our data-driven, nonparametric, hypothesis-generating, and automatic approach demonstrates the potential of foundational models in drug discovery and provides a scalable framework for drug repurposing that can be extended to other observational medical databases.