Abstract
High-throughput screening (HTS) remains a central pillar of small molecule discovery yet routinely fails for immune receptors and protein-protein interaction-driven targets. Here, we introduce HTS-Oracle, an experimentally validated AI system for prospective hit discovery that integrates molecular language modeling with cheminformatics to prioritize bioactive compounds at scale. We deploy HTS-Oracle across three clinically validated yet historically intractable immune targets, TREM2, CHI3L1, and CD28, representing cryptic binding pockets, intrinsically disordered proteins, and protein-protein interaction-driven immune checkpoint, respectively. Across the tested targets, HTS-Oracle reduces experimental screening requirements by up to >99% while increasing hit rates by up to 176-fold relative to traditional HTS. Notably, the platform remains predictive under extreme data sparsity, achieving an eightfold improvement for CD28 despite fewer than 2% actives in training. By consistently enriching for experimentally validated hits, HTS-Oracle establishes a new performance benchmark for hit discovery and unlocks small molecule access to immune targets long regarded as chemically inaccessible.