Abstract
Despite rapid advances in computational drug discovery, high-throughput screening (HTS) remains the primary method for identifying initial hits, particularly for targets with limited tractability to small molecules. Yet conventional HTS campaigns are costly and inefficient, often yielding hit rates below 2% and discarding valuable negative data. Here we present HTS-Oracle, a retrainable, deep learning-based platform that integrates transformer-derived molecular embeddings (ChemBERTa) with classical cheminformatics features in a multi-modal ensemble framework for hit prediction. We applied HTS-Oracle to the immune co-stimulatory receptor CD28, a prototypical difficult-to-drug target, and prioritized 345 candidates from a chemically diverse library of 1,120 small molecules. Experimental screening via temperature-related intensity change (TRIC) identified 29 hits (8.4% hit rate), representing an eightfold improvement over conventional methods such as surface plasmon resonance (SPR), TRIC, and affinity selection mass spectrometry (ASMS)-based HTS. By enriching true positives and filtering out non-binders upfront, HTS-Oracle streamlines the discovery pipeline and enables more focused, cost-effective screening. Two hit compounds disrupted the CD28-B7.1 interaction, with orthogonal validation provided by MST, ELISA, and molecular dynamics simulations. HTS-Oracle reduces screening burden and improves discovery efficiency, offering a powerful, scalable, and experimentally validated AI framework for accelerating hit identification across difficult-to-drug targets.