Deep contrastive feature compression with classical machine learning enables ligand discovery through efficient triage of large chemical libraries

Improving in silico compound-protein interaction (CPI) predictability is critical for productive drug discovery. Current deep learning approaches largely rely on end-to-end models trained on limited labeled CPI data, overlooking preexisting, specialized compound-protein input representations. We present Ligand Extra trees-Accelerated Docking (LEAD), a virtual screening framework that accelerates docking by integrating rapid first-pass CPI prediction via ET-Screen. Unlike end-to-end models, ET-Screen uses seven distinct representations, including embeddings from large-scale protein language models and molecular transformers, as well as an original CPI potential fingerprint. ET-Screen’s contrastive compression neural network maps the consolidated 2,457-dimensional compound-protein representation into a compact, discriminative form optimized for CPI classification by an ensemble of decision trees. ET-Screen outperforms state-of-the-art end-to-end approaches by up to 23.39% on diverse retrospective benchmarks while being ∼100× faster than standard-precision docking. Its speed enables triaging ∼10.8 million prospective drug candidates across five targets, reducing them to 10,000 for docking within the LEAD framework and ultimately yielding novel, experimentally-validated hits.