Multi-Granular Contrastive Alignment and Fusion for Fragment-Enhanced Virtual Screening

Virtual screening (VS) accelerates drug discovery by identifying bioactive molecules from large libraries. Recent deep learning methods treat VS as a dense retrieval task, embedding protein pockets and molecules into a shared space. However, these models rely on whole-molecule representations, limiting their ability to capture fine-grained, fragment-level interactions—despite the fragment-centric nature of the growing importance of fragment-based drug discovery (FBDD). We introduce FragCLIP, a fragment-centric, two-stage retrieval framework with multi-granular contrastive learning. Stage one learns to jointly embed protein pockets, molecules, and fragments, guided by non-covalent interaction (NCI) supervision. Stage two fuses molecule- and fragment-level embeddings into a unified representation. This design enables accurate alignment of interaction-relevant fragments with compatible pockets while preserving efficiency. FragCLIP boosts early enrichment (EF1) on DUD-E from 31.89 to 37.23 and outperforms docking and deep learning baselines on a new fragment-level benchmark FragBench. FragCLIP bridges molecular and substructure-level reasoning, offering a practical foundation for structure-based virtual screening in realistic FBDD workflows.