Drug-few: A few-shot learning method for target-specific drug virtual screening with interaction-informed adaptation

Ao Shen, Qiao Huang, Mingzhi Yuan, Jie Du, Yingfan MA, Manning Wang

Keywords: Drug virtual screening; Few-shot learning; AI for science

Abstract: Few-shot virtual screening (VS), which aims to identify active molecules for a target with only a few known ligands, is crucial for accelerating drug discovery in cases where experimental data are extremely limited. Traditional ligand-based few-shot learning methods rely on molecular similarity or latent embeddings to generalize from these limited examples, but they often fail to capture target-ligand interactions, leading to overlooked active molecules. Here, we present Drug-few, the first few-shot learning framework designed for strict target-specific VS that explicitly incorporates binding-relevant information. Drug-few introduces prompt tokens for rapid target-specific adaptation and incorporates lightweight adapter modules to refine pocket-ligand representations. These components are combined in a Gated Prompt Adapter (GPA), where the contribution of prompt tokens is dynamically modulated by interaction-aware signals. Extensive experiments on three benchmark datasets show that Drug-few consistently outperforms the zero-shot baseline, maintaining strong retrieval performance even when the target active molecules are dissimilar to the known ligands, demonstrating its ability to generalize to novel molecules.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 4047