SurfBind predicts epitopes from 3D molecular surfaces with SOTA accuracy
SurfBind reads 3D molecular fingerprints to predict antibody binding sites accurately.
SurfBind, developed by Fang Wu, Weihao Xuan, Jure Leskovec, Yejin Choi, and Li Erran Li, directly models the geometric and physicochemical fingerprints of 3D molecular surfaces to predict antibody-antigen binding sites (epitopes). Unlike prior methods that rely on sequences or backbone structures—which struggle with discontinuous surface-driven epitopes—SurfBind operates entirely on surface mesh representations. Its architecture combines a Transformer backbone with patch-level surface modeling, binder-aware cross-attention that accounts for the antibody partner, and a hierarchical coarse-to-fine prediction pipeline.
On standard benchmarks SAbDab and DB5.5, SurfBind achieves state-of-the-art results and demonstrates strong generalization to unseen antibodies and conformational states. The work highlights the value of interaction-aware surface modeling for understanding protein-protein interactions, with potential impacts on drug design and antibody engineering.
- SurfBind uses a Transformer architecture with patch-level surface modeling and binder-aware cross-attention for epitope prediction.
- Achieves state-of-the-art on SAbDab and DB5.5 benchmarks, generalizing to unseen antibodies and conformational states.
- Directly exploits 3D molecular surface geometry and physicochemical features, overcoming limits of sequence- or backbone-based methods.
Why It Matters
Enables more accurate antibody design and drug discovery by predicting binding sites directly from molecular surfaces.