Targeting protein–ligand neosurfaces with a generalizable deep learning tool

TL;DR

Presents a deep learning-based method to design proteins targeting neosurfaces of protein-ligand complexes, experimentally validated.

Problem

Designing new proteins that recognize neosurfaces generated by protein-ligand interactions is challenging with existing computational methods. Such chemically induced protein interactions are crucial for developing new therapeutic modalities like molecular glues and cell-based therapies.

Approach

The authors leveraged geometric deep learning to learn molecular surface representations. They designed proteins that bind to neosurfaces of protein-ligand complexes, experimentally validating against three drug-protein complexes: Bcl2-venetoclax, DB3-progesterone, and PDF1-actinonin.

Results & Contribution

Designed binders showed high affinity and accurate specificity, confirmed by mutational and structural characterization. Notably, surface fingerprints trained only on proteins generalized to neosurfaces induced by small molecules, demonstrating rare generalizability. This method can contribute to developing new therapeutic modalities such as drug-controlled cell-based therapies.