Deep learning molecular interaction motifs from receptor structures alone.

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in drug discovery

Background:

• Protein interactions rely on surface binding motifs.
• Traditional binder design is limited to known motifs, restricting novelty.
• Developing new methods for identifying potential binding sites is crucial.

Purpose of the Study:

• To present MotifGen, a deep learning network for predicting potential binder motifs from receptor structures.
• To generate human-interpretable motif profiles for functional groups and chemical interactions.
• To enable few-shot binder design applications and enhance novel binder discovery.

Main Methods:

• Developed MotifGen, a deep learning network.
• Generated motif profiles for 14 functional groups and 6 chemical interaction classes.
• Applied MotifGen to peptide binder design and small molecule binding site prediction.

Main Results:

• MotifGen predicts potential binder motifs directly from receptor structures.
• Generated motif profiles are human-interpretable and serve as pre-trained embeddings.
• Demonstrated effectiveness in peptide binder design and small molecule binding site prediction, outperforming or complementing existing methods.

Conclusions:

• MotifGen offers a novel, motif-centric approach for binder discovery.
• This method expands design strategies for challenging receptor targets.
• The approach facilitates the identification of novel binders without relying on pre-existing motif knowledge.