Mapping the space of protein binding sites with sequence-based protein language models

Area of Science:

• Computational biology
• Structural bioinformatics
• Drug discovery

Background:

• Protein binding sites are crucial for biological activity and are key targets for therapeutic interventions.
• Effective methods for detecting, comparing, and contextualizing binding sites are of significant interest in drug discovery.

Purpose of the Study:

• To present a novel approach for generating detailed representations of protein binding sites.
• To demonstrate the utility of these representations in downstream machine learning tasks for drug discovery.

Main Methods:

• Integration of protein language models with a 3D tessellation technique.
• Derivation of binding site representations combining functional, structural, and evolutionary information.
• Development of similarity metrics that balance local structure and global sequence effects.

Main Results:

• Generated rich and versatile binding site representations with unprecedented detail.
• Demonstrated that similarity metrics induce meaningful pocket clusterings.
• Showcased the simplification of downstream tasks including pocketome organization, druggability modeling, and benchmarking.

Conclusions:

• The developed method provides powerful embeddings for understanding and utilizing protein binding site information.
• This approach facilitates efficient contextualization of new binding sites and improves the development of machine learning models for drug discovery.