COLLAPSE: A representation learning framework for identification and characterization of protein structural sites

Area of Science:

• Computational biology
• Structural bioinformatics
• Protein science

Background:

• Identifying functional sites in proteins is vital for understanding biological mechanisms, disease risk, and targeted therapies.
• The rapid increase in protein structure data outpaces functional annotation capabilities.
• Current functional prediction methods have limitations, including lack of site-specificity, high error rates, and extensive dataset requirements.

Purpose of the Study:

• To develop a novel computational method for large-scale functional site annotation in proteins.
• To create a framework that learns effective representations of protein sites directly from structural and evolutionary information.
• To enable efficient and accurate prediction of protein function.

Main Methods:

• Developed COLLAPSE (Compressed Latents Learned from Aligned Protein Structural Environments), a deep learning framework operating on 3D atomic coordinates of protein sites.
• Utilized evolutionary relationships from homologous proteins as a self-supervision signal to learn site representations.
• Applied transfer learning to evaluate the generalizability of learned representations across various functional annotation tasks.

Main Results:

• COLLAPSE achieved state-of-the-art performance on benchmarks for protein-protein interactions and mutation stability.
• Accurate prediction of functional sites from the Prosite database was demonstrated.
• The framework successfully identified and annotated functional sites across large protein datasets, showcasing its efficiency and interpretability.

Conclusions:

• COLLAPSE provides a computationally efficient and interpretable platform for analyzing protein structures and annotating functional sites.
• The learned representations implicitly capture crucial structure-function relationships.
• This framework advances the ability to annotate protein function at scale, supporting biological discovery and drug development.