Protein functional site annotation using local structure embeddings.

Area of Science:

• Proteomics
• Bioinformatics
• Computational Biology

Background:

• Protein databases are rapidly expanding, leading to many proteins with unknown or ambiguous functions.
• Current machine learning methods struggle to link global protein function to specific responsible residues.

Purpose of the Study:

• To introduce PARSE (Protein Annotation by Residue-Specific Enrichment), a novel knowledge-based method for predicting protein function and annotating residue-level functional sites.
• To address the limitation of current methods in associating global function with specific residues.

Main Methods:

• PARSE combines pretrained embeddings of local protein structural environments with statistical techniques.
• It performs simultaneous prediction of global function and residue-level annotations.
• The method is knowledge-based and does not require supervised training, enabling one-shot predictions for rare functions.

Main Results:

• For enzyme catalytic function prediction, PARSE achieves performance comparable or superior to state-of-the-art methods (F1 score ~85%).
• PARSE provides significantly more precise residue-level annotations for identified functions.
• The method successfully predicted bacterial metalloproteases from the "dark proteome" using AlphaFold structures, revealing conserved catalytic sites despite sequence and fold divergence.

Conclusions:

• PARSE offers a powerful tool for annotating protein function and identifying functionally critical residues, particularly for proteins with unknown or rare functions.
• Its ability to leverage local structural information and operate without supervised training opens new avenues for function discovery in large-scale proteomic datasets.
• The method demonstrates the utility of local structure representations for discovering new functions, even in proteins with divergent sequences and folds.