Predictomes: A classifier-curated database of AlphaFold-modeled protein-protein interactions

Area of Science:

• Molecular Biology
• Structural Biology
• Bioinformatics

Background:

• Protein-protein interactions (PPIs) are crucial for biological processes, but comprehensive structural data is limited.
• Existing tools like AlphaFold-Multimer (AF-M) predict PPIs, but their confidence metrics often yield numerous false positives.
• A reliable method is needed to filter accurate PPIs from large-scale predictions for biological discovery.

Approach:

• Developed a machine learning model named Structure Prediction and Omics informed Classifier (SPOC).
• Trained SPOC on curated datasets to effectively differentiate true PPIs from false positives generated by AF-M.
• Applied SPOC to an extensive dataset of human genome maintenance proteins, generating thousands of high-confidence PPI predictions.

Key Points:

• SPOC demonstrates high performance in identifying genuine PPIs, even in large-scale proteome analyses.
• The study generated approximately 40,000 PPI predictions for human genome maintenance proteins.
• Predicted PPIs are accessible via predictomes.org, allowing users to evaluate their own predictions with SPOC.

Conclusions:

• SPOC provides a robust framework for interpreting large-scale AF-M predictions and filtering reliable PPIs.
• The identified high-confidence PPIs offer novel hypotheses for genome maintenance research.
• This work establishes a foundation for constructing a comprehensive proteome-wide structural interactome.