Predictomes, a classifier-curated database of AlphaFold-modeled protein-protein interactions

Area of Science:

• Structural biology
• Computational biology
• Bioinformatics

Background:

• Protein-protein interactions (PPIs) are crucial for cellular functions, but their structural basis remains incompletely understood.
• AlphaFold-Multimer (AF-M) predicts protein complex structures but often yields false positives, hindering large-scale analysis.
• Existing confidence metrics in AF-M are insufficient for distinguishing accurate PPI predictions from noise.

Purpose of the Study:

• To develop a reliable method for filtering true positive PPI predictions from AF-M.
• To enable accurate proteome-wide structural interactome mapping.
• To facilitate hypothesis generation in biological processes like genome maintenance.

Main Methods:

• Trained a machine learning model, Structure Prediction and Omics-informed Classifier (SPOC), using curated datasets.
• Integrated structural prediction with omics data for enhanced classification accuracy.
• Applied SPOC to an all-by-all interaction matrix of human genome maintenance proteins.

Main Results:

• SPOC effectively distinguishes true AF-M PPI predictions from false positives.
• Generated approximately 40,000 high-confidence PPI predictions for human genome maintenance proteins.
• Developed predictomes.org for accessing predictions and scoring custom AF-M outputs.

Conclusions:

• SPOC provides a robust framework for interpreting large-scale AF-M predictions.
• The identified high-confidence PPIs can drive new research in genome maintenance.
• This work advances the development of a comprehensive structural interactome.