Rēs ipSAE loquunt: What's wrong with AlphaFold's ipTM score and how to fix it

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Protein Interaction Prediction

Background:

• AlphaFold's TM-score is widely used to assess protein-protein interaction (PPI) structural prediction accuracy.
• Current TM-score formulations in AlphaFold 2/3 can be misleading when predicting PPIs involving proteins with disordered regions or accessory domains.
• This leads to inaccurate assessments of both prediction accuracy and the likelihood of interaction.

Purpose of the Study:

• To address the limitations of the AlphaFold TM-score in accurately evaluating PPIs.
• To develop a refined scoring metric that focuses on the interacting domains and is robust to non-interacting regions.
• To improve the discrimination between true and false positive protein-protein interactions predicted by AlphaFold.

Main Methods:

• Developed a new metric, interaction prediction Score from Aligned Errors (IPSAE).
• Modified the TM-score calculation to include only residue pairs with high predicted aligned error (PAE) scores.
• Adjusted the TM-score parameter to consider only residues with good inter-chain PAE, focusing on aligned residues.
• Utilized standard AlphaFold 2/3 JSON output files without altering the core AlphaFold code.

Main Results:

• IPSAE effectively calculates higher scores for domain-domain and domain-peptide interactions, even with extensive disordered regions.
• The new metric demonstrates improved ability to distinguish true protein complexes from false positives compared to the standard AlphaFold TM-score.
• The IPSAE program is implemented using only common AlphaFold output files, ensuring broad usability.

Conclusions:

• IPSAE provides a more accurate and reliable assessment of protein-protein interaction structural predictions.
• The method overcomes the limitations of the standard TM-score when dealing with complex protein constructs.
• IPSAE offers a valuable tool for researchers to better interpret AlphaFold predictions of protein interactions.