Improving AlphaFold2- and AlphaFold3-Based Protein Complex Structure Prediction With MULTICOM4 in CASP16.

Area of Science:

• Computational biology
• Structural biology
• Bioinformatics

Background:

• Accurate prediction of single-chain protein structures is largely solved by tools like AlphaFold.
• Predicting the structure of multichain protein complexes (multimers) remains a significant challenge.

Purpose of the Study:

• To develop and evaluate MULTICOM4, an advanced system for protein complex structure prediction.
• To improve the accuracy of modeling multichain protein assemblies.

Main Methods:

• Integration of AlphaFold2 and AlphaFold3 with novel in-house techniques.
• Development of protein complex stoichiometry prediction.
• Generation of diverse multiple sequence alignments (MSAs) using sequence and structure comparisons.
• Implementation of exception handling and deep learning-based model quality assessment.

Main Results:

• MULTICOM predictors achieved top performance in CASP16 Phase 0 (no stoichiometry info), with MULTICOM_human reaching an average TM-score of 0.752.
• In CASP16 Phase 1 (with stoichiometry info), MULTICOM_human remained a top predictor, achieving an average TM-score of 0.797.
• Demonstrated superior performance in protein complex structure prediction compared to other methods.

Conclusions:

• Combining AlphaFold2 and AlphaFold3 with enhanced inputs and ranking strategies improves multimer prediction.
• Accurate stoichiometry prediction and robust modeling techniques are crucial for advancing protein complex structure prediction.