Improving AlphaFold3 by Engineering MSA and Template Inputs

Area of Science:

• Structural biology
• Computational biology
• Biophysics

Background:

• AlphaFold3 offers a unified framework for predicting biomolecular structures and interactions.
• Its accuracy is contingent on the quality of multiple sequence alignment (MSA) and structural template inputs.
• Limited research exists on leveraging customized MSAs and templates to enhance AlphaFold3 performance.

Purpose of the Study:

• To systematically investigate the impact of diverse and engineered MSAs and templates on AlphaFold3 predictions.
• To evaluate the effectiveness of these customized inputs across protein monomers, multimers, and protein-ligand complexes.
• To compare AlphaFold3 performance with customized inputs against AlphaFold2 using identical inputs.

Main Methods:

• Systematic evaluation of AlphaFold3 performance using diverse and carefully engineered MSAs and templates.
• Benchmarking predictions for protein monomers, multimers, and protein-ligand complexes.
• Comparative analysis against default AlphaFold3 and AlphaFold2 with identical customized inputs.

Main Results:

• Consistent and substantial improvements in structure prediction accuracy were observed across all tested biomolecular types.
• Specific gains include higher TM-scores for monomers, improved DockQ scores for multimers, and lower ligand RMSD for protein-ligand complexes compared to default AlphaFold3.
• AlphaFold3 demonstrated significantly superior performance over AlphaFold2 when both utilized the same customized MSA and template inputs.

Conclusions:

• The study underscores the critical role of diverse and well-engineered MSAs and templates in enhancing AlphaFold3's predictive power.
• Customized inputs lead to significant accuracy gains, establishing a new state-of-the-art for AlphaFold3.
• This work provides a pathway for optimizing protein structure prediction through strategic input engineering.