Improved multimer prediction using massive sampling with AlphaFold in CASP15.

Area of Science:

• Computational biology
• Structural biology
• Bioinformatics

Background:

• AlphaFold2 significantly advanced protein structure prediction accuracy.
• Improving predictions for protein multimers remains a challenge.
• AlphaFold2's internal scoring function is key to ranking its predictions.

Purpose of the Study:

• To enhance protein multimer structure prediction by leveraging AlphaFold2's scoring function through massive sampling.
• To explore the impact of different sampling strategies on prediction quality.

Main Methods:

• Conducted extensive AlphaFold2 runs (274,289 models) across 38 CASP15 targets using six distinct settings.
• Enabled dropout layers during inference for uncertainty sampling and model diversity.
• Utilized both multimer v1 and v2 weights, with and without templates, and varied recycle counts.

Main Results:

• Achieved substantial improvement over the baseline NBIS-AF2-multimer, increasing mean DockQ from 0.43 to 0.56.
• Generated a high number of models per target (median 4810), leading to improved quality assessment.
• Identified multimer v1 as more susceptible to sampling improvements than v2.

Conclusions:

• Massive sampling, particularly with dropout and multimer v1, effectively boosts AlphaFold2's protein multimer prediction accuracy.
• The strategy significantly improved prediction quality for challenging targets.
• The method and code are publicly available for further research.