Improving protein structure prediction with extended sequence similarity searches and deep-learning-based refinement...

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Protein Structure Prediction

Background:

• The Critical Assessment of Structure Prediction (CASP) competition benchmarks protein structure prediction methods.
• DeepMind's AlphaFold2 and AlphaFold-Multimer represent state-of-the-art tools in protein structure prediction.
• Enhancing existing tools through input/output modifications is a viable strategy for improving prediction accuracy.

Purpose of the Study:

• To detail the methodology employed by the PEZYFoldings team in the CASP15 competition.
• To identify key factors contributing to the team's high performance in protein structure prediction.
• To analyze the impact of specific modifications on prediction accuracy and ranking.

Main Methods:

• Utilized AlphaFold2 and AlphaFold-Multimer as baseline prediction tools.
• Enhanced input data by collecting evolutionarily related sequences using extensive similarity searches.
• Incorporated a deep-learning-based refinement model and performed manual interventions, including domain parsing.

Main Results:

• Achieved 1st place (assessor's formulae) and 3rd place (Global Distance Test Total Score [GDT-TS]) in the single-domain category.
• Secured 10th place in the multimer category.
• Extensive sequence similarity searches were identified as a primary contributor to performance; manual interventions also showed significant impact on specific targets.

Conclusions:

• The combination of enhanced sequence data and manual interventions significantly boosted protein structure prediction accuracy.
• The refinement model had a minimal impact on absolute accuracy but improved self-estimated accuracy, aiding ranking.
• Future improvements lie in multimer prediction, database expansion, direct prediction from primary sequences, and automating manual interventions.