Petabase-Scale Homology Search for Structure Prediction

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• Multiple sequence alignments (MSAs) are crucial for protein structure prediction, as evidenced by AlphaFold2's success in the CASP15 competition.
• The effective utilization of MSAs remains a key area for advancing prediction accuracy.

Purpose of the Study:

• To investigate the impact of large-scale Sequence Read Archive (SRA) data on protein structure prediction accuracy.
• To evaluate the contribution of SRA-derived homologs and advanced ColabFold features to prediction performance.

Main Methods:

• A petabase-scale search of the SRA was performed to generate extensive aligned homologs for CASP15 targets.
• SRA-derived MSAs were merged with default ColabFold-search MSAs and used as input for ColabFold-predict.
• The effects of deep homology search and ColabFold's advanced features (e.g., increased recycles) were systematically tested.

Main Results:

• Using SRA data improved highly accurate predictions (GDT_TS > 70) to 66% for non-easy targets, compared to 52% with default ColabFold-search MSAs.
• The inclusion of SRA homologs was the primary factor in improving ColabFold's CASP15 ranking from 11th to 3rd place.
• Advanced ColabFold features also contributed to the overall improvement in prediction accuracy.

Conclusions:

• Expanding the scope of MSAs through large-scale data sources like the SRA is a powerful strategy for enhancing protein structure prediction.
• Combining SRA data with optimized computational approaches like ColabFold offers significant improvements in predicting protein structures.
• Further analysis of deep homology and advanced prediction features is warranted to fully leverage MSA data.