Improving deep learning protein monomer and complex structure prediction using DeepMSA2 with huge metagenomics data.

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• Accurate protein structure prediction is crucial for understanding biological function.
• Current methods rely heavily on the quality of multiple-sequence alignments (MSAs).
• Integrating diverse genomic and metagenomic data can improve MSA generation.

Purpose of the Study:

• To develop DeepMSA2, a novel pipeline for constructing uniform protein single- and multichain MSAs.
• To evaluate DeepMSA2's performance against state-of-the-art methods in protein structure prediction.
• To assess the impact of metagenomic data integration on MSA quality and downstream structure prediction.

Main Methods:

• Iterative alignment search across genomic and metagenome sequence databases.
• Development of the DeepMSA2 pipeline for MSA construction.
• Benchmarking DeepMSA2 MSAs against current methods for protein tertiary and quaternary structure prediction.
• Participation in the CASP15 experiment using an integrated pipeline with DeepMSA2.

Main Results:

• DeepMSA2 MSAs significantly increase the accuracy of protein tertiary and quaternary structure predictions.
• An integrated DeepMSA2 pipeline produced higher quality complex structural models than AlphaFold2-Multimer in CASP15.
• DeepMSA2's advantages stem from balanced alignment search, effective model selection, and integration of large metagenomic databases.

Conclusions:

• DeepMSA2 represents a significant advancement in MSA construction for protein structure prediction.
• Optimizing input data, such as MSAs, is as critical as predictor design for deep learning-based structure prediction.
• This work opens new avenues for improving deep learning protein structure prediction through enhanced MSA generation.