Improving deep learning protein monomer and complex structure prediction using DeepMSA2 with huge metagenomics data
Problem
Protein tertiary and quaternary structure prediction loses accuracy when multiple-sequence alignments (MSAs) are poorly constructed, especially for complexes.
Why it matters
Accurate protein structure prediction underpins drug design and understanding of disease mechanisms.
What the AI did
DeepMSA2 builds improved single- and multichain MSAs via iterative alignment search of genomic and metagenome databases, feeding deep learning structure predictors to raise accuracy.
Results
CASP15 complex models of considerably higher quality than AlphaFold2-Multimer server v.2.2.0; no numeric values stated in abstract
Impact
Independent CASP15 participation and large-scale benchmarks support the claimed improvement over state-of-the-art methods.
Record
| Venue | Nature Methods |
|---|---|
| Published | 2024 |
| Work type | article |
| Authors | Wei Hui Zheng, Qiqige Wuyun, Yang Li, Chengxin Zhang, Lydia Freddolino, Yang Zhang |
| Citations | 130 (OpenAlex, at import time) |
| References | 40 |
| Model or method | DeepMSA2 pipeline (with deep learning structure prediction) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1038/s41592-023-02130-4 |
| Open access | hybrid |
| Keywords | protein structure prediction, MSA construction, metagenomics, CASP15, AlphaFold2-Multimer, deep learning |
Importance rating
significant — thousands to millions affected
Protein tertiary/quaternary structure prediction is a foundational capability for biology and drug design; abstract shows state-of-the-art gains via CASP15 and large-scale benchmarks, but no scale numbers.