Integrating deep learning, threading alignments, and a multi‐MSA strategy for high‐quality protein monomer and complex structure prediction in CASP15
Problem
Predicting 3D structures of proteins and complexes, especially hard free-modeling targets, remains difficult.
Why it matters
Accurate protein structure prediction is a long-standing core problem in biology and medicine.
What the AI did
D-I-TASSER combined deep-learning restraints, threading alignments, multi-MSA ranking, and multi-domain modules for monomer prediction; DMFold-Multimer handled complexes, tested in the blind CASP15 experiment.
Results
Average TM-score 19% higher than standard AlphaFold2 on 47 CASP15 free-modeling targets
Impact
Outperformed standard AlphaFold2 on the hardest blind-assessed CASP15 targets.
Record
| Venue | Proteins Structure Function and Bioinformatics |
|---|---|
| Published | 2023 |
| Work type | article |
| Authors | Wei Hui Zheng, Qiqige Wuyun, Lydia Freddolino, Yang Zhang |
| Citations | 58 (OpenAlex, at import time) |
| References | 67 |
| Model or method | D-I-TASSER, DMFold-Multimer |
| Organization | UM-TBM and Zheng groups (institutions not stated) |
| Primary source | https://doi.org/10.1002/prot.26585 |
| Open access | hybrid |
| Keywords | protein structure prediction, CASP15, D-I-TASSER, DMFold-Multimer, AlphaFold2 |
Importance rating
significant — thousands to millions affected
Protein structure prediction is a core problem in biology and medicine; gains shown on 47 CASP15 free-modeling targets and 38 complexes; foundational field, but abstract states no impact scale.