Deep‐learning contact‐map guided protein structure prediction in CASP13
Problem
Predicting 3D protein structures for targets that lack homologous templates, where conventional pipelines performed poorly.
Why it matters
Protein structures underpin drug design and understanding of disease, but experimental determination is slow and costly.
What the AI did
Deep residual convolutional neural-network contact prediction (ResPRE) plus a NeBcon meta-predictor generated contact-maps that guided structure assembly simulations in C-I-TASSER and C-QUARK.
Results
Average TM-scores of first models were 28% (C-I-TASSER) and 56% (C-QUARK) higher than I-TASSER/QUARK on 50 CASP13 free-modelling domains.
Impact
Independent blind CASP13 assessment verified the deep-learning contact-map improvement over baseline pipelines.
Record
| Venue | Proteins Structure Function and Bioinformatics |
|---|---|
| Published | 2019 |
| Work type | article |
| Authors | Wei Hui Zheng, Yang Li, Chengxin Zhang, Robin Pearce, S. M. Mortuza, Yang Zhang |
| Citations | 208 (OpenAlex, at import time) |
| References | 54 |
| Model or method | ResPRE, NeBcon, C-I-TASSER, C-QUARK |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1002/prot.25792 |
| Open access | bronze |
| Keywords | protein structure prediction, deep learning, contact maps, CASP13, ResPRE |
Importance rating
significant — thousands to millions affected
Template-free protein structure prediction underpins drug design and disease understanding; blind CASP13 gains on 50 FM domains show a significant capability advance in a foundational field.