Protein structure prediction using deep learning distance and hydrogen‐bonding restraints in CASP14
Problem
Predicting 3D protein structures from sequence, especially hard free-modelling (FM) targets without templates, in CASP14.
Why it matters
Accurate protein structure prediction from sequence is a long-standing grand challenge in biology.
What the AI did
DeepPotential, a residual convolutional neural network, predicted spatial restraints from co-evolutionary MSA features; combined with DeepMSA2 and FUpred into D-I-TASSER and D-QUARK folding pipelines.
Results
For 37 FM targets, average TM-scores of first models by D-I-TASSER and D-QUARK were 96% and 112% higher than I-TASSER.
Impact
Blind CASP14 community benchmark showed substantial improvement over classical pipelines on hard targets.
Record
| Venue | Proteins Structure Function and Bioinformatics |
|---|---|
| Published | 2021 |
| Work type | article |
| Authors | Wei Hui Zheng, Yang Li, Chengxin Zhang, Xiaogen Zhou, Robin Pearce, Eric W. Bell, Xiaoqiang Huang, Yang Zhang |
| Citations | 61 (OpenAlex, at import time) |
| References | 58 |
| Model or method | D-I-TASSER, D-QUARK, DeepPotential (residual CNN), DeepMSA2, FUpred |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1002/prot.26193 |
| Open access | bronze |
| Keywords | protein structure prediction, CASP14, deep learning, TM-score, I-TASSER, QUARK |
Importance rating
major — millions of lives, an industry or a field
Protein structure prediction is a foundational biology challenge; abstract shows major blind-benchmark gains on hard FM targets but states no population/industry scale.