Fast and accurate Ab Initio Protein structure prediction using deep learning potentials
Problem
Predicting 3D protein structures ab initio remains inaccurate for proteins lacking sequence or structural homologs.
Why it matters
Structures of hard targets underpin drug design and disease understanding, yet such proteins are difficult to model.
What the AI did
DeepFold integrates spatial restraints predicted by multi-task deep residual neural networks with a knowledge-based energy function to guide gradient-descent folding simulations.
Results
On hardest targets, average TM-score 40.3% higher than trRosetta and 44.9% higher than DMPfold; folding simulations 262 times faster than fragment assembly.
Impact
Open-source DeepFold substantially outperforms leading methods on homolog-scarce protein targets in large-scale benchmarks.
Record
| Venue | PLoS Computational Biology |
|---|---|
| Published | 2022 |
| Work type | article |
| Authors | Robin Pearce, Yang Li, Gilbert S. Omenn, Yang Zhang |
| Citations | 44 (OpenAlex, at import time) |
| References | 43 |
| Model or method | DeepFold (multi-task deep residual neural networks) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1371/journal.pcbi.1010539 |
| Open access | gold |
| Keywords | protein structure prediction, deep learning, DeepFold, ab initio folding, TM-score |
Importance rating
significant — thousands to millions affected
Ab initio protein structure prediction for homolog-scarce proteins is a foundational structural biology problem; abstract shows large accuracy and 262x speed gains but no disease/industry scale.