Analysis of distance‐based protein structure prediction by deep learning in CASP13
Problem
Predicting 3D protein structures from sequence alone, especially for free-modeling targets without templates, was a major unsolved challenge.
Why it matters
Protein structures underpin understanding of function and drug design, but experimental structure determination is slow and costly.
What the AI did
RaptorX servers used deep convolutional residual networks (ResNet) to predict residue-residue contacts and build 3D models without time-consuming conformational sampling.
Results
Best contact prediction among 46 CASP13 groups; top L/5, L/2, L long-range precision of 70%, 58%, 45%; correct folds (TMscore > 0.5) for 18 of 32 FM targets
Impact
Blind community-wide CASP13 benchmark independently ranked RaptorX best for contact prediction among 46 groups.
Record
| Venue | Proteins Structure Function and Bioinformatics |
|---|---|
| Published | 2019 |
| Work type | article |
| Authors | Jinbo Xu, Sheng Wang |
| Citations | 132 (OpenAlex, at import time) |
| References | 50 |
| Model or method | RaptorX (deep convolutional ResNet) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1002/prot.25810 |
| Open access | green |
| Keywords | protein structure, CASP13, RaptorX, ResNet, contact prediction |
Importance rating
major — millions of lives, an industry or a field
Free-modeling protein structure prediction is a foundational scientific challenge; abstract reports benchmark results but no quantified real-world impact scale.