Single‐sequence‐based prediction of protein secondary structures and solvent accessibility by deep whole‐sequence learning
Problem
Predicting protein secondary structure and solvent accessibility from a single sequence alone, without relying on evolutionary information from multiple sequence alignments.
Why it matters
Protein structure prediction underpins drug design and understanding of disease-related proteins.
What the AI did
SPIDER3-Single uses Long Short-Term Bidirectional Recurrent Neural Networks (LSTM-BRNNs) to predict secondary structure, solvent accessibility, backbone angles and related properties from single sequences.
Results
Q3 accuracy of 72.5%; solvent accessibility correlation coefficient of 0.67; more accurate than evolutionary-based method for proteins with few homologs
Impact
The method improves single-sequence prediction, especially for proteins with few sequence homologs.
Record
| Venue | Journal of Computational Chemistry |
|---|---|
| Published | 2018 |
| Work type | article |
| Authors | Rhys Heffernan, Kuldip K. Paliwal, James Lyons, Jaswinder Singh, Yuedong Yang, Yaoqi Zhou |
| Citations | 118 (OpenAlex, at import time) |
| References | 32 |
| Model or method | LSTM-BRNN (SPIDER3-Single) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1002/jcc.25534 |
| Open access | closed |
| Keywords | protein structure, deep learning, LSTM, secondary structure, solvent accessibility |
Importance rating
real but narrower advance
Single-sequence protein structure prediction supports drug design and disease-protein study; a useful methodological advance in a foundational field, but abstract states no scale of impact.