Predicting protein–protein interactions through sequence-based deep learning
Problem
Experimental protein-protein interaction (PPI) data has low coverage and is noisy, so new PPIs and errors in existing data are hard to identify.
Why it matters
PPI knowledge underpins understanding of cellular function and disease, while experimental detection is costly and error-prone.
What the AI did
DPPI, a deep Siamese-like convolutional neural network with random projection and data augmentation, predicts PPIs from sequence information alone.
Results
Outperforms state-of-the-art methods on several benchmarks in auPR; computationally more efficient; predicts homodimeric interactions where other methods fail.
Impact
Benchmark evidence supports improved sequence-based PPI prediction over prior methods.
Record
| Venue | Bioinformatics |
|---|---|
| Published | 2018 |
| Work type | article |
| Authors | Somaye Sadat Hashemifar, Behnam Neyshabur, Aly Azeem Khan, Jinbo Xu |
| Citations | 366 (OpenAlex, at import time) |
| References | 46 |
| Model or method | DPPI (Siamese-like CNN) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1093/bioinformatics/bty573 |
| Open access | bronze |
| Keywords | deep learning, protein-protein interactions, sequence-based prediction, benchmarks, convolutional neural network |
Importance rating
significant — thousands to millions affected
PPI knowledge underpins cellular function and disease understanding, a broad biological field; abstract gives benchmark gains but no scale of impact.