Comprehensive evaluation of deep learning architectures for prediction of DNA/RNA sequence binding specificities
Problem
The relative merit of CNN, RNN, and hybrid deep learning architectures for predicting DNA- and RNA-binding specificity was unclear from existing studies.
Why it matters
Reliable prediction of DNA/RNA-binding specificities is important for biological and biomedical research.
What the AI did
Built deepRAM, an end-to-end deep learning tool with fully automatic model selection, enabling a fair, unbiased comparison of architectures for binding prediction.
Results
Deeper, more complex architectures give a clear advantage with sufficient training data; hybrid CNN/RNN architectures outperform others.
Impact
The systematic benchmark clarified which deep learning architectures best predict DNA/RNA binding specificity.
Record
| Venue | Bioinformatics |
|---|---|
| Published | 2019 |
| Work type | article |
| Authors | Ameni Trabelsi, Mohamed Amine Chaâbane, Asa Ben‐Hur |
| Citations | 153 (OpenAlex, at import time) |
| References | 39 |
| Model or method | deepRAM (CNN, RNN, hybrid CNN/RNN) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1093/bioinformatics/btz339 |
| Open access | hybrid |
| Keywords | deep learning, CNN, RNN, DNA/RNA binding, deepRAM, benchmark |
Importance rating
real but narrower advance
Benchmarking architectures for DNA/RNA-binding prediction aids biological research, but abstract states no scale of impact; a methodological comparison.