Comprehensive evaluation of deep learning architectures for prediction of DNA/RNA sequence binding specificities

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning

Background:

• Deep learning architectures are increasingly used for predicting DNA- and RNA-binding specificity.
• Current methods include convolutional neural networks (CNNs), recurrent neural networks (RNNs), and hybrid CNN/RNN models.
• The comparative performance of these architectures is not well-established.

Purpose of the Study:

• To systematically explore and compare deep learning architectures for DNA- and RNA-binding specificity prediction.
• To introduce deepRAM, a tool for implementing and evaluating various deep learning models.
• To provide guidelines for selecting appropriate network architectures.

Main Methods:

• Developed deepRAM, an end-to-end deep learning tool.
• Implemented a wide selection of deep learning architectures.
• Utilized a fully automatic model selection procedure for unbiased comparison.

Main Results:

• Deeper and more complex architectures yield better performance with adequate training data.
• Hybrid CNN/RNN architectures demonstrate superior accuracy compared to other methods.
• Recurrent networks enhance model accuracy but decrease feature interpretability.

Conclusions:

• Hybrid CNN/RNN architectures are recommended for DNA- and RNA-binding specificity prediction.
• Model complexity and data availability are key factors in performance.
• Practitioners can use these findings to select optimal architectures and understand model trade-offs.