Simple tricks of convolutional neural network architectures improve DNA-protein binding prediction

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Convolution neural network (CNN) models like DeepBind and DeepSEA excel at predicting DNA sequence functions.
• The utility of reverse complement and flanking DNA sequences in data augmentation is recognized but not fully understood.
• Understanding the role of these sequences in model training and testing is crucial for advancing predictive accuracy.

Purpose of the Study:

• To propose and evaluate novel CNN techniques for enhancing DNA sequence prediction tasks.
• To demonstrate the effectiveness of these techniques using DNA-protein binding prediction as a model system.
• To improve the predictive performance of computational models for genomic sequence analysis.

Main Methods:

• Developed CNN strategies including treating reverse complement DNA as a distinct sample to capture double-strand relationships.
• Implemented data augmentation by extending DNA sequences and cropping them into shorter segments to incorporate environmental information and regularization.
• Integrated predictions from multiple CNN models to maximize the predictive potential of DNA sequences.

Main Results:

• The proposed CNN tricks significantly improved DNA-protein binding prediction accuracy.
• Treating reverse complement DNA as a separate sample facilitated the use of deeper CNN architectures.
• Sequence augmentation, incorporating extended and cropped sequences, enhanced model performance by providing more contextual information and regularization.
• The integrated model achieved state-of-the-art results on 156 DNA-protein binding datasets, with an average AUC increase of 0.057 (P-value = 6 × 10⁻⁶²).

Conclusions:

• Novel CNN techniques, particularly those involving reverse complement and sequence augmentation, substantially enhance DNA sequence prediction.
• These methods enable the development of more powerful and accurate predictive models for genomic tasks.
• The findings provide a robust framework for improving computational predictions in genomics and molecular biology.