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

Abstract

With the accumulation of DNA sequencing data, convolution neural network (CNN) based methods such as DeepBind and DeepSEA have achieved great success for predicting the function of primary DNA sequences. Previous studies confirm the importance of utilizing the reverse complement and flanking DNA sequences, which has a natural connection with data augmentation. However, it is not fully understood how these DNA sequences work during model training and testing. In this study, we proposed several CNN tricks to improve the DNA sequence related prediction tasks and took the DNA-protein binding prediction as an illustrative task for demonstration. Different from the DeepBind, we treated the reverse complement DNA sequence as another sample, which enables the CNN model to automatically learn the complex relationships between the double strand DNA sequences. This trick promotes the using of deeper CNN models, improving the prediction performance. Next, we augmented the training sets by extending the DNA sequences and cropping each one to three shorter sequences. This approach greatly improves the prediction due to more environmental information from extending step and strong regularization effect of the cropping step. Moreover, this practice fits well with wider CNN models, which also increases the prediction accuracy. On the basis of DNA sequence augmentation, we integrated the results of different effective CNN models to mine the prediction potential of primary DNA sequences. On 156 datasets of predicting DNA-protein binding, our final prediction significantly outperformed the state-of-the-art results with an average AUC increase of 0.057 (P-value = 6 × 10-62). Source codes are available at https://github.com/zhanglabtools/DNADataAugmentation. Supplementary data are available at Bioinformatics online.