Identifying polyadenylation signals with biological embedding via self-attentive gated convolutional highway networks

Abstract Polyadenylation is responsible for regulating mRNA function and translation efficiency. The accurate polyadenylation signal (PAS) recognition is essential for understanding the genome annotation and function. Currently, computational methods mainly rely on human-designed features and do not extract complex discriminative patterns. In this work, we first adopt a shallow neural network to automatically learn biological embedding information with discriminative patterns from biological sequences alone. Next, we devise a hybrid deep learning framework termed PASNet to automatically extract underlying patterns by integrating gated convolutional highway networks with a self-attention mechanism and then identify PAS from genomic sequences. Specifically, we devise a gated convolutional highway unit by gated convolutional mechanisms, and the highway unit has significant paths to allow unimpeded information to pass between two adjacent layers. PASNet requires little prior knowledge and no laborious feature engineering. Besides, we evaluate PASNet on four different organism datasets through genome-wide experiments. Results show that our framework achieves the improvement with the error rate decreasing by 4% and 1.02% respectively on average compared to the published machine learning-based predictors and deep learning predictors for PAS identification.