Gated residual neural networks with self-normalization for translation initiation site recognition

Deep neural networks have contributed to significant progress in complex system modeling of biology. However, the existing computational methods cannot extract discriminative features for the translation initiation site (TIS) of complex biology systems, and feature representation methods heavily rely on statistical information, which is not informative enough, thus leading to unsatisfactory performance. To address the problems, we first pre-train and generate co-occurrence embedding of genomic sequences and then propose the competitive neural framework called TISNet by gating convolutional and attention mechanisms and identify TIS from genomic sequences alone. Specifically, we devise a novel gated residual network that contains a gated convolutional residual unit and a gated scaled exponential unit. The gating mechanism not only promotes the propagation of features but also alleviates gradient vanishing problems. Besides, to extract features at different scales, we further introduce a multiscale convolutional block and an attention block to directly learn local and long-distance patterns; then we use a special fusion block to combine information at the local and global levels. Extensive experiments indicate the superiority of TISNet over previous methods and show k-mer co-occurrence information can improve the performance which provides some biological insights into the regulatory mechanism of TIS in complex biology systems.