DeepATT DeepATT is a model for identifying functional effects of DNA sequences.

DeepATT is a model for identifying functional effects of DNA sequences. This is implemented by tensorflow-2.0. Our model has four built-in neural network constructions: convolution layer captures regulatory motifs, recurrent layer captures a regulatory grammar, category attention layer (improved from self-attention layer) selects corresponding valid features for different functions, and category dense layer (improved from local-connected dense layer) classifies the labels with feature vectors selected by the query vectors of the regulatory functions. We compare DeepATT with DeepSEA and DanQ, which are all implemented or replicated on our own platform. Comparison results demonstrate that DeepATT achieves state-of-the-art performance of 0.94519 AV-AUROC and 0.39522 AV-AUPR, which is far better than other non-coding DNA regulatory function prediction methods.