A deep learning model leveraging semantic features fusion for DNase I hypersensitive sites identification in the...

DNase I hypersensitive sites (DHSs) are chromatin regions that are extremely sensitive to the DNase I enzyme, increasing their accessibility for cellular processes. DHSs are crucial for understanding transcriptional regulation mechanisms and contain genetic variations linked to various diseases such as breast cancer, coronary artery disease, Alzheimer's disease, autoimmune disorders, and neurological conditions. However, conventional DHSs identification methods are labor-intensive and resource-heavy, necessitating the need for alternative cost-effective approaches with high performance.

In this study, we propose various computational models, namely the CNN model, the CNN-GRU fusion model, the CNN-kmer fusion model, and the CNN-GRU-kmer fusion model, to overcome the challenges associated with DHSs prediction. The CNN Model is based on a simple 1-dimensional convocational neural network (CNN). The CNN-GRU fusion model is based on a simple 1-dimensional CNN and gated recurrent unit (GRU) and then fuses the feature maps of CNN and GRU. The CNN-kmer fusion model is based on a simple 1-dimensional CNN and k-mer features. First, we input the k-mer features to a dense layer; the output of the dense layer is fused with CNN features. In the CNN-GRU-kmer fusion model, based on simple 1-dimensional CNN, GRU, and k-mer features, first we input the k-mer features to a dense layer; the output of the dense layer is fused with CNN features and GRU features and fed to a dense layer with a sigmoid function for prediction.

The proposed models were validated in the publicly available dataset, obtaining an accuracy of 0.8631, a sensitivity of 0.7209, a specificity of 0.9353, an MCC of 0.6468, an AUC ROC of 0.8528, and an AUC PR of 0.7530. These results surpass all performance evaluation metrics of state-of-the-art models.

This study presents that the model integrates semantic vector-based feature fusion representation, which effectively captures both local and global patterns with inherited spatio-temporal dependencies within complex DHSs sequences. The model's performance was validated both with and without semantic feature fusion, followed by quantitative and statistical analyses against individual models, significantly enhancing feature representation and classification performance. Source code and datasets are available at: https://github.com/malikmtahir/DNase/tree/main.