HDGS-Net: nucleosome occupancy prediction based on a hybrid dilated gated separable convolutional neural network.

Nucleosome positioning plays a central role in chromatin organization and gene regulation, yet its accurate computational prediction remains challenging. This study introduces a Hybrid Dilated Gated Separable Convolutional Neural Network (HDGS-Net), which integrates dilated convolution, gated convolution, and depthwise separable convolution to achieve continuous prediction of in vitro nucleosome occupancy at single-base resolution across the entire Saccharomyces cerevisiae genome. On benchmark datasets, HDGS-Net attained an average Pearson correlation coefficient of 0.87, outperforming conventional methods and demonstrating excellent cross-chromosome generalization capability. Sequence analysis confirms that DNA dinucleotide physical properties dominate nucleosome positioning, with AT-rich sequences inhibiting binding and GC-rich sequences promoting binding. Analysis of transcription start regions verifies that flanking nucleosome sequence features are highly conserved across different chromatin environments, supporting the universal regulatory role of sequence preference. Cross-species analysis demonstrates that the guiding efficacy of DNA sequence on nucleosome positioning varies among species, showing quantitatively decreasing contributions in Caenorhabditis elegans, Saccharomyces cerevisiae, and Schizosaccharomyces pombe. This study provides a high-accuracy predictive tool for investigating dynamic nucleosome positioning.