A hybrid deep neural network for robust single-cell genome-wide DNA methylation detection

DNA methylation is an epigenetic mechanism that occurs when methyl groups are added to the 5th carbons of DNA cytosine residues. The process primarily takes place at CpG sites within the genome for the purpose of gene expression. Most cancerous cells result from aberrant DNA methylation, and the process is also linked to neurological disorders such as Alzheimer's and Parkinson's diseases. To discern the link between DNA methylation patterns and diseases, the methylation status of CpG sites throughout the genome must be known. Existing practical sequencing techniques can only map out methylation statuses for 10% to 40% of CpG sites. To address this deficiency, we have developed a hybrid deep neural network to estimate missing methylation statuses across the entire genome. The network was built with convolutional neural network layers and bidirectional LSTM neural network layers. The network extracts features from raw DNA sequences and creatively utilizes information contained in neighboring CpG sites. Our network achieved accuracy rates of 91% to 93% on the task of DNA methylation status identification, which is a statistically significant improvement over existing leading computational methods.