Generating Genomic Maps of Z-DNA with the Transformer Algorithm

Z-DNA and Z-RNA were shown to play an important role in various processes of genome functioning acting as flipons that launch or suppress genetic programs. Genome-wide experimental detection of Z-DNA remains a challenge due to dynamic nature of its formation. Recently we developed a deep learning approach DeepZ, based on CNN and RNN architectures, that predicts Z-DNA regions using additional information from omics data collected from different cell types. Here we took advantage of the transformer algorithm that trains attention maps to improve classifier performance. We started with pretrained DNABERT models and fine-tuned their performance by training with experimental Z-DNA regions from mouse and human genome wide studies. The resulting DNABERT-Z outperformed DeepZ. We demonstrated that DNABERT-Z finetuned on human data sets also generalizes to predict Z-DNA sites in mouse genome.