UC Riverside

The prediction of DNA methylation from the primary DNA sequence allows one to impute the methylation status of cytosines with insufficient sequencing coverage. Various deep learning models have been proposed in the literature, including transformer-based models and convolutional neural networks. In this study, we investigate the performance of long-range state-space models based on the Hyena architecture on the task of DNA methylation prediction on six plant species. First, we train the HyenaDNA framework to obtain a genome-wide foundation model for each species. Then, we fine-tune these foundation models using the sequence data surrounding the methylated or unmethylated cytosines. Extensive experimental results show that our model predicts DNA methylation with higher accuracy than state-of-the-art methods in the literature.