Melody: Decoding the Sequence Determinants of Locus-Specific DNA Methylation Across Human Tissues

DNA methylation is a fundamental epigenetic modification that plays crucial roles in transcriptional regulation, cellular differentiation, and genome stability. However, how locus-specific DNA methylation is determined by intrinsic DNA sequence features remains poorly understood. Here, we introduce Melody, a deep learning framework designed to elucidate the sequence determinants underlying human DNA methylation landscapes across multiple tissues. Trained solely on genomic sequence, Melody accurately predicts cell-type-specific methylation profiles across 39 human tissues, markedly outperforming existing state-of-the-art approaches at the whole-chromosome, hypomethylated, and cell-type-specific levels. Moreover, Melody demonstrates strong generalization in meQTLs variant effect prediction, and Melody reveals key sequence motifs associated with methylation variability. Finally, to extend prediction to unseen cell types, we develop a sequence-based model augmented with scRNA-seq–derived embeddings, enabling accurate inference of cell-type-specific methylation states directly from transcriptomic data. Together, Melody provides a powerful framework for decoding the genomic and transcriptional logic that governs DNA methylation specificity in the human epigenome.