Researchers have developed a deep learning framework called Melody that predicts locus-specific DNA methylation from 10-kilobase genomic sequences, according to a paper published in Nature Communications. The model integrates both local and long-range sequence signals to model the epigenetic modification, which regulates transcription, cellular differentiation, and genome stability.
Tests across 39 human tissues showed Melody consistently outperforming existing state-of-the-art computational methods in whole-chromosome, hypomethylated-region, and cell-type-specific benchmarks. Beyond standard profile prediction, Melody generalizes to methylation quantitative trait locus (meQTL) effect prediction. The system also pinpoints regulatory sequence motifs linked to methylation variability across human tissues.
The researchers developed an extension named Melody-G to predict methylation in cell types not present in existing profiles. Melody-G incorporates single-cell RNA-seq foundation model embeddings, allowing it to infer methylation states in previously unprofiled cell types directly from transcriptomic data. The authors presented the tools as a unified framework to link intrinsic genomic sequence and cellular state to the human methylome.
The study was received on January 7, 2026, accepted on August 5, 2026, and published on August 17, 2026. Junru Jin, Ding Wang, and Jianbo Qiao contributed equally as first authors, alongside co-authors Wenjia Gao and Siqi Chen. Shu Wu, Ran Su, and Leyi Wei jointly supervised the project across institutions including Shandong University, the Chinese Academy of Sciences, Tianjin University, Macao Polytechnic University, and the University of Electronic Science and Technology of China. The National Natural Science Foundation of China funded the work under grant numbers 62322112 and 62222311.
