DeepMethylation delivers tissue-specific DNA methylation predictions

Researchers Wenran Li, Shijia Yu, Yingyu Cheng, and Sijia Wang, based at the Shanghai Institute of Nutrition and Health (Chinese Academy of Sciences), published DeepMethylation in PLOS Computational Biology on July 1, 2026. The framework uses a CNN-based module to read local DNA sequence and an MLP-based module for tissue-specific epigenomic annotations (chromatin accessibility, histone marks, transcription factor binding, and genomic position features), then combines both to predict CpG methylation status. Trained and tested on GTEx EPIC array data across nine tissues (blood, lung, breast, kidney, ovary, prostate, testis, colon, and skeletal muscle), the model achieved an average accuracy of 0.847 and AUROC of 0.909, which the authors report as outperforming the prior CNN-based baseline MRCNN (accuracy 0.819, AUROC 0.852) on blood data. The paper also introduces Delta DeepMethylation (DDM), which compares predicted methylation for reference versus alternate allele sequences around a CpG site to estimate a SNP's regulatory effect; the authors report DDM's predicted effects track known methylation QTLs (mQTLs) without the linkage-disequilibrium confound that affects standard mQTL association analysis.

The model was trained on 754,119 CpG sites from the Illumina EPIC array (GTEx v8 cohort) plus 25 engineered epigenomic features, and validated against roughly 24 million CpG sites from whole-genome bisulfite sequencing data. Cross-tissue transfer held up reasonably well (AUROC/AUPRC above 0.83 in all 81 tissue-pair comparisons), though performance dropped on CpG sites not covered by any array (whole-genome AUROC of 0.618 versus 0.712 on EPIC-covered sites), which the authors flag as a harder prediction regime. Code, a demo, and usage instructions are posted on the authors' GitHub repository.

Teams working on epigenetic biomarker discovery, aging clocks, or variant-to-function pipelines get an open, tissue-specific model that can be retrained or fine-tuned on GTEx-style epigenomic annotations without needing new sequencing. The DDM variant-scoring approach is a useful complement to existing sequence-based regulatory variant tools (the paper benchmarks it favorably against DeepSEA2.0/Beluga), particularly where LD is a concern in fine-mapping candidate CpG-modifying variants.

This is a single peer-reviewed paper published today with no independent replication or third-party coverage yet; the reported AUROC and benchmark comparisons come from the authors' own evaluation. Watch for independent reproduction using the public GitHub code, and for whether the model generalizes to tissues and populations beyond the GTEx cohort it was trained on.