Poster #61: Hongxiang Fu PhD

MethScope: Recurrent Pattern Encoding for Sparse DNA Methylomes

Hongxiang Fu1,2,*, Hao Xu1,2,*, Chin Nien Lee3, Cameron Cloud1,2, Yanxiang Deng3,4,5,6,†, Wanding Zhou1,3,† 1Center for Computational and Genomic Medicine, The Children’s Hospital of Philadelphia, PA, 19104, USA, 2Department of Bioengineering, University of Pennsylvania, Philadelphia, PA, 19104, USA, 3Department of Pathology and Laboratory Medicine, University of Pennsylvania, Philadelphia, PA, 19104, USA, 4Epigenetics Institute, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA, 5Institute of Aging, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA, 6Abramson Family Cancer Research Institute, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA.

DNA methylation is a well-established epigenetic marker with important applications in biological research and clinical studies. However, methylome analysis remains challenging because DNA methylation datasets are often highly dimensional and sparse compared with other omics data. Here, we demonstrate that Most Recurrent Methylation Patterns (MRMPs) identify discrete, long-range methylation correlations that are not readily captured by conventional window- or region-based approaches. MRMPs uncover higher-order principles of epigenetic organization while providing a compact and interpretable representation of genome-wide methylation variation. Building on additive combinations of MRMP-associated methylation signals, we developed MethScope, a fast and interpretable machine-learning framework for encoding sparse methylation profiles. MRMP-based encoding supports mini-bulk deconvolution of spatial methylation maps, whole-genome upscaling, rapid cell-type annotation, and cellular embedding with enhanced lineage resolution. Using decoder networks, MethScope accurately reconstructs cell-type-specific methylomes from random measurements covering only ~0.1% of genomic CpG sites, achieving an AUPRC of approximately 0.99, including for cell types and culture conditions absent from the training data. Compared with existing approaches, MethScope maintains strong performance under extreme sparsity, where other methods struggle to reliably perform annotation, deconvolution, or imputation. MethScope is also robust across sequencing and array platforms, providing a scalable and unified framework for characterizing cell identity, identifying cancer cell of origin, and recovering sparse methylation profiles across a broad range of data densities.