Epigenomic tensor predicts disease subtypes and reveals constrained tumor evolution

Understanding the epigenomic evolution and specificity of disease subtypes from complex patient data remains a major biomedical problem. We here present DeCET (decomposition and classification of epigenomic tensors), an integrative computational approach for simultaneously analyzing hierarchical heterogeneous data, to identify robust epigenomic differences among tissue types, differentiation states, and disease subtypes. Applying DeCET to our own data from 21 uterine benign tumor (leiomyoma) patients identifies distinct epigenomic features discriminating normal myometrium and leiomyoma subtypes. Leiomyomas possess preponderant alterations in distal enhancers and long-range histone modifications confined to chromatin contact domains that constrain the evolution of pathological epigenomes. Moreover, we demonstrate the power and advantage of DeCET on multiple publicly available epigenomic datasets representing different cancers and cellular states. Epigenomic features extracted by DeCET can thus help improve our understanding of disease states, cellular development, and differentiation, thereby facilitating future therapeutic, diagnostic, and prognostic strategies.