Epigenomic Tensor Predicts Disease Subtypes and Reveals Constrained Tumor Evolution [RNA-seq]

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 between 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. RNA was isolated from fresh-frozen tissues comprising uterine leiomyomas and matched myometrium from 20 patients. Basecalls were performed using RTA v.2.4.11 on Nextseq instrument. Reads were adapter trimmed using TrimGalore v0.4.4 with parameters --illumina --stringency 13 --paired. After adapter trimming reads were aligned to the hg19 genome with the UCSC knownGene annotation with STAR options --quantmode. The fourth column of the quantmode output was extracted to obtain reads aligning to features in the correct orientation. The principal components are obtained using DESeq2 plotPCA modified to return the first 10 PCs. Prior to PCA, a regularized log transform is applied to the count data (DESeq2 rlog option blind=False).