A systematic dissection of the epigenomic heterogeneity of lung adenocarcinoma reveals two different subclasses with distinct prognosis and core regulatory networks

Abstract

Background Lung adenocarcinoma (LUAD) is a highly malignant and heterogeneous tumor that involves various oncogenic genetic alterations. Epigenetic processes play important roles in lung cancer development. However, the variation in enhancer and super-enhancer landscapes of LUAD patients remains largely unknown. To provide an in-depth understanding of the epigenomic heterogeneity of LUAD, we investigate the H3K27ac histone modification profiles of tumors and adjacent normal lung tissues from 42 LUAD patients and explore the role of epigenetic alterations in LUAD progression. Results A high intertumoral epigenetic heterogeneity is observed across the LUAD H3K27ac profiles. We quantitatively model the intertumoral variability of H3K27ac levels at proximal gene promoters and distal enhancers and propose a new epigenetic classification of LUAD patients. Our classification defines two LUAD subgroups which are highly related to histological subtypes. Group II patients have significantly worse prognosis than group I, which is further confirmed in the public TCGA-LUAD cohort. Differential RNA-seq analysis between group I and group II groups reveals that those genes upregulated in group II group tend to promote cell proliferation and induce cell de-differentiation. We construct the gene co-expression networks and identify group-specific core regulators. Most of these core regulators are linked with group-specific regulatory elements, such as super-enhancers. We further show that CLU is regulated by 3 group I-specific core regulators and works as a novel tumor suppressor in LUAD. Conclusions Our study systematically characterizes the epigenetic alterations during LUAD progression and provides a new classification model that is helpful for predicting patient prognosis.

HyperChIP: identification of hypervariable signals across ChIP-seq or ATAC-seq samples

“…To evaluate the performance of HyperChIP, we collected three data sets from large-scale cancer studies (Table 1 ). The first data set comprised H3K27ac ChIP-seq profiles, a histone modification marking both active promoters and enhancers, of tumor tissues of 36 lung adenocarcinoma (LUAD) patients [ 6 ]. The second one consisted of ATAC-seq samples of 34 non-small cell lung carcinoma (NSCLC) patients [ 18 ].…”

“…Besides, assay for transposase-accessible chromatin using sequencing (ATAC-seq) has been widely adopted for the detection of open chromatin [ 3 ]. As a common computational task for learning from ChIP/ATAC-seq data, identifying genomic regions with significant changes of ChIP/ATAC-seq signal intensities across samples is essential to understanding the epigenetic mechanisms that orchestrate the variation of gene expression program [ 4 – 6 ]. For this task, there are two major analyses suited to distinct application scenarios.…”

“…The second analysis does not require prior knowledge regarding the labels of samples and aims at identifying hypervariable ChIP/ATAC-seq signals across samples, which can then be used as features for the classification of the samples. This analysis is of particular importance in classifying samples of different cancer patients and gaining insights into the epigenetic markers of different cancer subtypes/stages [ 6 , 9 , 10 ]. Since the second analysis is intrinsically an unsupervised one, it typically requires many more samples to achieve reliable results compared to the first analysis.…”

“…In recent years, with the decrease of sequencing costs, there are more and more large-scale studies in which tens or even hundreds of ChIP/ATAC-seq profiles for different human individuals are generated, and hypervariable analysis is becoming increasingly prevalent in the analysis of ChIP/ATAC-seq data [ 6 , 9 , 10 , 17 , 18 ]. In particular, hypervariable ChIP/ATAC-seq signals across cancer patients could be potential epigenetic markers of different cancer subtypes/stages, and these markers may serve as therapeutic targets and may contribute to the prognosis of patients [ 6 ]. To our best knowledge, however, there are currently no such computational tools that are specifically developed for hypervariable ChIP/ATAC-seq analysis.…”

“…On the other hand, using an ordinary model fitting framework will almost certainly lead to conserved p -values and low statistical power for identifying HVRs, since it is difficult to strictly avoid the influence of true HVRs on the fitting process. A previous study of ours has suggested that p -values derived from an ordinary parameter estimation framework can hardly bear the strength of multiple testing adjustment [ 6 ].…”