Bioinformatics analysis of thousands of TCGA tumors to determine the involvement of epigenetic regulators in human cancer

Abstract

Background: Many cancer cells show distorted epigenetic landscapes. The Cancer Genome Atlas (TCGA) project profiles thousands of tumors, allowing the discovery of somatic alterations in the epigenetic machinery and the identification of potential cancer drivers among members of epigenetic protein families.

Methods: We integrated mutation, expression, and copy number data from 5943 tumors from 13 cancer types to train a classification model that predicts the likelihood of being an oncogene (OG), tumor suppressor (TSG) or neutral gene (NG). We applied this predictor to epigenetic regulator genes (ERGs), and used differential expression and correlation network analysis to identify dysregulated ERGs along with co-expressed cancer genes. Furthermore, we quantified global proteomic changes by mass spectrometry after EZH2 inhibition.

Results: Mutation-based classifiers uncovered the OG-like profile of DNMT3A and TSG-like profiles for several ERGs. Differential gene expression and correlation network analyses revealed that EZH2 is the most significantly over-expressed ERG in cancer and is co-regulated with a cell cycle network. Proteomic analysis showed that EZH2 inhibition induced down-regulation of cell cycle regulators in lymphoma cells.

Conclusions: Using classical driver genes to train an OG/TSG predictor, we determined the most predictive features at the gene level. Our predictor uncovered one OG and several TSGs among ERGs. Expression analyses elucidated multiple dysregulated ERGs including EZH2 as member of a co-expressed cell cycle network.

Figure 1

Visualizing genomic alterations of bromo domain containing genes. The core of the plot reflects the phylogenetic relationships between bromo domain containing proteins estimated by the sequence similarity of their associated domains. The inner circle displays the expression fold-changes between tumors and healthy tissues. High expression in tumors is indicated in red, while low expression in tumors is shown in blue. The outer circle illustrates the proportion of tumors with 'deep loss' (blue) or 'high-level gain' (red) changes. Mutation rates are reflected by the outer stacked bar plots.

Figure 2

Selecting features for OG and TSG prediction. A) Box plots illustrate feature differences between OGs (red), TSGs (blue) and NGs (gray). Associated p-values on the top of each box plot are based on one-tailed Mann-Whitney U tests and reflect the differences between OGs and NGs, and TSGs and NGs. B) Dots reflect the frequencies of protein altering mutations in the combined set of tumors from seven cancer types. OGs (red), TSGs (blue) and NGs (gray) are sorted alphabetically on the x-axis. C) Proportions of loss of function (LOF) to benign mutations are plotted against the entropy based mutation selection scores for all human genes. Blue indicates high fractions of LOF mutations, while red indicates high mutation selection. D) Stacked bar plots present the relative frequencies of mutation classes in the combined tumor panel for OGs, TSGs and NGs.

Figure 3

Predicted TSGs and most recurrent mutations in the ERG family. A) Predicted TSGs are listed along with the proportions of mutated samples in each indication and overall frequencies of LOF to benign mutations. B) Most recurrent mutations within the ERG family (del: deletion, *: nonsense mutation, fs: frameshift). C) Mutation profiles of ARID1A and PBRM1. Non-synonymous mutations are represented as solid circles, with color distinguishing different cancer types. Circle sizes are proportional to the mutation frequencies.

Figure 4

Epigenetic regulators with significant gene expression profiles in cancer. Significantly (A) over- and (B) under-expressed ERGs are ranked according to the combined p-values (based on Fisher's probability test) over all cancer types. Numbers reflect log2-fold changes, while colors reflect associated p-values. ERGs with consistently over-expression in tumors included EZH2 (p_F = 3.2 × 10⁻¹¹²), ATAD2 (p_F = 1.9 × 10⁻⁷⁶), PRDM13 (p_F = 2.7 × 10⁻²⁷), DPF1 (p_F = 1.0 × 10⁻¹⁹), DNMT1 (p_F = 8.3 × 10⁻¹⁹), SUV420H2 (p_F = 1.7 × 10⁻¹⁵), WHSC1 (p_F = 3.3 × 10⁻¹⁵), TRIM28 (p_F = 1.2 × 10⁻⁸), BAZ1A (p_F = 2.2 × 10⁻⁷), PRMT1 (p_F = 9.6 × 10⁻⁶), and HDAC10 (p_F = 8.1 × 10⁻⁵). ERGs with consistently lower expression in tumors included KAT2B (p_m = 1.0 × 10⁻⁷⁴), EZH1 (p_m = 2.3 × 10⁻⁴²), SMARCA2 (p_m = 2.0 × 10⁻²⁵), NCOA1 (p_m = 1.2 × 10⁻¹⁰), ZMYND11 (p_m = 3.6 × 10⁻⁹), PRDM2 (p_m = 9.5 × 10⁻⁷), BAZ2B (p_m = 3.5 × 10⁻⁶) and SIRT1 (p_m = 8.1 × 10⁻⁶).

Figure 5

Expression plots of significantly expressed ERGs. Gene expressions of (A) EZH2, ATAD2, EZH1, and (B) BRDT, PRDM13 and PRDM9 are shown in RPKM units (black: healthy tissues, red: tumors). Gene expression levels are reflected by RPKM values.

Figure 6

Co-expression network analyses. A) Using Cytoscape co-expressed genes are visualized as networks with nodes representing genes and edges reflecting pairwise co-expression relationships in healthy tissues. B) Numbers of co-regulated cancer genes in healthy tissues (right panel of the plot) are shown along with the mutation frequencies (left panel of the plot). Mutation frequencies are presented as stacked bars with cancer type dependent coloration. ERGs are sorted on the y-axis by the overall mutation frequencies. C) A subnetwork within the main co-expression network contains 24 co-expressed ERGs. Colors indicate the corresponding ERG families. D) EZH2 (green) and 99 co-expressed genes form one co-regulated network that is significantly enriched for cell cycle regulators. Genes that are directly connected with EZH2, because they show a very high degree of co-expression (R > 0.85), are highlighted in orange. Genes that are present in the network, but not directly connected with EZH2 are shown in blue. F) Examples of positive correlations between EZH2 and co-expressed cell cycle regulators. Each dot reflects the gene expression level (represented by variance stabilized RNAseq count data) of EZH2 (x axis) and the co-expressed gene (y axis). Dots are colored according to tissue type.

Figure 7

Quantitative mass spectrometry based proteomic analysis after EZH2 inhibition. A) Protein expression differences of selected cell cycle regulators in EPZ-6438 (Epizyme^®, Cambridge, MA) versus DMSO treated lymphoma cells (WSU-DLCL2) are represented as log2 ratios. B) Global protein expression changes after 8 days of EPZ-6438 treatment. Down-regulated cell cycle regulators are highlighted in dark blue.