Epigenetic modifications in KDM lysine demethylases associate with survival of early-stage NSCLC

Abstract

Background: KDM lysine demethylase family members are related to lung cancer clinical outcomes and are potential biomarkers for chemotherapeutics. However, little is known about epigenetic alterations in KDM genes and their roles in lung cancer survival.

Methods: Tumor tissue samples of 1230 early-stage non-small cell lung cancer (NSCLC) patients were collected from the five independent cohorts. The 393 methylation sites in KDM genes were extracted from epigenome-wide datasets and analyzed by weighted random forest (Ranger) in discovery phase and validation dataset, respectively. The variable importance scores (VIS) for the sites in top 5% of both discovery and validation sets were carried forward for Cox regression to further evaluate the association with patient's overall survival. TCGA transcriptomic data were used to evaluate the correlation with the corresponding DNA methylation.

Results: DNA methylation at sites cg11637544 in KDM2A and cg26662347 in KDM1A were in the top 5% of VIS in both discovery phase and validation for squamous cell carcinomas (SCC), which were also significantly associated with SCC survival (HR_cg11637544 = 1.32, 95%CI, 1.16-1.50, P = 1.1 × 10^-4; HR_cg26662347 = 1.88, 95%CI, 1.37-2.60, P = 3.7 × 10^-3), and correlated with corresponding gene expression (cg11637544 for KDM2A, P = 1.3 × 10^-10; cg26662347 for KDM1A P = 1.5 × 10^-5). In addition, by using flexible criteria for Ranger analysis followed by survival classification tree analysis, we identified four clusters for adenocarcinomas and five clusters for squamous cell carcinomas which showed a considerable difference of clinical outcomes with statistical significance.

Conclusions: These findings highlight the association between somatic DNA methylation in KDM genes and early-stage NSCLC patient survival, which may reveal potential epigenetic therapeutic targets.

Keywords: DNA methylation; KDM; Lung cancer; Lysine demethylase; Survival.

Fig. 1

Analysis work flow. Adenocarcinoma and squamous cell carcinoma samples from Harvard, Spain, Norway, and Sweden cohorts were used for the discovery phase of analysis. Data from The Cancer Genome Atlas (TCGA) were used for validation. Ranger is a weighted version of random forest for controlling for the covariates including age, gender, smoking status, and histological stage. Variable importance score (VIS) was estimated for each CpG site and was ranked in descending order. CpG sites ranked in top 5% in both discovery and validation sets were selected for further evaluation by Cox regression. Multiple testing correction by false discovery rate (FDR) method was used if necessary

Fig. 2

Ranger in discovery (a) or validation (b) in adenocarcinomas. Weighted random forest (Ranger, where confounders like age, gender, smoking status, and stage are adjusted with given 100% weight) was employed in discovery phase (a) and validation set (b) to evaluate the importance of variables. Ranger provides VIS (variable importance score) for each methylation sites. Variables that were in top 5% (red lollipop) or top 10% (yellow lollipop, a flexible criterion) in both discovery phase (a) and validation (b) would be carried forward

Fig. 3

Ranger in discovery (a) or validation (b) in squamous cell carcinomas. Weighted random forest (Ranger, where confounders like age, gender, smoking status, and stage are adjusted with given 100% weight) was employed in discovery phase (a) and validation set (b) to evaluate the importance of variables. Ranger provides VIS (variable importance score) for each methylation sites. Variables that were in top 5% (red lollipop) or top 10% (yellow lollipop, a flexible criterion) in both discovery phase (a) and validation (b) would be carried forward

Fig. 4

Association between DNA methylation at sites cg11637544 in KDM2A and cg26662347 in KDM1A with overall survival of squamous cell carcinomas and correlation between these two sites and their corresponding gene expression. Fixed-effects meta-analysis was used to combine the results from discovery and validation sets for squamous cell carcinomas (SCC) (a cg11637544_KDM2A; b cg26662347_KDM1A). I²and corresponding P value were used to evaluate heterogeneity across studies. DNA methylation level was categorized to six quantiles, and box plot for gene expression was drawn for each quantile. Pearson correlation was used to estimate the correlation coefficient (r) and the P value; gene expression was log2 transformed before analysis (c cg11637544_KDM2A; d cg26662347_KDM1A)

Fig. 5

Survival classification tree for adenocarcinomas. Survival classification tree was built with seven CpG sites as well as covariates using the merged data of discovery and validation sets among adenocarcinoma cases (a), which identified five clusters with significantly different survival curves (b). Cox regression was used to compare the outcomes among clusters (cluster 4 as reference) and represented by hazard ratio (HR), 95% confidence interval (95%CI), and the P value (c)

Fig. 6

Survival classification tree for squamous cell carcinomas. Survival classification tree was built with five CpG sites as well as covariates using the merged data of discovery and validation sets among adenocarcinoma cases (a), which identified four clusters with significantly different survival curves (b). Cox regression was used to compare the outcomes among clusters (cluster 4 as reference) and represented by hazard ratio (HR), 95% confidence interval (95%CI), and the P value (c)