A two-phase epigenome-wide four-way gene-smoking interaction study of overall survival for early-stage non-small cell lung cancer

Abstract

High-order interactions associated with non-small cell lung cancer (NSCLC) survival may elucidate underlying molecular mechanisms and identify potential therapeutic targets. Our previous work has identified a three-way interaction among pack-year of smoking (the number of packs of cigarettes smoked per day multiplied by the number of years the person has smoked) and two DNA methylation probes (cg05293407_TRIM27 and cg00060500_KIAA0226). However, whether a four-way interaction exists remains unclear. Therefore, we adopted a two-phase design to identify the four-way gene-smoking interactions by a hill-climbing strategy on the basis of the previously detected three-way interaction. One CpG probe, cg16658473_SHISA9, was identified with FDR-q ≤ 0.05 in the discovery phase and P ≤ 0.05 in the validation phase. Meanwhile, the four-way interaction improved the discrimination ability for the prognostic prediction model, as indicated by the area under the receiver operating characteristic curve (AUC) for both 3- and 5-year survival. In summary, we identified a four-way interaction associated with NSCLC survival among pack-year of smoking, cg05293407_TRIM27, cg00060500_KIAA0226 and g16658473_SHISA9, providing novel insights into the complex mechanisms underlying NSCLC progression.

Keywords: NSCLC; epigenomics; four‐way interaction; gene–smoking interaction.

Fig. 1

Flow chart of study design and statistical analysis. Patients diagnosed with lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC) from the Harvard, Spain, Norway, and Sweden cohorts were utilized in the discovery phase for screening, whereas data from the Cancer Genome Atlas (TCGA) were employed for validation. Subgroup analyses were conducted based on quartiles of TRIUNE. FDR, false discovery rate; NSCLC, non‐small lung cancer.

Fig. 2

The 3D figure for visualization of four‐way interactions. Results of four‐way interaction among pack‐years of smoking and cg16658473 _SHISA9 , cg05293407 _TRIM27 , and cg00060500 _KIAA0226 . The varying effects of cg16658473 _SHISA9 measured using HR at different smoking and cg05293407 _TRIM27 methylation levels in the subgroup with low and high level of cg00060500 _KIAA0226 , respectively. HR, hazard ratio.

Fig. 3

Comparison of overall survival time of patients among TRIUNE subgroups. (A) Kaplan–Meier survival curves for patients grouped by TRIUNE. Patients were categorized into four subgroups by using the quartiles of TRIUNE as the cutoffs. The number of patients in Group 1 to Group 4 was 248. The gray lines indicate the groups of patients based on the quartiles of the risk scores derived from a clinical model without the addition of the TRIUNE. (B) Hazard ratios and P values were derived from the Cox proportional hazards model for four groups of patients, where Group 1 was set to be reference group. Error bars indicate 95% confidence interval (CI).

Fig. 4

Forest plots of results from stratification analysis of TRIUNE. Hazard ratio (HR) with 95% confidence interval (CI) of TRIUNE on non‐small‐cell lung cancer (NSCLC) survival in different subgroups stratified by clinical characteristics. LUAD, lung adenocarcinoma; LUSC, lung squamous cell carcinoma. HR, 95% CI, and P value was derived from a Cox proportional hazards regression model. Error bars indicate 95% CI.

Fig. 5

ROC curves were generated to compare the predictive performance of different models. Three models were: (a) one using only covariates (clinical variables); (b) one incorporating covariates and all elements of three‐way interactions (cg05293407 _TRIM27 , pack‐year of smoking and another CpG probe) with FDR‐q ≤ 0.05; and (c) another one including covariates and a four‐way interaction (pack‐year of smoking, cg05293407 _TRIM27 , cg00060500 _KIAA0226 and cg16658473 _SHISA9 ) with FDR‐q ≤ 0.05. (A) ROC for 3‐year survival prediction. (B) ROC for 5‐year survival prediction. AUC, area under the receiver operating characteristic curve; ROC, receiver operating characteristic. The dashed line represents the performance of a random classifier, corresponding to an AUC of 0.5. FDR‐q, false discovery rate‐adjusted q value.