DNA Methylation of LRRC3B: A Biomarker for Survival of Early-Stage Non–Small Cell Lung Cancer Patients

Lung cancer remains the leading cause of cancer-related mortality worldwide, with an estimated 224,390 new cases and 158,080 deaths in the United States alone in 2016 (1). Non–small cell lung cancer (NSCLC) comprises approximately 80% of lung cancers diagnosed in the United States (2, 3). Clinically relevant prognostic information for lung cancer is mainly obtained by staging. However, patients within a stage have differing survival outcomes, suggesting there are other factors affecting prognosis (4). Hundreds of studies have been published on discovering prognostic factors beyond staging, such as molecular prognostic markers, and these have provided some biological and clinical insights (4–6). For early-stage (stage I and II) NSCLC, such markers might help further identify patients with poor prognosis, who can be targeted for aggressive adjuvant treatments to improve survival outcomes.

Epigenetic alterations, such as DNA methylation, are regarded as innovative cancer biomarkers because of their stability, frequency, and reversibility (7–11). In recent years, several epigenetic markers of lung cancer have been identified by in vitro and animal studies (7–11). However, to demonstrate the clinical relevance of those potential markers, large multicenter studies based on clinical data remain necessary (7).

LRRC3B encodes leucine-rich repeat-containing 3B, which is an evolutionarily highly conserved leucine-rich repeat-containing protein (12). Aberrant DNA methylation of LRRC3B has been associated with several types of cancer. LRRC3B is a target of aberrant DNA methylation in gastric cancer and may function as a tumor suppressor (12). Furthermore, LRRC3B methylation intensity is significantly higher in cancer tissues than in the corresponding non-cancerous tissues of colorectal cancer (13). In addition, LRRC3B is methylated and/or deleted with high frequency in major epithelial cancers, including breast, cervical, lung, kidney, ovarian, colon, and prostate cancers (14). However, none of those studies provided strong clinical evidence by showing that DNA methylation of LRRC3B could predict lung cancer survival.

In this study, we performed a two-stage multicenter association analysis between lung tumor DNA methylation in LRRC3B and overall survival (OS) of 613 early-stage (stage I and II) NSCLC patients from an independent Caucasian population, followed by additional independent replication based on another cohort of 617 patients with early-stage lung cancer from The Cancer Genome Atlas (TCGA). We detected three specific DNA methylation probes in LRRC3B as biomarkers of OS and demonstrated clear associations between those methylation probes and LRRC3B expression.

We harmonized datasets from five international study centers. The discovery phase consisted of four cohorts: United States, Spain, Norway, and Sweden. The validation phase was based on a TCGA cohort.

152 newly diagnosed patients with early-stage NSCLC were recruited at Massachusetts General Hospital (MGH) from 1992 to 2005. During curative surgery, tumor specimens were collected with complete resection and snap-frozen. Tumor DNA was extracted from 5-μm-thick histopathologic sections. Amount (tumor cellularity >70%) and quality of tumor cells for each specimen were evaluated by an MGH pathologist. In addition, specimens were histologically classified using Word Health Organization criteria. The study protocol was approved by the institutional review boards at the Harvard School of Public Health and MGH. All patients provided written informed consent.

Descriptions of this study population were previously reported (15). In brief, tumors specimens were collected by surgical resections from 226 patients with early-stage NSCLC recruited during 1991 to 2009. Tumor DNA was extracted from fresh-frozen tumor specimens and further checked for integrity and quantity. The study was approved by the Bellvitge Biomedical Research Institute institutional review board. All patients provided written informed consent.

Descriptions of this study population were previously reported (16). In brief, tumors specimens were collected by surgical resection from 144 patients with early-stage NSCLC recruited at Oslo University Hospital-Rikshospitalet during 2006 to 2011. Tumor tissues obtained during surgery were snap-frozen in liquid nitrogen and stored at −80°C until DNA isolation. This study was approved by the institutional review board and regional ethics committee (S-05307). All patients received oral and written information and signed a written consent.

Descriptions of this study population were previously reported (17). In brief, specimens were collected by surgical resection of 104 tumors of patients with early-stage NSCLC recruited at Skåne University Hospital during 2004 to 2008. The study was approved by the regional ethical review board (registration no. 2004/762 and 2008/702). All patients provided written informed consent.

Data from 617 patients with early-stage NSCLC were included from The Cancer Genome Atlas (TCGA). OS times and covariates were included. HumanMethylation450 DNA methylation image data (IDAT files) were downloaded on October 1, 2015.

We also collected gene expression data of 437 patients with early-stage NSCLC from the Gene Expression Omnibus (GEO) database. Those patients had complete clinical information, including OS time, smoking status, clinical stage, and histology.

DNA methylation was assessed by Illumina Infinium HumanMethylation450 BeadChips (Illumina Inc.). GenomeStudio Methylation Module V1.8 (Illumina Inc.) was used to transform raw image data into beta values (continuous numbers ranging from 0 to 1). Probes meeting the following criteria were removed: (i) detection P > 0.05 in more than 5% of samples; (ii) coefficient of variance (CV) <5%; (iii) potentially contains or extends SNPs (for Infinium I probes, SNPs at the site of single base extension; for Infinium II probes, SNPs at CpG site; for both SNPs, SNPs located within 10 bp from site of single base extension) with MAF > 0.05 in 1000 Genomes Project 20110521 release for European population (18); (iv) cross-reactive probes (18) or cross-hybridizing probes (19); or (v) probes passing QC in only one center. Samples with >5% undetectable probes were excluded. Methylation signals were further processed for quantile normalization (20), type I and II probe design bias correction (21), and batch effects adjustment (22).

In the U.S. cohort, a whole-genome DASL HT assay (Illunina Corp.) was used to measure gene expression values from a subset of patients with NSCLC. Expression of all genes was normalized using dChip software before analysis.

In the Norway cohort, a subset of lung adenocarcinoma samples had both methylation and mRNA expression data. Gene expression was assessed through microarrays from Agilent Technologies (SurePrint G3 Human GE, 8 × 60 K). Gene expression data were log₂-transformed and normalized by Genespring GX analysis software v.12.1 (Agilent Technologies).

In the Sweden cohort, gene expression data were available on 117 tumor samples, measured by Illumina Human HT-12 V4 microarrays, and 97 samples had both methylation and expression data. Gene expression data were quantile-normalized and mean-centered for each probe across all samples. Probe sets without signal intensity above the median of negative control intensity signals in at least 80% of samples were excluded from analysis.

In the TCGA cohort, gene expression was measured by RNA sequencing (RNA-seq). Data preprocessing was done by the TCGA workgroup. Raw counts were normalized by expectation maximization (RSEM). Level-3 (gene level) gene quantification data were downloaded from the TCGA data portal and were further checked for quality. Expression of all genes was extracted and quantile-normalized before analysis.

After data preprocessing, there are 24 DNA methylation probes in LRRC3B, we performed survival analysis by applying univariate Cox proportional hazard models to test the association between each DNA methylation probe and OS time of 613 patients in the discovery cohorts (United States, Spain, Norway, and Sweden). To minimize the influence of potential outliers, DNA methylation level of each probe was dichotomized by median value. The Cox proportional hazard model was:

Where i = index of patients; j = index of probes, and

for the i^th patient.

Three out of 24 DNA methylation probes in LRRC3B were significantly associated with OS after adjusting for multiple testing (P ≤ 10⁻³). Hazard ratio (HR, between high methylation and low methylation) and 95% confidence interval (CI) were estimated for each probe. The same analyses were performed on 617 patients in the validation cohort to evaluate validity of the results. We further performed sensitivity analyses in both discovery and validation cohorts by building a multivariate Cox model adjusting for covariates, including age, gender, smoking status, cancer stage, and histology. The detailed model was:

Where i = index of patients; j = index of probes; and methylation_ij was defined the same as in (1).

The methylation level of each selected probe was dichotomized by median. Then, Kaplan–Meier curves were plotted to compare OS of patients with high methylation levels (above median) with those with low methylation levels (below median) separately in discovery and validation cohorts.

We separated patients into different groups based on methylation levels of three probes. The association between OS and methylation group was tested by univariate as well as multivariate Cox models, as shown below. Kaplan–Meier curves were separately plotted for different groups in the discovery cohort, validation cohort, and total population. Log-rank tests were performed to test the differences between the curves.

Univariate Cox model:

Multivariate Cox model:

Also, among 805 patients who had treatment data, we performed a sensitivity analysis to adjust for confounding effects of adjuvant treatment using the following model:

Where i = index of patients and

for i^th patient.

We performed Cox regressions that include all three selected methylation sites in it. For each person, a risk score was calculated by multiplying methylation levels and the correspondent Cox regression coefficients. People with high-risk scores (above median of discovery cohort) and low-risk scores were assigned to two groups. Kaplan–Meier curves were separately plotted for those groups in the discovery cohort, validation cohort, and total population. Log-rank tests were performed to test the differences between the curves.

In the discovery cohort, 216 patients with early-stage NSCLC had complete gene expression data and DNA methylation data. Gene expression data in different centers were measured by different platforms. Data were combined after being log₂-transformed and standardized (centralized by mean and scaled by standard deviance) in each center. Linear regression showed that there was no significant residual batch effect by center. Associations between expression of LRRC3B and DNA methylation level of the three selected probes were tested by linear regression.

In the validation cohort, 427 patients with early-stage NSCLC had complete mRNA sequencing data and clinical information. Of those, 344 also had complete DNA methylation data. Expression of LRRC3B was log₂-transformed and standardized before analysis. Associations between LRRC3B expression and DNA methylation level of the three selected probes were tested by linear regression.

In summary, 216 patients in the discovery cohort and 427 patients in the validation cohort had LRRC3B expression data and clinical information, including OS, smoking status, clinical stage, and histology. We also collected gene expression data for 437 patients with early-stage NSCLC from the GEO database. Those patients had complete clinical information, including OS, smoking status, clinical stage, and histology. We dichotomized LRRC3B expression by median in each cohort. Associations between expression and OS were tested using a multivariate Cox model, adjusted for age, gender, smoking status, clinical stage, and histology.

DNA methylation data of 69 pairs of tumor and adjacent non-tumor tissues and gene expression data of 53 pairs of tumor and adjacent non-tumor tissues were collected from TCGA. Student t tests were used to test the difference in methylation levels in 24 probes of LRRC3B between tumor and adjacent non-tumor tissues. Gene expression data were log₂-transformed and student t tests were then used to test the difference in LRRC3B expression between tumor and adjacent non-tumor tissues.

All statistical analyses were performed using R version 3.3.0 (The R Foundation).

The study design is shown in Fig. 1, and Supplementary Table S1 lists demographic and clinical pathology data for all populations. We first evaluated the association between each of 24 DNA methylation probes in LRRC3B and OS of 613 patients in the discovery cohort using a univariate Cox model. Three DNA methylation probes (cg13046257, cg17623116, and cg19600115) located at the 5′ untranslated region (5′ UTR) were significantly associated with OS after adjusting for multiple testing (P ≤ 10⁻³; Supplementary Table S2). We then tested the validity of associations in 617 patients in the validation cohort; associations remained significant with consistent effect size (Supplementary Table S3).

Methylation levels for these three probes were consistently distributed in both discovery and validation cohort samples (Fig. 2A). In addition, associations remained significant after adjusting for covariates, including age, gender, smoking status, cancer stage, and histology in the discovery cohort (cg13046257: HR, 0.76; 95% CI, 0.60–0.95; cg17623116: HR, 0.73; 95% CI, 0.59–0.91; cg19600115: HR, 0.72; 95% CI, 0.57–0.90), as well as in the validation cohort (cg13046257: HR, 0.53; 95% CI, 0.38–0.76; cg17623116: HR, 0.61; 95% CI, 0.43–0.85; cg19600115: HR, 0.54; 95% CI, 0.38–0.76; Fig. 2B and Supplementary Table S4).

For each of three selected probes, we plotted Kaplan–Meier curves for patients with high methylation levels (above median) at that probe against those with low methylation levels in discovery and validation cohorts respectively. Compared with those with low methylation levels, patients with high methylation levels had significantly longer OS in both discovery (cg13046257: HR, 0.66; 95% CI, 0.53–0.82; cg17623116: HR, 0.68; 95% CI, 0.55–0.85; cg19600115: HR, 0.66; 95% CI, 0.53–0.82) and validation cohorts (cg13046257: HR, 0.51; 95% CI, 0.36–0.72; cg17623116: HR, 0.62; 95% CI, 0.45–0.87; cg19600115: HR, 0.57; 95% CI, 0.41–0.80) for all three probes (Fig. 3A–C).

By considering methylation profiles of all three probes together, we separated patients into different groups. Patients with at least two hypermethylated sites had significantly longer OS than the others in the discovery cohort (HR, 0.62; 95% CI, 0.50–0.77; P = 2.02 × 10⁻⁰⁵), validation cohort (HR, 0.55; 95% CI, 0.39–0.77; P = 4.44 × 10⁻⁰⁴), and total population (HR, 0.61; 95% CI, 0.52–0.74; P = 2.80 × 10⁻⁰⁷; Fig. 4A–C and Supplementary Table S5). Patients with three hypermethylated sites also had significantly longer OS than the others in the discovery cohort (HR, 0.57; 95% CI, 0.43–0.75; P = 5.80 × 10⁻⁰⁵), validation cohort (HR, 0.45; 95% CI, 0.29–0.68; P = 1.74 × 10⁻⁰⁴), and total population (HR, 0.55; 95% CI, 0.43–0.69; P = 2.30 × 10⁻⁰⁷; Fig. 4A–C and Supplementary Table S5). After adjusting for age, gender, smoking status, clinical stage, and histology, the association remained statistically significant (Supplementary Table S5). Among 805 patients who had treatment information, a sensitivity analysis to adjust for confounding effects of adjuvant treatment (chemotherapy and/or radiation) showed that the association remained consistent and significant (Supplementary Table S20). For each person, we also built a risk score based on methylation levels (Materials and Methods). People with high-risk scores (above the median of discovery cohort) have significant shorter survival time than those with low risk scores, in both discovery and validation cohort. Notably, the risk scores of people in validation cohort were calculated based on Cox regression coefficients estimated in discovery cohort (Fig. 4A–C).

DNA methylation plays an important role in regulating gene expression (23). We tested the association between DNA methylation at the three probes and LRRC3B expression, as well as the association between gene expression and OS.

Linear regression showed that all three DNA methylation probes (cg13046257, cg17623116, and cg19600115) were positively associated with LRRC3B expression in both discovery and validation cohorts (Fig. 5A–C and Supplementary Table S6). In each cohort, regression coefficients for the three probes were similar, indicating potential coherent regulating effects of DNA methylation in that region. Note that expression profiles were measured by different platforms in the discovery and validation cohorts, which resulted in different regression coefficients across the two cohorts. The P value of probe cg13046257 in the discovery cohort was borderline, which was likely due to the small sample size. Furthermore, expression data from different centers in the discovery cohort were measured by different platforms, which likely induced variation in expression profiles.

We further tested the association between LRRC3B expression and OS among 216 patients in the discovery cohort, 427 patients in the validation cohort, and 437 patients in the GEO dataset. After adjusting for age, gender, smoking status, clinical stage, and histology, patients with high expression levels had longer OS than those with low expression levels (discovery: HR, 0.85; 95% CI, 0.57–1.27; validation: HR, 0.63; 95% CI, 0.41–0.98; GEO: HR, 0.59; 95% CI, 0.40–0.87; Supplementary Table S7). The association in the discovery cohort was not significant, which was likely due to small sample size and platform variations of expression profiles. The associations in the two larger cohorts (validation and GEO) were significant, and the effect sizes were similar.

We also analyzed differences in LRRC3B methylation and expression levels between 53 pairs of tumor and adjacent non-tumor tissues. Probes cg13046257 and cg17623116 were hypomethylated in tumor compared with adjacent non-tumor tissues (cg13046257: fold change = 0.92, FDR = 7.06 × 10⁻⁰⁸; cg17623116: fold change = 0.81, FDR = 1.14 × 10⁻¹⁴; Supplementary Table S8). In addition, LRRC3B had significantly lower expression in tumor tissues (fold change = 0.26; P = 1.11 × 10⁻⁰⁴; Supplementary Table S9).

We investigated putative implications of DNA methylation biomarkers in the tumor-suppressor gene LRRC3B on 1,230 patients with early-stage NSCLC. Patients could be distinguished into different risk groups based on their methylation levels at three loci in LRRC3B. The validity of these methylation biomarkers was demonstrated by: (i) associations between DNA methylation and OS were replicated in two large populations; (ii) coherence with results of potential downstream gene expression pathways in multiple independent large populations; and (iii) consistency with previously published studies across different types of cancer.

LRRC3B encodes leucine-rich repeat-containing 3B, which is an evolutionarily highly conserved leucine-rich repeat-containing protein (12). It covers about 88 kb at the human chromosome 3p24.1 locus, with two predicted CpG islands in the promoter region (chr3:26664104-26664796 and chr3:26665950-26666164). Previous studies showed LRRC3B is a potential tumor suppressor in gastric and colorectal cancers (12, 13). One study showed LRRC3B expression is reduced in 90.9% of gastric cancer cell lines and 88.5% of gastric tumor tissues (12), which is consistent with our finding that LRRC3B expression is repressed in 75.47% (40/53) of lung cancer tumor tissues.

LRRC3B hypermethylation was reported at six CpG sites around the first CpG island (chr3:26664104-26664796) in the promoter region in 78 gastric tumor tissues, comparing with non-tumor tissues (11), which is consistent with our results (Supplementary Table S8). However, we found that LRRC3B was hypomethylated in a non-CpG island region near the gene body (chr3: 26700158–26751141), which covered two of the three probes we identified (Supplementary Table S8). Methylation levels of three sites found in our study all show strong positive relationships with LRRC3B expression, consistent with previous results demonstrating that there is no inverse relationship between methylation of non-CGIs and gene expression (23, 24).

Solid evidence from functional and in vitro experiments supports a tumor-suppressor role for LRRC3B across various types of cancers (12–14). Studies have also demonstrated the potential of DNA methylation or expression biomarkers as prognostic biomarkers for some cancer types (12–14). However, most studies focused on in vitro experiments or small-scale population data. Our study was based on clinical data from large populations, thus building a bridge between previous experimental studies and clinical applications. Considering the convenience, feasibility, and economy of future applications of the biomarkers, we selected only three methylation probes in a restrictive way and tested robustness of their associations with OS in two independent populations. To further validate reliability of the biomarkers, we tested (i) the association between DNA methylation of the three probes and LRRC3B expression; and (ii) the association between gene expression and OS in at least two independent populations. The results of those analyses are robust and coherent with each other. We also performed sensitivity analysis to provide more information, which can be found in Supplementary Information (Supplementary Table S10–S20 and Supplementary Fig. S1A–S1C). However, we recognize some limitations of this study. Our study is based on observational data, and DNA methylation and gene expression were measured once at the same time point. The casual mechanism between DNA methylation and gene expression cannot be evaluated in our study.

In summary, we provide evidence of the potential development of a biomarker panel based on DNA methylation of LRRC3B for the OS of early-stage NSCLC, thus filling the gap between in vitro studies and clinical application. Future studies may focus on the additional implications of LRRC3B epigenetics, including early-diagnosis biomarkers and drug target discovery.

No potential conflicts of interest were disclosed.

The study was approved by institutional review boards at Harvard School of Public Health, Massachusetts General Hospital, Bellvitge Biomedical Research Institute, Oslo University Hospital, Lund University, and Skåne University Hospital.

Conception and design: Y. Guo, R. Zhang, L. Su, D.C. Christiani

Development of methodology: Y. Guo, L. Su, Å. Helland, M. Esteller, D.C. Christiani

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): Y. Guo, R. Zhang, S.M. Salama, M.M. Bjaanæs, M. Planck, L. Su, J. Staaf, Å. Helland, M. Esteller, D.C. Christiani

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): Y. Guo, R. Zhang, S. Shen, Y. Wei, L. Su, Z. Zhu, Å. Helland, M. Esteller, D.C. Christiani

Writing, review, and/or revision of the manuscript: Y. Guo, R. Zhang, Y. Wei, T. Fleischer, L. Su, Å. Helland, D.C. Christiani

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): Y. Guo, A. Karlsson, L. Su, D.C. Christiani

Study supervision: Y. Guo, R. Zhang, L. Su, D.C. Christiani

We thank all participants in this study. This work was supported by the National Institutes of Health (National Cancer Institute) grants (CA209414, CA092824, CA090578, CA074386, and ES00002; to D.C. Christiani) and the Raymond P. Lavietes Family Fund (to D.C. Christiani).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.