Pathway-analysis of published genome-wide association studies of lung cancer: A potential role for the CYP4F3 locus

Abstract

The Fatty acids (FAs) metabolism is suggested to play a pivotal role in the development of lung cancer, we explored that by conducting pathway-based analysis. We performed a meta-analysis of published datasets of six genome wide association studies (GWASs) from the Transdisciplinary Research in Cancer of the Lung (TRICL) consortium, which included 12,160 cases with lung cancer and 16,838 cancer-free controls. A total of 30,722 single-nucleotide polymorphisms (SNPs) from 317 genes relevant to FA metabolic pathways were identified. An additional dataset from the Harvard Lung Cancer Study with 984 cases and 970 healthy controls was also added to the final meta-analysis. In the initial meta-analysis, 26 of 28 SNPs that passed false discovery rate multiple tests were mapped to the CYP4F3 gene. Among the 26 top ranked hits, was a proxy SNP, CYP4F3 rs4646904 (P = 8.65×10⁻⁶, FDR = 0.018), which is suggested to change splicing pattern/efficiency and to be associated with gene expression levels. However, after adding data of rs4646904 from the Harvard GWAS, the significance in combined analysis was reduced to P=3.52×10⁻³ [odds ratio (OR)=1.07, 95% confidence interval (95%CI)=1.03-1.12]. Interestingly, the small Harvard dataset also pointed to the same direction of the association in subgroups of smokers (OR = 1.07) and contributed to a combined OR of 1.13 (95%CI = 1.06-1.20, P=6.70×10⁻⁵). The results suggest that a potentially functional SNP in CYP4F3 (rs4646904) may contribute to the etiology of lung cancer, especially in smokers. Additional mechanistic studies are warranted to unravel the potential biological significance of the finding.

1. Introduction

Lung cancer is the leading cause of cancer deaths in the United States, with an estimated 224,210 new cancer cases and 159,260 deaths in 2014 [1]. Surgery is more effective for the early-stage lung cancer without metastasis, when the tumor can be completely removed. Unfortunately, the disease is typically asymptomatic, until it is in the late stages, and the best chance for surgical treatment is usually lost. Consequently, the mortality is high: the five-year survival rate for lung cancer is only 4.0% for patients with distant tumors, compared with 54.0% for patients with localized tumors [2]. Chest radiograph is a common approach for detecting lung cancer, but the annual screening with chest radiograph has not reduced lung cancer mortality, compared with usual care [3]. Although low-dose computed tomography (CT) screening is a promising method for detecting lung cancer in individuals at high risk [4, 5], the treatment is expensive, and it exposes patients to some extra doses of radiation. Besides, the low-dose CT screening has been associated with a high rate of false positive results, with ~95% of benign lung nodules that are challenging for radiologists to identify [4]. Therefore, there is still a demand for new less-invasive but more efficient biomarkers in the risk assessment and early detection of lung cancer.

Cancer cells often share the attributes of metabolic abnormalities, such as perturbation in the energy metabolism of glucose, glutamine and lipids [6, 7]. Fatty acids (FAs), an important subgroup of lipids, have emerged as a recent focus of cancer research. FAs are synthesized de novo to continually provide lipids for energy production, cell membrane regeneration and lipid modification of proteins to meet the excessive bioenergetics and structural demands of highly proliferating cancer cells. Additionally, FAs and their derivatives are also important signaling molecules that may affect many fundamental cell processes, including cellular survival, proliferation, migration, angiogenesis and therapy resistance [8]. Furthermore, the metabolites of FAs could augment survival [9–11] and enhance adhesion [12] in lung cancer cells, and thus promote progression and metastasis and induce angiogenesis in human lung cancer [13–16].

More recently, increasing evidence suggests that the FAs are potential biomarkers for monitoring development and progression of lung cancer [17–20]. For instance, in serum samples from 55 patients with lung cancer and 165 similar pulmonary patients without known cancer, Liu et al. found that free FAs and their metabolites demonstrated good sensitivity and specificity for the identification of adenocarcinoma of the lung [17]. Similarly, other studies demonstrated that FAs from erythrocyte total lipids might be used as diagnostic biomarkers of lung adenocarcinoma, squamous cell carcinoma, and small cell lung cancer [18, 19]. Most recently, Zhang et al. revealed that serum unsaturated free FAs could be used as potential biomarkers for early detection and disease progression of lung cancer [20]. Consequently, we hypothesized that genetic variants in genes involved in the FA synthesis, degradation, metabolism and transport were potential susceptibility factors for lung cancer.

The aim of the present study was to assess the associations between single-nucleotide polymorphisms (SNPs) in the FAs metabolic pathways and risk of lung cancer. We first conducted a meta-analysis of six lung cancer genome wide association studies (GWASs) within the Transdisciplinary Research in Cancer of the Lung (TRICL) consortium [21–23], and we added additional data from another GWAS from Harvard University [24] for the identified significant SNPs that were corrected by the false discovery rate (FDR) for multiple testing correction and with potential functions in bioinformatics analyses [25, 26].

2. Methods and Materials

2.1 Study populations and genotyping

The detailed information about the study participants is presented in Supplemental Table 1, and all the participants were of European descendent. A written informed consent was obtained from each participant in the GWASs, and the present study followed the study protocols approved by the institutional review board for each of the participating institutions.

2.2 Initial meta-analysis

The present study started from the combined genotyping and imputation dataset of six previously published GWASs of lung cancer from the TRICL consortium. Within the TRICL consortium [21], 12,160 lung cancer cases and 16,838 controls from Europe and North America participated in the GWASs[22]. All the cases and controls were at least frequency-matched on age and sex. The participants were described in previous publications: The University of Texas MD Anderson Cancer Center (MDACC) GWAS [27], the Institute of Cancer Research (ICR) GWAS [28], the National Cancer Institute (NCI) GWAS [21], the International Agency for Research on Cancer (IARC) GWAS [27], Samuel Lunenfeld Research Institute study (SLRI) GWAS [22], Toronto, and The Helmholtz-Gemeinschaft Deutscher Forschungszentren Lung Cancer GWAS, Germany (GLC) [22].

For the TRICL GWASs, the overwhelming majority of patients had non-small cell lung cancer, leaving only a few with small cell lung cancer. Genotyping data were from combined datasets of different platforms of Illumina HumanHap 317, 317+240S, 370Duo, 550, 610 or 1M arrays [21]. These datasets were imputed for all original scans for over 10 million SNPs using the 1000 Genomes Project (phase I integrated release 3, March 2012) as the reference by using IMPUTE2 v2.1.1, MaCH v1.0 or minimac (version 2012.10.3) software. The quality control process was detailed in previously published reports [21, 22].

2.3 Additional data for meta-analysis

An additional independent GWAS dataset of a Caucasian population was provided by the Harvard Lung Cancer Susceptibility Study [24]. For the Harvard GWAS, for which details of participant recruitment have been described previously [24], genotyping data was derived from 1000 cases and 1000 controls using Illumina Humanhap610-Quad arrays. Cases were patients aged >18 years, with newly diagnosed, histologically confirmed primary non-small cell lung cancer. Controls were healthy non-blood-related family members and friends of patients with cancer or with cardiothoracic conditions undergoing surgery.

Unqualified samples were excluded, if they had (i) overall genotype completion rates <95%; (ii) gender discrepancies; (iii) unexpected duplicates or probable relatives (based on pairwise identity by state value, PI_HAT in PLINK>0.185); (iv) heterozygosity rates >6 times the standard deviation from the mean; or (v) individuals evaluated to be of non-Caucasians [using the HapMap release 23 including Japanese in Tokyo, Japan (JPT), Han Chinese in Bejing, China (CHB), Utah Residents (CEPH) with Northern and Western Ancestry (CEU) and Yoruba in Ibadan, Nigeria (YRI) populations as a reference]. Unqualified SNPs were excluded, when they (i) were not mapped on autosomes; (ii) had a call rate <95% in all GWAS samples; (iii) had minor allele frequency (MAF) <0.01; or (iv) had a genotype distribution deviated from those expected by Hardy-Weinberg equilibrium (P < 1.0×10⁻⁶). After applying pre-specified quality control, the Harvard genotype data were available for 984 cases and 970 controls.

2.4 Gene and variant selection

To identify relevant genes of interest, we used two electronic databases, Gene Cards [29] and MSigDB [30] to search for FA metabolism-related genes. Specifically, keyword searches using “fatty acid” in pathway & interaction section and “cancer” in disorder section in the Gene Cards and one heading “fatty acid” in the MSigDB. As a result, 317 genes located on autosomal chromosomes were identified from the FA biosynthesis, metabolism and degration pathways (Supplementary Table 2 and Supplementary Figure 1). Genotyped, imputed common SNPs (minor allele frequency ≥ 0.05) within these genes or their ± 2kb flanking regions were selected for association analysis. Hence, 30,722 SNPs in the FA metabolic pathways had been extracted from the TRICL GWASs and used for further analysis.

2.5 Statistics

Meta-analysis was first performed among the six GWASs from the TRICL consortium. Briefly, the association between each SNP and lung cancer risk was assessed by unconditional logistic regression using an additive genetic model of the risk allele. The Cohran’s Q statistic to test for heterogeneity and the I² statistic to quantify the proportion of the total variation due to heterogeneity were calculated [31]. Fixed-effects models were applied, when there was no heterogeneity among the datasets (based on the criteria P > 0.10 and I² < 25%); otherwise, random-effects models were applied [32]. The Benjamini and Hochberg’s false discovery rate (FDR) method was used for correction of the multiple comparisons[33]. Associations were considered significant, if an FDR value was less than 0.05. For the top-hit variants of interest, we then focused on those SNPs with potential functional as predicted by the online prediction tools: SNPinfo [25] and pfSNP [34]. Linear regression analysis was used to test expression quantitative trait loci (eQTL) associations with data obtained from the RNA sequencing project of the 1000 Genomes Project samples [conducted by Genetic European Variation in Health and Disease Consortium (GEUVADIS) [35]. Gene expression values that were more than three standard deviations from the mean were considered as outliers[36]. Finally, we tested the association between predicted functional SNP and lung susceptibility in the Harvard GWAS [24], and subgroup analyses stratified by histology types (squamous and adenocarcinoma) and smoking status (smoker and non-smoker) were also performed. LocusZoom was used to produce regional association plots [37]. All statistical analyses were carried out by SAS software (version 9.1.3; SAS Institute, Cary, NC, USA) or R (2.6.0) unless specified otherwise. The analysis flow chart is present in Figure 1.

Figure 1.

Analysis flow chart of the present study.

3. Results

The six GWASs from the TRICL consortium consisted of 12,160 cases and 16,838 controls of European ancestry (Supplementary Table 1). The associations between SNPs of genes involved in the FA metabolic pathways and lung cancer risk in the TRICL consortium are shown in Supplementary Figure 2 (the Manhattan plot). Among 30,722 SNPs, 1,609 were nominally associated with lung cancer susceptibility at P < 0.05. Of these, as detailed in Table 1, 28 SNPs in three genes reached the preset statistical thresholds (FDR < 0.05). Among these 28 SNPs, two SNPs in TNF (rs1800628 G>A) and GPX5 (rs116260720 C>A) were located in the previously identified and reported lung cancer susceptibility major histocompatibility complex (MHC) region; therefore, they were not further pursued in further analyses. The other 26 SNPs were mapped within the CYP4F3 gene region on chromosome 19. Supplementary Figure 3 shows the regional association plots of CYP4F3. All the top-hits (FDR < 0.05 and P < 10⁻⁴) in the CYP4F3 region were in high linkage disequilibrium (LD, r² > 0.8, Figure 2) based on the hg19/1000 Genome European populations.

Table 1.

Associations between SNPs in the fatty acid pathways and NSCLC risk with FDR-P<0.05

SNP	Gene	Chr	Position	Allele*	TRICL							Harvard
					EAF	Q†	I2	Effects ‡	OR (95%CI)	P	FDR	EAF	OR (95%CI)	P
rs1800628	TNF	6	31546850	G/A	0.11	0.541	0	++++++	1.18 (1.11–1.24)	5.46×10−9	0.0002	0.09	1.14 (0.99–1.32)	0.071
rs4646521	CYP4F3	19	15769872	G/T	0.35	0.672	0	++++++	1.09 (1.05–1.13)	3.83×10−6	0.018	0.37	0.99 (0.91–1.07)	0.774
rs10405638	CYP4F3	19	15764870	C/T	0.36	0.478	0	++++-+	1.09 (1.05–1.12)	7.59×10−6	0.018	0.38	0.99 (0.92–1.08)	0.888
rs4646904§	CYP4F3	19	15763721	G/A	0.36	0.488	0	++++-+	1.09 (1.05–1.13)	8.65×10−6	0.018	0.37	0.99 (0.92–1.08)	0.884
rs2102959	CYP4F3	19	15767392	A/G	0.36	0.573	0	++++-+	1.08 (1.05–1.12)	1.07×10−5	0.018	0.38	0.99 (0.91–1.07)	0.779
rs10401516	CYP4F3	19	15760577	C/T	0.36	0.552	0	++++-+	1.08 (1.05–1.12)	1.17×10−5	0.018	0.38	0.99 (0.92–1.08)	0.895
rs4646514	CYP4F3	19	15769270	A/C	0.36	0.562	0	++++-+	1.08 (1.04–1.12)	1.27×10−5	0.018	0.38	0.98 (0.90–1.06)	0.633
rs4646515	CYP4F3	19	15769379	G/C	0.36	0.562	0	++++-+	1.08 (1.04–1.12)	1.27×10−5	0.018	0.38	0.98 (0.90–1.06)	0.615
rs718258	CYP4F3	19	15766419	G/A	0.36	0.564	0	++++-+	1.08 (1.04–1.12)	1.28×10−5	0.018	0.38	0.99 (0.91–1.07)	0.807
rs4646520	CYP4F3	19	15769775	A/G	0.36	0.558	0	++++-+	1.08 (1.04–1.12)	1.33×10−5	0.018	0.38	0.98 (0.90–1.06)	0.599
rs10404163	CYP4F3	19	15769805	A/T	0.36	0.558	0	++++-+	1.08 (1.04–1.12)	1.33×10−5	0.018	0.38	0.98 (0.90–1.06)	0.599
rs4646513	CYP4F3	19	15769253	C/A	0.36	0.561	0	++++-+	1.08 (1.04–1.12)	1.34×10−5	0.018	0.38	0.98 (0.90–1.06)	0.639
rs2073600	CYP4F3	19	15757643	C/T	0.36	0.475	0	++++-+	1.08 (1.04–1.12)	1.51×10−5	0.018	0.37	1.00 (0.92–1.08)	0.983
rs2072598	CYP4F3	19	15757528	A/G	0.36	0.472	0	++++-+	1.08 (1.04–1.12)	1.53×10−5	0.018	0.37	1.00 (0.92–1.08)	0.990
rs2280750	CYP4F3	19	15757428	T/A	0.36	0.473	0	++++-+	1.08 (1.04–1.12)	1.54×10−5	0.018	0.37	1.00 (0.92–1.08)	0.992
rs2283611	CYP4F3	19	15759828	A/G	0.36	0.461	0	++++-+	1.08 (1.04–1.12)	1.54×10−5	0.018	0.37	1.00 (0.92–1.08)	0.990
rs2283609	CYP4F3	19	15759764	C/T	0.36	0.462	0	++++-+	1.08 (1.04–1.12)	1.54×10−5	0.018	0.37	1.00 (0.92–1.08)	0.990
rs2077040	CYP4F3	19	15755840	T/C	0.36	0.474	0	++++-+	1.08 (1.04–1.12)	1.55×10−5	0.018	0.37	1.00 (0.92–1.08)	0.990
rs2077080	CYP4F3	19	15755392	G/C	0.36	0.474	0	++++-+	1.08 (1.04–1.12)	1.55×10−5	0.018	0.37	1.00 (0.92–1.08)	0.990
rs4808346	CYP4F3	19	15758931	T/C	0.36	0.462	0	++++-+	1.08 (1.04–1.12)	1.55×10−5	0.018	0.37	1.00 (0.92–1.08)	0.990
rs8107976	CYP4F3	19	15753683	A/G	0.36	0.481	0	++++-+	1.08 (1.04–1.12)	1.55×10−5	0.018	0.37	1.00 (0.92–1.08)	0.988
rs2203998	CYP4F3	19	15754893	A/G	0.36	0.477	0	++++-+	1.08 (1.04–1.12)	1.56×10−5	0.018	0.37	1.00 (0.92–1.08)	0.990
rs8106799	CYP4F3	19	15753759	G/A	0.36	0.48	0	++++-+	1.08 (1.04–1.12)	1.56×10−5	0.018	0.37	1.00 (0.92–1.08)	0.990
rs2204000	CYP4F3	19	15755112	G/A	0.36	0.479	0	++++-+	1.08 (1.04–1.12)	1.56×10−5	0.018	0.37	1.00 (0.92–1.08)	0.986
rs2203999	CYP4F3	19	15754975	C/T	0.36	0.479	0	++++-+	1.08 (1.04–1.12)	1.56×10−5	0.018	0.37	1.00 (0.92–1.08)	0.986
rs4646512	CYP4F3	19	15764220	C/T	0.36	0.454	0	++++-+	1.08 (1.04–1.12)	1.58×10−5	0.018	0.38	1.00 (0.92–1.08)	0.948
rs116260720	GPX5	6	28502794	C/A	0.22	0.335	12.5	+++++-	1.10 (1.05–1.14)	1.65×10−5	0.018	0.21	1.11 (1.00–1.22)	0.043
rs4646516	CYP4F3	19	15769449	A/G	0.36	0.568	0	++++-+	1.08 (1.04–1.12)	1.67×10−5	0.018	0.38	0.98 (0.90–1.06)	0.639

NSCLC: non-small cell lung cancer; EAF, effect allele frequency; FDR, false discovery rate; TRICL, Transdisciplinary Research in Cancer of the Lung;

^*reference allele / effect allele;

^†fixed effect models were used when no heterogeneity was found between studies (Q>0.10 and I² <25.0%); otherwise, random effect models were used;

^‡means positive association, and-means negative association;

^§online function prediction tool, SNPinfo [25], found that rs4646904 might influence splicing efficiency of CYP4F3;

SNPs in bold were putative functional and were not located in the major histocompatibility complex region.

Figure 2.

Gene structure of the CYP4F3 gene and linkage disequilibrium plot of 26 SNPs mapped to CYP4F3 and passed false discovery rate multiple tests.

Using the online tool SNPinfo and pfSNP, the CYP4F3 rs4646904 G>A SNP was putatively functional, because it is predicted to influence exonic splicing efficiency of CYP4F3. We then performed an eQTL analysis to evaluate the mRNA expression levels of CYP4F3 by the genotypes of rs4646904. The GEUVADIS RNA sequencing of the 1000 Genomes Project has only a combined normalized transcriptome and genome sequencing data by performing mRNA and small RNA sequencing on 465 lymphoblastoid cell lines derived from five populations: CEU, Finns (FIN), British (GBR), Toscani (TSI) and Yoruba (YRI). To keep populations consistent and comparable with the present study, we only used data of 373 samples from European descendants (i.e., TSI, GBR, FIN and CEU). Besides, six values for CYP4F3 mRNA expression levels were considered an outlier and removed (Supplementary Figure 4). Gene expression differences among genotypes were examined using a regression model in which the association between gene expression and genotypes was considered additive, assuming that a trend by the number of variant alleles exists. As shown in Figure 3, the number of the variant A allele was shown to be associated with higher expression levels of CYP4F3 (P_additive = 0.001), compared with the common G allele.

Figure 3.

Expression quantitative trait loci (eQTL) association of CYP4F3 rs4646904. The eQTL analyses were preforemed in addtive model. We used RNA-seq data from 368 non-hispanic European individuals, which are part of the 1000 Genome Project.

Therefore, we chose rs4646904 as the tagSNP of the CYP4F3 region. In the meta-analysis of the TRICL consortium, there was no heterogeneity observed among the six GWASs, with I² of 0 and the Q-test P value of 0.488. In combined analysis, we found that the per-unit increase of the variant A allele was associated with 1.09-fold increased risk of lung cancer [95% confidence interval (95% CI) = 1.05-1.13, P = 8.65×10⁻⁶, Table 1].

The genotyping data of the associated rs4646904 SNP was available in the Harvard GWAS (984 patients and 970 controls, minor allele frequency = 0.37, Table 1). When this dataset was added to the TRICL datasets, the summary effect estimate obtained from the expanded meta-analysis was 1.07 (95% CI =1.03-1.12, P = 3.52×10⁻³). For this analysis, we used in a random-effects model, because the effect estimate displayed a moderate degree of heterogeneity with I² = 30.9% and the Q test P = 0.192 (Figure 4). When stratified by lung tumor types, the combined effect was 1.10 (95% CI = 1.04-1.15, P = 8.55×10⁻⁴I² = 1.5%, Q-test P = 0.413) in a fixed-effects model (with a smaller I² value) for adenocarcinomas and 1.07 (95% CI = 0.97-1.19, P = 0.212, I² = 61.0%, Q-test P = 0.017) for squamous carcinomas in a random-effects model (with a larger I² value). We also examined the expected association of rs4646904 with risk of lung cancer in the 10,059 smokers (5078 cases and 4981 controls) from MDACC, IARC, SLRI, GLC and Harvard GWASs. Consistently, a significant effect of the rs4646904 A allele on lung cancer risk was observed in the combined meta-analysis (OR = 1.13, 95%CI= 1.06-1.20, P = 6.7×10⁻⁵I² = 0.0%, Q-test P = 0.706). However, among 2232 non-smokers (381 cases and 1851 controls), the result failed to reach significance (OR = 0.84, 95%CI= 0.71-1.01, P = 0.063, I² = 0.0%, Q-test P = 0.911). Tests for heterogeneity indicated that the carcinogenesis effect of the rs4646904 A allele was predominantly limited to smokers (I² = 89.7%, Q-test P = 0.002).

Figure 4.

Forest plot of A allele effect of CYP4F3 rs4646904 in all cases (Panel A), adenocarcinoma (Panel B), Squamous cell carcinoma (Panel C), smokers (Panel D), non-smokers (Panel E) from GWASs [the Institute of Cancer Research (ICR) GWAS, the MD Anderson Cancer Center (MDACC) GWAS, the International Agency for Research on Cancer (IARC) GWAS, the National Cancer Institute (NCI) GWAS, the Samuel Lunenfeld Research Institute study (SLRI) GWAS, German Lung Cancer Study (GLC) and Harvard lung cancer study (Harvard)].

3. Discussion

Deregulation of the FA metabolic pathways has been implicated in many cancers [38–40]. In the present study, we explored associations between genetic variants in the FA metabolic pathways and risk of lung cancer. We examined all typed and imputed SNPs of the genes involved in the FA metabolic pathways from six published GWASs of lung cancer within the TRICL consortium. We also tried to find additional support from the Harvard GWAS. In the meta-analyses of the TRICL GWASs, the predicted functional SNP, CYP4F3 rs4646904G > A, showed a significant association with lung cancer risk, but this association was not significant in the Harvard Lung Cancer Susceptibility Study; however, this SNP remained significant in the final combined analysis as well as in subgroups of squamous cancer and smokers.

The CYP4F3 gene contains 14 exons and 13 introns and encodes a member of the cytochrome P450 superfamily of enzymes. The CYP4F3 pre-messenger RNA is spliced into two mature transcripts (i.e., CYP4F3B and CYP4F3A). Human CYP4F3s are the main catalysts in the oxidation of FAs: they can ω-hydroxylate a variety of long chains and very long chains as well as saturated, unsaturated and branched chain FAs, vitamins with long alkyl side chains, and the physiologically important prostaglandins and hydroeicosatetraenoic acids[41]. Additionally, the ability of CYP4F3 to ω-hydroxylate both pro- and anti-inflammatory leukotrienes (i.e., LTB4) indicates that they may function in both the activation and resolution phases of the inflammatory response[42]. The CYP4F3 rs4646904G > A variant was a novel, potentially functional SNP that had the smallest P value in the present meta-analysis of the TRICL GWASs. Although CYP4F3 rs4646904G>A is a synonymous variant, it is located within an exon splicing enhancer (ESE)[25], which has discrete sequences within exons that promote both constitutive and regulated splicing[43]. Efficient splicing has limited tolerance of mutations in the ESEs, even if they have no effect on protein coding[44]. Additionally, ChIP-seq data indicate that rs4646904 is located in the enhancer region containing histone modification marks of H3k09me3, H3k27me3 and H3k09me3[45]. Cigarette smoking is the major cause of cancers of the respiratory tract, and CYP4F3 was found to be up-regulated in human airway epithelium in healthy current smokers, compared with those of never smokers [46–48], which suggests that CYP4F3 may be induced by smoking and thus contributes to carcinogenesis. The finding that the rs4646904 A variant allele was associated with significant higher gene expression levels than the G allele is consistent with the findings in other cancers, in which unregulated expression levels of CYP4F3s were found in pancreatic ductal adenocarcinoma, compared to benign lesions[49], as well as related with hepatocyte differentiation status [50] and progression of HCV-associated hepatocellular carcinoma [51].

Despite the underlying biological plausibility supporting this lung cancer-associated rs4646904 in the meta-analysis of the TRICL consortium with 12,160 cases and 16,838 controls, the relatively small Harvard GWAS dataset with 984 cases and 970 controls did not provide a further support for the association, but the combined meta-analysis revealed a slight effect with P < 0.05. Nevertheless, according to a PASS [52] power calculation, the Harvard GWAS had a mere power of 0.18 to detect an OR of 1.1. In contrast, the combined meta-analysis of the TRICL and Harvard GWASs with a total of 13,144 cases and 17,808 controls achieved a power of 0.98 for detecting an OR of 1.1, which suggest that rs4646904 may have a very small but genuine effect on lung cancer risk. Interestingly, the rs4646904-cancer risk association was significant in subgroups with squamous cancer and smokers but not adenocarcinoma and non-smokers, and there were high heterogeneity among the results of these subgroups, indicating that there might be an interaction between smoking and rs4646904 [53]. It may be because the exposure to cigarette smoke may negatively affect the synthesis of n−3 long-chain polyunsaturated fatty acid from the precursor in mammary gland cells [54]. Hence, rs4646904 might be biomarker for lung cancer risk in smokers.

It should be noted that there were limitations of the present study. First, none of the CYP4F3 variants was identified in previous TRICL GWASs as significant at a GWAS threshold (P < 10⁻⁷). Indeed, recent GWASs have identified at least 10 independent loci, and, as a result, only a small fraction of heritability could be explained by these SNPs. The challenge remains to identify the many additional common risk loci that are expected to have smaller genetic effects [55]. The pathway-based analysis, such as in the present study, reduces the significance level of P values based on the number SNPs examined in genes in the studied pathway. Therefore, such a pathway analysis with integration of association results with gene expression should be considered a complementary approach [56, 57] to the GWAS analyses. Second, we are aware that the observed effects may vary by body mass index or sex, but such data were not available from the TRICL consortium for us to further examine the risk modification. Thus, further analyses including body mass index and sex are warranted to improve our understanding of the apparent heterogeneity of effects. Third, because we only included non-Hispanic white populations, the generalizability to other ethnic populations needs further investigation.

In conclusion, this combined meta-analysis of six GWASs from the TRICL consortium and Harvard GWAS identified a potentially predictive functional marker (CYP4F3 rs4646904) for lung cancer risk in Caucasian populations, especially in smokers.

Abbreviations

FA

Fatty acid

GWAS

genome-wide association study

TRICL

Transdisciplinary Research in Cancer of the Lung

CYP4F3

cytochrome P450, family 4, subfamily F, polypeptide 3

SNPs

single nucleotide polymorphisms

FDR

false discovery rate

OR

odds ratio

CI

confident interval

eQTL

expression quantitative trait loci