Genetically determined elevated C-reactive protein associated with primary colorectal cancer risk: Mendelian randomization with lifestyle interactions

Abstract

Systemic inflammation-related etiologic pathways via inflammatory cytokines in the development of colorectal cancer (CRC) have not been convincingly determined and may be confounded by lifestyle factors or reverse causality. We investigated the genetically predicted C-reactive protein (CRP) phenotype in the potential causal pathway of primary CRC risk in postmenopausal women in a Mendelian randomization (MR) framework. We employed individual-level data of the Women’s Health Initiative Database for Genotypes and Phenotypes Study, which consists of 5 genome-wide association (GWA) studies, including 10,142 women, 737 of whom developed primary CRC. We examined 61 GWA single-nucleotide polymorphisms (SNPs) associated with CRP by using weighted/penalized MR weighted-medians and MR gene-environment interactions that allow some relaxation of the strict variable requirements and attenuate the heterogeneous estimates of outlying SNPs. In lifestyle-stratification analyses, genetically determined CRP exhibited its effects on the decreased CRC risk in non-viscerally obese and high-fat diet subgroups. In contrast, genetically driven CRP was associated with an increased risk for CRC in women who smoked ≥ 15 cigarettes/day, with significant interaction of the gene-smoking relationship. Further, a substantially increased risk of CRC induced by CRP was observed in relatively short-term users (< 5 years) of estrogen (E)-only and also longer-term users (5 to > 10 years) of E plus progestin. Our findings may provide novel evidence on immune-related etiologic pathways connected to CRC risk and suggest the possible use of CRP as a CRC-predictive biomarker in women with particular behaviors and CRP marker-informed interventions to reduce CRC risk.

Introduction

Colorectal cancer (CRC) in postmenopausal women ages 50 years and older accounts for the majority (approximately 90%) of newly diagnosed female CRC cases and related deaths [1], thus mainly contributing to CRC’s third ranking in cancer incidence and mortality in women of the United States and Westernized countries [2,3]. Of the established risk factors, the inflammatory process plays a critical role in promoting CRC development [4]. In particular, inflammation-induced colorectal carcinogenesis can be driven by local inflammation of the colorectal mucosa owing to inflammatory bowel disease [5-7]. Further, as supported by evidence that reduction of systemic inflammation has the tissue-specific effect of reduced inflammation in the colorectal mucosa [8], systemic inflammation is considered one of the driving mechanisms behind CRC development [9,10]. However, whether systemic inflammation has a similar carcinogenic effect via inflammatory cytokines produced in an innate immunity response to the inflammation is less certain. In particular, C-reactive protein (CRP), a major acute-phase reactant synthesized in response to acute and chronic low-grade inflammation, is postulated as a key pro-inflammatory biomarker that has been implicated in colorectal carcinogenesis [11-15]. The molecular carcinogenic mechanism by which an elevated circulating level of CRP influences the risk for CRC is only partially understood. For example, CRP acts as a lectin-like oxidized low-density lipoprotein receptor-I (LOX-I) ligand, elevating the mRNA expression of the LOX-I gene that is important to CRC growth in CRC cell lines [12]. In addition, CRP directly interacts with higher levels of lipopolysaccharide (LPS) immunoglobulin and LPS-binding protein (LBP), biomarkers of gut hyperpermeability, inducing increased systemic exposure to intestinal bacterial products [16,17]; CRP then alters the gut microbiome, gradually forming a microenvironment that is an important step in colorectal carcinogenesis [13-15].

Higher levels of CRP due to local inflammation have been associated with a high risk of CRC [5,18]. However, findings from previous observational and genetic epidemiologic studies of elevated circulating CRP concentrations not directly related to local inflammation, and CRC risk are mixed: some positive associations with CRC [19-21] and advanced colorectal adenoma [22-24], null results [10,25,26], and inverse relationships [27-29]. Of note, few studies [10,21] of the association between the plasma level of CRP and CRC risk found significant relationships among only subjects with a short period of follow-up (< 2 and 5 years), suggesting the potential existence of reverse causality between CRP and CRC risk. Overall, the conflict between those findings may be partly explained by selection bias, confounding and/or interaction effects of lifestyle factors, relatively short-time exposure to the biomarker, and reverse causation; it thus calls for in-depth research, for example, Mendelian randomization (MR) study, which can address those challenges and potentially provide improved causal inference of the immune-related etiologic pathways connected to CRC risk.

MR examines the exposure (e.g., CRP) on an outcome (e.g., CRC risk) using genetic variants as an “instrumental” variable [30]. This approach may establish a relatively unbiased causal inference between CRP and CRC risk by first reducing potential biases and confounding given random assignment of exposure owing to randomly assorted genetic alleles at the time of gamete formation; the overall observed allelic effects are thus less likely to be influenced by environmental factors. Second, MR can address short-term exposures to inflammatory biomarkers by examining the associated alleles as a proxy for lifelong exposure [30,31]. Third, MR may also reduce the potential of reverse causation because alleles precede the phenotype and relevant clinical outcomes [31]; thus, in an MR framework, the elevated level of CRP is less likely to result from the immune response caused by premalignant or preclinical tumor growth. Finally, an MR study can be comparable to randomized clinical trials in inferring a causal relationship when the genetic instruments are fulfilled the conditions required for genomic validity in MR analysis [30]. To date, few MR studies have evaluated genetically determined CRP in relation to CRC risk, with inconsistent reports: 1 null result [32] and 2 positive relationships [21,33].

In the study detailed here, we conducted MR analysis with a focus on postmenopausal women, a vulnerable population with a high incidence of inflammation [34] and CRC. We incorporated a large number of genome-wide single-nucleotide polymorphisms (SNPs) obtained from our earlier and other genome-wide association studies (GWASs) [35-38]. As in previous MR studies, we used a traditional MR estimate from the inverse-variance weighed (IVW) method. Further, we performed more recently developed analyses, including weighted median (WM) and penalized weighted median (PWM) estimates [39,40], and also estimated corrected MR that accounts for gene-environment (G×E) interactions [41]. These approaches allow some relaxation of the strict rules for MR instrumental variables by down-weighting the contribution of heterogeneity of genetic variants to analysis and by incorporating interactions between genes and lifestyles into our MR analysis. We ultimately tested the hypothesis that genetically predicted CRP that interacts with lifestyle factors is causally associated with CRC risk.

Materials and methods

Study population

The data of our study population were obtained from the Women’s Health Initiative (WHI) Database for Genotypes and Phenotypes (dbGaP), which encompasses the 2 WHI study arms, Clinical Trials and Observation Studies, representing one of the largest studies to date on postmenopausal women in the U.S. In particular, we used the biggest substudies of the WHI dbGaP, the Harmonized and Imputed GWASs, which contribute to a joint genomic effort across the 5 WHI GWASs (Table S1; AS264, GARNET, GECCO, HIPFX, and WHIMS). Details on the study designs and rationale are available elsewhere [42]. Briefly, the WHI is a long-term prospective cohort study focusing on strategies for preventing diseases among postmenopausal women. Healthy women were enrolled in the WHI study between 1993 and 1998 at more than 40 clinical centers across the U.S. if they were 50-79 years old, postmenopausal, expected to stay near the clinical centers for at least 3 years after enrollment, and able to provide written informed consent. Participants were eligible for the WHI dbGaP study if they had met the eligibility requirements for data submission to dbGaP and had provided DNA samples. Of 16,088 women in total who reported their race/ethnicity as non-Hispanic white, we applied multilevel exclusion criteria: in the GWAS for CRP, genomic data quality control (QC), resulting in 10,798 women (Table S1); and in the associations between GWAS CRP-SNPs and CRC risk, < 1 year follow-up for cancer outcomes and a diagnosis of any cancer type at screening, finally leading to a total of 10,142 women. They had received follow-ups through August 29, 2014, with a mean follow-up of 16 years, and 737 (7% of the eligible 10,142 women) had developed primary CRC. The studies were approved by the institutional review boards of each WHI clinical center and the University of California, Los Angeles.

Lifestyle variables and CRC outcome

Participants had completed self-administered questionnaires at screening, providing demographic and lifestyle information, and data collection was periodically monitored by the WHI coordinating clinical centers for data QC. With 41 variables initially selected from literature review [11,43-61] for their association with inflammation and CRC, we conducted preliminary analyses such as univariate and multivariate/stepwise regressions and a multicollinearity test, resulting in a final selection of the following 16 variables for our analysis: demographic and socioeconomic factors (age, education, and family income); a family history of CRC; comorbidity (cardiovascular disease ever and high cholesterol requiring medication ever); behaviors (exercise and cigarette smoking); dietary factors (daily alcohol and saturated fatty acids [SFA] intake); and reproductive history (age at menopause, durations of past oral contraceptive [OC] use and of exogenous estrogen [E]-only and E plus progestin [E+P] use). Trained staff had measured anthropometric characteristics, including height, weight, and waist and hip circumferences.

For our lifestyle-stratified analyses for the G×E tests, the following variables as potential effect modifiers on the basis of previous studies [12,22-24,29,57,62-70] and relevant cutoff values were used: 30 kg/m² body mass index (BMI), 88 cm waist circumference, 0.85 waist-to-hip ratio (WHR), 10 hours/week metabolic equivalents (METs), and 9% calories/day from SFA on the basis of obesity and relevant lifestyle guidelines [71-73]; at least 15 cigarettes/day; 14 g alcohol intake/day (1 drink standard for women) [74]; a 5-year interval of E-only/E+P use ranging from never used to used 15 years or longer; and a 7-year median of OC use.

CRC outcomes were determined by a committee of physicians’ review of medical records, pathology and cytology reports, and tumor registry abstracts, and finally coded into the central WHI database according to the National Cancer Institute’s Surveillance, Epidemiology, and End-Results guidelines [75]. The time from enrollment to CRC development, censoring, or study end point was calculated and presented in years for the purpose of analysis.

Genotyping and instrumental variables

Genotyped data were derived from the WHI GWASs database in dbGaP, which had been normalized to the Genome Reference Consortium Human Build 37 and imputed using the 1000 Genomes reference panel [76]. Details of the data-cleaning process have been discussed previously [35,42,77]. In brief, DNA was obtained from blood samples at baseline and genotyped using several GWAS platforms [42]. The SNPs’ harmonization was checked via pairwise concordance among all samples across the GWASs. Through the first and second data QC steps, SNPs were filtered on a missing call rate of < 2%, a Hardy-Weinberg equilibrium of P ≥ 1E-04, and Ȓ ² ≥ 0.6 imputation quality [78]. Individuals with unexpected duplicates, first- and second-degree relatives, and outliers on the basis of genetic principal components (PCs) were excluded.

We selected GWA CRP-SNPs from 4 GWASs. Our earlier GWAS [35] used the WHI Harmonized and Imputed GWASs data for the association with CRP as a binary outcome (> 3.0 mg/L, reflecting chronic low-grade inflammation status [79,80]). We used results from this earlier GWAS; of 82 SNPs, 5 index/independent SNPs not in linkage disequilibrium (i.e., LD < 0.3) were selected (Table S1). Analyzing the same population for genes/SNPs in association with CRP as well as CRC risk can reduce bias derived from different population structures between phenotype and outcomes, one of the biases which have frequently been issued in MR analysis. In addition, we included the 3 outside GWASs [36-38] that recently reported GWA CRP-SNPs as a naturally log-transformed continuous variable (mg/L). They used different genotype and analytic strategies: HapMap-based 1000 Genomes imputed data analysis [36], exome-wide common and low/rare coding variants search [37], and genome-wide analysis of discovery panel combined with replication panel [38]. Of 89 SNPs obtained from those 3 studies, 16 SNPs overlapped, so findings from the most recent study were selected. In total, 61 independent SNPs (5 from our study plus 56 from others; LD < 0.3) were finally included in our analysis (Table S1). For all SNPs, the allele associated with a higher CRP level was assigned as an alternative (risk) allele, with the other as a reference allele.

Statistical analysis

The effect of each SNP on CRC risk was calculated in our study population via Cox regression with the adjustment in the first stage by age and 10 genetic PCs and in the second stage by lifestyle covariates in addition to age and 10 genetic PCs. The proportional hazard assumption was checked via a Schoenfeld residual plot and ρ evaluation. The Cox results from each of the 5 GWASs were combined via fixed-effect meta-analysis, by which the potential winner’s bias can be reduced [81].

We first checked MR assumptions to see whether our data met the conditions required for a valid causal inference. Three basic assumptions are essential for our genetic instruments to be valid for MR analysis: the genetic variants are (i) solidly predictive of the phenotype, (ii) independent of confounding factors that connect the phenotype-outcomes, and (iii) independent of the outcomes, conditioning on the phenotype and the confounding factors linked to phenotype-outcome relations (i.e., no pleiotropic pathways other than the phenotype) [82]. The first assumption was initially addressed with the selection of CRP SNPs at genome-wide significance. The inter-individual variability explained by the selected SNPs was approximately 6% [35,37,38], and the F-statistic was 107.80 on the basis of our sample size and number of instrumental variables [83]; with the traditional threshold of the 10 F-statistic [84], the strength of our genetic instruments are considered sufficient. Unlike the first assumption, the second and third may not be fully empirically tested because they rely on all the confounders, both measured and unmeasured [40]. For the horizontal pleiotropic effect on the association between CRP and CRC risk, we excluded pleiotropic SNPs (Table S2) that were associated with obesity (BMI and WHR), diabetic syndromes and diabetes (fasting glucose and insulin, post 2-hour glucose, and type 2 diabetes [T2DM]), and dyslipidemia (low-/high-density lipoproteins, total cholesterol, and triglycerides) [30,85]. While our 5 GWA CRP-SNPs did not overlap, 4 SNPs (with BMI, post 2-hour glucose, T2DM, and dyslipidemia) of the other 56 GWA CRP-SNPs were excluded. Also, we accounted for the covariates (listed above in the Lifestyle variables subsection) in the analyses of SNPs and CRC risk to remove the potential confounding effect. Further, we performed an MR-Egger analysis for assessing vertical directional pleiotropy (i.e., the third assumption) and checked whether the pleiotropic effects were skewed in one direction rather than being balanced [86].

Having demonstrated that our genetic instruments met the basic MR conditions, we conducted the MR analysis separately by analytic type of CRP phenotype - binary or continuous. In addition to the IVW method [87], we employed recent MR methods, including WM/PWM estimates [39,40] and MR G×E interactions [41]. Of note, the WM/PWM estimates allow up to 50% of genetic variants’ invalidity. Specifically, the WM method assigns a weight to the ordered estimate and establishes linearity between neighboring estimates, yielding a more stable estimate than the IVW does if the precision of the individual estimates varies substantially [40]. When the estimates from invalid instruments are not balanced about the true effect, use of the WM is inappropriate, but the PWM method can minimize this issue by down-weighting outlying genetic variants with heterogeneous estimates [40]. The PWM can also give better results if there is directional pleiotropy. In each of the separate MR analyses (i.e., CRP as binary or continuous), we conducted hypothesis-driven lifestyle-stratified analyses, and using the MR G×E method [41], calculated a corrected MR estimate that takes into consideration the interactions between CRP SNPs and selected lifestyles (BMI, waist, WHR, MET, % calories from SFA, family history of CRC, smoking, alcohol, and hormone therapy, such as OC, E-only, and E+P use). For this G×E analysis, we created a weighted genetic score (GS) using a polygenic additive model [88] with our 5 and the 56 outside GWA CRP-SNPs. Next, we rescaled the GS to the unit of CRP by conducting a linear regression; using β0 (slope) and β1 (intercept), we computed the scaled CRP-GS (= β0 + β1×GS), where 2 GSs were perfectly correlated (r = 1.0) [39,88].

In the MR analysis of our 5 GWA SNPs, we further adjusted for a correlation between CRP phenotype and CRC in which exposure and outcome were evaluated within the same population. For parameters necessary for the MR analysis, the change in CRP (> 3.0 vs. ≤ 3.0 mg/L) in log-odds and the mean change in log-transformed CRP per allele were obtained from our WHI GWAS and the 3 other earlier ones, respectively. The final MR results were reported as risk ratios (i.e., hazard ratios [HRs] and 95% confidence intervals [CIs]) for the change in CRC risk per unit increase in log-odds or log-transformed CRP.

The heterogeneity among MR estimates, reflecting additional pleiotropic evidence, was estimated via Cochran’s Q test. A 2-tailed P < 0.05 was statistically significant. R3.6.3 with survival, metafor, forestplot, ggplot2, and ggthemes packages was used.

Results

The associations between each of the 61 GWA CRP-SNPs (5 for binary, reflecting chronic inflammation, and 56 for the naturally log-transformed continuous CRP variable) and CRC risk from the meta-analyses of the 5 WHI GWASs are shown in Table 1; the analytic results from the 2-stage adjustment are presented, including the first stage of adjustment for age and 10 genetic PCs and the second stage of adjustment for lifestyle covariates in addition to age and 10 genetic PCs. The pooled analysis for the genetic instruments combining 5 SNPs (Figure 1A) and 56 SNPs (Figure 1B) each did not yield statistically significant associations (the 5 SNPs’ PWM-HR_2nd-stage = 1.20, 95% CI: 0.85-1.68; the 56 SNPs’ PWM-HR_2nd-stage = 1.29, 95% CI: 0.83-1.99). Removing pleiotropic SNPs did not apparently change the pooled estimate.

Table 1.

Two-stage multiple Cox regressions of genome-wide SNPs associated with CRP predicting colorectal cancer risk

SNP	Chr	Position¥	Gene	Allele		Stage 1				Stage 2
						Adjustment for age and 10 PCs				Adjustment for covariates*
										in addition to age and 10 PCs
				Ref	Alt	HR	(95% CI)	p	p-het†	HR	(95% CI)	p	p-het†
GWAS examining CRP as a binary outcome reflecting high immune response and chronic inflammation (CRP > 3.0 mg/L)
rs2794520	1	159678816	CRPP1/CRP	C	T	0.97	(0.87-1.08)	0.555	0.271	0.95	(0.86-1.06)	0.361	0.563
rs2243458	12	121424490	HNF1A	C	T	1.02	(0.91-1.13)	0.792	0.047	1.06	(0.95-1.19)	0.276	0.047
rs1169311	12	121440731	C12orf43	C	T	1.02	(0.91-1.13)	0.775	0.168	1.05	(0.94-1.17)	0.398	0.098
rs429358	19	45411941	APOE	T	C	0.95	(0.80-1.13)	0.554	0.745	0.95	(0.80-1.13)	0.568	0.511
rs5117	19	45418790	APOC1	T	C	1.02	(0.72-1.45)	0.893	0.216	1.05	(0.74-1.49)	0.793	0.157
GWASs analyzing CRP as a continuous variable that was naturally log-transformed (mg/L)
rs75460349	1	27180088	ZDHHC18	C	A	1.16	(0.78-1.73)	0.459	0.970	1.16	(0.78-1.74)	0.469	0.877
rs2293476	1	40036847	PABPC4	G	C	1.04	(0.92-1.18)	0.525	0.954	1.05	(0.92-1.19)	0.490	0.904
rs1805096	1	66102257	LEPR	A	G	1.03	(0.92-1.15)	0.608	0.298	1.03	(0.92-1.15)	0.619	0.376
rs469772	1	91530305	ZNF644	T	C	1.12	(0.99-1.28)	0.081	0.239	1.08	(0.95-1.23)	0.257	0.257
rs4129267	1	154426264	IL6R	T	C	0.92	(0.83-1.02)	0.122	0.467	0.90	(0.81-1.00)	0.051	0.587
rs2794520	1	159678816	CRPP1/CRP	T	C	1.03	(0.93-1.15)	0.555	0.271	1.05	(0.95-1.17)	0.361	0.563
rs1800947	1	159683438	CRP	G	C	1.30	(0.88-1.92)	0.180	0.422	1.37	(0.92-2.05)	0.125	0.401
rs1417938	1	159684186	CRP	T	C	1.02	(0.91-1.14)	0.694	0.576	1.03	(0.92-1.15)	0.657	0.539
rs10925027	1	247612562	NLRP3	C	T	1.15	(0.97-1.36)	0.117	0.725	1.21	(1.01-1.46)	0.043	0.576
rs12995480	2	629881	TMEM18	T	C	1.07	(0.94-1.22)	0.328	0.897	1.01	(0.88-1.16)	0.888	0.969
rs1260326	2	27730940	GCKR	C	T	0.97	(0.87-1.09)	0.636	0.785	1.00	(0.89-1.12)	0.999	0.765
rs4246598	2	88438050	FABP1	C	A	0.97	(0.87-1.07)	0.529	0.526	0.98	(0.88-1.09)	0.757	0.597
rs9284725	2	102744854	IL1R1	A	C	1.04	(0.92-1.18)	0.525	0.778	1.04	(0.91-1.18)	0.576	0.779
rs13409371	2	113838145	IL1F10	G	A	1.03	(0.93-1.15)	0.559	0.834	1.02	(0.92-1.14)	0.691	0.782
rs1441169	2	214033530	IKZF2	G	A	1.05	(0.95-1.16)	0.363	0.399	1.07	(0.96-1.19)	0.215	0.741
rs2352975	3	49891885	TRAIP	T	C	1.16	(0.95-1.41)	0.144	0.842	1.17	(0.95-1.43)	0.130	0.966
rs1514895	3	170705693	EIF5A2	A	G	1.06	(0.94-1.18)	0.354	0.934	1.08	(0.96-1.21)	0.212	0.934
rs4705952	5	131839618	IRF1	A	G	1.07	(0.93-1.23)	0.335	0.216	1.03	(0.89-1.18)	0.709	0.083
rs17658229	5	172191052	DUSP1	T	C	0.97	(0.72-1.32)	0.868	0.055	0.93	(0.68-1.27)	0.652	0.166
rs9271608	6	32591588	HLA-DQA1	A	G	1.13	(0.71-1.80)	0.606	0.699	1.20	(0.74-1.94)	0.453	0.567
rs12202641	6	116314634	FRK	T	C	0.91	(0.82-1.01)	0.087	0.167	0.90	(0.80-1.00)	0.050	0.124
rs6901250	6	117114025	GPRC6A	G	A	1.02	(0.91-1.15)	0.679	0.716	1.06	(0.95-1.20)	0.292	0.656
rs1490384	6	126851160	CENPW	T	C	0.94	(0.85-1.05)	0.295	0.507	0.96	(0.86-1.07)	0.420	0.596
rs9385532	6	130371227	L3MBTL3	T	C	0.96	(0.86-1.07)	0.485	0.614	1.00	(0.90-1.12)	0.974	0.860
rs1880241	7	22759469	IL6	G	A	0.94	(0.85-1.04)	0.234	0.016	0.99	(0.89-1.10)	0.830	0.013
rs2710804	7	36084529	EEPD1	T	C	0.93	(0.83-1.04)	0.184	0.468	0.93	(0.83-1.04)	0.194	0.279
rs13233571	7	72971231	BCL7B	T	C	1.01	(0.87-1.19)	0.857	1.000	0.94	(0.80-1.11)	0.478	0.993
rs4841132	8	9183596	PPP1R3B	A	G	0.85	(0.70-1.03)	0.098	0.824	0.88	(0.72-1.06)	0.175	0.598
rs2064009	8	117007850	TRPS1	C	T	0.96	(0.86-1.07)	0.444	0.550	0.95	(0.85-1.05)	0.317	0.526
rs2891677	8	126344208	NSMCE2	C	T	1.00	(0.90-1.11)	0.975	0.843	1.03	(0.93-1.14)	0.582	0.992
rs643434	9	136142355	ABO	G	A	0.93	(0.84-1.04)	0.201	0.541	0.96	(0.86-1.07)	0.449	0.489
rs1051338	10	91007360	LIPA	T	G	1.02	(0.90-1.14)	0.802	0.581	1.02	(0.91-1.15)	0.731	0.452
rs10832027	11	13357183	ARNTL	G	A	0.96	(0.85-1.07)	0.427	0.540	0.95	(0.85-1.07)	0.416	0.835
rs10838687	11	47312892	MADD	G	T	0.99	(0.87-1.13)	0.894	0.811	0.98	(0.86-1.12)	0.807	0.891
rs1582763	11	60021948	MS4A4A	A	G	1.04	(0.93-1.16)	0.490	0.952	1.00	(0.89-1.11)	0.955	0.955
rs7121935	11	72496148	STARD10	A	G	0.99	(0.88-1.11)	0.836	0.330	1.02	(0.90-1.14)	0.803	0.184
rs11108056	12	95855385	METAP2	G	C	0.99	(0.83-1.19)	0.951	0.131	1.02	(0.84-1.23)	0.842	0.199
rs10778215	12	103537266	C12orf42	A	T	0.96	(0.87-1.06)	0.430	0.922	0.99	(0.90-1.10)	0.904	0.740
rs7310409	12	121424861	HNF1A	A	G	0.99	(0.89-1.11)	0.910	0.159	0.98	(0.88-1.09)	0.653	0.103
rs2239222	14	73011885	RGS6	A	G	1.03	(0.91-1.18)	0.603	0.835	1.05	(0.92-1.19)	0.476	0.526
rs112635299	14	94838142	SERPINA1/SERPINA2P	T	G	0.88	(0.57-1.36)	0.570	0.199	0.95	(0.62-1.48)	0.828	0.364
rs4774590	15	51745277	DMXL2	A	G	1.17	(1.05-1.30)	0.004	0.139	1.17	(1.05-1.30)	0.004	0.229
rs1189402	15	53728154	WDR72	G	A	0.93	(0.83-1.05)	0.241	0.831	0.95	(0.84-1.07)	0.372	0.822
rs340005	15	60878030	RORA	G	A	1.02	(0.92-1.14)	0.703	0.342	0.96	(0.86-1.07)	0.464	0.634
rs10521222	16	51158710	SALL1	T	C	1.41	(0.86-2.30)	0.172	0.510	1.16	(0.68-1.98)	0.587	0.559
rs1558902	16	53803574	FTO	T	A	1.01	(0.91-1.13)	0.834	0.791	1.02	(0.91-1.14)	0.732	0.851
rs178810	17	16097430	NCOR1	C	T	0.95	(0.86-1.06)	0.388	0.457	0.97	(0.87-1.08)	0.546	0.357
rs10512597	17	72699833	CD300LF/RAB37	T	C	0.95	(0.82-1.09)	0.443	0.097	0.97	(0.84-1.12)	0.682	0.142
rs2852151	18	12841176	PTPN2	G	A	0.96	(0.87-1.07)	0.491	0.276	0.95	(0.85-1.06)	0.358	0.430
rs4092465	18	55080437	ONECUT2	A	G	0.95	(0.75-1.19)	0.646	0.915	0.93	(0.74-1.18)	0.564	0.957
rs12960928	18	57897803	MC4R	T	C	0.99	(0.88-1.12)	0.915	0.814	0.99	(0.88-1.11)	0.818	0.875
rs4420638	19	45422946	APOC1	G	A	1.03	(0.70-1.50)	0.892	0.766	0.93	(0.63-1.38)	0.731	0.669
rs1800961	20	43042364	HNF4A	T	C	0.90	(0.53-1.52)	0.687	0.372	0.89	(0.51-1.55)	0.681	0.392
rs2315008	20	62343956	ZGPAT	T	G	1.08	(0.96-1.20)	0.195	0.620	1.10	(0.98-1.22)	0.096	0.961
rs2836878	21	40465534	PSMG1	A	G	1.11	(0.99-1.24)	0.078	0.288	1.11	(0.99-1.24)	0.080	0.319
rs6001193	22	39074737	TOMM22	G	A	1.01	(0.91-1.13)	0.846	0.320	1.04	(0.93-1.17)	0.445	0.606

Alt, alternative; Chr, chromosome; CI, confidence interval; CRP, C-reactive protein; GWAS, genome-wide association study; HR, hazard ratio; PCs, principal components; Ref, reference; SFA, saturated fatty acids; SNP, single-nucleotide polymorphism; WHI, Women’s Health Initiative studies.

^*Covariates adjusted in the analyses include education; annual family income; family history of colorectal cancer; cardiovascular disease ever; high cholesterol requiring medication ever; body mass index; waist-to-hip ratio; physical activity; % calories from SFA/day; dietary alcohol in g/day; number of cigarettes/day; age at menopause; duration of oral contraceptive use; and durations of exogenous estrogen (E)-only and E plus progestin use.

^¥GRCh 37 coordinated.

^†Heterogeneity in estimates among the 5 WHI sub-GWASs was evaluated by Cochran’s Q test with fixed effects.

Figure 1.

Scatter plots for the effects of individual CRP-genetic instrumental variables on colorectal cancer risk. Each black dot reflects a genome-wide CRP-raising genetic variant. The blue lines indicate penalized weighted median estimates and 95% CIs. (CI, confidence interval; CRP, C-reactive protein; HR, hazard ratio; IVW, inverse-variance weighted; MR, Mendelian randomization; PWM, penalized weighted median; SNP, single-nucleotide polymorphism). A. Five genome-wide SNPs associated with high immune response and chronic inflammation (CRP > 3.0 mg/L) (PWM HR_2nd-stage = 1.20, 95% CI: 0.85-1.68; MR-Egger intercept p value = 0.673). B. Fifty-six genome-wide SNPs associated with CRP phenotype that was naturally log-transformed (mg/L) (PWM HR_2nd-stage = 1.29, 95% CI: 0.83-1.99; MR-Egger intercept p value = 0.824).

Obesity-related lifestyle factors and smoking

We further performed lifestyle-specific stratification analyses. In the SFA-stratified analysis of our 5 CRP-SNPs (Table 2), a 1-unit increase in the genetically predicted chronic inflammation (defined as > 3.0 mg/L of CRP) was associated with approximately 3 times higher risk for CRC in the low fat-diet subgroup (SFA < 9%), particularly in the MR stage 2 of adjustment for multiple lifestyle variables (WM-HR_2nd-stage = 2.88, 95% CI: 1.24-6.68; PWM-HR_2nd-stage = 3.06, 95% CI: 1.21-7.72). In contrast, the inverse relationship between genetically determined chronic inflammation and CRC risk was observed in the high fat-diet subgroup (SFA ≥ 9%), specifically in the first MR stage (IVW-HR_1st-stage = 0.79, 95% CI: 0.64-0.96).

Table 2.

Mendelian randomization analysis: the effect of genetically predicted chronic inflammation status on colorectal cancer risk by dietary fat intake

GWAS examining CRP as a binary outcome reflecting high immune response and chronic inflammation (CRP > 3.0 mg/L)
Analytic method	Stage 1				Stage 2
	Adjustment for age and 10 PCs				Adjustment for covariates*
					in addition to age and 10 PCs
	HR¶	(95% CI)	p	p-het†	HR¶	(95% CI)	p	p-het†
<% calories from SFA < 9.0%>
Inverse-variance weighted	1.53	(0.78-2.98)	0.156	0.009	1.92	(0.47-7.89)	0.270	0.102
Weighted median	1.69	(0.80-3.53)	0.168	0.089	2.88	(1.24-6.68)	0.014	0.003
Penalized weighted median	1.68	(0.82-3.47)	0.158	0.083	3.06	(1.21-7.72)	0.018	0.004
MR-Egger: slope	2.96	(0.01-762.22)	0.578		14.75	(0.01-1.12e+06)	0.502
intercept	0.87	(0.27-2.76)	0.726		0.65	(0.06-6.77)	0.600
<% calories from SFA ≥ 9.0%>
Inverse-variance weighted	0.79	(0.64-0.96)	0.029		0.80	(0.52-1.24)	0.235
Weighted median	0.82	(0.57-1.18)	0.292		0.70	(0.47-1.03)	0.067
Penalized weighted median	0.82	(0.57-1.18)	0.290		0.70	(0.47-1.03)	0.072
MR-Egger: slope	0.38	(0.14-1.06)	0.057		0.25	(0.01-4.96)	0.238
intercept	1.17	(0.94-1.44)	0.106		1.28	(0.69-2.38)	0.301

CI, confidence interval; CRP, C-reactive protein; GWAS, genome-wide association study; HR, hazard ratio; MR, Mendelian randomization; PCs, principal components; SFA, saturated fatty acid; SNP, single-nucleotide polymorphism. Numbers in bold face are statistically significant.

^*Covariates that were adjusted for in analyses of the association between GWA CRP-SNPs and colorectal cancer risk include education; annual family income; family history of colorectal cancer; cardiovascular disease ever; high cholesterol requiring medication ever; body mass index; waist-to-hip ratio; physical activity; % calories from SFA/day; dietary alcohol in g/day; number of cigarettes/day; age at menopause; duration of oral contraceptive use; and durations of exogenous estrogen (E)-only and E plus progestin use. Variable tested for stratification was not included as a covariate in the stage 2 analysis.

^¶The MR estimate (except weighted/penalized weighted medians) was adjusted for a correlation between CRP phenotype and colorectal cancer within the same population.

^† p value for heterogeneity test in MR estimates across strata (< 9.0% from SFA vs. ≥ 9.0% from SFA) was evaluated with Cochran’s Q test.

Different patterns were observed in the waist circumference-stratified analysis (Figure 2A). Genetically determined chronic inflammation was associated with 20% reduced risk for CRC among the non-viscerally obese group, whereas a proportionate relationship (15% increased CRC risk) was observed with limited statistical power among the viscerally obese group.

Figure 2.

Forest plots of MR estimates. The plots show the effects of genetically predicted chronic inflammation status (CRP > 3.0 mg/L) on colorectal cancer risk in obesity/hormone use-specific strata, presented as 95% CIs (red lines) of the estimates and the inverse-variance weights and the weighted and penalized weighted medians (percentages proportional to the size of the blue squares). The MR estimates were calculated in the 2 stages on the SNP-colorectal cancer association: in the first stage, only age and 10 PCs and in the second stage, lifestyle covariates in addition to age and 10 PCs were adjusted. p-het is from the heterogeneity test in MR estimates across strata. (CI, confidence interval; CRP, C-reactive protein; E, exogeneous estrogen; HR, hazard ratio; IVW, inverse-variance weight; MR, Mendelian randomization; SNP, single-nucleotide polymorphism; P, progestin; PC, principal component; PWM, penalized weighted median; WM, weighted median). A. IVW MR estimates by waist circumference. B. WM/PWM MR estimates by E plus P use.

In the smoking-stratified analysis of our 5 and the outside 56 GWA CRP-SNPs (Table 3), a 1-unit increase in the log-transformed genetically predicted CRP was associated with 50% or greater risk of CRC among never smokers, but only in the first MR stage. The elevated effect on CRC risk was more profound in the first stage of MR analysis after excluding pleiotropic SNPs. Similarly, among smokers with 15 or more cigarettes/day, genetically driven chronic inflammation was associated with 60% increased risk for CRC in the first MR stage. By taking into account our 5 GWA CRP-SNPs’ interactions with cigarette smoking, the corrected MR estimate indicated a greater risk for CRC in relation to genetically elevated chronic inflammation (MR G×E HR = 2.59, 95% CI: 1.13-5.91). The MR-Egger tests showed no significant evidence of directional pleiotropy across the tested associations.

Table 3.

Mendelian randomization analysis: the effect of genetically predicted CRP (as log-transformed or categorized) on colorectal cancer risk by number of cigarettes per day

Analytic method	All SNPs								After exclusion of pleiotropic SNPs
	Stage 1				Stage 2				Stage 1							Stage 2
	Adjustment for age and 10 PCs				Adjustment for covariates*				Adjustment for age and 10 PCs							Adjustment for covariates*
					in addition to age and 10 PCs											in addition to age and 10 PCs
	HR	(95% CI)	p	p-het†	HR	(95% CI)	p	p-het†	HR		(95% CI)		p		p-het†	HR	(95% CI)	p	p-het†
GWASs analyzing CRP as a continuous variable that was naturally log-transformed (mg/L)
<Never smokers>
Inverse-variance weighted	1.48	(1.01-2.18)	0.045	0.795	1.51	(0.97-2.37)	0.069	0.309	1.53	(1.02-2.30)		0.041		0.742		1.57	(0.97-2.54)	0.063	0.235
Weighted median	1.82	(1.02-3.26)	0.044		1.59	(0.85-2.96)	0.146		1.82	(0.97-3.44)		0.064				1.75	(0.91-3.34)	0.092
Penalized weighted median	1.82	(1.02-3.25)	0.044		1.65	(0.87-3.12)	0.125		1.82	(1.004-3.31)		0.049				1.79	(0.94-3.42)	0.079
MR-Egger: slope	1.58	(0.88-2.84)	0.122		1.44	(0.73-2.86)	0.286		1.59	(0.85-2.95)		0.140				1.46	(0.71-3.01)	0.300
intercept	1.00	(0.96-1.03)	0.766		1.004	(0.97-1.04)	0.853		1.00	(0.96-1.03)		0.876				1.01	(0.97-1.05)	0.780
GWAS examining CRP as a binary outcome reflecting high immune response and chronic inflammation (CRP > 3.0 mg/L)
<≥ 15 cigarettes/d>
Inverse-variance weighted¶	1.58	(1.03-2.41)	0.041	0.908	1.63	(0.91-2.93)	0.083	0.788
Weighted median	1.52	(0.74-3.12)	0.256		1.62	(0.76-3.46)	0.210
Penalized weighted median	1.52	(0.72-3.19)	0.269		1.62	(0.77-3.41)	0.201
MR-Egger: slope	0.36	(0.04-3.26)	0.236		0.29	(0.01-10.57)	0.351
intercept	1.37	(0.86-2.18)	0.119		1.44	(0.68-3.06)	0.219
<MR G×E interaction: scaled GWA CRP weighted-genetic score€ × number of cigarettes/d>
G×E univariate slope§	2.59	1.13-5.91	0.032

CI, confidence interval; CRP, C-reactive protein; GWAS, genome-wide association study; HR, hazard ratio; MR, Mendelian randomization; PCs, principal components; SFA, saturated fatty acid; SNP, single-nucleotide polymorphism. Numbers in bold face are statistically significant.

^*Covariates that were adjusted for in analyses of the association between GWA CRP-SNPs and colorectal cancer risk include education; annual family income; family history of colorectal cancer; cardiovascular disease ever; high cholesterol requiring medication ever; body mass index; waist-to-hip ratio; physical activity; % calories from SFA/day; dietary alcohol in g/day; number of cigarettes/day; age at menopause; duration of oral contraceptive use; and durations of exogenous estrogen (E)-only and E plus progestin use. Variable tested for stratification was not included as a covariate in the stage 2 analysis.

^† p value for heterogeneity test in estimates among GWA CRP-SNPs was evaluated with Cochran’s Q test with fixed effects.

^¶The MR estimate was adjusted for a correlation between CRP phenotype and colorectal cancer within the same population.

^§The MR G×E slope reflects an MR estimate from univariate analysis that was corrected for G×E interaction with a modeled (cigarette) variable.

^€The scaled weighted-genetic score was estimated with 5 CRP SNPs whose effect size reflects chronic inflammation (CRP > 3.0 mg/L) per allele.

Hormone therapy

Whereas genetically predicted chronic inflammation was not associated with CRC risk among E+P nonusers and relatively short-term users (i.e., < 5 years), long-term E+P users showed strongly positive relationships (Figure 2B). For example, among long-term E+P users (10 years or longer), women with genetically elevated chronic inflammation had 27 times higher risk for CRC. Similar null associations with CRC risk were observed among E+P nonusers and short-term users in the MR analysis for genetically elevated CRP phenotype (Table S3). Further, among long-term E+P users (5 to < 10 years), a 1-unit increase in the log-transformed genetically determined CRP was associated with 20 times or higher risk for CRC (PWM-HR_1st-stage = 28.60, 95% CI: 1.63-501.89; PWM-HR_{1st-stage after excluding pleiotropy} = 21.24, 95% CI: 1.30-346.88).

Similar patterns were observed in the E-only subgroups (Figure 3). Among nonusers, no association between genetically elevated CRP and CRC risk was observed. Notably different from the E+P subgroups, genetically elevated CRP with increased CRC risk was found only among relatively shortterm users (< 5 years). No directional pleiotropic effects across the analyses were observed.

Figure 3.

Forest plot of MR estimates by E-only use. The plot shows the effects of genetically predicted CRP phenotype on colorectal cancer risk in E-only subgroups, presented as 95% CIs (red lines) of the estimates and the inverse-variance weights (percentages proportional to the size of the blue squares). The MR estimates were calculated in the 2 stages on the SNP-colorectal cancer association: in the first stage, only age and 10 PCs and in the second stage, lifestyle covariates in addition to age and 10 PCs were adjusted. p-het is from the heterogeneity test in MR estimates across strata. (CI, confidence interval; CRP, C-reactive protein; E, exogeneous estrogen; HR, hazard ratio; IVW, inverse-variance weight; MR, Mendelian randomization; OC, oral contraceptive; SNP, single-nucleotide polymorphism; PC, principal component).

No other subgroups presented a statistically significant increased risk for CRC in association with genetically determined chronic inflammation or CRP levels (Tables S3 and S4). Further, in the MR G×E interactions, MR estimates corrected for cigarette smoking, breast feeding, and depressive symptoms reached statistical significance, and no pleiotropic effect of the estimates was detected (Table S5).

Discussion

We examined the genetically predicted effect of CRP levels (a natural log-transformed 1 mg/L increase or > 3 mg/L vs. ≤ 3.0 mg/L, indicating chronic low-grade inflammation) on postmenopausal CRC risk by performing MR analysis and found that genetically determined CRP levels and chronic inflammation status were associated with increased risk for CRC in subgroups of women with particular lifestyle behaviors. The MR findings, if the modeled genetic instruments are not affected by vertical and horizontal pleiotropic effects, can be comparable with those from randomized clinical trials, thus providing a fairly robust causal inference [30,87]. Our MR analysis removed pleiotropic SNPs in relation to obesity and glucose and lipid metabolism, which may confound the CRP-cancer risk association, and further identified a wide range of confounding factors that were accounted for in the genetic analysis with CRC outcomes. Our MR estimates, such as WM and PWM, allow some relaxation of the rigid conditions for MR instruments, thus yielding a more robust estimate of the causal effect than a traditional MR estimate. In addition, the MR approach can examine the lifelong effect of exposure (CRP) on CRC risk, and it is less susceptible to reverse causality. Thus, although our MR study was not designed to directly evaluate biologic mechanisms, our findings indicate that the long-standing effect of CRP lies in the causal pathway of CRC development. To the best of our knowledge, this study is the first to report the causal relationship between genetically determined CRP phenotype and CRC risk within lifestyle-specific subgroups in an MR framework.

In the stratification analyses by obesity, we detected a reduced risk of CRC in relation to CRP in non-viscerally obese women, whereas increased CRP-associated CRC risk was observed even with a lack of statistical power in viscerally obese women. In line with results from previous studies [62-65], our finding indicates that larger waist circumference, a biomarker of visceral fat accumulation, is more important than overall obesity (as indicated by a high BMI) in CRC risk in women. Our results further support the hypothesis [66,67] that chronic inflammation is one of the interconnected pathways that mediate the adiposity-CRC association, but additional studies on larger populations are warranted to confirm our results.

In obesity-related lifestyle-specific analyses, particularly in the dietary-SFA stratification, CRP-increased risk for CRC in the low fat-intake subgroup was observed, whereas CRP-reduced risk for CRC in the high fat-intake subgroup. Previous studies [89-91] showed that intake of n-3 polyunsaturated fatty acids improved intestinal barriers, leading to reduced exposure to intestinal bacterial products, thus contributing to less inflammatory potential and inhibition of production of pro-inflammatory cytokines such as CRP. Also, in contrast to the results from the aforementioned studies, one study [12] demonstrated that CRP together with oxidized low-density lipoprotein interact with LOX-I receptors, elevating the mRNA expression of genes (e.g., LOX-I, CEA, and MMP1/2) important in colorectal tumorigenesis. However, no published studies to date have examined the effect of SFA intake on CRP levels in relation to colorectal carcinogenesis. Our finding of the protective effect of CRP on CRC risk when the intake of SFAs is high suggests that different fat types of lipid metabolism can be implicated in the immune-related CRC pathways and calls for further biologic-mechanism studies.

Cigarette smoking is an established risk factor for CRC, accounting for 15-20% of CRC cases [92,93] with a dose-response relationship of daily cigarette consumption [94]. Tobacco-derived carcinogens reach the colorectal mucosa through the digestive tract and the circulatory system, causing the potential carcinogenesis in this target organ [95]. Circulating levels of inflammatory cytokines such as CRP have been directly associated with smoking in CRC case-control studies [29,67]. Our data, with an average 15-year induction period from onset of smoking to cancer formation, despite relatively short-term exposure [96,97], showed strong risk for CRC in relation to CRP in women who smoked ≥ 15 cigarettes/day. Moreover, consistent with other studies that reported an interaction of smoking with the CRP-CRC risk association [22-24], our study indicates the presence of a significant CRP-induced causal pathway for CRC development, which was corrected for the gene-smoking interaction. Altogether, the inflammatory pathways that interact with smoking can be postulated to be involved in colorectal carcinogenesis. Interestingly, our MR analysis of the CRP-increased CRC risk was also observed in never-smokers to a slightly lesser extent; this may be the result from unmeasured confounders or the existence of other epigenetic molecular pathways, thus further studies addressing these issues with a more comprehensive molecular dataset are required.

In the analyses stratified by hormone therapy, a substantially increased risk of CRC in relation to CRP was found in the E+P users. In particular, CRP-induced high risk for CRC was found in longer-term E+P users (≥ 5 years) while no association of the CRP-CRC risk was found in nonusers and relatively short-term E+P users (< 5 years). Lifetime cumulative exposure to estrogen, particularly to the opposed (E+P) hormone use, has been considered a protective factor for CRC risk [62,68,69], as shown in in vitro and case-control studies revealing that introduction of an E-receptor gene resulted in marked growth suppression in CRC cells and was associated with low risk for microsatellite instability-positive CRC [98,99]. However, this protective effect was observed only in relatively short-term E+P users (< 5 years) [68,69]. Further, unlike natural progesterone, the synthetic progestin contained in the E+P combination has an affinity for androgen and mineralocorticoid receptors, leading to cell proliferation and anti-apoptosis, which in turn may contribute to colorectal carcinogenesis [100]. Taken together, that evidence and our findings suggest that longer-term cumulative exposure to the opposed estrogen may be implicated in the colorectal tumorigenic pathway that is presumably mediated by inflammation. In contrast, only the relatively short-term E-only users (< 5 years) in our MR analysis had an increased risk of CRC in relation to CRP, a finding supported by another WHI study that had a 7-year follow-up, [70]; this further warrants studies on a larger population with longer-term follow-up of biologic-function investigation for more conclusive results.

We tested for MR directional pleiotropy and reduced any potential pleiotropic effect by using multifaceted approaches: removal of pleiotropic SNPs, adjustment of a wide range of confounding factors, corrections of MR estimates for gene-lifestyle interactions, and use of WM/PWM estimates. Nonetheless, our findings could be biased due to residual and unmeasured confounding factors. Also, potential nonlinearity in the SNP-exposure and exposure-outcome relationships would tend toward a null result, so generating a spurious association is less likely [101]. Our results could have inflated false-positive rates due to multiple comparisons. Finally, our findings should not be extrapolated to other races or ethnicity, in whom the interrelationships between genetic instruments, CRP, and CRC risk may be different.

In summary, we attempted to improve the causal inference between genetically elevated CRP and CRC risk in postmenopausal women by quantifying the association in an MR framework. We obtained evidence of the potential causal association between lifelong exposure to CRP and CRC risk, particularly in women with low intake of SFAs, regular smokers, and relatively short-term users of unopposed and long-term users of opposed estrogen. Further biologic-mechanism studies may help to elucidate the inflammatory-induced colorectal carcinogenesis pathways that interact with fatty acids, smoking, and hormone therapy. Our findings may yield novel evidence on immune-related etiologic pathways connected to CRC risk, and suggest the possible use of CRP as a CRC-predictive biomarker in women with particular behaviors and CRP marker-informed interventions to reduce CRC risk.

Disclosure of conflict of interest

None.