TGF-β-pathway-based polygenic risk score modifies the association between red meat intake and colorectal cancer risk: Application of a novel pathway-based PRS method

Abstract

Background:

Red and/or processed meat are established colorectal cancer (CRC) risk factors. Genome-wide association studies (GWAS) have reported over 200 variants associated with CRC risk. We used functional annotation data to identify subsets of variants within known pathways to construct pathway-based Polygenic Risk Scores (pPRS) to assess interactions with meat intake.

Methods:

A pooled sample of 30,812 cases and 40,504 CRC controls from 27 studies were analyzed. Quantiles for red and processed meat intake were constructed. 204 GWAS variants were annotated to genes with AnnoQ and assessed for overrepresentation in PANTHER-reported pathways. pPRS’s were constructed from significantly overrepresented pathways. Covariate-adjusted logistic regression models evaluated interactions between pPRS and red or processed meat intake in relation to CRC risk.

Results:

A total of 30 variants were overrepresented in four pathways: Presenilin-Alzheimer disease, Cadherin/WNT-signaling, Gonadotropin-releasing hormone receptor, and TGF-β signaling. We found a significant interaction between TGF-β-pPRS and red meat intake (OR_int = 0.95; 95% CI = 0.92–0.98; p = 0.003). When variants in the TGF-β pathway were assessed, we observed significant interactions of red meat with rs2337113 (intron SMAD7 gene, Chr18), and rs2208603 (intergenic region BMP5, Chr6) (p = 0.0005 & 0.036, respectively). There was no evidence of pPRS x red meat interactions for other pathways or with processed meat

Conclusions:

This pathway-based interaction analysis revealed a statistically significant interaction between variants in the TGF-β pathway and red meat consumption that impacts CRC risk.

Impact:

These findings shed light into the possible mechanistic link between red meat consumption and CRC risk.

INTRODUCTION:

Colorectal cancer (CRC) has the third highest cancer incidence, and the second highest cancer mortality worldwide (1). It is estimated that if current trends are maintained, 3.2 million new cases of CRC will occur in 2040, with a concomitant 73.4% increase in deaths to over 1.6 million per annum (2). An ever-expanding body of evidence has identified several modifiable/lifestyle risk factors for CRC. Examples that have been associated with differential CRC risk include smoking, alcohol consumption, decreased dietary consumption of fiber, whole grains, vegetables, and fruits, living with obesity or being overweight, low physical activity, inflammatory bowel disease, and increased intake of red meat and processed meat (3–11). Red meat and processed meat have been labeled as “probably carcinogenic to humans” (group 2a carcinogen), and “carcinogenic to humans” (group 1 carcinogen), respectively, by the International Agency for Research on Cancer (IARC) due to the evidence vis-à-vis CRC (12). Different putative causal pathways have been proposed for the association between red meat and/or processed meat with CRC. Mainly, the presence of heme iron combined with long-term intestinal dysbiosis, in addition to the neoformation of mutagenic compounds (N-nitroso compounds, heterocyclic amines, polycyclic aromatic hydrocarbons) which have been reported to act both systemically, and locally at the mucosa of the colon and rectum (12–16); thus, potentially mediating the association between red meat and/or processed meat consumption and CRC risk.

The majority of CRC cases are considered sporadic, as over three-quarters of the patients do not have a family history (17). However, genome-wide association studies (GWAS) have informed a heritable basis for the disease, with over 200 single-nucleotide polymorphisms (SNPs) showing a marginally statistically significant association with the trait (18). Moreover, non-additive estimates of heritability such as gene-environment (GxE) interactions, explain some of the variance in the heritability of CRC (19). GxE interactions refers to the scenario where the effect of a genetic variant differs across levels of the environmental exposure, or vice versa. That is, the susceptibility of a person to a given environmental exposure can be influenced by a genetic variant. Previously, most studies that evaluated GxE interactions in CRC were considered underpowered to detect these interactions (20). To address this, we recently published a genome-wide GxE interaction analysis that reported two significant novel interactions for red meat consumption: the rs4871179 SNP in chromosome (chr) 8 (downstream of HAS2) and the rs35352860 SNP in chr18 (SMAD7 intronic region) (21). However, traditional candidate GxE analyses, which sequentially evaluate variants associated with the trait, prohibit the adequate identification of individuals at elevated CRC risk, and do not explain the underlying mechanisms of this highly polygenic carcinogenic process.

Polygenic risk scores (PRS) aggregate multiple SNPs reported through GWAS with a typically small effect to estimate heritability or genetic liability to a particular phenotype. In a classic PRS, a sum of the effect alleles that an individual carries, weighted by the effect sizes of the variants (e.g., GWAS summary statistics), is computed (22). This enables the output of a risk estimate at the individual level that could be leveraged to stratify patients and provide targeted screening, with the potential to guide precise CRC prevention (23). However, when modeled in relation to a phenotype, this canonical PRS approach assumes that every SNP that is an element of the PRS (i.e., ${SNP}_i\in PRS$) exhibits the same estimate of association (i.e., ${\beta }_{{SNP}_i}=\dots ={\beta }_{{SNP}_n}$ ). Moreover, a loss of information ensues, as the PRS groups SNPs that are associated with the phenotype via dissimilar biological processes. To address this challenge, pathway-based PRS (pPRS) construct gene sets that aggregate risk alleles across a priori defined pathways via existing annotated databases (e.g., REACTOME, PANTHER Pathway, KEGG (24–26)). This strategy provides functionally informed risk scores that have enabled detection of putative disrupted biological pathways for various polygenic traits (e.g., Alzheimer’s disease, schizophrenia, coronary artery disease (27–29)) but has yet to be deployed for CRC risk. Furthermore, pathway-based PRS can yield substantially greater power to test pPRS x E interactions relative to traditional PRS x E evaluations (30).

In this study, we applied functional annotation data to identify subsets of GWAS variants that are components of biological pathways to construct pathway-based PRS. Additionally, we evaluated multiplicative interactions between pPRS and red meat or processed meat intake and CRC risk in a pooled dataset to ascertain whether genes within specific pathways that exhibit a statistically significant interaction with the environmental factor influence CRC risk.

MATERIALS and METHODS:

Study participants

Details on study participants, genotyping information, and environmental exposure data have been previously published (31,32). In brief, 27 studies from the Genetics and Epidemiology of Colorectal Cancer Consortium (GECCO), the Colorectal Cancer Transdisciplinary Study (CORECT), and the Colon Cancer Family Registry (CCFR) were utilized (Supplementary Table S1). These studies have been previously described (21,33). The outcome/case status was defined as a positive diagnosis of invasive colorectal adenocarcinoma (International Classification of Diseases codes (ICD)153–154), and confirmed via medical records, pathology reports, or death certificate information. Nested case-control sets were assembled via risk-set sampling in cohort studies, and population-based controls were matched on age, sex, race, and enrollment date/trial group (in the SELECT trial) in case-control studies. Cases were further stratified by tumor location as follows: proximal colon (ICD-9 codes: 153.0, 153.1, 153.4, and 153.6), distal colon (ICD-9 codes: 153.2, 153.3, 153.7), and rectum (ICD-9 codes: 154.0, 154.1). Given the very small number participants of non-European descent across studies we only included individuals of European ancestry according to self-reported race and ethnicity and confirmation via genetic data. Exclusion criteria included: 1) Duplicate participants or cryptic relatedness; 2) Diagnosis of advanced adenoma (only), defined as an adenoma 1cm or larger in diameter and/or with tubulovillous, villous, or high-grade dysplasia/carcinoma-in-situ histology; and 3) Missing data for red meat and processed meat intake. All participants gave written informed consent, and studies were approved by their individual institutional review boards.

Genotype data

Genotyping arrays utilized by each study are described in Supplementary Table S1. Quality control and genotyping metrics were conducted as previously described (31,32). Briefly, samples were excluded with the following criteria: 1) Single nucleotide polymorphisms (SNPs) with missing call rate >2–5%; 2) Departure from Hardy-Weinberg equilibrium (threshold = P<1×10–4); 3) Mismatch between genotyped and reported sex; 4) Discordant genotype calls with duplicate samples. The Haplotype Reference Consortium panel of 39.1 million variants was used to impute to the genotypes with the University of Michigan Imputation Server (34). We used the Binarydosage R package (https://cran.r-project.org/web/packages/BinaryDosage) to process and manage the imputed genotypes in our sample. Imputed SNPs with a minor allele frequency (MAF) <1% and/or imputation accuracy of R² ≤ 0.8 were dropped. PLINK 1.9 was utilized to compute principal components on 30,000 randomly sampled SNPs with MAF >5% to control for population stratification (35). For use in polygenic risk score analyses, we leveraged data from 204 autosomal SNPs that have been previously reported to showcase a marginal genome-wide statistically significant association with CRC (P<5×10–8) (18).

Environmental exposure data

In-person interviews, phone interviews, or structured self-administered questionnaires were utilized to collect environmental risk factors and sociodemographic variables as previously described (33). After quality-control checks in each study were conducted, common data elements (CDEs) were defined. Study-specific data elements were mapped to the CDEs through an iterative process with outreach to data contributors. Via SAS and a T-SQL script, all CDEs were pooled into a dataset that included common definitions, standardized permissible values, and standardized coding. The dataset was evaluated for quality assurance, with the distribution of the data within and between studies assessed for outlier observations. Red meat intake was defined as consumption of beef, pork, and/or lamb, while processed meat included intake of bacon, sausages, luncheon/deli meats, and/or hot dogs. Both variables were expressed as servings per day (i.e., servings/day), with one serving equivalent to 70.9 grams or 2.5 ounces (36). Heterogeneity in the red meat/processed meat definition for individual studies caused overlap in some instances, with studies including processed meat under the red meat definition. We evaluated red meat and processed meat intake by constructing sex-study-specific quartiles, with the median of intake for each quartile utilized to define consumption values, fitted as continuous predictors in regression models. This approach increased our ability to detect underlying dose-response relationships, relative to models that parametrize meat intake using categorical quartiles. Food frequency questionnaires were utilized to compute total energy intake (kcal/day). Mean imputation was utilized for studies with partial missingness. In studies that did not report energy intake (ASTERISK, DACHS, PHS, UKB), the value was set to a constant (zero). Body mass index (BMI) was computed as weight (kg)/height(m)². The predefined World Health Organization BMI cut-off points were used to generate an ordinal categorical variable: underweight (<18.5 kg/m²), normal weight (18.5 to <25 kg/m²), overweight (≥25.0 to <30 kg/m²), and obese (≥30 kg/m²). Physical activity was estimated in metabolic equivalents as task hours per week (MET-h/wk) for participants in studies with available information to approximate average amount of time per week spent on leisure activities, or all activities when leisure was not specified. We further dichotomized the physical activity variable into active (≥ 8.75 MET-h/wk) and inactive (< 8.75 MET-h/wk), in line with guidelines that recommend a minimum of 150 minutes of moderate or 75 minutes of vigorous activity per week for individuals living with cancer and for cancer prevention (37) (Supplementary Table S2).

Pathway overrepresentation

From the 204 genome-wide significant autosomal SNPs reported by Fernandez-Rozadilla et. al (18), rsIDs were extracted in accordance with the genome reference consortium (GRC) human build 37 (GRCh37). These SNPs were then annotated using the Annotation Query (AnnoQ) platform, which provides integrated functional annotation supported by a database of ~39 million pre-annotated human variants from the Haplotype Reference Consortium (38). Annotations to Ensembl and Reference sequence (RefSeq) genes were derived using inferences from Annotate Variation (ANNOVAR), SnpEff, and Ensembl Variant Effect Predictor (VEP) (39–43). SNPs residing in enhancer regions were linked to their target genes via the Predicted by Experimental Results: Enhancer-Gene Relationships Illustrated by a Nexus of Evidence (PEREGRINE), which leverages experimental data from chromatin interaction analysis by paired-end tag sequencing, expression quantitative trait locus, and Hi-C across 78 cell and tissue types (44). Duplicate gene identifiers were removed, and because annotations from multiple tools, gene models, and enhancers may map a SNP to more than one gene, the union of all unique gene assignments was retained without applying prioritization among annotation sources; with only protein-coding genes carried forward to enrichment analysis. The resultant unique genes were uploaded into the Protein ANalysis THrough Evolutionary Relationships (PANTHER) classification system v.18.0. (45,46) which is built on 143 annotated complete genomes and assigns sequences to PANTHER Pathways (v3.6.7; 177 pathways and 3092 components) (25,45,46). The set of genes falling within each pathway were tested for overrepresentation relative to the PANTHER Pathway annotation sets using Fisher’s exact test, with false discovery rate correction applied for multiple testing (p < 0.05) (45). Each significantly overrepresented pathway was subsequently used for downstream analyses. Code to conduct pathway overrepresentation analysis available at: https://github.com/quemeb/Annoq_Overrepr_Workflow.

Statistical analyses

Standard polygenic risk score (PRS) methods (22) were used to compute pathway-based polygenic risk scores (pPRS). SNP-specific weights were extracted after regressing CRC on the set of 204 variants in our sample with a logistic regression model that adjusted for study, sex, age, and three principal components of ancestry. A complete list of variants and weights can be found in Supplementary Table S3. The subset of weights corresponding to each of the overrepresented pathways were used to compute the pPRS. To identify potential interactions between pPRS and red meat/processed meat intake (pPRSxE), a 1 degree-of-freedom (1-df) test of gene-environment (GxE) interaction was utilized to determine statistical significance of the interaction term pPRSxE in multivariable logistic regression. All models were adjusted for study, age, sex, and the first three principal components to account for population structure. Simulations have shown that Type I error rates are conserved when the same sample is used to both generate PRS weights and to evaluate potential pPRSxE interactions (30). As variants can simultaneously participate in multiple pathways, pPRS subsets that included unique variants, and variants that showed overlap to other pathways were generated. To contrast the utility of deploying pPRS, we generated a PRS that included all 204 variants (overall PRS), and a PRS that contained the subset of these variants that were not overrepresented among PANTHER-defined pathways (no pathway PRS, n = 174). Potential effect modification of the association between red meat/processed meat intake and CRC risk by pPRS was evaluated by further stratifying the models by quartiles of the corresponding pPRS. In addition, we conducted confounding analyses as described by Keller (47), to evaluate potential confounding of the pPRSxE interactions by smoking (never, former current), BMI (kg/m²), intake of alcohol (g/day), fiber (g/day), fruits (servings/day), vegetables (servings/day), total caloric intake (kcal/day), and physical activity (MET-h/wk). Interaction analyses of individual variants overrepresented in a single pathway were controlled for multiple testing with Bonferroni-adjusted p-values presented. We generated Locus Zoom plots for select variants to inspect the strength of the association, linkage disequilibrium (LD), and position of findings relative to other variants in the region. Plots were generated with LocusZoom (v1.3.32), LD was estimated with 1000G EUR study population controls. All analyses were performed using the R programming language (v.4.2.1). All p-values are two-sided, with a value < 0.05 deemed statistically significant.

Data availability

Data utilized to generate results in this study are available in the database of Genotypes and Phenotypes (dbGAP) (http://view.ncbi.nlm.nih.gov/dbgap-controlled) accession numbers: phs001415.v1.p1, phs001315.v1.p1, phs001078.v1.p1, phs001903.v1.p1, phs001856.v1.p1 and phs001045.v1.p1.

RESULTS:

A total of 30,812 cases and 40,504 controls met the inclusion criteria. Of these, 138 had missing information on red meat intake (cases = 69, controls = 69) and 5,764 had missing processed meat intake information (cases = 2,634, controls = 3,130). Tumors were more commonly located in the proximal colon (35.2%), followed by the distal colon (30.4%), and the rectum (27.9%); 6.5% of the cases had a missing tumor location. Compared to controls, CRC cases, on average, had a higher BMI, were more likely to be male, and reported a higher total caloric intake. Cases had a higher mean intake of red meat (0.60 servings/day ± 0.46), and processed meat (0.35 servings/day ± 0.34) compared to controls (0.54 servings/day ± 0.44 & 0.28 ± 0.29, respectively, P-values < 0.001). Participant characteristics are described in detail in Supplementary Table S2.

Pathway overrepresentation analysis and pathway-based polygenic risk scores

Pathway overrepresentation analysis revealed that 30 of the 204 SNPs that were a priori marginally associated with CRC were statistically significantly overrepresented across four biological pathways (p-values <0.001): 1) Gonadotropin-releasing hormone receptor pathway, n = 16 SNPs; 2) Presenilin-Alzheimer disease pathway, n = 8 SNPs; 3) Cadherin/WNT signaling pathway, n = 7 SNPs; and 4) Transforming Growth Factor – beta (TGF-β) signaling pathway, n = 14 SNPs (Supplementary Table S4). Some variants were simultaneously overrepresented in multiple pathways (Figure 1). The biggest overlap was reported between SNPs in the Gonadotropin-releasing hormone receptor pathway and TGF-β signaling pathway (n = 9). The rs35470271 SNP, which maps to chromosome 3 in the intergenic region of the ZNF621 (Zinc Finger Protein 621), was the most commonly overrepresented SNP, with involvement in Gonadotropin-releasing hormone receptor pathway, presenilin-Alzheimer disease pathway, and Cadherin/WNT signaling pathway (Figure 1).

Figure 1.

Overlap and set size of known colorectal cancer risk loci overrepresented across biological pathways.

Cadherin/WNT, Cadherin/WNT signaling pathway; Presenilin, Presenilin-Alzheimer disease pathway; TGF-B, Transforming Growth Factor – beta signaling pathway; Gonadotropin, Gonadotropin-releasing hormone receptor pathway.

Note: Set size corresponds to the number of single nucleotide polymorphisms contained in each pathway-based Polygenic Risk Score (pPRS).

Overrepresented and uniquely overrepresented SNPs in the four biological pathways were leveraged to construct pPRS. As expected, moderate to strong correlations were found between pPRS’s and their corresponding pPRS that contained only uniquely overrepresented SNPs (Figure 2). A strong correlation was found between the presenilin-Alzheimer disease-pPRS and Cadherin/WNT signaling-pPRS (r² = 0.69), and between the TGF-β signaling-pPRS (TGF-β-PRS) and Gonadotropin-releasing hormone receptor-pPRS (r² = 0.62) (Figure 2), driven by variants that overlapped across said pathways (Figure 1).

Figure 2.

Correlogram for pathway-based polygenic risk scores (pPRS) constructed from known colorectal cancer risk loci overrepresented in biological pathways.

Note: The numbers and gradient scale correspond to Pearson’s correlation coefficients. Overall PRS contains 204 SNPs that have been previously associated with colorectal cancer risk.

Interaction between pPRS with red meat or processed meat intake in CRC risk

We observed a statistically significant interaction for the overall PRS with red meat intake (p-value = 0.0048). Thus, suggesting that the effect of red meat intake on CRC risk, is modified by genetic susceptibility as captured by the PRS. However, when we deployed the novel pPRS’s to evaluate the putative multiplicative interaction with meat intake (pPRSxE), the interaction was significant for the TGF-β-PRS with red meat intake (p-value = 0.0031), with no evidence of an interaction in the other three biological pathways (Figure 3A). Moreover, the interaction remained significant when assessed in the TGF-β-PRS with only uniquely overrepresented variants for the TGF-β signaling pathway (p-value = 0.0007). No statistically significant interactions were identified for processed meat intake (p-values > 0.05) (Figure 3B).

Figure 3.

Odds Ratios and 95% Confidence Intervals for interactions between pathway-based Polygenic Risk Scores with (A) red meat, and (B) processed meat intake on colorectal cancer risk.

pPRS, pathway-based Polygenic Risk Scores; OR, Odds Ratio; CI, Confidence Interval.

Note: The No Pathway PRS contains variants that have previously shown a genome-wide significant association with colorectal cancer but were not overrepresented in biological pathways. For variants that are exclusively overrepresented in a single pathway, a second subset is generated (i.e., “Uniques”) and the appropriate pPRS computed. P-value corresponds to a 1-degree-of-freedom test of pPRSxE.

All models were adjusted for study, age, sex, and the first three principal components.

When the pPRSxE interactions were examined in a multivariable multinomial logistic regression model with tumor location as the outcome variable (i.e., proximal colon, distal colon, and rectum), a statistically significant interaction was observed between the overall PRS, and the TGF-β-PRS and red meat intake in relation to CRC risk in the proximal colon and in the rectum (all p-values < 0.05) (Supplementary Table S5). In the TGF-β-PRS with only uniquely overrepresented variants for the TGF-β signaling pathway, a statistically significant interaction with red meat intake was identified for tumors in the proximal and distal colon (p-value = 0.0152 & p-value = 0.0245, respectively). Furthermore, we observed a statistically significant interaction between the TGF-β-PRS and processed meat intake for proximal colon adenocarcinomas (p-value = 0.0488) (Supplementary Table S5).

We stratified the association between red meat/processed meat intake in relation to CRC risk by quartiles of the distribution of the pPRS’s (Supplementary Table S6). Individuals in the ≤25^th percentile of the overall PRS stratum were 1.32 (95% CI = 1.21–1.44; p-value = 1.22×10–9), times as likely to be diagnosed with CRC associated with an increase red meat intake (median of servings/day from study-sex-specific quartiles), by contrast, individuals in the ≥75^th percentile of the overall PRS were only 1.17 times as likely (OR = 1.17; 95% CI = 1.07–1.28; p-value = 0.0005) (Table 1). This pattern is consistent with a quantitative interaction, whereby the direction of the association is the same across strata, but the magnitude of the association differs. Specifically, we observed that the CRC risk due to red meat intake monotonically decreases with increasing pPRS’s. Similarly, the association between red meat intake and CRC risk was stronger for individuals with lower genetic risk based on the TGF-β-PRS distribution. Concretely, individuals in the first (Q1), second (Q2), third (Q3), and fourth quartile (Q4) of the TGF-β-PRS had a 37%, 33%, 27%, and 17% higher CRC risk, respectively (Q1 OR = 1.37; 95% CI = 1.26–1.49; p-value = 1.20×10–13, Q2 OR = 1.33; 95% CI = 1.22–1.45; p-value = 4.39×10–11, Q3 OR = 1.27; 95% CI = 1.16–1.38; p-value = 5.45×10–08, Q4 OR = 1.17; 95% CI = 1.07–1.27; 4.41×10–04) (Table 1).

Table 1.

Association between red meat/processed meat intake^a with Colorectal Cancer risk stratified by quartiles of pPRS.

	pPRS Q1		pPRS Q2		pPRS Q3		pPRS Q4
Red Meat	OR	p-value	OR	p-value	OR	p-value	OR	p-value
	(95% CI)		(95% CI)		(95% CI)		(95% CI)
TGF-β-PRS	1.37	1.20E-13	1.33	4.39E-11	1.27	5.45E-08	1.17	4.41E-04
	(1.26–1.49)		(1.22–1.45)		(1.16–1.38)		(1.07–1.27)
TGF-β-PRSuniquesb	1.36	8.65E-13	1.31	8.26E-10	1.32	1.19E-10	1.36	8.64E-04
	(1.25–1.48)		(1.20–1.42)		(1.21–1.44)		(1.06–1.26)
Overall PRS	1.32	1.22E-09	1.32	1.14E-10	1.35	2.83E-12	1.17	5.02E-04
	(1.21–1.44)		(1.21–1.44)		(1.24–1.47)		(1.07–1.28)
No Pathway PRSc	1.29	3.09E-08	1.3	1.92E-09	1.33	1.99E-11	1.24	1.61E-06
	(1.18–1.41)		(1.19–1.41)		(1.23–1.45)		(1.13–1.35)
Processed Meat
TGF-β-PRS	1.39	3.68E-06	1.5	9.67E-09	1.3	1.93E-04	1.42	9.85E-07
	(1.21–1.6)		(1.31–1.73)		(1.13–1.5)		(1.23–1.63)
TGF-β-PRSuniquesb	1.47	5.56E-08	1.32	1.00E-04	1.48	2.44E-08	1.31	1.21E-04
	(1.28–1.69)		(1.15–1.52)		(1.29–1.7)		(1.14–1.51)
Overall PRS	1.34	8.40E-05	1.48	5.43E-08	1.43	5.04E-07	1.27	1.09E-03
	(1.16–1.55)		(1.28–1.70)		(1.24–1.65)		(1.10–1.47)
No Pathway PRSc	1.34	8.15E-05	1.56	5.63E-10	1.39	3.68E-06	1.27	1.05E-03
	(1.16–1.55)		(1.35–1.79)		(1.21–1.60)		(1.10–1.47)

pPRS, Pathway-based Polygenic Risk Score (PRS); Q1, First Quartile; Q2, Second Quartile; Q3, Third Quartile; Q4, Fourth Quartile; OR, Odds Ratio; CI, Confidence Interval.

^aIntake was modeled as the per-quartile increase in the median of study-specific quartiles fitted as a continuous predictor.

^bpPRS constructed with variants that are only overrepresented in the TGF-β signaling pathway.

^cPRS constructed with 174 SNPs that did not show pathway overrepresentation.

Note: All models were adjusted for study, sex, age and the first three Principal Components.

TGF-β-PRS and red meat intake interactions in CRC risk

We further interrogated the robustness of the statistically significant interaction between the TGF-β-PRS and red meat intake, upon further adjustment of the model with PRS’s that contained variants not overrepresented in the TGF-β signaling pathway, but that had been previously associated with CRC risk.

The interaction between the TGF-β-PRS and red meat intake remained significant after we included a fixed term with a PRS constructed from variants that showed no overlap with the TGF-β signaling pathway (n = 16) (p-value = 0.0040), with no change in the estimate of association for the interaction term (OR_{int =} 0.951; 95% CI = 0.919–0.984). Similarly, when we included PRS’s as model covariates for SNPs overrepresented in other pathways regardless of overlap with the TGF-β signaling pathway (n = 25), or with variants that are not overrepresented in the TGF-β signaling pathway, but are marginally associated with CRC risk (n = 190), the interaction between TGF-β-PRS and red meat intake was still significant (p -value = 0.0039 & p-value = 0.003, respectively) and the estimates of association remained consistent (OR_int = 0.951; 95% CI = 0.919–0.984 & OR_int = 0.949; 95% CI = 0.916–0.982). Furthermore, when we fitted a PRS constructed with the variants in the intersection between the TGF-β signaling-pPRS and the Gonadotropin-releasing hormone receptor-pPRS (i.e., overlap between both pathways, n = 9), no evidence of an interaction with red meat intake was observed (OR_int = 0.987; 95% CI = 0.954–1.021; p-value = 0.457), which suggests that the significance of the TGF-β-PRS interaction with red meat is driven by variants uniquely overrepresented in the TGF-β signaling pathway (Table 2). Sensitivity analyses confirmed the TGF-β-PRS interaction with red meat was not confounded by smoking (never, former, current), BMI (kg/m²), consumption of alcohol (g/day), fiber (g/day), fruits (servings/day), vegetables (servings/day), total caloric intake (kcal/day) and physical activity (MET-h/wk) (Supplementary Table S7).

Table 2.

Multiplicative interactions between TGF-β-PRS and red meat intake in association with colorectal cancer risk.

	TGF-β-PRS x Red Meat
	ORint	95% CI	P-interactiona
Model 1	0.950	0.918–0.983	0.0031
Model 2	0.951	0.919–0.984	0.0040
Model 3	0.951	0.919–0.984	0.0041
Model 4	0.949	0.916–0.982	0.0030
Model 5	0.987	0.954–1.021	0.4570

PRS, Polygenic Risk Score; OR, Odds Ratio; CI, Confidence Interval.

Model 1: Adjusted for study, sex, age and the first three Principal Components.

Model 2: Adjusted for covariates in Model 1, plus PRS that contains variants that were not overrepresented in the TGF-β pathway, but that showed statistically significant overrepresentation in other pathways (n = 16)

Model 3: Adjusted for covariates in Model 1, plus PRS that contains SNPs that do not exhibit overrepresentation for the TGF-β pathway but demonstrate overrepresentation in other pathways (n = 25).

Model 4: Adjusted for covariates in Model 1, plus a PRS constructed with all SNPs that do not showcase overrepresentation for the TGF-β pathway but have been marginally associated with the trait (n = 190).

Model 5: pPRS constructed with common variants between TGF-β, and Gonadotropin releasing hormone pathways (n = 9). Adjusted for covariates in Model 1.

^aComputed with a 1 degree-of-freedom test.

When the variants that were uniquely overrepresented in the TGF-β signaling pathway were assessed separately, we found a statistically significant interaction with red meat consumption for rs2337113 which maps to chromosome 18 in an intronic region for the gene SMAD7 (SMAD family member 7) (p-value = 0.0005 ), and for rs2208603, which maps to chromosome 6 in the intergenic region for BMP5 (bone morphogenetic protein 5) (p-value = 0.036) (Table 3). Regional plots show that the rs2337113 and rs2208603 align with variants in LD previously reported (21,48), which highlights the relevance of these regions to understand CRC risk (Supplementary Figure 1A-B). Albeit after controlling for multiple testing, only the interaction between rs2337113 and red meat consumption retained statistical significance (Bonferroni-adjusted p-interaction = 0.0025) (Table 3). Furthermore, when the five uniquely overrepresented SNPs in TGF-β signaling pathway were removed, the estimated OR_int between the overall PRS and red meat intake was 0.96 (95% CI = 0.93–1.00) but was not statistically significant (p-value = 0.0513).

Table 3.

Gene-environment interactions (GxE) between single nucleotide polymorphisms (SNP) uniquely overrepresented in the TGF-β pathway^a and red meat intake in relation to colorectal cancer risk.

								SNP x Red Meat
SNP	Chr	BP Position	Effect	Gene	Ref	Alt	Alt Allele Freq (1000G EUR)	OR (95% CI)	Raw Pinteractionb	Bonferroni Adjusted Pinteractionc
rs3809570	15	67000117	Intronic/Upstream	SMAD6	C	A	0.249	1.013 (0.958–1.071)	0.6580	1
rs2337113	18	46452327	Intronic	SMAD7	A	G	0.450	1.088 (1.037 –1.142)	0.0005	0.0025
rs2208603	6	55577214	Intergenic	BMP5	T	C	0.708	0.946 (0.898–0.996)	0.0360	0.180
rs62404966	6	55712124	Intronic	BMP5	C	T	0.245	0.962 (0.909–1.019)	0.1830	0.915
rs6912214	6	55721302	Intronic	BMP5	T	C	0.458	1.019 (0.971–1.070)	0.4430	1

SNP, single nucleotide polymorphism; Chr, chromosome; BP Position, base pair position based on GRCh37; OR, Odds Ratio; CI, Confidence Interval.

^aSNPs that do not show overlap with any other pathway, except TGF-β.

^bComputed with a 1 degree-of-freedom test.

^cAdjusted for M = 5 independent tests.

Models adjusted for study, age, sex, and the first three principal components.

DISCUSSION:

In this comprehensive analysis of more than 70,000 individuals, we report a statistically significant interaction between a PRS with red meat intake in relation to CRC risk. Furthermore, we constructed four pathway-based PRS’s which when deployed, suggested for the first time, that the significance of the PRS and red meat interaction was specifically driven by variants overrepresented in the TGF-β signaling pathway (Figure 3A).

The TGF-β signaling pathway plays a biological role in cell proliferation, differentiation, migration, and apoptosis (49). Many of these roles are context-dependent and appear paradoxical. In benign tissues, TGF-β signaling inhibits epithelial growth and cell proliferation, while promoting differentiation, and apoptosis (49,50). Classically associated with epithelial cancers, TGF-β signaling acts as a tumor suppressor via promotion of cell cycle arrest and apoptosis during the tumor initiation phase (50). However, it facilitates tumor cell proliferation, epithelial-mesenchymal transition, fibrosis, inflammation, and cellular stemness via activation of the canonical and noncanonical TGF-β signaling pathways during tumor progression (51). Moreover, autocrine TGF-β signaling activation can change the architecture of the tumor microenvironment and suppress anti-tumor immune-mediated responses, enabling cell invasion, dissemination, and therapeutic resistance (52).

In CRC, an extensive body of work has documented the involvement of TGF-β signaling at multiple stages of disease. Elevated TGF-β levels in plasma and primary tumor tissue correlate with higher tumor stage and recurrence (53,54), while its inhibition can prevent metastasis by inducing a potent cytotoxic T-cell response in murine models (55). Moreover, the consensus molecular subtype (CMS) classification of CRC identified, a distinctive transcriptional signature for TGF-β signaling (CMS-4), which conveyed the worst relapse-free survival of all 4 CMS (56,57). The potential for an interaction between the TGF-β signaling pathway with red meat intake in relation to CRC risk remains understudied. Nevertheless, it has been suggested that oxysterols and aldehydes, formed through the oxidation of saturated fat in red meat, may upregulate TGF-β1 expression in macrophages and fibroblasts in the tumor vicinity (58). This sustained exposure to elevated concentrations of TGF-β1 in the colonic mucosa could facilitate clonal proliferation. However, this theory is not specific to red meat intake, as other sources of saturated fat exist (59). Additionally, we found no evidence to suggest that the TGF-β- PRS x red meat interaction was confounded by body mass index or total caloric intake (Supplementary Table S7). Therefore, it is unlikely that our findings are solely driven by increased saturated fat intake.

CRC exhibits biological heterogeneity by tumor subsite (60). Consistent with this, we observed subsite-specific variation in the TGF-β–PRS x red meat interaction, with statistically significant associations for proximal colon and rectal cancers, but not for distal colon tumors (Supplementary Table S5). However, further studies with tumor molecular profiling and environmental exposure data are required to understand which distinct elements of the subsite-specific tumor profile may modify the relationship between meat intake and CRC risk.

It is possible, that the underlying mechanism by which the TGF-β signaling pathway predisposes to an increased CRC risk rests in downstream components. Thus, in our study, we propose that polymorphisms targeting key elements of the canonical TGF-β signaling pathway interact with red meat intake to affect CRC risk (Table 3).

The rs2337113 SNP, which maps to the SMAD7 gene, codes for an intracellular protein that is considered a negative regulator of TGF-β1 (61). Located in a region (long arm) of chromosome 18 that is commonly deleted in CRC (62,63), multiple SMAD7 polymorphisms have been associated with CRC risk in Asian and European populations (48,64–66). In addition, deletion of SMAD7 was associated with a protective dose effect in overall survival and disease-free survival in 264 tumor biopsies of colorectal cancer patients, while amplification was associated with poor prognosis (67). The involvement of the rs2337113 variant in CRC risk is supported by the work of Tian et al (68). Furthermore, this consortium previously published a genome-wide statistically significant interaction between another SMAD7 polymorphism (rs35352860, chr18:48927384, C>T) with red meat intake in relation to CRC risk (21), which showcases a strong correlation (ρ = 0.642; 95% CI = 0.638–0.646; p-value = 2.2E-16) with the rs2337113 SNP (A>G). In our previous publication, we hypothesized that inhibition of the liver-derived peptide hepcidin via overexpression of rs35352860 could prevent the internalization of the transmembrane receptor ferroportin with a concomitant increase in iron output into the bloodstream, which in turn, could facilitate free radical formation and a proinflammatory, pro-carcinogenic state. This aligns with the work by Stolfi et al. (69), which demonstrated that SMAD7 silencing inhibited the growth of CRC cell lines both in vitro and in vivo after transplantation into immunodeficient mice. Additionally, a significant increase in SMAD7 expression was observed in 14 matched pairs of CRC and adjacent tissues (69). Thus, our finding of an interaction between rs2337113 and red meat, provides additional evidence to support the potential involvement of SMAD7 in CRC risk.

The other interaction between red meat and individual SNPs in the TGF-β signaling pathway was for a variant in the 6p12.1 region (rs2208603) which resides in the intergenic region of the gene BMP5. This gene codes for a cytokine member of the bone morphogenetic proteins (BMPs)/growth differentiation factors; a protein class within the TGF-β superfamily (70). In CRC, deep sequencing of sporadic CRC samples identified loss of BMP5 as an early and CRC-specific event, while high expression correlated with longer survival. Moreover, knockdown of BMP5 in the SW480 cell line (adenocarcinoma of the colon, Dukes’ type B) promoted proliferation. Thus, suggesting its relevance as a tumor suppressor gene (71). However, mutational events are uncommon; only <4% prevalence of non-silent somatic mutations in CRC was found in TCGA samples (72). Further research is required to evaluate the potential interaction between BMP5 and red meat intake in relation to CRC. However, Xiao et al. (73) highlighted the role of BMP5 as a regulator of hepcidin and systemic iron homeostasis. In BMP5-inactive murine models, exposure to an iron-poor or high-iron diet decreased hepcidin levels in young mice. This transcriptional inactivation of hepcidin was exacerbated when concomitant BMP6 inactivation ensued, leading to hepatic and extrahepatic iron overload. Notably, the iron overload became more severe in animals exposed to a high-iron diet (73). This further supports our a priori mentioned hypothesis in which functional inhibition of hepcidin could foster a pro-carcinogenic state, increasing CRC risk.

Although previous literature has published on the association between the Gonadotropin-releasing hormone pathway, presenilin-Alzheimer disease pathway, and Cadherin/WNT signaling pathway with CRC risk (74–76), we found no evidence of a statistically significant interaction with red meat or processed meat intake on CRC risk (Figure 3A-B). However, the observed overlap of SNPs overrepresented across multiple pathways suggests crosstalk and co-regulation between biological pathways (Figure 1). Prime examples are the multilevel processes jointly regulated by the TGF-β signaling pathway and Cadherin/WNT signaling. In conjunction, these pathways control extracellular gradients of morphogens during embryonic development, regulate target gene expression at the nucleus, and protein-protein interactions in the cytoplasm. Moreover, TGF-β and WNT synergistically promote tumorigenesis (77). Thus, further research is required to elucidate the potential antagonistic or synergistic effects of multilevel, multifactor interactions across these biological pathways in relation to red meat and/or processed meat intake for CRC.

Our study has several strengths. We leveraged harmonized pooled data that underwent systematic quality control, enabling the consideration of tumor localization, study design, and potential confounders of the exposure and outcome in the putative causal pathway. Additionally, the implementation of the novel pathway-based PRS x E approach by Gauderman et al. (30) allowed us not only to detect an interaction between SNPs that previously showed a marginal association with CRC with red meat intake in a single score, but also to identify the variants driving this signal and underpin the pathway leading to increased CRC risk. This approach supports the construction of constrained PRSs with variants selected according to a priori established criteria (e.g., pathway overrepresentation) (Figure 2) and improves statistical power to detect PRSxE interactions relative to traditional methods. Consistent with this rationale, Xuechen Chen and colleagues reported no interaction between a 140-variant CRC PRS and red and processed meat intake (78), however, a conventional PRS may dilute interaction signals from exposure-relevant variants by the inclusion of a larger subset of variants that influence CRC risk via exposure-independent pathways. This gain in power may be independent of the number of variants included in the PRS, as a prior study showed no linear relationship between PRS predictive accuracy and the number of SNPs included (79). Moreover, the interactions between the overall PRS, TGF-β-PRS and TGF-β-PRS_uniques with red meat remained significant after a Bonferroni adjustment for multiple comparisons (${M=8;adjustedsignificancelevel=\alpha }_A=\frac{\alpha }{M}=\frac{0.5}{8}=0.00069)$ A key strength of our approach lies in its comprehensive nature: we integrate annotations from SnpEff, ANNOVAR, and VEP, drawing from both Ensembl and RefSeq databases to enhance the robustness of SNP-to-gene mapping. To further account for regulatory variants, we incorporated PEREGRINE-based enhancer–gene link annotations, enabling the capture of potential functional effects from non-coding SNPs. Although we recognize that incorporating study-specific eQTL or chromatin interaction data would offer an even finer resolution of regulatory effects, our multi-tool strategy substantially reduces annotation discrepancies and improves the accuracy of linking non-coding variants to their target genes. Nonetheless, we also recognize some limitations. Alternative tools and databases, such as Reactome (reactome.org) or Gene Ontology (geneontology.org), offer different pathway definitions that may not fully overlap with those in PANTHER, albeit no gold standard exists. However, when Reactome was utilized to annotate variants, the signaling by TGF-β family members pathway showed evidence of activation (FDR p-value = 0.04). In addition, different methods to explore novel GxE interactions and subsequently conduct pathway analyses have been reported (80). Nevertheless, our approach, which leverages a priori reported variants that have been marginally associated with the trait of interest, allowed us to propose a more parsimonious set of putative causal SNPs that, in conjunction with red meat intake, may be acting through similar biological processes to influence CRC risk. Predominantly, questionnaire data was utilized to construct the red and processed meat intake variables and given that many of the included studies have a case-control design, participants were asked to report on intake 1–2 years before study selection. Thus, we cannot rule out the potential for misclassification bias of the exposure and/or recall bias. In cohort studies, exposures were assessed at the time of blood draw or buccal collection at study-specific times (81). Also, we utilize the median of study-specific quartiles to evaluate meat intake, which does not account for absolute differences and may miss non-linear dose-response relationships. However, this approach is a valid method in pooled analysis of nutritional exposures (82). Furthermore, total caloric intake (kcal/day) was not reported in four studies and set to zero (ASTERISK, DACHS, PHS, UKB). Albeit sensitivity analyses confirmed that, adjustment with a fixed indicator for study in the interaction models, prevented those four studies to contribute information for estimating the energy effect. However, they did contribute to the estimation of other effects and the pPRSxE interaction term. In addition, the interaction between TGF-β-PRS and red meat was not confounded by total caloric intake (kcal/day). Also, we recognize that additional unmeasured confounders may still impact estimates of association. Finally, our findings may not be generalizable to other racial and ethnic populations besides those of European ancestry, which contributed to the individual studies.

CONCLUSION:

In summary, this pathway-based interaction analysis proposes, for the first time, that a statistical interaction between SNPs overrepresented in the TGF-β signaling pathway and red meat intake impacts CRC risk. This finding provides a potential mechanistic explanation linking red meat consumption and CRC risk and highlights the importance of dietary guidance and personalized prevention strategies.

Impact statement:

In this work, we developed pathway-based Polygenic Risk Scores which, for the first time, suggested that red meat intake interacts with variants overrepresented in TGF-β signaling pathway to impact colorectal cancer risk.