Association of germline variants with KRAS-mutation status in colorectal cancer

Abstract

Somatic mutations in KRAS are a common driver of colorectal cancer (CRC) and present at different frequencies by race, sex, tumor site, ethnicity, and genetic similarity. Inherited germline variants may influence tumor somatic mutation frequency by altering mutation or DNA repair processes or altering cellular, immunological and/or microenvironmental responses after a mutation. We hypothesized that the germline genetic background modifies somatic KRAS mutation frequency in CRC. To test this, we performed a genome-wide association study (GWAS) in 7071 individuals with CRC, using KRAS mutation status as the phenotype. Single-nucleotide variants were chosen for validation analyses based on P values from the discovery GWAS, predicted in silico functional effects, and proximity to genes with potential cancer relevance. A validation analysis of 101 SNVs of interest was performed in 2482 individuals. No SNVs were significantly associated with KRAS-mutant CRC (P value < 0.0005). One variant rs73067863-T showed a non-significant exploratory association with fewer KRAS-mutant tumors in the combined sample (P value = 9.7 × 10^–7, OR = 0.75). Follow-up studies are needed to determine if these or other germline variants impact population differences in KRAS mutations in CRC.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-39644-8.

Introduction

Somatic mutations in KRAS are a common driver observed in 30–50% of colorectal cancers (CRC), causing extended tumor cell survival, elevated proliferation, altered metabolism, and metastasis through multiple downstream effector pathways¹. CRCs with KRAS mutations are more likely to metastasize², show decreased overall survival³, and be resistant to EGFR inhibitors⁴. Although CRCs with KRAS mutations were once considered undruggable, newer therapeutics that directly inhibit mutant KRAS have shown some success in combination with other drugs in treating CRC^5,6.

Multiple studies have evaluated the frequency of somatic mutations in populations of different self-reported race/ethnicity (e.g. Black) and different genetic similarity (e.g. genetic ancestry). KRAS mutations in CRC are more common in Black individuals compared to Hispanic and non-Hispanic White individuals. In a study of KRAS mutations by self-reported race, non-Hispanic White individuals were 36% less likely to have KRAS-mutant CRC tumors compared to non-Hispanic Black individuals⁷. By genetic similarity, a cohort study of 18,761 individuals with CRC and available KRAS mutation data found that individuals with genetic similarity adjacent to African reference populations (AfA) were significantly more likely to have KRAS mutations compared to individuals with genetic similarity adjacent to European reference populations (EuA)⁸. KRAS mutations in CRC also vary in frequency by sex (more common in females⁹) and tumor location (more common in proximal CRCs¹⁰). Biological mechanisms resulting in the observed associations between KRAS mutations in CRC and race, genetic similarity, sex, and tumor location remain unknown.

Although most somatic driver mutations in cancer occur by chance, some inherited germline variants are associated with tumor mutation burden (TMB) and mutational signatures. For example, Lynch Syndrome is driven by pathogenic germline variants in DNA mismatch repair genes, leading to CRC tumors with high TMB and high microsatellite instability (MSI-H)¹¹. After a somatic mutation event, germline effects on the cellular microenvironment and immune activity can influence whether the cell with the mutation continues to divide and whether it progresses to become a tumor^12,13. A pan-cancer study of human leukocyte antigen (HLA) variants identified specific HLA genotypes that were predictive of tumor oncogenic mutations; this observation was likely driven by neoantigens and inherited differences in antigen presentation and immune destruction of cells after a mutation^12,13. With the increasing availability of paired germline genomic and somatic mutation data, identification of interactions between germline variants and specific somatic mutations is now possible. Indeed, associations between germline variants and somatic TP53 mutations in breast cancer have been identified¹⁴, as well as associations between local genetic similarity and EGFR and KRAS mutations in lung cancers¹⁵, and other somatic mutations and cancers^12,13,16,17.

The objective of our study was to identify inherited variants associated with somatic KRAS mutations in CRC. We hypothesized that germline variants associate with KRAS-mutant CRC and that the associated germline variants may have higher risk allele frequency in individuals of AfA or higher effect size compared to individuals of EuA. To identify associated variants, we performed a discovery analysis in individuals of EuA and a multi-ancestry validation analysis. Studying the mechanisms leading to these associations may lead to a better understanding of the events after the occurrence of a somatic mutation that lead to tumor development; this may lead to new biological insights to prevent or treat CRC.

Results

Genome-wide discovery analysis

From the discovery datasets, 7071 individuals were defined as EuA through principal component analysis (PCA) of germline genotyping (Fig. 1A). The dataset included 3328 females and 3743 males (Supplemental Table 1). Individuals with KRAS mutation status and non-EuA genetic similarity were saved for validation analyses and combined with additional validation samples. PCA for genetic similarity in the validation analysis is shown (Fig. 1B). In EuA individuals included in the discovery analysis, KRAS mutations were present in 33% of tumors (Table 1). After variant filtering of 39,128,432 SNVS, 6,861,047 SNVs were included in the genome-wide analyses. No variants were associated with KRAS tumor mutations meeting a genome-wide significance level (P value < 5 × 10^–8) (Fig. 2). Three SNVs had P values < 1 × 10^–6 and 50 had P values 1 × 10^–6–1 × 10^–5 for association with KRAS mutation status (Supplemental Table 2).

Fig. 1.

PCA for genetic similarity analysis. Principal component analysis was performed to approximate genetic similarity to reference populations. (A) PCA was performed from genome-wide data in 10,124 individuals to identify individuals of EuA for discovery analyses (blue). (B) PCA was performed for subsequent validation analyses using genotypes from a targeted ancestry genotyping array of 95 SNVs in 2482 individuals to identify individuals of EuA and AfA. Self-reported race/ethnicity is indicated by color (Asian, red; African American/Black, green; Latino, light blue; South Asian, purple; and European/White, dark blue), and individuals without self-reported race/ethnicity information are in black. PC, principal component; EuA, European reference population adjacent; AfA, African reference population adjacent; PCA, principal component analysis; SNV, single nucleotide variant.

Table 1.

Study demographics and KRAS mutation status.

	Discovery dataset		Validation dataset		Combined data		P value
	KRAS-mut (%)	Total n	KRAS-mut (%)	Total n	KRAS-mut (%)	Total n
All individuals	2352 (33)	7071	858 (35)	2482	3210 (34)	9553
AfA	NA	NA	73 (42)	172	73 (42)**	172	0.02**
EuA	2352 (33)	7071	702 (33)	2109	3054 (33)**	9180
H	NA	NA	43 (46)	94	43 (46)	94
Other ancestry	NA	NA	40 (37)	107	40 (37)	107
Females	1099 (33)	3328	411 (36)	1142	1510 (34)	4470
Males	1253 (33)	3743	447 (33)	1340	1700 (33)	5083
Proximal CRC	850 (38)	2252	281 (45)	629	1131 (39)*†	2881
Distal CRC	825 (31)	2652	228 (31)	742	1053 (31)*	3394	< 0.0001*
Rectal CRC	480 (32)	1516	174 (30)	587	654 (31)†	2103	< 0.0001†
Transverse colon	113 (27)	414	33 (32)	104	146 (28)	518
Unspecified site	84 (35)	237	142 (34)	420	226 (34)	657

Fisher’s exact test was used to compare KRAS mutations by proximal versus distal tumor location (*) and proximal versus rectal tumor location  (†), as well as KRAS mutations in AfA individuals versus EuA individuals (**).. KRAS-mut, Tumor with KRAS mutation; CRC, colorectal cancer; EuA, European reference population adjacent; AfA African reference population adjacent; H, Hispanic reference population adjacent. N, number; %, percent.

Fig. 2.

Manhattan plot of discovery GWAS for CRC tumor KRAS mutation status. Discovery GWAS data comparing individuals of EuA with CRC and a KRAS mutation (n = 2352) and individuals without a KRAS mutation (n = 4719) are plotted by − log₁₀(P value). Chromosome numbers are indicated. The dotted line indicates genome-wide significance (P value < 5 × 10^–8). GWAS, genome-wide association study; CRC, colorectal cancer; EuA, European reference population adjacent.

KRAS mutation status by genetic similarity and tumor site

Consistent with other reports, we found KRAS mutations to be significantly more common in tumors from AfA individuals (42%) compared to EuA individuals (33%) (P value = 0.02) (Table 1). KRAS mutations were also more common in proximal colon tumors (39%) compared to distal colon and rectal tumors (31%) (P value < 0.0001) (Table 1). No significant differences in KRAS mutation were observed by sex. Specific amino acid changes were described in 630 of 3210 individuals with KRAS mutations, including 132 in the discovery and 498 in the validation analyses. The most common mutations were p.G12D (34%) and p.G12V (25%) (Supplemental Table 3).

A total of 101 SNVs were prioritized for validation in a set of individuals of multiple ancestries with CRC (Supplemental Table 4) as described through a workflow chart (Fig. 3A) and sectioned Manhattan plot (Fig. 3B). The prioritized SNVs included 20 SNVs with sex-specific risk effects, 15 SNVs with tumor location-specific risk effects, and 6 SNVs with both sex and location-specific risk effects (P values < 1 × 10^–4 for specific covariate). Validation analysis of 101 SNVs was performed in 2482 individuals (Supplemental Table 5). These included 1127 individuals with targeted array genotypes that passed overall quality control (QC) standards (Supplemental Table 6). Validation studies included individuals with genome-wide genotype data who did not have EuA by PCA in the discovery analysis or only had genotype or KRAS mutation data available after initial discovery analyses. The availability of genotyping data differed by SNV based on source or array QC. Using PCA to measure genetic similarly, we defined 172 individuals as AfA, 2109 individuals as EuA, 94 individuals as having genetic similarity adjacent to a Hispanic reference population (H), and 107 as undetermined (Table 1). The H cluster showed high admixture. Findings in AfA and H groups were considered exploratory due to small sample sizes.

Fig. 3.

Workflow for identifying variants for validation analysis. The filtering steps for identifying SNVs to include in the validation analysis from the discovery analysis results are depicted as a workflow (A) and as sections of the graphed discovery data (B). Filtering steps included P value for association, ORs, MAF, redundancy as determined by linkage disequilibrium, and in silico and other functional data. A final determination was whether a Fluidigm genotyping assay could be designed for the SNV. SNV, single nucleotide variant; MAF, minor allele frequency; EuA, European reference population adjacent; AfA, African reference population adjacent; LD, linkage disequilibrium; OR, odds ratio.

None of the SNVs in the multi-ancestry validation analyses were significantly associated with somatic KRAS mutation status after Bonferroni adjustment for multiple comparisons of 101 SNVs (P value < 0.0005). A heterogeneity analysis identified four validation SNVs (rs2298437, rs74030413, r116195080, rs726800) with evidence of heterogeneity (Cochran’s Q < 0.05) by genetic similarity groups. Two SNVs showed suggestive associations of interest (P value < 0.005) (Table 2). This included rs73067863 in EuA (P value = 0.0016, OR 0.67) and combined (P value = 0.0036, OR 0.73) analyses. rs2298437, which showed evidence of heterogeneity by genetic similarity groups (Cochran’s Q < 0.003) showed evidence of association in AfA (P value = 0.0014, OR 0.4) but not other populations. In a combined analysis of both discovery and validation samples (n = 9553) neither rs73067863 (P value = 9.7 × 10^–7, OR = 0.75) nor rs2298437 (P value = 7.9 × 10^–4, OR = 1.11) were associated with KRAS mutations at genome-wide significance (P value < 5 × 10^–8). Additional SNVs with P values < 0.05 were identified, including one observed in the combined analysis, six in AfA individuals and one in H individuals (Table 2). Ten SNVs that were incompatible with array design were analyzed for association with tumor KRAS mutation using existing genome-wide data (n = 1274 individuals). From this analysis, three SNVs showed P values < 0.05, including one in AfA individuals and two of an SNV in LD with rs73067863. One SNV had a P value < 0.05 by tumor location. None of the SNVs tested showed sex-specific associations with tumor KRAS mutation.

Table 2.

SNVs with evidence of association in combined and validation analyses.

SNV	Chr	Position	Ref	Alt	Pop	N (KRAS-mut)	OR (95% CI)	P value	Data type
rs73067863	3	36489472	C	T	All	2299 (793)	0.726 (0.584–0.898)	0.004	Combined
					EuA	2019 (676)	0.672 (0.523–0.856)	0.002	Combined
rs2298437	21	30496968	C	T	AfA	160 (70)	0.399 (0.221–0.687)	0.001	Combined
rs16966448	16	9830940	C	T	All	2252 (776)	0.868 (0.762–0.989)	0.033	Combined
rs116195080	3	141700851	A	G	AfA	148 (69)	0.306 (0.107–0.811)	0.021	Combined
rs197110	1	57183103	G	T	AfA	151 (64)	0.478 (0.242–0.893)	0.026	Combined
rs74030413	16	80740298	C	T	AfA	160 (71)	0.444 (0.209–0.905)	0.028	Combined
rs726800	17	13597287	A	G	AfA	140 (62)	1.833 (1.076–3.214)	0.029	Combined
rs9892103	17	13563369	A	G	AfA	142 (63)	0.486 (0.236–0.953)	0.041	Combined
rs8131342**	21	30501326	A	G	AfA	162 (70)	0.601 (0.359–0.984)	0.047	Combined
rs4798561	18	7366503	A	G	Latino/Admixed	91 (42)	2.607 (1.277–5.742)	0.012	Combined
rs1814562*	3	36453242	C	T	All	1246 (403)	0.661 (0.487–0.886)	0.007	Genomic
					EuA	1097 (348)	0.599 (0.425–0.832)	0.003	Genomic
rs522098	11	116455969	A	G	AfA	71 (26)	0.06 (0.003–0.408)	0.014	Genomic

*in LD with rs73067863. **LD (R^2 = 0.61) with rs2298437. † Not on validation array, GWAS data only; SNV, single nucleotide variant; Chr, chromosome; GRCh38, Human genome assembly GRCH38 position, Ref, reference allele; Alt, alternate allele; Pop, population; N, number; EuA, European reference population adjacent; AfA, African reference population adjacent; KRAS-mut, KRAS-mutant tumor; OR, odds ratio; CI, confidence interval; FDR, false discovery rate; LD, linkage disequilibrium; Array, targeted genotyping, GWAS, genome-wide association study array.

.

Discussion

Here, we describe a GWAS to identify germline variants that associate with the presence of somatic KRAS mutations in CRC with the aim of determining if germline variants may explain population differences in KRAS mutation frequency. Although this was a study of over 7000 individuals, no variants were identified that met genome-wide significance for association with KRAS-mutation status in CRCs. Two variants of interest were identified, which do not meet genome-wide significance and should be interpreted with caution as they could be false positive findings.

rs73067863 is located in intron 5 of STAC (SH3 and cysteine-rich domain)^18,19. This SNV has been identified as an eQTL of MLH1 in GTEx Version 8, though the more recent version of GTEx, Version 10, does not indicate any eQTL function of this SNV or any other SNV in linkage disequilibrium with rs73067863²⁰. We found that the rs73067863-T allele was associated with lower likelihood of a KRAS mutation with an allele frequency of 0.07 in individuals with KRAS-mutant CRC compared to 0.10 in those with KRAS-WT CRC. rs73067863 alleles differ in frequency by genetic similarity groups, but there was no evidence of heterogeneity (Cochran’s Q = 0.37). The T-allele has a higher frequency in individuals of EAsA (0.29) compared to EuA (0.09) and AfA (0.10)¹⁸. EAsA individuals were not included as a separate study population due to low numbers. Thus, future multi-ancestry studies of individuals with CRC are needed to determine if rs73067863-T is similarly associated with KRAS-WT CRC in EAsA or other populations not evaluated here.

No consistent predicted functional effects of rs73067863 were identified through in silico analyses. Additional in vitro studies such as DNA accessibility, DNA protein binding, splicing, and local methylation analyses are needed to determine if this SNV impacts expression or regulation of MLH1 or another nearby gene. Of potential importance for result interpretation, we excluded individuals with known MSI-H tumors in our analyses for multiple reasons. These include the high overall somatic mutation rates in MSI-H tumors due to defective DNA repair which could mask effects of low impact variants on KRAS mutations, the low overall frequency of MSI-H CRCs (~ 15%), and the small proportion of MSI-H tumors with KRAS mutations (~ 5–14%)^21,22. MSI-H tumors can be driven by deficiency of MLH1 which is essential to DNA mismatch repair^23,24. MLH1 deficiency and MSI-H status are more common in KRAS-WT CRC compared to KRAS-mutant CRC²⁴. Further evaluation of this locus and studies which include MSI-H CRCs resulting from MLH1 methylation or germline/somatic mutations are needed to determine if rs73067863 associates with MSI in KRAS-WT CRC and/or alters MLH1 protein expression.

rs2298437 was not significantly associated with KRAS mutations in a combined analysis (OR = 1.1, raw P value = 7.9 × 10^–4) but showed a suggestive protective effect of rs2298437-T for KRAS-mutant CRC only observed in an exploratory analysis of data from AfA individuals (OR = 0.4, raw P value = 0.0014). The rs2298437-C allele is more common in AfA individuals (MAF = 0.76) compared to EuA individuals (MAF = 0.57)¹⁸. This SNV encodes a p.Tyr48Cys missense variation in KRTAP19-4, a keratin-associated protein^18,19. Little is known about the functions of KRTAP19-4, but it is predominantly expressed in skin and hair²⁰. Interestingly, the effect may differ by genetic similarity because an opposite directional effect was observed in EuA individuals in the discovery and validation analyses (OR = 1.15, raw P value = 8.7 × 10^–5 and OR = 1.08, raw P value = 0.23, respectively) compared to AfA individuals (OR = 0.4); the variant showed significant evidence of heterogeneity by genetic similarity groups using Cochran’s Q analysis. Risk alleles have previously been identified that have different frequencies by genetic similarity to reference populations or that only convey risk in individuals of certain ancestries²⁵. However, the AfA sample size was small, so caution in overinterpretation of these findings should be taken. Further studies of rs2298437 in a larger sample of AfA individuals with CRC are needed to determine if the observed association and evidence of heterogeneity by genetic similarity groups is genuine.

To consider the larger context of race and colorectal cancer, Black individuals have 20% greater CRC incidence and 31–44% greater mortality compared to Non-Hispanic White individuals²⁶. In the United States, current CRC screening rates are similar in Blacks and Non-Hispanic Whites, (65% vs 68%, respectively), though historically, screening differences have contributed to an estimated 19% of CRC mortality disparities between groups^26,27. The effects of social determinants of health and systemic racism are likely contributors to the observed elevated CRC mortality in Black individuals^26,28. Nonetheless, the observed differences in KRAS mutation by genetic similarity and race remain unexplained.

It is possible that the higher KRAS mutation frequency in AfA individuals with CRC is driven by germline genomic effects that were not measured by this study, such as copy number variants, rare SNVs, genomic regions not included on microarrays, or regions difficult to genotype (such as the HLA-locus). It is also possible that there may be variants associated with KRAS mutation status unique to tumor subtype that could have been missed in this study. Studies within tumor subtype may provide additional insights. The discovery GWAS analysis was conducted only in EuA individuals as the numbers of individuals of other genetic similarity groups were too small to split into discovery and validation groups, and we wanted to maximize diversity in the validation analyses. Some variants important for KRAS mutation may only be present or enriched in certain populations and these would have been missed as they were not included in our validation study. This study was likely underpowered to identify associations with low effect size and those enriched in non-EuA populations.

We did not attempt to identify germline variants associated with specific mutational changes of KRAS or codon sites of mutation. This type of association is plausible if a germline variant increases risk of a mutational signature or alters the immune response to a specific neoantigen. However, we did not have access to specific mutational changes for many of the contributing studies and, as such, the overall sample size was underpowered for mutation or codon-specific association studies. The contributing studies also used different KRAS genotyping methods; it is possible that rare oncogenic mutations may have been missed, leading to some individuals being misclassified as wildtype for KRAS. Associations between germline variants and specific mutational changes of KRAS may be identified in future studies with a larger sample.

Somatic mutations in BRAF may be driven by germline effects, which could indirectly affect rates of KRAS mutations. Somatic mutations in BRAF are twice as likely in non-Hispanic White individuals with CRC (12%) compared to non-Hispanic Black (6%) or Hispanic/Latino (7%) individuals and tend to be mutually exclusive with KRAS mutations⁹. We did not include/exclude participants based on tumor BRAF mutations because many of our participants did not have this data, and BRAF-mutant tumors are more likely to be MSI²², which are excluded. Future studies with consideration for BRAF mutations may clarify the interactions between KRAS, BRAF, and germline variants.

This was a study of genetic factors. Non-genetic effects may contribute to the observed population differences in KRAS mutation frequency. One possible non-genetic driver is tumor site location, which also varies by population²⁹. KRAS mutations are more common in proximal CRC and are highest in cecal tumors among all colorectal cancers^10,30,31. Another possible driver is gene-environment interactions between KRAS mutations and social determinants of health, such as diet³². Low levels of plasma adiponectin (an adipokine inversely linked with obesity and excess energy balance) are associated with increased incidence of KRAS-mutated (but not KRAS-wild-type) colorectal cancer^33,34. Associations between neighborhood deprivation index and tumor genomics have been identified, such as epigenetic alterations in breast cancer³⁵ and KRAS mutations in non-small cell lung cancer³⁶, though these studies were based on small sample sizes. Improving enrollment of underrepresented populations in genomic studies and considering the effects of social determinants of health should be high priorities for cancer genomics research.

Conclusion

We conducted a GWAS of > 7000 individuals to identify inherited variants associated with KRAS mutations in CRC. No variants were found to be associated with KRAS mutation status at genome-wide significance. One variant, rs73067863, showed suggestive evidence of association in meta-analyses and ancestry-specific analyses and therefore warrants additional studies. It remains unknown if and how germline variants influence KRAS mutations in CRC and their impact on observed population differences in KRAS mutation frequency in CRC.

Methods

Study participants

All participants in the discovery GWAS (n = 7071) (Supplemental Table 1, Supplemental “Methods”) or validation studies (n = 2482) (Supplemental Table 5, Supplemental Methods) were diagnosed with invasive CRC. Individuals were eligible if they had existing germline genome-wide genotyping data or germline DNA for genotyping as well as tumor KRAS mutation status or a tumor DNA sample available for KRAS mutational analysis. Individuals with a hereditary CRC syndrome, such as familial adenomatous polyposis or Lynch syndrome, were excluded, as somatic events may be affected by the germline pathogenic variant. Due to high somatic mutation burden and low rates of concurrent KRAS mutation, individuals with MSI-H tumors were also not included. Individuals with MSI-H tumors were also excluded because a subset of these tumors are found in individuals with Lynch syndrome which could have other mechanisms for selection of tumors with KRAS somatic mutations. We did not exclude individuals with POLE, POLD, or MUTYH somatic or germline pathogenic variants as these data were not available for most individuals, and the incidence of individuals germline pathogenic variants in these genes was expected to be small (< 1%)³⁷. DNA extraction and sequencing to identify KRAS mutations are as described (Supplemental “Methods”).

Germline genomic data analysis

All analyses of genomic datasets were performed using PLINK (version 1.9, RRID:SCR_001757) (Purcell Lab, available at: https://www.cog-genomics.org/plink/)³⁸. For the discovery analysis, participant genetic similarity was categorized through PCA. Participants with self-reported White, Black, or Asian race and reference samples from the 1000 Genomes Project were used to identify clusters for genetic similarity analysis³⁹. All clusters available at the time of analysis were included, namely Yoruba in Ibadan (YRI) to estimate AfA, Northern Europeans from Utah (CEU) to estimate EuA, and Japanese in Tokyo (JPT) and Han Chinese in Beijing (CHB) to estimate similarity adjacent to East Asian reference populations (EAsA). SNVs in the reference samples were compared pairwise by similarity clusters and minor allele frequency (MAF). The 5000 most cluster-specific variants were compiled and used for genetic similarity PCA. The centroid for each of 3 groups was computed. Individuals were categorized into a cluster if they were within two standard deviations (Euclidean distance) of the centroid of that group. All who were not categorized into any clusters were classified as Admixed. For individuals lacking genome-wide data, genetic similarity was determined using genotypes from 95 population-specific SNVs (Supplemental Tables 7–8) identified from our PCA. Genetic similarity analysis was completed in 10,124 individuals considered for discovery analysis to identify participants of EuA.

In individuals of EuA, we evaluated the association of 39,128,432 germline SNVs with tumor KRAS mutation status (yes or no) using logistic regression. SNVs were included if they were biallelic, had a minor allele frequency (MAF) ≥ 1%, had an imputation score (info/R²) > 0.4, and data missingness < 10%⁴⁰. We performed log-additive analyses for each SNV and KRAS mutation status, considering study of origin as a covariable. A stratified analysis of the discovery data was performed to identify SNVs that were associated with tumor KRAS mutation and had a different effect by patient sex or tumor location in the proximal colon, distal colon, or rectum.

Prioritization of variants for validation analysis

Variants of interest from the discovery analysis were prioritized for validation studies (Fig. 3A and B). In brief, all SNVs associated with tumor KRAS mutation with P values < 10^–4 and a MAF ≥ 10% using GnomAD (version 3.0) (Broad Institute, available at: https://gnomad.broadinstitute.org/) were considered for inclusion (n = 472)⁴¹. These SNVs were further evaluated for linkage disequilibrium (LD) using LDLink software and the SNPClip tool (version 5.1) (National Cancer Institute, available at: https://ldlink.nih.gov/) for high-throughput evaluation⁴².

All SNVs that had a P value < 10^–5 for association and met MAF criteria were considered high priority, and 1–2 SNVs per LD locus were included for validation (n = 21) (Supplemental Table 4). From the LD loci with high association, in silico findings were used to select SNVs for validation analysis that were most likely to alter transcription of nearby genes, including identifying overlap with previous GWAS findings in the NHGRI-EBI GWAS Catalog (version 1.0) (NHGRI-EBI, available at: https://www.ebi.ac.uk/gwas/)⁴³. SNVs within 50 kb of genes were identified and prioritized using dbSNP (National Library of Medicine, available at: https://www.ncbi.nlm.nih.gov/snp/)⁴⁴. We also prioritized SNVs based on predicted regulatory functions. The SNVs were evaluated for overlap with expression quantitative trait loci (eQTLs) identified in Gene Tissue Expression (GTEx) (version 8) (Broad Institute, available at: https://gtexportal.org/home/) transverse and sigmoid colon datasets and BarcUVa-Seq, using the CoTrEx 2.0 browser (version 2.0) (CoTrEx, available at: https://barcuvaseq.org/cotrex/)^20,45,46. We prioritized SNVs that were possible eQTLs (nominal P value < 0.05) through any of these databases. Additional predicted regulatory functions were identified from Roadmap and ENCODE epigenomic data using three browsers; RegulomeDB included eQTL and motif data (version 2.2) (Stanford University, available at: https://regulomedb.org/regulome-search/), HaploReg validated eQTL and GWAS hits (version 4.2) (Broad Institute, available at: https://pubs.broadinstitute.org/mammals/haploreg/haploreg.php), and the Epilogos browser included additional published epigenomic data (version 1.0) (Altius, available at: https://epilogos.net/)^47–49. Variants were prioritized if they were ranked 5 or lower in RegulomeDB with a chromatin accessibility peak or transcription factor binding. Variants that were within 25 kb of an enhancer, transcription start site, or showed active transcription in any tissue in Epilogos were also prioritized. We also prioritized SNVs with suggestive association in analyses within sex or tumor location (P values < 10^–4).

SNVs with P values ranging from 10^–5 to 10^–4 that met MAF criteria received lower priority and were included for validation only if there were one or more likely in silico findings for potential biological significance. This potential biological significance was evaluated using the same tools previously described to prioritize SNVs in LD and we included 1–2 SNVs per LD locus.

Multi-ancestry validation analysis

Additional genome-wide genotyping data came from multiple studies through the Cancer Genome Atlas (TCGA), the Genetics and Epidemiology of Colorectal Cancer Consortium (GECCO) (n = 370) the Ohio Colorectal Cancer Prevention Initiative (OCCPI) (n = 904), and the Hispanic Colorectal Cancer Study (n = 81)^50–52. Individuals were grouped by genetic similarity based on PCA of ancestry array SNVs as previously described. To identify H participants, an additional 901 samples were used for PCA, but were not included in the validation due to lack of KRAS mutation status. We did not perform population-specific analyses for self-reported Asian or South Asian individuals due to small numbers (n = 19).

For targeted germline genotyping of 1127 individuals for whom genotype data was not previously generated (Supplemental Table 5), two custom real-time PCR arrays were used: one array with ancestry-specific SNVs comparable to another previously described¹⁴ and a second array with KRAS-mutation associated SNVs identified from the discovery. If SNVs were incompatible with the genotyping array design or failed all array experiments, SNVs were replaced with others in LD. SNVs that could not be directly replaced were analyzed only in samples with genome-wide data. Genotyping arrays were completed by GSR using a Biomark HD high-throughput real-time PCR system on a Fluidigm 96 × 96 array. Array quality control was performed using methods previously described¹⁴.

To determine if any SNVs chosen for validation studies associated with tumor KRAS mutations, we used logistic regression analysis using adjustment for genetic similarity, study, sex, age, and tumor location. After Bonferroni adjustment for multiple comparisons, P values < 0.0005 were considered significant. For SNVs with sex-specific or side-specific risk effects in the discovery analysis (OR > 1.2 or < 0.7, P value < 1 × 10^–4), we repeated analysis of potential sex-specific effects or side-specific effects by colon location in the validation analysis.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Nijole Pollock Tjader: conceptualization, funding acquisition, investigation, methodology, validation, visualization, writing—original draft, writing—reviewing. Johnny Ramroop: investigation, methodology, writing—review and editing. Tanish Gandhi: investigation, writing—review and editing. Cara Dauch: investigation, writing—review and editing. Owen Meadows: investigation, writing—review and editing. Patrick Stevens: data curation, writing—reviewing and editing. Rachel Pearlman: resources, writing—review and editing. Heather Hampel: resources, writing—review and editing. Elom K Aglago: resources, writing—review and editing. Sonja I Berndt: resources, writing—review and editing. Amanda Bloomer: resources, writing—review and editing. Hermann Brenner: resources, writing—review and editing. Daniel D Buchanan: resources, writing—review and editing. Peter T Campbell: resources, writing—review and editing. Yin Cao: resources, writing—review and editing. Andrew T Chan: data curation, resources, supervision, writing—review and editing. Iona Cheng: resources, writing—review and editing. Niki Dimou: resources, writing—review and editing. David A Drew: resources, writing—review and editing. Amy J French: methodology, resources, writing—review and editing. Peter Georgeson: writing—review and editing. Marios Giannakis: resources, writing—review and editing. Graham G Giles: resources, writing—review and editing. Maria Gomez: resources, writing—review and editing. Stephen B Gruber: data curation, resources, writing—review and editing. Michael Hoffmeister: funding acquisition, data curation, resources, supervision, writing—review and editing. Wen—Yi Huang: resources, writing—review and editing. Meredith AJ Hullar: writing—review and editing. Jeroen R Huyghe: resources, writing—review and editing. Nicole Loroña: writing—review and editing. Victor Moreno: resources, writing—review and editing. Christina C Newton: writing—review and editing. Jonathan A Nowak: resources, writing—review and editing. Mireia Obón-Santacana: resources, writing—review and editing. Shuji Ogino: resources, writing—review and editing. Andrew Pellatt: resources, writing—review and editing. Anita R Peoples: resources, writing—review and editing. Jennifer B Permuth: resources, writing—review and editing. Stephanie L Schmit: resources, writing—review and editing. Robert E Schoen: resources, writing—review and editing. Erin M Siegel: data curation, funding acquisition, resources, writing—review and editing. Robert S Steinfelder: resources, writing—review and editing. Wei Sun: resources, writing—review and editing. Jamie K Teer: data curation, validation, writing—review and editing. Claire E Thomas: resources, writing—review and editing. Quang M Trinh: data curation, resources, writing—review and editing. Konstantinos Tsilidis: resources, writing—review and editing. Tomotaka Ugai: resources, writing—review and editing. Caroline Y Um: resources, writing—review and editing. Bethany Van Guelpen: resources, writing—review and editing. Syed H Zaidi: resources, writing—review and editing. Jane Figueiredo: data curation, funding acquisition, resources, writing—review and editing. Ulrike Peters: data curation, funding acquisition, project administration, resources, supervision, writing—review and editing. Amanda I Phipps: project administration, writing—review and editing. Joseph Paul McElroy: conceptualization, data curation, formal analysis, writing—review and editing. Amanda Ewart Toland: conceptualization, supervision, funding acquisition, investigation, visualization, methodology, writing—review and editing, project administration.

Data availability

Data generated or analyzed during this study are included in this published article in Supplementary Tables, in dbGAP and/or the following data repositories as listed below. The Cancer Genome Atlas (TCGA), HPFS, NHS, and some of the GECCO somatic mutation data and germline genotyping data are available in dbGAP under accession numbers (phs000178.v11.p8, phs001111.v1.p1, phs000722.v3.p2, phs002050.v1.p1 [since updated to phs002050.v2.p1]). Access to other data from international sites can be facilitated by the corresponding author upon request.

Declarations

Competing interests

Heather Hampel consults for LynSight, Exact Sciences, Genome Medical, GI OnDemand, and Carelon, and holds stock/stock options for Genome Medical and GI OnDemand. Jonathan Nowak receives research support from Natera and is a consultant for Leica Biosystems. Shuji Ogino served as a consultant for Sanofi Pasteur S.A. Ulrike Peters was a consultant with AbbVie and her husband holds individual stocks for the following companies: Amazon, ARM Holdings PLC, BioNTech, BYD Company Limited, Crowdstrike Holdings Inc, CureVac, Google/Alphabet, Microsoft Corp, NVIDIA Corp, and Stellantis. The other authors declare no conflicts of interest.