Assessment of genetic susceptibility to multiple primary cancers through whole-exome sequencing in two large multi-ancestry studies

Our study included ancestrally diverse individuals with multiple primary cancers or no cancer from two large study populations, the Kaiser Permanente Research Bank (KPRB) [10] and the UK Biobank (UKB) [11]. It additionally included individuals with a single cancer in the UKB study population only. From the KPRB, we included individuals who were previously genotyped through the Research Program on Genes, Environment and Health (RPGEH) and the ProHealth Study. For the UKB, we specifically studied participants from the 200K release of WES data [11, 12].

For both study populations, ascertainment of cancer diagnoses has been previously described [7, 13]. Both studies included prevalent and/or incident diagnoses of malignant, borderline, and in situ primary tumors [13]. ICD codes indicating non-melanoma skin cancer or metastatic cancer were not considered primary tumors. Cancers were primarily defined according to the SEER site recode paradigm [14]. However, for hematologic cancers, we incorporated morphology following WHO classifications [15], placing cancers into three major subtypes: lymphoid neoplasms, myeloid neoplasms, and NK- and T-cell neoplasms (Additional file 1: Table S1). Cases were individuals with ICD-9 or ICD-10 codes for primary tumors at two or more distinct organ sites. In the KPRB, controls without a cancer diagnosis, at the last follow-up, were matched 1:1 to cases on age at specimen collection, sex, genotyping array, and reagent kit. In the UKB, controls included all individuals without a cancer diagnosis at the last follow-up.

In both study populations, we excluded duplicates/twins and first-degree relatives, retaining the individual from each related pair who had higher coverage at targeted sites. Following quality control (QC) of WES data (described below), the KPRB and UKB study populations used in this project included 3111 and 3318 cases with multiple primary cancers and 3136 and 162,717 cancer-free controls, respectively. The UKB also contributed 29,091 individuals with a single cancer diagnosis. While our study was primarily unselected for cancer type, prostate cancer cases were oversampled in the KPRB due to the inclusion of individuals from the ProHealth Study.

Genetic ancestry was defined using genome-wide, imputed array data that underwent extensive QC, as previously described [13]. Ancestry principal components (PCs) were computed using flashPCA2 [16] by projecting our study samples onto PCs defined by 1000G phase 3 reference populations [17]. Individuals were assigned to the closest reference population using distance from the top 10 PCs. Individuals with ancestral PCs greater than five standard deviations from the reference population mean were excluded. The final analytic dataset included individuals of European, African, East Asian, South Asian, and Hispanic/Latino ancestry; however, the analysis was largely biased toward individuals of European ancestry as they were overrepresented (Additional file 2: Fig. S1). A total of N = 646 (10.2%) and N= 8739 (5.26%) individuals were of non-European ancestry in the KPRB and UKB, respectively (Table 1).

The Regeneron Genetics Center used the Illumina NovaSeq 6000 platform to perform WES for both study populations where the source of DNA was saliva for the KPRB and blood for the UKB. Sample preparation and QC were performed using a high-throughput, fully automated process that has been previously described in detail [18]. Briefly, following sequencing, reads were aligned to the GRCh38 reference genome and variants were called with WeCall [18] for the KPRB and DeepVariant [19] for the UKB. WeCall is a fast, accurate algorithm that jointly identifies and infers genotypes at sites relative to a reference genome. DeepVariant is a computationally scalable deep neural network approach to calling variants [20]. WeCall was first used to call variants in the initial 50K release of the UKB whole-exome sequence data and in our KPRB sequence data. Later, DeepVariant was applied to the 200K release of the UKB WES data we use here after we had processed the KPRB data. Regardless, both algorithms have high sensitivity and specificity for calling genetic variants, so their findings should be comparable across the two studies. Finally, samples with gender discordance, 20× coverage at less than 80% of targeted sites, and/or contamination greater than 5% were excluded.

Additional QC was applied to filter low-quality variants and related individuals. First, genotype calls with low depth of coverage (DP) were updated to missing (DP < 7 for SNPs and DP < 10 for indels). Then, sites with low allele balance (AB) were removed. Specifically, variants without at least one sample having AB ≥ 15% for SNPs or AB ≥ 20% for indels were excluded. Following previous studies [18], we excluded variants with missingness > 10% and HWE p-value < 10⁻¹⁵, computed across all individuals in each study population. After these steps, a total of ~3.51M high-quality sites were retained for the KPRB and ~15.92M were retained for the UKB; excluding singletons, there were ~1.36M and ~8.22M variants, respectively. In the UKB, the larger number of variants observed was due to rare variation present in the larger sample size; when restricting to common variants (MAF > 1%), there were ~186K and ~137K variants, respectively, for the KPRB and UKB.

Genetic association analyses of single variants and genes investigated the following cancer phenotypes: (1) diagnosis with at least two primary cancers across any of the 36 organ sites (“any 2+ primary cancers”) and (2) groupings of individuals defined by a shared index cancer at one of 16 organ sites with at least 50 cases from each study population (“cancer-specific analyses”). Primary analyses compared multiple cancer cases to cancer-free controls. Within our cancer-specific analyses of 16 organ sites, there were cases shared across our index cancer groupings. For example, the set of individuals with at least one diagnosis of breast cancer overlaps with those having at least one ovarian cancer diagnosis.

Single-variant and gene-based association analyses were performed using REGENIE v2.2.4, a machine-learning approach for performing whole-genome regression to correct for cryptic population structure, as well as adjust for case-control imbalance by applying saddlepoint approximation when the standard case-control p-value is less than 0.05 [21]. We assessed single-variant associations for high-quality variants shared across both populations with minor allele count (MAC) > 2 across cancer phenotype cases and controls within each study. The number of variants tested in our single-variant analyses varied by cancer phenotype (~337K [other female genital cancer-specific analysis] to ~722K [any 2+ primary cancers]). WES variants were functionally annotated using SnpEff v5.0 [22] and dbNFSP v3.5 [23] accessed through ANNOVAR [24]. Missense variants were classified using five algorithms: (1) SIFT (“D”), (2) HDIV from Polyphen2, (3) HVAR from Polyphen2, (4) LRT (“D”), and (5) MutationTaster (“A” or “D”). For our gene-based burden analyses, we restricted to rare variants with a MAF < 0.5%, including singletons, computed across all individuals within each study population. Following previous work, three gene-based models were evaluated and the model with the lowest p-value was selected [25]: (1) all rare variants with predicted loss of function (pLOF) by SnpEff, (2) pLOF and missense rare variants predicted to be deleterious by the above five classification algorithms, and (3) pLOF and missense rare variants predicted to be deleterious by at least one algorithm. In our gene-based and single-variant analyses, we adjusted for covariates including age, top 10 PCs, and sex (except for sex-specific index cancers of the breast, cervix, ovary, uterus, other female genital organs, and prostate). In the KPRB population, we additionally adjusted for genotyping array and reagent kit, as they were used to perform case-control matching. In the UKB, we adjusted for flow cell (S2 vs S4), which differed for the initial 50K and subsequent 150K release of WES samples.

Single-variant and gene-based burden analyses for each phenotype were combined across study populations in a fixed-effects meta-analysis using METASOFT [26] and metafor v3.0.2 [27], respectively. For our single-variant analyses, we report all suggestive, independent [linkage disequilibrium (LD) r² < 0.2] associations with p < 5 × 10⁻⁶. For our gene-based analyses, we report all associations adjusted for the number of genes tested (p < 2.65 × 10⁻⁶ = 0.05/18,842). In both analyses, we report meta-analysis p-values.

We also evaluated whether the variants and genes associated with the diagnosis of multiple primary cancers (versus non-cancer controls) remained associated when comparing individuals with multiple cancers to those diagnosed with a single cancer. These analyses assessed whether the variants or genes were pleiotropic for developing multiple cancers or general markers of susceptibility to a specific cancer. We undertook these analyses in the UKB sample only, since individuals diagnosed with a single primary cancer were not sequenced in the KPRB. Single-variant and gene-level analyses were implemented as described above. For each variant or gene of interest identified in our case-control analyses, we performed a case-case analysis comparing individuals diagnosed with multiple cancers to those diagnosed with a single cancer. For our cancer-specific analyses, we compared individuals diagnosed with the index cancer plus any other cancer to those diagnosed with the index cancer only. For example, for a finding discovered in our cancer-specific analysis of prostate cancer, we performed a case-case analysis comparing individuals diagnosed with prostate cancer plus any other cancer to individuals with only a prostate cancer diagnosis.

Our meta-analyses included 6429 cases with multiple primary cancers and 165,853 cancer-free controls (Table 1). All cases had at least two independent primary cancer diagnoses, and 656 cases had more than two diagnoses (Additional file 2: Fig. S2). In the KPRB, the maximum number of cancer diagnoses for an individual was 6 (n = 1), and in the UKB, the maximum number was 5 (n = 2). Overall, 36 unique cancer sites were represented across multiple cancer cases in the two study populations, with 180 unique pairs of sites (e.g., breast and melanoma) and 298 unique ordered pairs of sites by diagnostic sequence (e.g., breast followed by melanoma) (Additional file 1: Table S2). Only 51 of the 298 ordered pairs had at least 25 cancer cases when grouping individuals by first and second cancer diagnosis (i.e., ignoring any subsequent cancer diagnoses; Additional file 1: Table S2, Fig. 1). The top ordered pairs represented in the combined study populations were prostate then melanoma (N = 221), cervix then breast (N = 202), melanoma then prostate (N = 180), breast then melanoma (N = 174), and prostate then colorectal (N = 170). Prostate, breast, melanoma, colorectal, and cervix were the most common sites of first cancer diagnoses (Fig. 1). The prevalence of each cancer pair was similar in the KPRB and UKB (Additional file 2: Fig. S3). As most individual cancer pairs were underpowered for downstream analysis, we considered all multi-cancer cases combined, as well as groupings of individuals with a shared index cancer (16 cancers) (Additional file 2: Fig. S4, Additional file 1: Table S3). Among those with multiple cancers, the cancers with the largest number of cases were prostate (N = 1977; oversampled in KPRB), breast (N = 1874), melanoma (N = 1443), colorectal (N = 1324), and urinary bladder (N = 829).

We found two independent, genome-wide significant associations (p < 5 × 10⁻⁸) and 20 suggestive associations (p < 5 × 10⁻⁶) between individual variants and the multiple cancer phenotypes (i.e., either any 2+ primary cancers or cancer-specific analyses) (Fig. 2, Additional file 1: Table S4). We found an additional two significant and two suggestive associations (Additional file 2: Fig. S5) in our cancer-specific analyses of lymphoid and myeloid neoplasms; however, we assumed them to represent somatic alterations in the blood as they had low allele balance across our heterogenous samples (Additional file 2: Fig. S6) and occur in genes known to be impacted by clonal hematopoiesis of indeterminate potential (CHIP) [28]. Results were relatively homogeneous across the KPRB and UKB study populations (Additional file 1: Table S4). When stratifying by sex, there were no clear material or statistically significant differences in the results; the associations remain in the same direction and were homogeneous across sex subgroups (Additional file 1: Table S4). Additionally, when restricting analyses to European-only individuals, we found 17 (of the 22) associations had only minor changes (<10%) in their effect estimates and corresponding slight decreases in their p-values (Additional file 1: Table S4). Thus, a large majority of our findings likely have similar effects across ancestries and including individuals of all ancestries improves statistical power. The five SNPs with ≥ 10% changes in their corresponding effects when restricting to the European population may have been driven in part by the non-European ancestry individuals.

Of our 22 findings, two variants were suggestively associated with any 2+ primary cancers, rs555607708 (OR [95% CI] = 2.72 [1.79, 4.15], p = 3.10 × 10⁻⁶), a frameshift variant in CHEK2 known to be associated with risk at many cancer sites [29], and rs146381257 (OR [95% CI] = 7.82 [3.28, 18.62], p = 3.45 × 10⁻⁶), a 5′upstream variant in ZNF106. The risk-increasing allele for rs555607708 (CHEK2) was most commonly found among individuals with at least one breast cancer (41.9%), prostate cancer (30.6%), melanoma (22.6%), or cervical cancer (16.1%) (Fig. 2). For rs146381257 (ZNF106), frequencies were increased in prostate cancer (33.3%), lung cancer (28.6%), breast cancer (28.6%), lymphoid neoplasms (23.8%), urinary bladder cancer (19.0%), pancreatic cancer (14.3%), and kidney cancer (14.3%).

An additional 10 of our findings were previously reported risk variants for a single cancer (Fig. 2). Notably, we detected an association with the MC1R variant rs1805008 for melanoma [30] (OR [95% CI] = 1.56 [1.35, 1.81], p = 2.73 × 10⁻⁹), when comparing all individuals with at least one melanoma diagnosis plus any other cancer diagnosis to cancer-free controls. We also replicated the previously associated prostate-specific antigen (PSA) variant, rs17632542 [31] (KLK3, OR [95% CI] = 1.49 [1.28, 1.73], p = 3.87 × 10⁻⁷) in individuals with at least one prostate cancer diagnosis. In addition, we replicated associations between missense risk variant rs6998061 (8q24 locus, POU5F1B) and multiple tumor types in both our prostate cancer-specific analysis [32] (OR [95% CI] = 1.23 [1.13, 1.33], p = 4.39 × 10⁻⁷) and our colorectal cancer-specific analysis [33] (OR [95% CI] = 1.25 [1.15, 1.37], p = 1.06 × 10⁻⁷).

The remaining variants demonstrating associations with multiple cancer phenotypes were not previously associated with any single cancer (Fig. 2). They included a variant discovered in our breast cancer-specific analysis, rs143745791 (NCBP1, OR [95% CI] = 5.95 [2.79, 12.67], p = 3.76 × 10⁻⁶), for which 16.2% of carriers, restricted to cases, had a breast and cervical cancer diagnosis, and a variant discovered in our urinary bladder cancer-specific analysis, rs141647689 (SDK1, OR [95% CI] = 9.29 [3.63, 23.80], p = 3.45 × 10⁻⁶), for which 14.3% of carriers also had prostate cancer (Fig. 2). Three variants found in our lymphoid neoplasm-specific analysis had increased frequencies in cases who also had a diagnosis of prostate cancer: rs535484207 (RANBP2, OR [95% CI] = 256.01 [26.82, 2442.95], p = 1.46 × 10⁻⁶), rs139586367 (UFL1, OR [95% CI] = 284.06 [27.95, 2886.15], p = 1.79 × 10⁻⁶), and rs191064896 (ADGRB1, OR [95% CI] = 108.36 [15.02, 781.08], p = 3.32 × 10⁻⁶), where 21.4%, 40.0%, and 25.0% of carriers for the risk-increasing allele, for each respective variant, had both cancers. The ADGRB1 variant was also present at increased frequencies among individuals with a lymphoid neoplasm and breast cancer diagnosis (25.0%, Fig. 2). Additionally, we identified a single variant in our head and neck cancer-specific analysis, rs12253181 (RTKN2, OR [95% CI] = 1.99 [1.67, 2.37]). Colocalization analyses, within a 500-kb region of the risk SNP, with ezQTL [34] detected a negative correlation between ARID5B expression in whole blood and effects on cancer risk (Additional file 2: Fig. S7). However, these findings should be interpreted with caution since R2 may not adequately control for LD between rare variants and only captures cis-eQTLs in coding regions.

Out of 18,842 genes tested, we found 10 significant associations (p < 2.65 × 10⁻⁶) across our analyses of any 2+ primary cancers and our cancer-specific analyses (Fig. 3, Additional file 1: Table S5). An additional four CHIP genes (ASXL1, TET2, JAK2, and DDX41) were significantly associated with myeloid neoplasms and are likely driven by somatic alterations (Additional file 2: Fig. S8).

In our analyses of any 2+ primary cancers and our breast cancer-specific analysis, we replicated associations for known pleiotropic genes, BRCA2 (pLOF, p = 3.76 × 10⁻¹¹ and 1.91 × 10⁻⁹) and CHEK2 (pLOF + missense, p = 2.95 × 10⁻¹¹ and 1.67 × 10⁻⁸) (Fig. 3). BRCA2 also emerged in our ovarian cancer-specific analysis (pLOF, p = 1.91 × 10⁻⁹). We found associations between the known prostate cancer gene ATM and any 2+ primary cancers and in our prostate cancer-specific analysis (pLOF + missense, p = 9.84 × 10⁻⁷ and 2.56 × 10⁻⁶). Additional associations were observed between SAMHD1 and SLC642 and any 2+ primary cancers (pLOF + missense, p = 2.40 × 10⁻⁷ and p = 5.44 × 10⁻⁷, respectively). BRCA1 also surfaced in the breast cancer-specific analysis (pLOF, p = 6.68 × 10⁻⁸).

Predicted loss of function variants in BRCA1 and BRCA2 were present at increased frequencies in individuals with a breast cancer diagnosis and ovary as an additional cancer site (Fig. 3), such that 28.6% and 13.6% of individuals, respectively, were a carrier for at least one variant in the burden set. For BRCA1, there was also an increase of carriers with an additional melanoma (9.52%) or lung cancer (9.52%) diagnosis. For BRCA2, there was an increase of carriers with an additional uterine (8.47%), lung (6.78%), or colorectal cancer (6.78%).

Out of the 22 associated variants (Fig. 2), 10 remained associated when comparing individuals with multiple cancers to those with single cancers (Additional file 1: Table S6; p < 0.05). Two of these variants were positively associated in our analysis of any 2+ primary cancers: rs555607708 (CHEK2; OR [95% CI] = 1.57 [1.09, 2.25], p = 0.015) and rs146381257 (ZNF106; OR [95% CI] = 5.38 [1.07, 27.18], p = 0.042). The other eight variants were positively associated with the diagnosis of a specific index cancer plus any other cancer versus the specific cancer alone (Additional file 1: Table S6). Two of these eight variants were associated in our breast cancer-specific case-case analysis: rs7872034, a missense variant in SMC2 (OR [95% CI] = 1.16 [1.05, 1.27], p = 0.0025), and rs143745791, a missense variant in NCBP1 (OR [95% CI] = 3.71 [2.08, 6.61], p = 8.37 × 10⁻⁶).

Of the 10 findings from the gene-level burden analyses (Fig. 3), eight remained positively associated with multiple cancers in comparison with single cancers (p < 0.05; Additional file 1: Table S7). Five of these genes were discovered in our case-case analysis of any 2+ primary cancers: SLC6A2 (OR [95% CI] = 1.86 [1.42, 2.41], p = 3.90 × 10⁻⁶), ATM (OR [95% CI] = 1.42 [1.15, 1.77], p = 1.10 × 10⁻³), CHEK2 (OR [95% CI] = 1.56 [1.23, 1.98], p = 2.31 × 10⁻⁴), SAMHD1 (OR [95% CI] = 1.56 [1.14, 2.13], p = 5.34 × 10⁻³), and BRCA2 (OR [95% CI] = 1.86 [1.31, 2.65], p = 5.42 × 10⁻⁴). ATM (OR [95% CI] = 1.82 [1.20, 2.75], p = 4.64 × 10⁻³) was positively associated in our prostate cancer-specific case-case analysis, and the two remaining genes were positively associated in our breast cancer-specific case-case analysis: BRCA1 (OR [95% CI] = 2.38 [1.07, 5.30], p = 0.0340) and BRCA2 (OR [95% CI] = 1.97 [1.22, 3.18], p = 5.50 × 10⁻³).