Quantifying the genetic correlation between multiple cancer types

Abstract

Many cancers share specific genetic risk factors including both rare high-penetrance mutations and common single nucleotide polymorphisms (SNPs) identified through genome-wide association studies (GWAS). However, little is known about the overall shared heritability across cancers. Quantifying the extent to which two distinct cancers share genetic origin will give insights to shared biological mechanisms underlying cancer and inform design for future genetic association studies. In this study, we estimated the pair-wise genetic correlation between six cancer types (breast, colorectal, lung, ovarian, pancreatic and prostate) using cancer-specific GWAS summary statistics data based on 66,958 case and 70,665 control subjects of European ancestry. We also estimated genetic correlations between cancers and 14 non-cancer diseases and traits. After adjusting for 15 pair-wise genetic correlation tests between cancers, we found significant (p<0.003) genetic correlations between pancreatic and colorectal cancer (r_g=0.55, p=0.003), lung and colorectal cancer (r_g=0.31, p=0.001). We also found suggestive genetic correlations between lung and breast cancer (r_g=0.27, p=0.009), and colorectal and breast cancer (r_g=0.22, p=0.01). In contrast, we found no evidence that prostate cancer shared an appreciable proportion of heritability with other cancers. After adjusting for 84 tests studying genetic correlations between cancer types and other traits (Bonferroni-corrected p-value: 0.0006), only the genetic correlation between lung cancer and smoking remained significant (r_g=0.41, p=1.03x10^-6). We also observed nominally significant genetic correlations between body mass index (BMI) and all cancers except ovarian cancer. Our results highlight novel genetic correlations and lend support to previous observational studies that have observed links between cancers and risk factors.

Introduction

In the US, cancer remains the second leading cause of death with an estimated 1.69 million new cancer diagnoses and 600,000 cancer deaths in 2017 (1). Six cancer types—breast, colorectal, lung, ovarian, pancreatic and prostate cancer—together constitute more than 50% of annual cancer diagnoses (1). The etiologies of these cancers are complex and associated heritability estimates (2) from twin studies range between 0.15 (colon) and 0.57 (prostate), indicating genetic components. For pancreatic cancer, data have been too sparse to estimate heritability based on twin studies. However, Mucci and colleagues (2) did observe significant, albeit lower relative to other cancers, familial risks for pancreatic cancer. These results are supported by a meta-analysis based on more than 6,500 cases, where having a relative diagnosed with pancreatic cancer was associated with a 1.8-fold (95% CI: 1.5-2.1) risk increase of pancreatic cancer (3).

Various cancers share both environmental and genetic risk factors including rare high-penetrant mutations in genes such as BRCA2 which predisposes to breast, ovarian, lung, prostate and pancreatic cancers (4). Genome-wide association studies (GWAS) have identified more 350 distinct (reported SNPs > 500kb apart) genomic regions that are associated with cancer (5) of which several are shared between cancer types (6). For example, the 8q24 (7–9) and TERT (10–13) regions have been associated with multiple cancer types (pleiotropy), although specific alleles often differ. In contrast to pleiotropy, which does not take the direction of association into account, genetic correlation describes the genome-wide correlation in allele effects and thus, considers the allele-specific direction of association between two traits. Thus, the shared genetic etiology between two traits can be either due to a shared genetic variant (or variants) with non-equal non-zero effect sizes (pleiotropy) or via a correlation between effect sizes for causal variants (genetic correlation) (14). Identifying genetic regions that are associated with multiple cancer types may be useful for determining mechanisms involved in global carcinogenesis. However, the benefit of simultaneously studying multiple cancer types relies on their genetic correlation.

With the introduction of GWAS, it is now possible to quantify the phenotypic variance explained by genotyped single nucleotide polymorphisms (SNPs) in single trait $\left(h_g^2\right)$ (15–17) and two trait (r_g) (18) settings by using variance component methods. Lu and colleagues (19) used this approach to estimate $h_g^2$ for 12 cancers and found significant non-zero genetic contribution for eight of them. However, their average sample size for each cancer was only 1,793 cases (range: 564-2,848) and 2,200 controls (range: 574-3,159), leading to imprecise heritability estimates. Sampson and colleagues (20) estimated the genetic correlation across 13 different cancers in an average sample size across cancers of 3,807 (range: 535-5,942) cases and 2,625 (range: 1,056-10,857) controls. Although no genetic correlation withstood multiple testing after adjusting for 91 tests, they observed the strongest genetic correlations (all p<0.01) between kidney and testis cancer (r_g=0.73, SE=0.28), diffuse large B-cell lymphoma and osteosarcoma (r_g=0.53, SE=0.21), diffuse large B-cell lymphoma and chronic lymphocytic leukemia (r_g=0.51, SE=0.18), and bladder and lung (r_g=0.35, SE=0.14) cancer.

The sample sizes of these prior studies make interpretation of their findings difficult. Indeed, a drawback with the variance component approach is its requirement for individual-level data, which prohibits researchers from leveraging GWAS results based on meta-analyses which are often based on much larger sample sizes. The recently developed cross-trait linkage disequilibrium (LD) score regression approach overcomes this limitation by estimating the proportion of phenotypic variance explained by common SNPs (21) and the genetic correlation between two traits (22) using summary statistics only. Here, we set out to quantify the pair-wise genetic correlation between across breast, colorectal, lung, ovarian, pancreatic and prostate cancer, capitalizing on summary statistics obtained from GWAS data in 66,958 case and 70,665 control subjects obtained from the GAME-ON, PanScan and GECCO consortia. In addition, we estimated the genetic correlation between each of these cancers and 14 non-cancer traits which have all been suggested to be linked to cancer and for which we had access to GWAS summary statistics.

Material and Methods

The GAME-ON network of consortia for post-GWA research, PanScan and GECCO

We utilized three large-scale cancer genetic epidemiological consortia: GAME-ON, PanScan III and GECCO. The Genetic Associations and Mechanisms in Oncology (GAME-ON) consortia is a network of cancer-specific post-GWAS initiatives. The five GAME-ON sites are breast (DRIVE), colorectal (CORECT), lung (TRICL-ILLCO), ovarian (FOCI) and prostate (ELLIPSE) (23). One of the main goals with GAME-ON was to test hypotheses across the cancer types that might illuminate common mechanisms of susceptibility. PanScan is a part of The Pancreatic Cancer Cohort Consortium with the goal of conducting GWAS to identify susceptibility markers for pancreatic cancer. For this study, we utilized genome-wide summary statistics from PanScan III (24). The Genetic Epidemiology of Colorectal Cancer Consortium (GECCO) is a large collaborative consortium evaluating genetic and environmental risk factors for colorectal cancer (25). Details of GAME-ON, PanScan, GECCO and the participating studies are available at http://epi.grants.cancer.gov/gameon/, http://epi.grants.cancer.gov/PanScan/ and https://www.fredhutch.org/en/labs/phs/projects/cancer-prevention/projects/gecco.html. Sample sizes for each cancer is listed in Table 1. These studies have been described in detail previously (23). For each cancer type, genotyping was performed using Illumina or Affymetrix arrays of varying densities described elsewhere (23,26,27). For all studies except GECCO, imputation was performed using the 1,000 Genomes Project Phase 1 version 3 reference haplotypes, resulting in up to ~10 million SNPs available for the analysis for each cancer type. For GECCO, data was imputed using an in-house reference panel of 2,159 whole-genome sequenced European-ancestry GECCO participants (28). Since imputation quality scores were not readily accessible across all GWAS, we only included HapMap 3 SNPs as a proxy for well-imputed SNPs (29).

Table 1. Overview of cancers analyzed.

Number of subjects, observed heritability (standard error) and heritability (95% CI) on the liability scale explained by studied SNPs for each cancer. Cumulative risks used for calculating $h_g^2$ on the liability scale were obtained from SEER.

Cancer Type	Cases	Controls	$h_g^2$(se) – observed scale	$h_g^2$ (95% CI)– liability scale
Breast	15,748	18,084	0.12 (0.02)	0.14 (0.09-0.18)
Colorectal	15,716	18,154	0.13 (0.02)	0.11 (0.07-0.14)
Lung	12,160	16,838	0.14 (0.03)	0.13 (0.08-0.19)
Ovarian	4,369	9,123	0.10 (0.04)	0.07 (0.02-0.12)
Pancreatic	5,107	8,845	0.07 (0.04)	0.05 (0-0.10)
Prostate	14,160	12,724	0.25 (0.03)	0.27 (0.21-0.33)

Non-Cancer Traits

We also estimated the genetic correlations between the six cancer types and 14 non-cancer traits (29) for which GWAS summary statistics were publicly available (Supplementary Table S1). The included traits were schizophrenia, bipolar disorder, coronary artery disease, type 2 diabetes, Crohn’s disease, ulcerative colitis, rheumatoid arthritis, ever/never smoked tobacco, height, BMI, fasting serum glucose, triglycerides, LDL cholesterol and HDL cholesterol. The average sample size for the non-cancer traits was 70,488 subjects.

Statistical Methods

Associations between SNPs and cancer risk were estimated by unconditional logistic regression adjusted for age, sex (when applicable), and top principal components (ranging from two to six across contributing GWAS) to adjust for potential population stratification. We used a newly developed method for estimating heritability due to common SNPs and genetic correlations that employs only summary statistics together with linkage disequilibrium (LD) information from a reference panel (21,22). This method, known as LD score regression, relies on the fact that SNP-specific association statistics reflect the associations of all SNPs in LD with that SNP. Thus, for a polygenic trait, SNPs in high-LD regions will on average have higher χ² statistics than SNPs in low-LD regions and similarly, for two polygenic, genetically correlated traits with z-scores z₁ and z₂, the product z₁z₂ will on average be higher for SNPs with high LD than SNPs with low LD. Formally, the relationship between the expected χ² statistic for SNP j and the LD score l(j) for SNP j can be described by $E\left[{\chi }_j^2\right]\approx \frac{N_j{h}_g^2}{M}{l}_j+1,$ where N_j is the sample size, $h_g^2$ is the heritability due to included SNPs, M is the number of SNPs and l(j) := ∑_kr²(j, k) where r²(j, k) is the correlation between SNP j and all other SNPs i. Since the observed SNP heritability estimates for binary traits are not directly comparable to more traditional estimates of heritability (e.g. from twin studies), we transformed the observed heritability to the liability scale which takes both ascertainment and disease prevalence into account as previously described (30). To obtain estimates of cancer-specific prevalence, we used SEER cumulative risks (31). We can extend the calculations of single-trait heritability to include two traits: $E\left[z_{1j}{z}_{2j}\right]=\frac{\sqrt{N_1{N}_2}{r}_g}{M}{l}_j+\frac{N_s\rho }{N_1{N}_2},$ where r_g is the genetic covariance, N₁ and N₂ are sample sizes for trait 1 and 2 respectively, N_s is the number of overlapping samples and ρ is the phenotypic correlation in the overlapping samples. We estimate r_g by the slope of the regression of z_1jz_2j on the LD score and tested for its difference from 0, as described previously (22). In total, we conducted 15 pair-wise genetic correlation tests between cancers types at a significance threshold of p=0.05/15=0.003. In our secondary analyses, estimating genetic correlations between cancer and non-cancer traits, we conducted 84 tests and for these analyses, we considered p-values less than p=0.05/84=0.0006 significant. For all analysis, we only included SNPs from HapMap3.

Results

Cancer-specific GWAS summary statistics from data of 66,958 case and 70,665 control subjects across six cancer types were shared between the cancer sites. We first estimated the heritability due to common SNPs $\left(h_g^2\right)$ on the observed and liability scale (Table 1) using cancer-specific cumulative risks (Supplementary Table S2) as reported in SEER (31). Heritability estimates on the liability scale ranged between 0.04 (pancreatic cancer) and 0.27 (prostate cancer). Using cross-trait LD score regression, we quantified the pair-wise genetic correlations between breast, colorectal, lung, ovarian, pancreatic and prostate cancer. A summary of the results is displayed as a 6x6 matrix (Table 2) with the estimated pair-wise genetic correlations and associated standard errors in the upper right corner of the matrix, and corresponding p-values for each genetic correlation can be seen in the lower left corner of the matrix. We found significant positive genetic correlations between pancreatic and colorectal cancer (r_g=0.55, p=0.003), lung and colorectal cancer (r_g=0.31, p=0.001), as well as suggestive positive genetic correlations between lung and breast cancer (r_g=0.27, p=0.009), and colorectal and breast cancer (r_g=0.22, p=0.01).

Table 2. Genetic correlations between cancers.

Genetic correlations with standard errors are in the upper right part of the table, corresponding p-values are in the lower left part of the table.

	Breast	Colorectal	Lung	Ovarian	Pancreatic	Prostate
Breast	1	0.22 (0.091)	0.27 (0.11)	0.26 (0.20)	0.17 (0.19)	0.06 (0.09)
Colorectal	0.014	1	0.31 (0.097)	-0.08 (0.13)	0.55 (0.19)	0.09 (0.07)
Lung	0.009	0.001	1	-0.17 (0.17)	0.32 (0.19)	0.095 (0.08)
Ovarian	0.18	0.57	0.32	1	-0.40 (0.29)	0.02 (0.14)
Pancreatic	0.37	0.003	0.08	0.17	1	-0.06 (0.14)
Prostate	0.52	0.2	0.25	0.89	0.68	1

We also estimated the genetic correlation between each of these six cancer types and 14 other diseases and traits implicated to be linked to cancer (schizophrenia, bipolar disorder, coronary artery disease, type 2 diabetes, Crohn’s disease, ulcerative colitis, rheumatoid arthritis, ever/never smoked tobacco, height, body mass index (BMI), fasting serum glucose, triglycerides, LDL cholesterol and HDL cholesterol (Figure 1, Table 3, Supplementary Tables S1 and S3). The strongest positive genetic correlation was, as expected, between lung cancer and smoking status (never/ever; r_g=0.41, p=1.03x10^-6), which was the only correlation to remain significant after adjusting for multiple testing. Other nominally significant genetic correlations with lung cancer were with BMI (r_g=0.12, p=0.03) and an inverse genetic correlation with HDL levels (r_g=-0.15, p=0.02). For breast cancer, we observed positive genetic correlations with ulcerative colitis (r_g=0.24, p=0.002), schizophrenia (r_g=0.14, p=0.004) and height (r_g=0.14, p=0.01). In contrast, triglycerides (r_g=-0.13, p=0.03) and BMI (r_g=-0.11, p=0.04) both had inverse genetic correlations with breast cancer. For colorectal cancer, we observed positive genetic correlations with BMI (r_g=0.16, p=6.2x10^-4) and triglycerides (r_g=0.13, p=0.03) and an inverse genetic correlation with schizophrenia (r_g=-0.09, p=0.05). Ovarian cancer showed positive genetic correlations with type 2 diabetes (r_g=0.47, p=0.01) and ulcerative colitis (r_g=0.29, p=0.03). For pancreatic cancer, we observed a positive genetic correlation with BMI (r_g=0.24, p=0.04) and an inverse genetic correlation with HDL levels (r_g=-0.41, p=0.02). Finally, prostate cancer showed an inverse correlation with BMI (r_g=-0.08, p=0.04).

Figure 1.

Genetic correlations between cancers and non-cancer traits. Nominally significant genetic correlations (p<0.05) are highlighted with *.

Table 3. Nominally significant genetic correlations between cancers and non-cancer traits.

For each cancer-trait genetic correlation observed, current evidence from observational studies and corresponding reference are also listed. P values are not corrected for multiple testing.

Cancer	Non-cancer trait	rg	se	p	Epidemiological observations	Reference
Breast Cancer	Ulcerative Colitis	0.243	0.08	0.002	+	(55)
Breast cancer	Schizophrenia	0.142	0.049	0.004	+	(56)
Breast cancer	Height	0.143	0.055	0.01	+	(57)
Breast cancer	Triglycerides	-0.13	0.061	0.033	-	(58)
Breast cancer	BMI	-0.112	0.053	0.035	- (premenopausal) + (postmenopausal)	(35)
Colorectal cancer	BMI	0.157	0.046	6.22E-04	+	(35)
Colorectal cancer	Triglycerides	0.126	0.058	0.029	+	(59)
Colorectal cancer	Schizophrenia	-0.091	0.046	0.048	No association	(56)
Lung cancer	Ever/Never Smoked	0.412	0.084	1.03E-06	+	(60)
Lung cancer	HDL	-0.151	0.063	0.017	-	(59)
Lung cancer	BMI	0.116	0.054	0.032	- (smokers) No association (non-smokers)	(35)
Ovarian cancer	Type-2 Diabetes	0.469	0.191	0.014	+	(61)
Ovarian cancer	Ulcerative Colitis	0.291	0.137	0.034	No association	(55)
Pancreatic cancer	HDL	-0.405	0.167	0.015	Unknown	N/A
Pancreatic cancer	BMI	0.243	0.117	0.038	+	(35)
Prostate cancer	BMI	-0.083	0.04	0.039	-	(35)

Discussion

It is well known that cancer tends to cluster in families, which has been attributed to shared environmental factors and genetics (2). Many lifestyle factors associated with cancer including smoking, obesity and alcohol intake have been shown to influence risk of multiple cancer types, implying that different cancer types share an underlying biological mechanism. Similarly, genetic variation in genes such as BRCA1/2 and TERT has been associated with risks of multiple different cancer types, providing empirical support that there are specific regions in the genome that harbor genetic variation influencing risk of multiple cancer sites. However, although many cancer types might share susceptibility loci (i.e. pleiotropy), their genetic correlation, which depends on specific risk alleles and their direction of associations, might not be strong. Here, we aimed to assess the latter among six cancers, breast, colorectal, lung, ovary, pancreatic and prostate, as well as between cancer types and seven additional disease and seven non-disease traits that have all been found to have heritable components.

We used SEER estimates (31) to obtain cancer-specific cumulative risks, recognizing that not all study subjects came from US populations. For comparison, we also calculated the heritability on the liability scale using cumulative risks obtained from Mucci et al (2). No qualitative difference was observed for the liability estimates using the two different sources of cumulative risks. In concordance with previous studies, the cancer-specific heritability estimates observed here were lower than what has been observed in twin studies. This is not unexpected given that here, we are only estimating the additive genetic component based on common SNPs captured by GWAS, and thus, any contribution to the heritability based on factors such as gene-gene interactions, gene-environment interactions, structural variants or rare variants, will not be captured by our analysis. Among the studied cancers, we observed the largest heritability for prostate cancer ($h_g^2$=0.27) in agreement with previous twin studies (2). We also compared our cancer-specific heritability results to previous studies estimating heritability based on GWAS data. In general, our results were comparable with previous studies. Lu and colleagues (19) estimated $h_g^2$ for breast cancer to be 0.13 (95% CI: 0-0.56) compared to our estimate of 0.14 (95% CI: 0.09-0.18). For lung cancer, previous estimates have varied with Lu et al(19) estimating $h_g^2$ to 0.10 (95% CI: 0-0.24) in European populations while Sampson et al (20) estimated $h_g^2$ to be 0.21 (95% CI: 0.14-0.27), compared to our estimate of 0.13 (95% CI: 0.08-0.19). For ovarian cancer, we observed a small heritability ($h_g^2$=0.07, 95% CI: 0.02-0.12) compared to Lu and colleagues (19) ($h_g^2$=0.30, 95% CI: 0.18-0.42). It is not clear why we observe this discrepancy in results. We also observed lower $h_g^2$ of pancreatic cancer (0.05, 95% CI: 0-0.10) than previously observed (0.18, 95% CI: 0.06-0.30 for Lu and colleagues (19) and 0.10, 95% CI: 0.04-0.16 for Sampson et al (20)) but we note that the confidence intervals are wide and overlap. For prostate cancer, we observed a heritability of 0.27 (95% CI: 0.21-0.37) compared to Lu (19) (0.81, 95% CI: 0.32-1), and Sampson (20)(0.29, 95% CI: 0.15-0.42). We note that heritability estimates reported here are all on the liability scale and were calculated using SEER rates for all three studies including ours.

We found that colorectal cancer showed significant genetic correlations with pancreatic and lung cancers, with the largest genetic correlation observed for the two gastrointestinal tract cancers: the r_g for colorectal and pancreatic cancer was 0.55 (p=0.003). Amundadottir and colleagues studied cancer risk for first up to fifth degree relatives in an Icelandic population and observed an increased risk for pancreatic cancer among colon cancer patients (and vice versa) for first degree relatives but not beyond (32). Colorectal cancer patients have been observed to have a higher incidence of pancreatic cancer than the general population (33). Further, Lynch syndrome, the most common hereditary colorectal cancer syndrome, has also been shown to increase risk for pancreatic cancer (34). Obesity is a well-established risk factor for both colorectal and pancreatic cancer (35,36) and we observed nominally significant (p<0.05) genetic correlations between BMI and both colorectal and pancreatic cancer.

Colorectal cancer also showed suggestive genetic correlation with breast cancer in agreement with the study from Amundadottir and colleagues (32). A recent study found that women diagnosed with breast cancer have a 1.59-fold (95% Confidence Interval (CI): 1.53-1.65) increased risk of developing colorectal cancer compared to the general population (37).

Breast and lung cancer showed a suggestive positive genetic correlation (r_g=0.27, p=0.009), supported by observational studies finding familial co-segregation of the two cancers (38,39) as well as overlap in multiple susceptibility genes such as BRCA2, CHEK2 (40) and LSP1 (7). In contrast, Amundadottir and colleagues did not observe a significant co-occurrence among relatives (32). A recent cross-cancer GWAS based on the GAME-ON data (23) identified a pleiotropic locus at 1q22 that was associated with both breast and lung cancer.

We expect these results to generate testable hypotheses about mechanisms. There are data to support inflammation response, DNA repair and stress responses, to list just a few. For example. the genetic correlations between breast, colorectal and lung cancer might be driven in part by genetic variants in the inflammation pathway. Indeed, a recent analysis of genetic variation in the inflammation pathway from the GAME-ON consortium identified SH2B3, a key negative regulator of cytokine signaling to be associated with all three cancers (41). Further, the authors found no evidence that genetic variation in inflammation-related genes was associated with prostate cancer risk. This observation corroborates our findings that prostate cancer does not share an appreciable genetic component with other cancers and supports inflammation to be a pathway in which genetic variation affects the risk of breast, colorectal and lung cancer. Removing all SNPs with a chi-square test statistic >25 in the individual cancer GWAS, did not change our results: Colorectal-breast cancer (r_g=0.22 for all SNPs and r_g=0.23 excluding significant SNPs); Colorectal-lung cancer (r_g=0.31 for all SNPs and r_g=0.34 excluding significant SNPs); Colorectal-pancreatic cancer (r_g=0.55 for all SNPs and r_g=0.58 excluding significant SNPs); Breast-lung cancer (r_g=0.27 for all SNPs and r_g=0.33 excluding significant SNPs). Thus, it is likely that the genetic correlations that we observe are due to yet unidentified SNPs, and future studies should focus on simultaneously study genetically correlated cancers with the goal of identifying SNPs that are associated with multiple cancer types.

We did not observe evidence that either ovarian or prostate cancer shared an appreciable amount of heritability with other cancers, although our sample size for ovarian cancer was relatively small, leading to wide confidence intervals. We had higher statistical power to detect genetic correlations involving prostate cancer, but no estimate was >0.1, suggesting that prostate cancer has a genetic contribution that is distinct from that of other cancer types studied here. Witte and Hoffmann used polygenic risk scores to investigate a potential shared heritability between breast and prostate cancer, and although they observed a potential common polygenic model between non-aggressive prostate cancer and breast cancer, they observed no evidence of a common model between overall prostate cancer and breast cancer, consistent with our results here (42). It is important to note that our analysis does not capture rare higher penetrance mutations, and thus, the observed increased risk of multiple cancers among relatives to prostate cancer cases (32) is likely to at least in part be attributable to rare variants such as BRCA2.

In addition, we examined the genetic correlation between these cancer types and 14 non-cancer diseases and traits. The only genetic correlation between cancer and non-cancer traits that withstood correction for multiple testing was smoking status and lung cancer. The strongest lung cancer susceptibility locus is in the 15q25 region which contains genes encoding the nicotinic acetylcholine receptor subunits CHRNA5, CHRNA3 and CHRNB4 and has also been associated with smoking behavior with associations in the same direction (43–45). We found that BMI showed nominally significant genetic correlations with all cancers except ovarian cancer. While BMI showed positive genetic correlation with colorectal, lung and pancreatic cancer, it showed negative genetic correlations with breast and prostate cancer. These results mirror recent Mendelian randomization (MR) studies of BMI and colorectal (46), breast and lung cancer (47,48), providing further evidence that obesity is involved in cancer development. An MR study of prostate cancer found a non-significant lower risk associated with a BMI genetic score (Odds ratio (OR)=0.98; 95% CI: 0.96-1.00; p = 0.07) (49). Although the positive genetic correlation between BMI and lung cancer seems to contradict results from observational studies, the observational association between BMI and lung cancer might be due to residual confounding by smoking (35). BMI and smoking behavior have been shown to share a genetic basis (r_g=0.20, p=8.3x10^-7) (22) and further, cell-type enrichment heritability analysis have shown that both smoking behavior and BMI are enriched for central nervous system (CNS)-related cell types. Therefore, it is possible that smoking and BMI to some extent affect lung cancer risk through the same biological mechanism.

We observed very few genetic correlations with prostate cancer compared with the other cancers of comparable sample size and no genetic correlation with non-cancer traits were >0.1. We also note the lack of genetic correlation between prostate cancer and type 2 diabetes (r_g=0.03, 95% confidence interval: -0.10-0.15, p=0.67). The epidemiological inverse association between prostate cancer and type 2 diabetes is well-documented (50–53) and a previous study showed that 10 out of 36 type 2 diabetes SNPs were associated (2 with increased risk and 8 with decreased risk) with advanced prostate cancer (54). The different directions of significant prostate cancer associations across type 2 diabetes SNPs are consistent with the lack of genetic correlation (which is sensitive to direction of effects) observed in this study.

Sampson and colleagues previously used individual-level GWAS data to estimate genetic correlations between 13 cancers in 49,492 cancer cases and 34,131 controls, including estrogen receptor negative (ER-) breast cancer, lung cancer, pancreatic and prostate cancer (20). Although, they did not observe the statistically significant genetic correlations between the cancers studied here, their standard errors were in general large, making it difficult to compare the results. We note that the ER- breast cancer and pancreatic datasets they used are a subset of the data analyzed here. While we had access to GWAS summary statistics based on cancer subtypes including ER- breast cancer, squamous cell lung cancer, lung adenocarcinoma, serous, clear cell and endometrioid ovarian cancer and aggressive prostate cancer, sample sizes for these subsets were too small for meaningful analysis. Based on our experience, LD score regression requires at least 10,000 cases for adequately stable estimates at these heritabilities. We note a few limitations with only having access to summary statistics data compared to individual-level data. Most importantly, the standard errors associated with the estimated genetic correlations based on LD score regression are larger than what is obtained by similar methods using individual-level data. In addition, we are not able to conduct any subtype analysis on the original traits that might be of interest, for example, it might have been of interest to study the genetic correlation between BMI and breast cancer stratified by menopausal status. LD score regression leverages summary statistics rather than individual-level data and thereby overcome many of the issues associated with relying on individual-level data. Moreover, appropriate QC steps were conducted as part of the cancer-specific GWAS meta-analysis. Further, we limited our analysis to HapMap 3 SNPs to ensure well-imputed data.

In summary, our results indicate that some cancers show modest genetic correlations; in particular, breast, colorectal and lung cancer share some degree of genetic basis. In contrast, prostate cancer appears to have a unique genetic architecture that is not shared with breast, lung, ovarian and pancreatic cancer. Further, a number of cancer types show genetic correlations with obesity, highlighting the involvement of adiposity-related processes in cancer. As GWAS sample sizes continue to increase and GWAS summary statistics from other traits become available, we will be able to additionally characterize the shared heritability between cancer types including histologic subtypes as well as with non-cancer traits.