Leveraging genome-wide association studies to better understand the etiology of cancers.

The standard GWAS analytical procedure consists of several steps, starting with the collection of DNA samples and corresponding phenotypic information (Figure  1 ). After collection, the DNA samples are genotyped using SNP microarrays or other sequencing technologies (depending on the study design). Subsequently, the obtained genotyping data are evaluated and filtered using several quality control criteria (such as genotype call rate per sample or variant, departures from Hardy–Weinberg equilibrium, and genotyping technology‐related metrics). Genotype imputation is then carried out to computationally infer untyped variant genotypes using haplotype reference panels, particularly when SNP arrays are adopted for genotyping. Finally, genotype–phenotype associations are statistically tested for each genotyped or imputed variant. Although the early GWAS adopted the χ 2 ‐test or Cochran–Armitage trend test for the genotype–phenotype association testing, the standard choices for current studies are regression analysis‐based tests that can account for covariates, such as population stratification, age, and sex. When multiple series of association datasets are available, a meta‐analysis combining these datasets boosts statistical power. Overview of a genome‐wide association study (GWAS) workflow. (A) Individuals with a specific phenotypic trait (e.g., case/control) are recruited, and their genomic DNA is collected. (B) Genome‐wide genotype data is obtained from the collected genomes using single nucleotide polymorphism (SNP) microarray or high‐throughput sequencing. (C) The obtained genotype data is evaluated according to several quality‐control criteria to filter out low‐quality data; this step also includes the detection of sample swaps and population stratification in the sample. The panel depicts quality control by SNP array intensity signal cluster separation. (D) The imputation step computationally infers untyped genetic variants in the sample based on a reference haplotype panel, which is typically constructed by whole‐genome sequencing. (E) Statistical association testing with the phenotype is repeated for each (typed or imputed) variant. The association statistics for every variant in the GWAS are aggregated and referred to as GWAS summary statistics. (F) Meta‐analysis by combining summary statistics of multiple GWAS improves statistical power. In most cases, disease‐associated variants identified by GWAS show odds ratios of <1.5. These relatively limited effect sizes are because of the evolutionary/selective pressures; if a disease‐associated allele confers severe risk, natural selection will eliminate the allele from the population. Logically, one could wonder if such small‐effect variants have significant biological impacts. However, importantly, a weak effect size of a variant does not mean limited clinical relevance of the mapped susceptibility gene. For example, TP53 is a well‐known cancer suppressor gene, which is identified as a significantly mutated gene by somatic mutation analyses. 20 , 21 The oncogenic effect of TP53 deficiency in the development of melanoma has been experimentally demonstrated. 22 GWAS of melanoma identified rs1641548, an intronic SNP of TP53 , as an associated variant. 23 This SNP is, however, commonly seen in general populations of European ancestry, with a minor allele frequency of approximately 18%, and shows an odds ratio of 1.08, exemplifying that the clinical relevance of the detected loci is independent of the effect sizes estimated in the GWAS. One important concept for correctly interpreting GWAS results is LD. When a duly conducted GWAS reveals a significant association signal, in addition to the most significant marker, the nearby markers usually also exhibit similar association p values. Statistically, this is due to the correlation between the genotyped nearby markers, referred to as LD, which reflects that, evolutionarily, recombination events are infrequent between physically close genetic markers (Figure  2 ). Due to LD, GWAS do not require direct genotyping of all common variants present in the population. Genotyping only a subset of them, called “tagging” SNPs, which partially or completely predict the genotype of other nearby SNPs through genotype imputation methods, is enough. 24 Thus, for cost‐effectiveness, the GWAS study design exhaustively exploits LD for association mapping. However, LD also imposes a challenge in interpreting the detected association signals. The existence of a significant association between the tested variant and a phenotype does not always imply that it is the causal variant (even if it shows the most significant association in the region, called the lead variant). The association may reflect LD between the tested variant and another variant having the true causal effect. Linkage disequilibrium (LD) and its influence on genome‐wide association study results. An ancestral chromosome harboring the disease causal variant is inherited for generations, during which recombination events occur in the chromosome. As recombination events do not frequently occur between the causal variant and its physically close genetic variants, descendent chromosomes are likely to coharbor these proximal and causal variants. As a result, these proximal‐variant genotypes show a correlation with the causal‐variant genotypes; and a disease‐association signal similar to that of the causal‐variant. Moreover, these noncausal but correlated variants can show a stronger association than the causal variant due to statistical noise.

Kyuto Sonehara: Writing – original draft; writing – review and editing. Yukinori Okada: Writing – review and editing.

Approval of the research protocol by an Institutional Review Board: N/A. Informed Consent: N/A. Registry and the Registration No. of the study/trial: N/A. Animal Studies: N/A.

K.S. was supported by the Japanese Society for the Promotion of Science (JSPS) KAKENHI (23K14451). Y.O. was supported by the JSPS KAKENHI (2H00476), the Japan Agency for Medical Research and Development (AMED) (JP22km0405211, JP22ek0410075, JP22km0405217, JP22ek0109594, JP223fa627002, JP223fa627010, JP233fa627011, and JP23zf0127008), JST Moonshot R&D (JPMJMS2021 and JPMJMS2024), the Takeda Science Foundation, Bioinformatics Initiative of Osaka University Graduate School of Medicine, Institute for Open and Transdisciplinary Research Initiatives, Center for Infectious Disease Education and Research (CiDER), and Center for Advanced Modality and DDS (CAMaD), Osaka University.

While GWAS reveal numerous genotype–phenotype associations, nearly 90% of the associated variants are located within noncoding regions of the genome, 25 making the functional interpretation challenging. This hampers the elucidation of the underlying molecular mechanisms mediating the association with the disease. This problem can be circumvented by using molQTL mapping (Figure  3A ). Molecular quantitative trait loci are genetic variants that are associated with intermediate molecular phenotypes, such as gene expression (eQTL), 26 , 27 splice junction usage (splicing QTL), 28 , 29 protein expression (pQTL), 30 , 31 and chromatin accessibility (chromatin accessibility QTL). 32 , 33 For example, if a disease‐associated variant is characterized as an eQTL, then the genetic predisposition is potentially mediated by the modulation of regulation of expression of the respective gene (Figure  3B ). However, the existence of LD can confound these findings. For instance, even when the causal variants in the GWAS and those causing the eQTL are different, LD can cause them to appear associated; that is, the genetic predisposition indicated by GWAS can be accidentally attributed to variants responsible for the eQTL effect. To avoid this confounding of results, colocalization analyses are carried out, which evaluate the probability that the causal variants for eQTL and those in the GWAS are the same. 34 Although most molQTL studies were undertaken in samples of normal tissues, mainly blood or blood‐derived cells due to sample accessibility, some studies leveraged existing large sample collections of tumors, such as TCGA, for molQTL mapping in tumor tissues. 29 , 33 , 35 , 36 , 37 , 38 , 39 Which type of molecular phenotypes is specifically promising for functional interpretation of a given disease GWAS is an open question. Also, novel molQTL types are still being explored. 40 , 41 Functional characterization of genome‐wide association study (GWAS) associations. (A) Representative examples of molecular quantitative trait loci (QTL) are illustrated. (B) If a GWAS variant associated with a specific phenotype is characterized as an expression QTL (eQTL) for a specific gene, the gene expression change can be a candidate mediator of the GWAS association. (C) In the transcriptome‐wide association study framework, by constructing gene expression prediction models using multiple variants, we can impute the genetically regulated gene expression only based on the GWAS genotype data. caQTL, chromatin accessibility QTL; pQTL, protein QTL; sQTL, splicing QTL. To prioritize candidate causal genes, TWAS is proposed as a straightforward framework as it links the genotype effects on gene expression to disease susceptibility. 42 In TWAS, gene expression prediction models are trained using reference samples comprising both genotype and gene expression data. The prediction models so obtained are then applied to the GWAS data, and the association between the genetically predicted gene expression and the disease risk is explored (Figure  3C ). Transcriptome‐wide association study has the following advantages: (1) the multiple testing burden is reduced compared to GWAS because statistical testing is carried out only for genes whose expression can be significantly explained by genetic variants; (2) the analysis directly implicates candidate genes with the disease, making the targets of functional follow‐up experiments clearer; (3) TWAS can be undertaken if the gene expression prediction models and GWAS summary statistics are available. Statistical techniques and population‐matched reference LD information make the individual‐level genotype data unnecessary for TWAS. 43 , 44 Despite having these advantages, TWAS can allow spurious links caused by LD between different causal variants in a manner similar to how it affects the eQTL results (discussed above). This problem, specifically referred to as LD contamination, 44 is addressed by performing colocalization analyses as a follow‐up. Transcriptome‐wide association study and colocalization analyses play complementary roles in the functional interpretation of GWAS associations; TWAS screens for candidate molecular phenotypes mediating GWAS associations, whereas colocalization analyses filter out spurious associations caused by LD contamination. Transcriptome‐wide association study has been applied to GWAS of cancers, such as ovarian cancer, 28 breast cancer, 45 pancreatic cancer, 46 colorectal cancer, 47 and lung adenocarcinoma. 16 For example, in the TWAS of ovarian cancer, Gusev et al. applied prediction models of gene expression levels and splice junction levels to a GWAS of ovarian cancer ( n  > 50,000). 28 The TWAS analysis identified at least one target gene for 6 out of 13 genome‐wide significant regions of the original GWAS data. Moreover, the application of the TWAS framework is not limited to protein‐coding genes. We reported a miRNA‐based TWAS; miRNA are small noncoding RNAs of 21–25 nucleotides, which are key modulators of the posttranscriptional gene regulation. 48 In the study, we undertook small RNA sequencing on 141 Japanese individuals and constructed a miRNA‐eQTL catalog. We trained miRNA expression prediction models using this dataset and then applied the models to publicly available Japanese GWAS 49 (mean n  > 190,000). We found several significant associations, including that between miR‐1908‐5p and colorectal cancer.

Generally, only the genetic loci surpassing the genome‐wide significant threshold—typically set as p  = 5.0 × 10 −8 , accounting for the genome‐wide multiple testing correction 50 —are considered important in GWAS. However, the GWAS data contains important information beyond those genome‐wide significant loci. Indeed, in most GWAS to date, the genome‐wide significant loci explain only a fraction of the total heritability estimated from genome‐wide genotype data, 51 suggesting that a substantial proportion of phenotype‐associated variants have too small individual effects to reach the stringent genome‐wide significance threshold. These subthreshold variants are informative to gain insights into etiological backgrounds (Figure  4A ). For example, by undertaking a comparative analysis of GWAS results of IGCTs 52 and that of TGCTs, 53 we found that IGCTs and TGCTs had similar genetic etiologies. Intracranial germ cell tumors are a group of pediatric brain tumors whose etiology is largely unknown due to their extremely low incidence. 54 Histologically, IGCTs are like TGCTs. Thus, we hypothesized that they could have similar genetic etiologies. Indeed, the TGCTs' 60 risk loci 53 (identified in a large‐scale GWAS of TGCTs among Europeans) showed similar trends in our IGCT GWAS. 52 This indicates a shared genetic basis of the IGCT and TGCT development, which may provide a rationale for devising TGCT treatment‐inspired therapies for IGCTs. Utilization of the genome‐wide polygenic signals. (A) Polygenic inheritance consists of numerous risk variants with small effects. Current genome‐wide association study (GWAS) sample sizes are insufficient to detect all of them at the stringent genome‐wide significance level. (B) Genetic correlation refers to the correlation of genetic effects on a pair of traits across the genome. (C) Polygenic risk score (PRS) is calculated by summing up the genotype × effect size of the single nucleotide polymorphisms (SNPs) in the PRS model across the genome. Incorporating subthreshold variants into PRS models improves their predictive performance. Generally, we can estimate genetic correlations between a given pair of traits based on the polygenic signals derived from the genome‐wide SNP genotypes (Figure  4B ). One such representative method is modeling phenotype pairs using bivariate linear mixed models. When genome‐wide SNP genotype data is available, the genetic correlation, the correlation of phenotypic effects of all causal genetic variants in the genome between a pair of traits, can be estimated through the application of appropriate statistical methods such as restricted maximum likelihood estimation. 55 , 56 Using this approach and two biobank datasets from Japan (BioBank Japan 57 ) and Europe (UK Biobank 58 ), we estimated the genetic correlations across 13 cancers, including biliary tract, breast, cervical, colorectal, endometrial, esophageal, gastric, hepatocellular, lung, non‐Hodgkin's lymphoma, ovarian, pancreatic, and prostate cancers. 59 Among these cancers, breast and prostate cancers showed a significant positive correlation both in the Japanese and European biobank datasets, which was further validated in Finnish biobank data (FinnGen). The subsequent meta‐analysis and pathway enrichment analysis of these two cancers revealed that the shared genetic susceptibility was caused by pathways related to sexual hormone responses. This approach can also be utilized to analyze the correlation between cancers and noncancerous traits. In another study, we undertook GWAS of five gynecologic diseases (namely uterine fibroid, endometriosis, ovarian cancer, uterine endometrial cancer, and uterine cervical cancer). 60 Genetic correlation analyses among these five diseases revealed that, except for uterine cervical cancer, the other four diseases showed positive correlations. Particularly strong genetic correlations were observed between endometriosis and ovarian cancer, ovarian cancer and uterine endometrial cancer, and uterine fibroid and ovarian cancer. These results are consistent with the etiology that the primary cause of uterine cervical cancer is infection of human papillomaviruses. Genome‐wide polygenic signals can be used to estimate the PRS. 61 The PRS is a score calculated by combining the effects of trait‐associated variants across the genome. During PRS estimation, the trait‐associated variants can be defined by a more relaxed significance threshold than the conventional genome‐wide significant threshold to better reflect overall genetic susceptibility (Figure  4C ). Indeed, in our study on PRS of 12 cancers (breast cancer, prostate cancer, colorectal cancer, skin cutaneous melanoma, lung adenocarcinoma, lung squamous carcinoma, uterine endometrial carcinoma, ovarian serous carcinoma, esophageal adenocarcinoma, glioblastoma multiforme, head and neck squamous cell carcinoma, and cervical cancer), incorporating genetic variants not reaching the commonly used genome‐wide significance level (i.e., p  < 5.0 × 10 −8 ) into PRS calculation models improved cancer risk stratification performances compared with only using genome‐wide significant variants. 62 In the study, we applied these PRS calculation models to TCGA data, calculated PRS values for the patients, and then evaluated the association of PRS with somatic alterations and clinical features. Higher PRS values were associated with earlier cancer onset and a lower burden of somatic alterations, which are most pronounced with prostate cancer. Recently, as the increase in GWAS sample sizes and power improves PRS accuracy, PRS is being increasingly regarded as a promising clinical instrument for cancer. 63 , 64 , 65 However, there is an important caveat to note here. Because most of the large‐scale GWAS currently available were undertaken in populations of European ancestry, the predictive value of PRS is much greater in European than non‐European people. 66 This low portability of PRS is due to the difference in LD structures and allele frequencies as well as the distribution of other trait‐specific factors between ancestries. 67 , 68 Whereas technical efforts, such as incorporating functional annotation into the model construction, 69 have been undertaken, increasing non‐European sample size in GWAS is the primary key to enabling equitable PRS portability. 70 , 71

Genome‐wide association study enables the identification of small‐effect but prevalent susceptibility variants throughout the human genome in a hypothesis‐free manner. These small‐effect, highly prevalent variants shape the polygenic inheritance of cancer, which can explain a substantial proportion of cancer heritability. As highlighted in this review, leveraging GWAS—at times in combination with multiple GWAS summary data and/or other omics layered analysis (such as transcriptomic and epigenomic analysis)—to estimate polygenic signals in various cancers will improve our understanding of their underlying genetic etiology. While existing GWAS predominantly focus on the binary status of disease onset (e.g., cancer development), analyzing genetic association with more refined or context‐specific phenotypes will be the key to developing better insights. For example, elucidating the genetic underpinnings of why patients respond differently to treatments may aid clinical decision‐making. 72 Similarly, constructing fine‐grained omics data will help in functional annotation of the GWAS‐implicated variants. Single‐cell sequencing technologies facilitate delineating transcriptomic and epigenomic diversity in the constituent cells of the tissue, leading to the characterization of cell type‐specific gene regulation. 73 Rare genetic variants must be considered while unraveling the pathophysiology of diseases and while predicting disease risks. For example, a Finnish study reported that estimated PRS and the rare pathogenic frameshift variants of PALB2 and CHEK2 additively modify the breast cancer risk. 74 Future large‐scale whole‐genome sequencing studies, which enable the comprehensive detection of both common and rare variants throughout the genome, will further clarify the utility of combining common and rare variants in the risk prediction of diseases. In future, increasing the breadth and depth of phenotype and genotype data will synergistically accelerate advancements in understanding disease biology.

Cancer has heritable components, as evidenced by epidemiological studies. 1 , 2 In the 1980–1990s, studies using candidate gene approaches and genetic linkage analyses of familial cancer successfully identified several cancer predisposition genes with high penetrance, such as BRCA1 / 2 3 , 4 and APC . 5 , 6 While pathogenic variants in these genes serve as a substantially important biomarker for the variant carriers, 7 the frequency of these high‐penetrance variants in the general populations is typically low (<0.5%), thereby explaining a limited proportion of the familial cancer risks in total. For instance, the presence of pathogenic variants of known cancer predisposition genes explained only 16% of the studied 5552 familial colorectal cancer cases. 8 Hypothetically, the remaining 84% of those cases were ascribed to polygenic inheritance—the inheritance pattern that is regulated by many genes with individually small effects. The advent of GWAS 9 has enabled one to map such small‐effect genetic variants; hitherto, researchers have performed several GWAS of various cancer types. 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 Genome‐wide association study is an agnostic experimental design that detects genotype–phenotype associations by comparing the allele frequency of genetic variants across the whole genomes of many individuals. Since its advent in the early 2000s, GWAS has identified more than 600,000 human genotype–phenotype associations, 18 and this number still continues to grow due to the decreasing costs of genotyping arrays and high‐throughput sequencing. These genotype–phenotype associations insights contribute to not only elucidating previously uncharacterized disease‐relevant genes and biological pathways but also facilitating successful drug discovery for a broad range of human diseases. 19 While early GWAS reports were often focused on identifying disease‐susceptibility loci, with the maturation of methodology and progress in data sharing, analytical methods that leverage GWAS results to derive further findings beyond individual susceptibility loci have been developed. Here, we present an overview of recent advancements in GWAS‐derived analytical methods and illustrate their applications to GWAS of cancers.

The authors declare no conflict of interest.