We used the most recent GWAS summary-level data from the Breast Cancer Association Consortium (BCAC) for breast cancer (122,977 cases and 105,974 controls) (Michailidou et al., 2017), the Ovarian Cancer Association Consortium (OCAC) for ovarian cancer (25,509 cases and 40,941 controls) (Phelan et al., 2017), the Prostate Cancer Association Group to Investigate Cancer Associated Alterations in the Genome (PRACTICAL) consortium for prostate cancer (79,148 cases and 61,106 controls) (Schumacher et al., 2018), the International Lung Cancer Consortium (ILCCO) for lung cancer (11,348 cases and 15,861 controls) (Wang et al., 2014), and the Oncoarray oral cavity and oropharyngeal cancer consortium for head/neck cancer (2,497 cases and 2,928 controls) (Lesseur et al., 2016).

To measure genome-wide genetic correlations for each cancer pair, we used the linkage disequilibrium (LD) score regression (LDSC) method (Schizophrenia Working Group of the Psychiatric Genomics Consortium et al., 2015). We applied pre-computed LD scores derived from ~1.2 million imputed variants from European populations that did not include the HLA region in the HapMap3 reference panel. LDSC controls for population structure using GWAS summary statistics without individual-level data.

We evaluated the partitioned genetic correlation across the five solid cancers within functional categories by using partitioned LDSC (ReproGen Consortium et al., 2015). We chose 11 functional categories as previously recommended (Zhu et al., 2019), including the DNase I digital genomic footprinting (DGF) region, DNase I hypersensitivity sites (DHSs), fetal DHS, intron, super-enhancer, transcription factor-binding sites (TFBS), transcribed region, and the histone markers H3K9ac, H3K4me1, H3K4me3, and H3K27ac from the Roadmap Epigenomics Project (Bernstein et al., 2010). Re-computed LD scores for variants classified in each particular annotation were used for estimating the cross-cancer genetic correlation within that functional group.

We estimated local genetic correlations between each pair of cancers in 1,703 pre-specified LD-independent regions using ρ-HESS (Shi et al., 2017). The goal of this method was to detect small contiguous regions of the genome in which the genetic associations of two traits are locally concordant, and to measure the local genetic correlation and p-values (p_ρ−HESS) between pairs of traits at local regions. Cancer pairs were considered to have genetic correlation at the local region if p_ρ−HESS passed the multiple testing correction (p_ρ−HESS < 0.05/1703).

For the cancer pairs with significant genome-wide genetic correlation, we conducted a pairwise cross-cancer meta-analysis by using PLEIO (Lee et al., 2021). The approach is based on a random-effect model, which can not only model genetic correlations across pairs of traits but can also correct for environmental correlations. It can seamlessly test multiple traits with various types by standardizing the effect sizes. Moreover, it maps pleiotropic loci through a variance component test and calculates statistical significance through an important sampling method. It overcomes the drawback of fixed-effect model methods such as ASSET (association analysis based on subsets) (Bhattacharjee et al., 2012). We conducted the cross-cancer meta-analysis on an Intel Xeon E5-2695 computer with the CPU operating at 2.10 GHz. This wastes ~10 min for each pair of cancers.

To separate the independent loci from the significant loci (p < 5 × 10⁻⁸), we used the clumping function in PLINK software (Purcell et al., 2007). SNPs with p < 1 × 10⁻⁵, an LD statistic r²>0.05, and a distance from the peak < 1,000 kb were assigned to the clump of that peak. Moreover, we set the NCBI human genome build 37 as the reference gene list.

We performed TWAS to identify gene–tissue pairs for each of the five solid cancers and used FUSION software based on the pre-computed 48 GTEx (version 7) tissue expression reference weights (Gusev et al., 2016). LD-reference data were derived from European descendants from the 1,000 Genomes Project. For each cancer, we conducted 48 TWASs, one tissue-cancer pair at a time. The false discovery rate (FDR) Benjamin–Hochberg procedure correction was used, and a result with an FDR < 0.05 was considered to be significant.

To validate our findings, we further conducted genome-wide genetic correlation analysis and cross-cancer meta-analysis of the five solid cancer GWAS datasets with the UK Biobank cohort from the IEU GWAS database project (Matthew et al., 2021): breast cancer (ID: ieu-b-4810), ovarian cancer (ID: ieu-b-4963), prostate cancer (ID: ieu-b-4809), lung cancer (ID: ieu-b-4954), and head/neck cancer (ID: ieu-b-4912). We applied the 1,000 Genomes Project variants (Phase 3) as the reference panel. The cross-cancer meta-analysis between each pair of replication datasets was implemented using the R software RE2C (Lee et al., 2017), which is another classical random-effect model-based method that tests heterogeneous effect size between individual summary statistics.

To gain biology insights from the shared risk genes, we performed Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis using the Enrichr web server (Kuleshov et al., 2016), which is a comprehensive resource for curated gene sets and a search engine that accumulates biological knowledge for further biological discoveries. The significant criterion is that the adjusted p-value is < 0.05.

We used STRING v10 (Szklarczyk et al., 2015) to analyze the PPI network. The basic assumption is that if two proteins are functionally associated, they may contribute to a common biological purpose. The interaction scores were derived from different sources, including experimentally determined interaction, database annotated information, and automated text mining knowledge.

A schematic overview of the present study is shown in Figure 1, that is, we estimated genome-wide, partitioned, and local genetic correlations of the five solid cancers. For the cancer pairs with significant genome-wide genetic correlation, we performed a cross-cancer meta-analysis to identify shared genetic loci. Finally, we conducted TWAS, pathway enrichment analysis, and PPI network analysis of the shared risk genes.

Among pairs of solid cancers, we found three pairs with positive genetic correlations at a significant threshold of p = 0.05: breast and ovarian cancers (r_g = 0.221, p = 0.0003), breast and lung cancers (r_g = 0.234, p = 7.6 × 10⁻⁶), and lung and head/neck cancers (r_g = 0.652, p = 0.010). The remaining pairs do not show significant genetic correlations (Table 1).

In the partitioned genetic correlation analysis, we observed significant genetic correlation in all 11 functional categories for the breast–lung cancer pair, with only two exceptions: Intron and SuperEnhance for the lung–head/neck cancer pair. As to the breast–ovarian cancer pair, there is no significant signal in H3K27ac, H3K4me3, H3K9ac, and SuperEnhance. The partitioned genetic correlations range from 0.033 to 0.546 (Figure 2; Supplementary Table S1).

We conducted ρ-HESS to investigate whether specific regions had a genetic correlation between each pair of the five solid cancers. The results show that the breast–ovarian cancer pair has a strong local genetic correlation in the 2q33 region (chromosome 2: 201576284-202818637, p = 8.83 × 10⁻⁶) (Figure 3A). In addition, three regions, including the 9p21 region (chromosome 9: 20463534-22206559, p = 6.71 × 10⁻⁶), 10q26 region (chromosome 10: 123231465-123900545, p = 4.26 × 10⁻⁷), and 11q13 region (chromosome 11: 68005825-69516130, p = 4.90 × 10⁻⁶), are found to have strong local genetic correlations in the breast–prostate cancer pair (Figure 3B). We did not observe any significant local genetic correlations for the other cancer pairs.

To assess the association of gene expression in specific tissue between each pair of the five solid cancers, we performed TWAS. A total of 1,669 gene–tissue pairs are significantly associated with breast cancer after Benjamini–Hochberg correction (Supplementary Table S2), in addition to 418 gene–tissue pairs with ovarian cancer (Supplementary Table S3), 1,116 gene–tissue pairs with prostate cancer (Supplementary Table S4), 155 gene–tissue pairs with lung cancer (Supplementary Table S5), and 15 gene–tissue pairs with head/neck (Supplementary Table S6). Among them, 306 gene–tissue pairs are overlapped for the breast–ovarian cancer pair, and the tissues involved are scattered; however, a number of genes are almost concentrated in the clumping region of rs4277389 on chromosome 17, such as CRHR1, LRRC37A, and MAPT (Supplementary Table S7). Moreover, 23 gene–tissue pairs are overlapped for the breast–lung cancer pair, and most of the gene signals are observed in the 1q22 region, especially gene GBAP1, which is simultaneously significant in eight tissues (adipose, artery, breast, fibroblast cell, sigmoid colon, transverse colon, esophagus, and vagina) (Supplementary Table S7). In addition, one gene–tissue pair (CFB-pituitary) is overlapped for the lung–head/neck cancer pair (Supplementary Table S7).

In the replication analysis, we confirmed the significance of the genetic correlation between the breast and ovarian cancer pair (r_g = 0.175, p = 0.0061), the breast and lung cancer pair (r_g = 0.125, p = 0.0018), and the lung and head/neck cancer pair (r_g = 0.506, p = 0.0005) in the UK Biobank. Then, we used cross-cancer meta-analysis (RE2C) to identify the shared genes between each of the three cancer pairs. For the breast–ovarian cancer pair, nine loci showed genome-wide significance. Of these, genes FGFR2, BNC2, ADAM29, ESR1, ATAD5, and TEFM were replicated when compared with their specific consortium results (Supplementary Table S8). Moreover, six loci demonstrated significance in the breast–lung cancer pair. Some genes were found to be replicated, such as MAP3K1 (rs12653202, pmeta=4.34×10-23), HIST1H family (rs13214023, pmeta=2.83×10-14), ASH1L (rs4971059, pmeta=5.47×10-9), and ZMIZ1 (rs7904249, pmeta=1.22×10-8) (Supplementary Table S9). In addition, we identified two loci shared in the lung–head/neck cancer pair, but neither was replicated (Supplementary Table S10).

We observed shared genes enriched in human T-cell leukemia virus 1 infection (HTLV-1) and antigen processing and presentation (APP) pathways. HTLV-1 was the first retrovirus discovered to cause adult T-cell leukemia (ATL), a highly aggressive blood cancer (Matsuoka and Jeang, 2011). The APP pathway is a key element for an efficient response to immune checkpoint inhibitor therapy, which can be exploited to enhance tumor immunogenicity and to increase the efficacy of immunotherapy. The use of immune checkpoint inhibitors has already shown significant clinical advances in a wide range of patients with cancer (D'Amico et al., 2022).

In total, we found 849 pairs of interaction in the PPI network (Supplementary Table S11). A total of 44 gene pairs have combined scores >0.95, in which the ESR1-NRIP1 pair has the highest score of 0.999. HIST1H family genes around the 6p22.1 region show strong interactions with high scores. We observed 26 genes with degrees >20, most of which are HIST1H family genes, in addition to ESR, HSPA4, TNF, and EHMT2 genes. HIST1H gene set expression was reported to be positively correlated with large tumor size, high grade, metastasis, and poor survival in patients with breast cancer (Liao et al., 2021), which were used as prognostic factors for survival prediction among patients with cervical cancer (Li et al., 2017). The PPI network for shared risk genes is shown in Figure 4.