Background

Cancer is an important cause of worldwide morbidity and mortality. Breast cancer (BC) accounted for approximately one in four cancer cases among women in 2018 [1], and ovarian cancer (OC) is the second most common form and leading cause of death due to cancer in the female reproductive system [2]. Although hereditary factors can explain 5–10% of the risk for breast or ovarian cancer [1], non-hereditary factors remain the major drivers. One-third to two-fifths of new cancer cases could be avoided by eliminating or reducing exposure to known risk factors [1, 3,4,5]. Thus, primary preventive measures that can reduce risk of BC and OC by targeting intervention on complex causal biomarkers or pathways are of increasing interest.

Unobserved confounding and reverse causality limit the ability of epidemiological observational studies to identify causalities. Observational studies are also often limited in acquiring large-scale exposure measurements, whereas randomised clinical trials are not widely available because of ethical concerns, high cost and long duration [6]. Recently, thousands of summary-level statistics from genome-wide association studies (GWASs) have provided great opportunities to support wide-range causal findings. Mendelian randomisation (MR) uses genome-wide significant genetic variants as instrumental variables (IVs) that can accurately assess the causal effect and direction of one exposure on a specific outcome after ruling out unobserved confounders in theory (“causal” represents the causality in statistics). Genetic correlation (GC) analysis can identify the common genetic risk between two specific traits [7]. Identifying GCs can provide useful etiological insights and help prioritise likely causal relationships [7]. Combining GC with MR can be used to identify direct causal relations and shared genetic risks for an exposure–outcome pair.

To date, MR has been used to explore the effects of alcohol consumption and glycemic, lipids and obesity traits on ovarian and breast cancers [6, 8,9,10,11]. Although many causal biomarkers have been identified, focusing on a specific exposure can provide only limited evidence for primary prevention, compared with focusing on a wide range of potential biomarkers. In addition, genome-wide correlations between BC and OC and complex exposures remain unclear. There is intense interest identifying these associations because the shared genetic architecture and causal associations can provide valuable references for joint, precise and priority intervention targets. Although information has been obtained on many risk factors, prioritising these factors remains essential. Complex network pathways may also occur that lead from extensive exposures to the occurrence of cancer, which may involve mediators as potential targets. Si et al. used network-MR to study biomarkers and complex metabolic pathways [12]. With a network, upstream or downstream targets for a specific risk factor can be identified and also contribute to primary prevention.

In this study, large-scale genomic summary-level statistics were used in a comprehensive analysis to gain insight into the complex relations of 95 cancer-related factors and nine cancer types. The aim of the research was to screen for robust causal biomarkers of BC and OC and then identify shared or distinct risk factors, stratify risks by priority and construct visual causal networks to guide prevention measures.

Methods

Determination of exposures

Risk factors associated with human cancers were hypothesised to also have potential carcinogenic mechanisms in the occurrence of breast and ovarian cancers. To determine candidate exposures, all MR analyses (Supplementary Table S1) were reviewed for any cancer types, and only the factors with available summary-level datasets were used in this research. Details about the process are described in Supplementary Text 1. Ultimately, 95 complex traits were included in the study. Except for the established birth length (BIRL), birth weight (BIRW) and age at menarche (AAM), most factors could be modified. Data sources of the 95 exposures are listed in Supplementary Table S2.

Among the factors, literature review showed that only 32 factors had been explored for association with both BC and OC, only 18 for BC and only 1 for OC. The other 44 exposures have not been studied previously in BC and OC (Supplementary Table S2). Previous MR studies have not explored GCs, risk stratification and network pathways for these risk factors on BC or OC. Because of the lack of information, the exposures were examined as candidate factors for BC and OC in this research [13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48]. The 95 factors were further divided into the following categories: anthropometry traits (16), blood biochemistry traits (8), disease traits (5), lifestyle traits (14), lipids/glycemic traits (11), metabolites (13), nutrients (23) and sex-related traits (5) (see Supplementary Text 1 for details).

Data sources of breast and ovarian cancers

Summary statistics for 122,977 cases of breast cancer and 105,974 controls of European ancestry were acquired from a combined study including the Breast Cancer Association Consortium (BCAC), Discovery, Biology and Risk of Inherited Variants in Breast Cancer Consortium (DRIVE), Collaborative Oncological Gene-environment Study (iCOGS) and several other GWAS meta-analyses [49]. Summary statistics of two BC subtypes, oestrogen receptor-positive (ER + BC, 69,501 cases) and oestrogen receptor-negative (ER − BC, 21,468 cases), were also included.

Genetic associations with OC were obtained from the Ovarian Cancer Association Consortium using an Illumina Custom Infinium array (OncoArray) including 25,509 epithelial OC cases and 40,941 controls [50]. The OC cases were further divided into five major invasive histotypes: high-grade serous OC (HGSOC, 13,037 cases), low-grade serous OC (LGSOC, 1012 cases), invasive mucinous OC (IMOC, 1417 cases), endometrioid OC (EOC, 2810 cases) and clear cell OC (CCOC, 1366 cases). Detailed information about BC and OC research populations and sample sizes can be found in Supplementary Text 2 and Supplementary Table S3, as well as the original publications [49, 50].

Statistical analysis

Two-sample Mendelian randomization

The study graph model and flowchart are shown in Fig. 1. The univariate two-sample MR method was used to determine the causal effect of each exposure on a target outcome (Fig. 1a). In the MR analysis, genome-wide significant genetic variants were used as instrumental variables (IVs) to examine causal associations of exposures with OC or BC. The MR approach assumes that IVs (1) are associated with the candidate exposure, (2) are not associated with confounders (upper red cross) and (3) are associated with the outcome only through the candidate exposure but not through other pathways (lower red cross) [51].

Fig. 1: Study design and framework of this research.
figure 1

Panel a was the causal graph of Mendelian randomization (MR), this method should satisfy three assumptions: (1) the IV is associated with the exposure, (2) the IV affects outcome only through the exposure (lower red cross), (3) the IV is not associated with the confounders (upper red cross). Panel b was a framework of multivariable MR, which used a pairwise design to measure exposure Xm–Xn pair on the risk of the outcome. The causal effects of candidate exposure Xn on outcome were calculated in turns. Panel c showed the framework of genetic correlation, the correlation between factors (Xn) and ovarian/breast cancer was measured. Panel d indicated a network-MR framework. This framework utilised the IVs of Xm to acquire the effect of Xm on Xn (potential mediator) and outcome, then, the IVs of Xn were utilised to acquire the effect of Xn on the outcome. Panel e was the flowchart of these analyses. LDSC trait linkage disequilibrium score regression, OC ovarian cancer, BC breast cancer, MR Mendelian randomization, IVW inverse variance weighted, WME weighted median estimator, MVMR multivariable Mendelian randomization.

Full size image

Candidate IVs associated with a specific risk factor were determined at the standard threshold of genome-wide significance (P < 5 × 10−8). Furthermore, candidate IVs within the threshold of linkage disequilibrium (LD, r2 < 0.01) were pruned to keep nearly independent IVs. The proportion of variance in the exposure explained by the IVs (R2) was calculated as 2β2 × MAF × (1 − MAF), where β is the association of single-nucleotide polymorphisms (SNP) with the exposure and MAF is the minor allele frequency. The F statistic was calculated from the R2 statistic as F = [(N − K − 1) / K] × [R2 / (1 − R2)], where N is the sample size and K is the number of IVs [52]. Generally, an F value > 10 indicated a strong IV [12, 53].

To acquire the causal estimator, first, the Wald ratio (WRO) was used to estimate causal effects for each SNP. Then, the conventional inverse variance weighted (IVW) MR method was used to aggregate causal estimators of all SNPs for the principal analyses in this research [54]. Furthermore, the weighted median estimator (WME) method was applied simultaneously as one of the sensitivity analyses to assess the robustness of causal findings. The method produces robust estimates in the presence of some invalid genetic instruments (when the number of invalid IVs < 50%) [55]. For those exposures with one IV, only WRO results were reported. Additionally, testing for the intercept of the MR-Egger regression was used to assess horizontal pleiotropy. Significant results in both IVW and WME analysis were viewed as robust associations in this research.

Pairwise multivariate Mendelian randomization

The multivariable MR (MVMR) approach [56] is designed to assess the robustness of causality under the possible horizontal pleiotropy and acquire the direct effects of the interested exposure on the outcome. Currently, a consistent standard for covariates selection from large-scale exposures has not been determined. Because adjusting more covariates causes a sharp decline in the number of IVs, and to avoid unknown bias caused by overadjustment, the causal effect of a specific exposure was assessed by adjusting other exposures one by one (here, pairwise MVMR) (Fig. 1b). The exposures (or covariates) that were selected from the above univariate MR analysis were significant for BC, OC or their subtypes. Adjusted effects of the exposures were estimated by the standard IVW–MVMR method [56]. The pairwise MVMR adjusted for a wide range of covariates was also considered a sensitivity analysis in this research.

Genetic correlation

Cross-trait linkage disequilibrium score regression (LDSC) is a useful epidemiological tool to estimate the GC of two traits (Fig. 1c) [7, 57]. An LDSC analysis can rapidly screen for correlations among a diverse set of traits, without needing to measure multiple traits on the same individuals [7]. Genetic variants used in LDSC usually required whole genome-wide SNPs. To keep a consistent number of SNPs for traits from different consortiums, they were matched with a common SNP list that was used in previous work [58]. The SNP list was a file called “w_hm3.noMHC.snplist” that included ~1.2 million SNPs based on the HapMap 3 reference panel stored in “ldsc” software. It can also be download from the LD Hub website (http://ldsc.broadinstitute.org). In addition, LD Hub website also recommend using this SNP list to reduce the number of SNPs to improve computing performance. In this research, genetic variants of the candidate traits were extracted from the MR Base (https://www.mrbase.org/) or the corresponding consortiums based on the SNP list to retain the common 1.2 million recommended SNPs. Then, the direction of effect values (Z-values) of each SNP across all traits was adjusted to ensure they corresponded to consistent-effect alleles. In GC analysis, the genetic dataset of Europeans from 1000 Genomes was used as a reference to compute LD scores. The genetic correlation coefficient was termed rg, which ranged from −1 to 1. Details about this method are introduced elsewhere [7].

Network Mendelian randomization

Network-MR was used to investigate the intermediate phenotypes in causal pathways to help to construct causal networks from the detected risk factors to outcomes (Fig. 1d) [59]. Network-MR was based on the univariate MR approach to achieve point-by-point analyses for each component (exposure, mediator and outcome), which were robust biomarkers (also include the results of WRO for biomarkers with only one IV) in the principal MR analysis. The initial network-MR framework was composed of three separate two-sample MRs that included pairwise analysis of one trait on another trait, as described elsewhere [12]. The network was organised according to the following steps: (1) keep all significant factors (robust association) of BC and OC (or subtypes) from the univariate MR analysis, (2) perform MR analysis for these factors with one another, (3) select the robust associations (both IVW and WME, or WRO) among the factors and (4) connect the significant exposure–exposure pairs and the previous exposure–outcome pairs to construct the final network.

Study procedures

The study flowchart is shown in Fig. 1e. First, the univariable MR method (including IVW, WME and WRO methods) was used to determine the causal effect of each exposure on a target outcome. Simultaneously, cross-trait LDSC was used to measure GCs across OC, BC and large-scale exposures. Then, pairwise MVMR was used to detect potential pleiotropy. Furthermore, the shared and specific risk factors for the outcomes were summarised, and the priority and rank of risk factors were defined. Finally, according to the network-MR design, causal networks for different outcomes were constructed to guide prevention practice.

Risk factors were prioritised at five levels: (1) level 1, robust MR evidence for both BC and OC plus GC evidence; (2) level 2, only robust MR results for both BC and OC; (3) level 3, robust MR evidence plus GC evidence for either BC or OC; (4) level 4, factors only robustly associated with BC or OC but without GC and (5) level 5, remaining MR evidence (only significant in one method), which was suggestive. In addition, a score was defined for each factor to indicate its importance, and then, all significant factors were ranked according to their importance. The score was acquired by adding the number of times a specific exposure was significant for the nine outcomes in the three univariate MR methods and the GC evidence. The level of a factor and its rank indicated the degree of robustness and universality, respectively.

All MR results were reported as odds ratios (ORs) and 95% confidence intervals (CIs) for genetically predicted per standard deviation or per unit increment of each risk factor. Bonferroni-corrected P-values (two-tailed) were used to show the significance of multiple testing. Two Bonferroni thresholds were used. The “moderate” one was P < 5.26 × 10−4 (0.05/95), which considered only the number of candidate exposures, whereas the “strict” one was P < 5.85 × 10−5 (0.05/855), which considered both the number of exposures and cancer subtypes. We picked the strict one for standard multiply testing. P-values that exceeded the strict Bonferroni threshold but were less than 0.05 indicated a suggestive association. All statistical analyses for MR were performed in the R software v 3.6.2. R package “TwoSampleMR” was used for two-sample MR analysis. Large-scale LDSC analyses were performed using the ‘ldsc’ software with the Linux system.

Results

Figure 2 shows the associations of 95 genetically determined risk factors with BC in the IVW method or WRO analysis. Twenty-three exposures were significantly associated with overall BC. Positive associations [OR (95% CIs)] were with adult height (ADUH) [1.06 (1.02–1.10)], C-reactive protein (CRP) [1.09 (1.03–1.15)], platelet count (PLT) [1.04 (1.00–1.08)], schizophrenia [1.07 (1.03–1.10)], chronotype [1.19 (1.08–1.31)], high-density lipoprotein cholesterol (HDL-C) [1.10 (1.04–1.15)], apolipoprotein A1 (Apo A1) [1.07 (1.02–1.11)], insulin-like growth factor-1 (IGF-1) [1.08 (1.03–1.13)], omega-6:total fatty acids (O6:TFA) [1.06 (1.01–1.12)] and age at menopause (ANM) [1.05 (1.03–1.07)]. Negative associations were with body fat mass (BFM) [0.92 (0.87–0.98)], childhood obesity (COBE) [0.82 (0.76–0.88)], hip circumference (HC) [0.79 (0.65–0.94)], body mass index (BMI) [0.72 (0.65–0.80)], waist circumference (WC) [0.72 (0.60–0.86)], waist-to-hip ratio (WHR) [0.72 (0.55–0.94)], comparative body size at age 10 (CBS-10) [0.60 (0.54–0.67)], average acceleration (AVEA) [0.93 (0.89–0.98)], education attainment (EDUA) [0.91 (0.85–0.98)], age of smoking initiation (AOSI) [0.67 (0.51–0.89)], overall activity (OACT) [0.46 (0.24–0.91)], interleukin-6 receptor subunit alpha (IL-6 sRa) [0.98 (0.96–1.00)] and IGF-1 receptor (IGF-1R) [0.82 (0.76–0.88)]. Of the significant factors, COBE, BMI, CBS-10, IGF-1R and ANM passed the strict Bonferroni correction, whereas WC and schizophrenia also passed the correction when only considering the number of exposures. Supplementary Figures S1 and S2 show the results for ER + BC and ER − BC.

Fig. 2: Association of 95 genetically determined risk factors with breast cancer.
figure 2

The odds ratio (OR) represented the effect of genetically predicted per unit increase in the risk factor. The majority of units were standard deviation (SD) and part of them were original units since the SD values were not available. Several R2 and F statistics were calculated by using the effect allele frequency in the outcome dataset since the missing allele frequency of exposure. For the number of SNP more than 1, the results were from the inverse variance weighted (IVW) Mendelian randomization; for the number of SNP equal to 1, the results were from the Wald ratio. * denotes P < 0.05, † denotes P < 5.26 × 10−4 (0.05/95), ‡ denotes P < 5.85 × 10−5 (0.05/855), # denotes potential pleiotropy in testing for the intercept of MR-Egger regression.

Full size image

The relations between 95 traits and overall OC in the IVW method or WRO analysis are shown in Fig. 3. Fourteen exposures were significantly associated with overall OC. Positive associations [OR (95% CIs)] were with HC [1.24 (1.07–1.44)], WC [1.24 (1.05–1.46)], CBS-10 [1.20 (1.04–1.39)], BMI [1.19 (1.05–1.35)], body fat percentage (BFP) [1.18 (1.06–1.31)], BFM [1.14 (1.06–1.24)], basal metabolic rate (BMR) [1.13 (1.03–1.23)], body fat-free mass (BFFM) [1.12 (1.02–1.23)], schizophrenia [1.07 (1.02–1.12)] and omega-6:omega-3 fatty acids (O6/O3) [1.12 (1.02–1.24)]. Negative associations were with thyroid-stimulating hormone (TSH) [0.82 (0.67–0.99)], EDUA [0.83 (0.73–0.94)], adiponectin [0.87 (0.76–0.99)] and IGF-1R [0.64 (0.55–0.74)]. Of the factors, IGF-1R passed the Bonferroni correction. Supplementary Figures S3 to S7 show results for OC subtypes.

Fig. 3: Association of 95 genetically determined risk factors with ovarian cancer.
figure 3

The odds ratio (OR) represented the effect of genetically predicted per unit increase in the risk factor. The majority of units were standard deviation (SD) and part of them were original units since the SD values were not available. Several R2 and F statistics were calculated by using the effect allele frequency in the outcome dataset since the missing allele frequency of exposure. For the number of SNP more than 1, the results were from the inverse variance weighted (IVW) Mendelian randomization; for the number of SNP equal to 1, the results were from the Wald ratio. * denotes P < 0.05, † denotes P < 5.26 × 10−4 (0.05/95), ‡ denotes P < 5.85 × 10−5 (0.05/855), # denotes potential pleiotropy in testing for the intercept of MR-Egger regression.

Full size image

Forty-eight factors were significant for at least one cancer type in the IVW method or WRO analysis. Supplementary Figures S8 and S9 show the results of pairwise MVMR for BC and OC and the 48 factors. Overall, the associations were relatively robust after adjusting for other potential pleiotropy factors in turn. Effects of OACT, diet fat (D-Fat), folate and N3 docosapentaenoic acid (N3-DPA) could change to some extent when adjusted for other factors, which suggested potential horizontal pleiotropy.

Figure 4 summarises the robust associations (significant in both IVW and WME methods) for BC, OC and their subtypes. In general, 16 factors were identified as robust traits for overall BC, including chronotype, HDL-C, IGF-1, Apo A1, schizophrenia, O6:TFA, ANM, PLT, AVEA, BFM, EDUA, COBE, HC, BMI, WC and CBS-10 (Fig. 4a). In addition, CRP, O6:TFA, Apo A1, ADUH, IGF-1, schizophrenia, ANM, monounsaturated fatty acids (MUFAs), BFM, AVEA, COBE, WC, BMI, WHR and CBS-10 were robustly associated with ER + BC (Fig. 4b). Associated with ER − BC were HDL-C, schizophrenia, ANM, BMR, cognitive performance (CP), BFFM, COBE, EDUA, HC, BMI, WC and CBS-10 (Fig. 4c). For ER + BC, ANM, BFM, WC, BMI and CBS-10 passed the strict multiple testing, as did COBE, EDUA, BMI, WC and CBS-10 for ER − BC. For overall OC, WC, CBS-10, BMI, BFP, BFM, O6/O3 and EDUA remained significant after the sensitivity analysis by the WME method (Fig. 4d). For OC subtypes, CBS-10, linoleic acid (LA), O6FA, BFM, zinc and BIRW had robust associations with IMOC. Causally associated with EOC were HC, BMR, BFFM, O6/O3, BFM, ANM, BIRW and N3-DPA. The factors BMR and BFFM also passed multiple testing when considering the number of exposures. In addition, the BFP, BFM and adiponectin were robustly associated with HGSOC, and the BFFM, BMR, sex hormone-binding globulin (SHBG) and O6/O3 were also robust factors for CCOC.

Fig. 4: Summarised robust associations for overall BC, OC and their subtypes.
figure 4

The odds ratio (OR) represented the effect of genetically predicted per unit increase in the risk factor. Only factors both significant in IVW and WME method could be shown. The panel a-h showed the results of overall BC, ER+ BC, ER- BC, overall OC, IMOC, EOC, HGSOC, and CCOC, respectively. The results for low-grade serous ovarian cancer were not shown since no robust risk factors were detected. * denotes P < 0.05, † denotes P < 5.26 × 10−4 (0.05/95), ‡ denotes P < 5.85 × 10−5 (0.05/855). IVW inverse variance weighted, WME weighted median estimator.

Full size image

Figure 5 summarises results for MR, GC and levels and rank of risk factors. Fifty-eight risk factors were significant for at least one type of cancer with at least one MR method (IVW, WRO or WME) (Fig. 5a). Thirty-one of the risk factors were robust factors for at least one BC/OC type (dark red). Moreover, many traits were shared risk factors across BC, OC and their subtypes. For example, BMI, BFM, CBS-10, HC, WC, schizophrenia, EDUA and IGF-1R were shared risk factors for overall BC and OC. In addition, 13 factors that shared genetic risk with these outcomes were also combined in the heat map, including OACT, COBE, BMI, BFFM, BMR, AVEA, CBS-10, schizophrenia, ANM, ADUA, CP, EDUA and insomnia (solid triangle). All significant results of GC analysis are shown in Fig. 5b. According to the results of MR and GC, five levels of risk factors were defined for BC and OC (Fig. 5c). Level 1 represented the robust MR plus the GC evidence for both BC and OC, which included only EDUA. Level 2 comprised only the robust MR results for both BC and OC, which included BMI, CBS-10, WC and BFM. In level 3, only COBE, AVEA, ANM, and schizophrenia were robustly associated with BC, and no factors were associated with OC. The above OACT, D-Fat, folate and N3-DPA, which showed potential horizontal pleiotropy in MVMR analysis, were mainly classified into level 5. Factors in level 5 were relatively unimportant compared with factors in other levels. Figure 5d shows the rank of the 58 significant factors according to their number of significant results (scores). The top 10 factors in order were CBS-10, EDUA, schizophrenia, BFM, BMI, ANM, WC, COBE, HC and O6/O3.

Fig. 5: Summarised results for MR, GC and the levels of identified risk factors.
figure 5

Panel a was the heat map of MR results. The coloured region represented the significant result in different methods including the IVW, WME and IVW + WME, while the white represented insignificant results. The direction of the triangle represents the direction of causal estimators where the upward triangle indicates positive association and the downward triangle indicates negative association. The solid triangle indicates a significant genetic correlation. Panel b was the result of genetic correlation (GC) analysis. The GC was quantified by the statistic of genetic correlation coefficient r, ranged from −1 to 1, and presented from blue to red colour. GC more than 0 represented a positive association and smaller than 0 represented a negative association. The asterisks (*) in the figure represent statistically significant results (P < 0.05). The results of low-grade serous ovarian cancer were not shown in GC analysis because the ldsc software failed to perform GC analysis on account of low h2 statistics or a small sample size. The insignificant factors for any outcomes were also not shown in this heat map. Panel c was the stratification of putative risk factors according to our criteria. The left one showed the factors for overall BC and OC only. The right one showed the factor for any BC and OC types. Panel d was the rank of these putative causal risk factors according to their scores defined by our criteria in the methods. The full annotation of the abbreviations in this Figure could be found in panel a or supplement Table S2.

Full size image

Causal pathways and networks were developed from identified risk factors to BC (Fig. 6a) and to OC (Fig. 6b). In the networks, HC, WC, O6:TFA, IGF-1R, BFM, CBS-10, chronotype, ANM, IGF-1, Apo A1, PLT, AVEA, COBE, EDUA and BMI had both direct and indirect effects on BC, whereas BFM, CBS-10, BFP, O6/O3, EDUA, BMI, WC, IGF-1R and TSH affected the risk of OC through their respective pathways. For example, CBS-10 could affect the risk of OC by acting on BFM, WC, BMI and BFP (yellow pathways), suggesting that early-life body status could act on later-stage stature and lead to the risk of OC. In addition, the effect of EDUA on OC could also be mediated by obesity-related traits (grey pathways), indicating education could drive health-related behaviour to control obesity and ultimately modify the risk of OC. Causal pathways of BC and OC subtypes are shown in Supplementary Figs. S10 and S11. Causal estimators of exposure–outcome pairs between each node in the causal network diagrams are shown in Supplementary Table S4. The other independent factors were not shown in networks because they had no identifiable mediators that could be explained as having only direct causal effects on cancer.

Fig. 6: Causal pathways and networks of identified risk factors for breast and ovarian cancer.
figure 6

The left panel was the network for breast cancer and the right one was for ovarian cancer. The arrows represented the direction of causal effect from one causal biomarker to another one, and finally, to ovarian/breast cancer (yellow highlight) from the network-MR analysis. This network only showed the factors with identifiable mediators. Each arrow represents a significant IVW-MR result.

Full size image

Discussion

In this study, genetic statistical methods were used to identify the relations between large-scale cancer-related exposures and breast and ovarian cancers. Thirty-one exposures were robust risk factors for at least one type of BC or OC. Among them, BMI, BFM, CBS-10, WC and EDUA were shared robust risk factors for overall BC and OC, which implied potential joint intervention targets. In addition, 13 shared factors were detected in both GC and MR analyses, including OACT, COBE, BMI, BFFM, BMR, AVEA, CBS-10, schizophrenia, ANM, ADUA, CP, EDUA and insomnia. Furthermore, risk factors were stratified into five levels and ranked in order to prioritise future intervention measurements. Finally, visual causal networks were constructed that showed potential causal pathways from identified large-scale exposures to target outcomes to guide primary prevention practices.

Of the 31 putative robust factors from MR analysis, 16 factors were new causal biomarkers, including WC, BMR, BFP, BFFM, HC, COBE, BFM, PLT, CP, Apo A1, O6FA, O6:TFA, O6/O3, LA, MUFAs and N3-DPA. The other 15 factors, including BMI, WHR, ADUH, CBS-10, BIRW, CRP, schizophrenia, AVEA, EDUA, chronotype, HDL-C, adiponectin, IGF-1, ANM and SHBG, have been previously reported to be causal biomarkers. [8, 9, 11, 60,61,62,63,64,65,66,67,68,69,70,71,72,73] Consistent with previous studies, in this study, AUDH, schizophrenia, HDL-C, and IGF-1 were positively associated with BC and BMI and schizophrenia were positively associated with OC [8, 60, 63, 65, 67, 69]. In addition, BMI, CBS-10, AVEA and EDUA were negatively associated with BC, also consistent with previous studies [11, 62, 70, 73]. Same with previous studies [9, 61, 65, 68, 69], significant causal associations of BIRW and adiponectin with BC and WHR, ADUH, BIRW, CRP, adiponectin, IGF-1, ANM and SHBG with OC were not detected in this study. However, additional negative associations were detected, including WHR with BC and ER + BC, BIRW with IMOC and EOC and adiponectin with ER − BC, OC and HGSOC. Additional positive associations included ADUH with CCOC; CRP with BC and ER + BC; ANM with BC, ER + BC, ER − BC, and EOC and SHBG with CCOC, although some associations were only significant in the IVW method.

Of the robust risk factors, COBE, BMI, BFFM, BMR, AVEA, CBS-10, schizophrenia, ANM, ADUA, CP and EDUA also showed significant GCs, which suggested a higher level of causal evidence for which they could be stratified as priority intervention targets (all of these factors were classified into levels 1, 2 and 3). A factor that has a GC with cancer is more worthy of attention because it shares genetic risks with the outcome, and thus, an individual with such a trait would have a higher cancer risk than other people at the genetic level, even if it was not a causal risk factor. For a causal risk factor with GC, a comprehensive intervention should be implemented that intervenes not only with the particular factor but also with other risk factors (including those upstream, downstream and in other pathways in a network). For example, schizophrenia was a causal risk factor and had strong GC with BC (Fig. 5a), indicating that an individual with schizophrenia was at a genetically higher risk of BC. In this situation, intervention for only schizophrenia could not rule out additional genetic risk of BC, compared with other conventional risk factors. Therefore, a comprehensive intervention should target alternative pathways, such as BMI, AVEA, IGF-1, WC, HC and O6:TFA et al (Fig. 6). Moreover, although GC analysis showed significant results for coffee consumption (CC), childhood BMI (CBMI), alcohol consumption (AC), BIRL, Apo B, allergies, cigarettes per day (CPD), low-density lipoprotein cholesterol (LDL-C), sleep duration (SDU), triglycerides (TG) and depression with several outcomes, MR studies did not support a significant causal effect of those factors. Therefore, they were not included as intervention targets. Although GC could increase the priority of a causal risk factor, intervention could fail to decrease cancer risk because the factor did not have a causal effect.

The newly identified risk factors O6FA, O6:TFA, O6/O3 and MUFAs had a strong causal effect on OC, BC or several subtypes. These results suggested an adverse effect of O6FA and a potential protective effect of omega-3 fatty acids (O3FA). Causal evidence of fatty acids on cancer risk is limited, but in their review, Saini et al. [74] noted that O3FA and O6FA suppressed and induced inflammation, respectively. Diets enriched in O6FA are associated with inflammation, which provides an ideal tumour microenvironment and is linked to cancer risk and metastasis [74]. By contrast, O3FA help to resolve inflammation and alter the function of vascular and carcinogen biomarkers and thus reduce cancer risk [74, 75]. These differences may explain why different fatty acids can increase or decrease BC or OC risks. In addition, the results in study are consistent with the previous opinion that the ratio of O6FA to O3FA is more crucial than the absolute amounts [74].

Notably, there was dimorphism in obesity-related traits for OC and BC risks. This causal evidence suggests double-sided effects of obesity traits for different female cancers and indicates the importance of maintaining a moderate body shape. The factor CBS-10 reflects the early-life status of an individual and also had strong causality and GC with cancer (ranked No. 1 among the factors), indicating early childhood intervention is key. The protective effect of education on lung cancer has been previously verified [76], and the results of this study further extend the protective effect of education to OC and BC. A relatively high level of education is more likely to drive positive health-related behaviours and thus reduce the risk of cancer. The robust inverse associations combined with the extensive GCs with both BC and OC indicate education is one of the top factors to be addressed in intervention. The protective effect of IGF-1R may be due to negative feedback regulation of IGF-1, which has been reported as a risk factor for breast cancer [64]. The risk of Apo A1 for BC was consistent with the reported positive effect of HDL-C [8], likely because Apo A1 is a transporter of HDL-C.

Some important results that are not consistent with those of other studies should also be noted. For example, with AAM, whereas negative associations with ER + BC and IMOC were indicated, the association was not significant with BC. By contrast, Day et al. [77] reported a strong negative relation with BC after removing or adjusting BMI-related SNPs. To attempt to reproduce that result, additional analyses were conducted by using different AAM datasets (versions 2014 and 2017), different BC datasets (Oncoarray, iCOGS, GWAS and combined), different LD thresholds (0.01 and 0.001), and different methods (IVW, WME and BMI-adjusted and BMI-excluded SNPs) (Supplementary Fig. S12). Several results in only specific conditions were consistent the conclusion of Day et al. The possible difference in conclusions could be because the BMI GWAS dataset could not be acquired that, together with the reported AMM 61 BMI-related SNPs (AAM increasing/BMI decreasing) produced the strong positive association with BC. However, the results in this study are consistent with their unadjusted results and are also consistent with those of another MR study by Qi et al. [78]. Therefore, future research should focus on the mechanism by which AAM combined with BMI affects the risk of BC and distinguish the data sources in MR studies.

Finally, based on the above evidence, risk factors were stratified and ranked in order to provide a reference for future primary prevention targets. A complex network for BC, OC and each subtype was constructed based on network-MR analysis. This study was the first to summarise and stratify causal biomarkers and construct causal networks for BC and OC using biostatistics and data-driven evidence. When using a network, stratified risk factors can be referenced to identify more important targets. For each factor, primary prevention interventions can be directed not only at direct effects but also to interrupt its pathway in other related nodes. For example, CBS-10 could be altered in childhood, which would benefit downstream adult BMI and ultimately decrease the risk of OC (Fig. 6). Furthermore, when a risk factor cannot be easily modified, such as lower EDUA for risk of BC, alternative downstream targets (e.g. Apo A1 and obesity-related traits) can be identified to implement primary prevention interventions.

Compared with previous studies, this study had the advantage of examining large-scale factors (the largest to our knowledge) associated with OC and BC under a comprehensive framework in order to detect causalities, genetic correlations, and shared or distinct factors and to prioritise risk factors and develop causal networks. The study also had limitations that should be noted. First, some of the candidate factors did not have enough IVs and might suffer bias from that weakness. Second, pleiotropy is a dilemma in MR. Fortunately, the WME was more robust to IV assumptions, and the pairwise MVMR design found that few of the identified risk factors were affected by pleiotropy. Third, in large-scale exploratory research of observational datasets, there are natural limitations to detailed exploration of mechanisms of each factor, as in most similar studies. Further research is required to determine biological mechanisms of the risk factors newly identified in this research, as well as to explore the suggestive evidence. The practical value of risk stratification and a causal network needs to be verified in further public health intervention practices.

Conclusions

Sixteen new factors associated with OC or BC were identified, including WC, BMR, BFP, BFFM, HC, COBE, BFM, PLT, CP, Apo A1, O6FA, O6:TFA, O6/O3, LA, MUFAs and N3-DPA. In addition, BMI, BFM, CBS-10, WC and EDUA were shared robust risk factors for overall BC and OC. Thirteen factors were significant in both GC and MR analyses, including OACT, COBE, BMI, BFFM, BMR, AVEA, CBS-10, schizophrenia, ANM, ADUA, CP, EDUA and insomnia. The risk factors were stratified into five levels and ranked to prioritise for future intervention measurements. Causal networks were developed that show pathways from putative factors to cancer in order to guide primary prevention.