Patterns of Associations with Epidemiologic Factors by High-Grade Serous Ovarian Cancer Gene Expression Subtypes

Among the 2,070 HGSC cases, 556 (27%) were classified as C1.MES, 340 (16%) as C5.PRO, 538 (26%) as C2.IMM, and 636 (31%) as C4.DIF ( Table 1 ). The distribution of HGSC subtypes was similar across study sites; however, there was larger variation among study sites with relatively few ( 2), OC use (ORs ∼ 0.6), 10+ years of OC use (ORs ∼ 0.45), 3+ full-term pregnancies (ORs = 0.63–0.78), >2 years breastfeeding (ORs = 0.40–0.61), and lifetime ovulatory years (OR ∼ 1.0; Table 2 ). We observed heterogeneity in associations across HGSC subtypes for several factors. A one-SD increase in the PRS was associated with increased odds of all subtypes (OR = 1.4) but most pronounced among C5.PRO (OR = 1.58; 95% CI = 1.41, 1.77). Having a first-degree relative with a history of breast cancer only was associated with higher odds of C4.DIF (OR = 1.44; 95% CI = 1.16, 1.78). Age at last pregnancy of 35+ was associated with lower odds of all subtypes (ORs ranged 0.72–0.85); however, for C2.IMM, ages at last pregnancy of 25+ were also associated with reduced odds (ages 25–29: OR = 0.73; 95% CI = 0.54, 0.97; ages 30–34: OR = 0.69; 95% CI = 0.51, 0.94). For menopausal hormone therapy use, ≥10 years of E-only hormone therapy was associated with increased odds of C1.MES (OR = 2.14), C2.IMM (OR = 2.00), and C4.DIF (OR = 1.90) but was attenuated for C5.PRO (OR = 1.18). History of endometriosis was associated with increased odds of C5.PRO (OR = 1.46; 95% CI = 0.98, 2.16) and C4.DIF (OR = 1.27; 95% CI = 0.94, 1.71) but not C1.MES or C2.IMM (OR = 0.92; 95% CI = 0.64, 1.32 and OR = 1.03; 95% CI = 0.72, 1.48, respectively). Associations between epidemiologic factors and HGSC C1.MES, C5.PRO, C2.IMM, and C4.DIF subtypes among the 2,070 cases and 16,633 controls from 11 case–control studies. NOTE: All models are adjusted for age at reference, number of live births, ever used OCs, and study site unless otherwise noted. Models are adjusted for age at reference, number of live births, and study site. Models are adjusted for age at reference, ever used OCs, and study site. Restricted to those who have had a pregnancy. Models are adjusted for age at reference, number of pregnancies, ever used OCs, and study site. Restricted to those who had a live birth. Reference defined as 5 years prior to diagnosis/interview date or 1 year prior to interview date. The mean age at diagnosis was similar for C1.MES, C5.PRO, and C2.IMM, ranging from 60.0 to 61.4 years, whereas C4.DIF tended to be younger at diagnosis (57.5 years; Table 3 ). In case-only analyses with C1.MES as the comparison group, results were similar to those from the case–control analyses. Having a first-degree history of breast cancer only was associated with increased odds of C4.DIF (OR = 1.39; 95% CI = 1.01, 1.90). One SD unit increase in the PRS was associated with higher odds of C5.PRO (OR = 1.14; 95% CI = 0.99, 1.31). Ages at last pregnancy of 25 to 29 and 30 to 34 were associated with lower odds of C2.IMM (OR = 0.65; 95% CI = 0.42, 0.99 and OR = 0.69; 95% CI = 0.44, 1.08, respectively) and a less pronounced but similar pattern for C5.PRO (ORs ranged 0.83–0.86). E-only hormone use for 10+ years was associated with lower odds of C5.PRO (OR = 0.49; 95% CI = 0.25, 0.95), with less pronounced associations for C2.IMM (OR = 0.84; 95% CI = 0.48, 1.45) and C4.DIF (OR = 0.81; 95% CI = 0.47, 1.40). History of endometriosis was associated with increased odds of C5.PRO (OR = 1.40; 95% CI = 0.81, 2.44) and C4.DIF (OR = 1.23; 95% CI = 0.76, 1.97). Case-only associations between epidemiologic factors and HGSC subtypes (C1.MES referent) accounting for the probability of subtype classification (entropy) among the 2,290 cases from 12 studies. NOTE: Models include SEA (case-only study). All models are adjusted for age at diagnosis, number of live births, ever used OCs, entropy, and study site unless otherwise noted. Models are adjusted for age at diagnosis, number of live births, and study site. Models are adjusted for age at diagnosis, ever used OCs, and study site. Restricted to those who have had a pregnancy. Models are adjusted for age at diagnosis, number of pregnancies, ever used OCs, and study site. Restricted to those women who had a live birth. Reference defined as 5 years prior to diagnosis date or 1 year prior to interview date. Models are adjusted for number of live births, ever used OCs, and study site. Regular reported aspirin use was associated with lower odds of C5.PRO (OR = 0.72; 95% CI = 0.48, 1.06) and was stronger for daily aspirin use (OR = 0.54; 95% CI = 0.32, 0.93) but was not associated with C1.MES, C2.IMM, or C4.DIF ( Table 2 ). Genital powder use was associated with increased odds of C2.IMM (OR = 1.42; 95% CI = 1.08, 1.86) but not C1.MES, C4.DIF, or C5.PRO. Current smoking was associated with increased odds of C4.DIF only (OR = 1.40; 95% CI = 1.11, 1.76). Former smoking was associated with higher odds of C2.IMM (OR = 1.25; 95% CI = 1.03, 1.51). Ever alcohol use was associated with lower odds of all subtypes (ORs = 0.42–0.69). Results from the case-only analyses with C1.MES as the comparison group were generally consistent with the case–control analyses. Regular aspirin use was associated with lower odds of C5.PRO (OR = 0.70; 95% CI = 0.43, 1.15) and was stronger for daily use (OR = 0.48; 95% CI = 0.25, 0.91). Genital powder use was associated with increased odds of C2.IMM (OR = 1.42; 95% CI = 0.94, 2.14). Current smoking was associated only with C4.DIF (OR = 1.34; 95% CI = 0.94, 1.91), and former smoking was associated with higher odds of C5.PRO and C2.IMM (OR = 1.29; 95% CI = 0.95, 1.76 and OR = 1.52; 95% CI = 1.16, 2.00, respectively). Ever use of alcohol was associated with increased odds of C5.PRO, C2.IMM, and C4.DIF (ORs = 1.31–1.46). The proportion of HGSC cases that were assigned their subtype with >80% probability varied by primary assignment; 74% of C1.MES, 60% of C4.DIF, 50% of C2.IMM, and 49% of C5.PRO cases reached that threshold. The mixtures of subtype proportions also differed by primary assignment (Supplementary Fig. S1). For example, for C1.MES and C5.PRO, the next most common subtype proportion was C2.IMM, and C1.MES samples had very low proportions of C4.DIF. C4.DIF samples had larger proportions of C1.MES and also substantial proportions of C2.IMM. C2.IMM tended to have a higher proportion of C5.PRO than the other subtypes. Results from sensitivity analyses restricting to HGSC cases to those assigned a subtype with >80% are presented in Supplementary Table S2 and were generally comparable with those presented in Table 2 , albeit with less statistical precision due to reduced sample size, suggesting that intratumoral heterogeneity of subtypes had minimal effect on the results. The results restricted to White individuals were unchanged (Supplementary Table S3). Figure 1 illustrates a heatmap for patterns of associations of dichotomized epidemiologic risk factors by HGSC subtype. Hierarchical clustering split the four subtypes into two major groups. C2.IMM and C4.DIF subtypes clustered together most closely (Pearson correlation of 0.87). C5.PRO showed the most different pattern of associations compared with the other subtypes, although the overall Pearson correlation between C5.PRO and C2.IMM/C4.DIF was still high at 0.82. Heatmap associating epidemiologic factors with HGSC subtypes among 2,070 cases and 16,633 disease-free controls in the 11 pooled case–control studies. The dichotomized variables included OC use (ever vs. never), number of full-term pregnancies (3+ vs. none), age at last pregnancy (35+ vs. < 25 years), breastfeeding (2+ years vs. never), age at menarche (14+ vs. ≤11 years), E-only hormone therapy (10+ years vs. never), endometriosis (yes vs. no), tubal ligation (yes vs. no), BMI at 18 years of age (≥25 vs. < 25 kg/m 2 ), regular aspirin use (any vs. none), genital powder use (any vs. none), current smoking (vs. never), former smoking (vs. never), alcohol use (ever vs. never), first-degree family history of ovarian cancer (vs. none), and first-degree family history of breast cancer only (vs. none). BrCa, breast cancer; HRT, hormone replacement therapy; Hx, history; OvCa, ovarian cancer. The heatmap shows consistent, strong increased odds of all subtypes associated with family history of ovarian cancer and strong decreased odds of all subtypes associated with OC use, three or more pregnancies, having breastfed for more than 2 years, and ever use of alcohol. For C1.MES, the only additional factor associated with increased odds was E-only hormone therapy use for 10+ years. For C5.PRO, duration of breastfeeding and regular aspirin use were associated with decreased odds and endometriosis was associated with increased odds. The association between E-only hormone therapy and C5.PRO was much weaker than for the other subtypes. Genital powder use and former smoking were uniquely associated with increased odds of C2.IMM, and E-only hormone therapy was also strongly associated with increased odds of C2.IMM. For C4.DIF, current smoking and family history of breast cancer were uniquely associated with increased odds and similar to C5.PRO, endometriosis was associated with increased odds. Like C1.MES and C2.IMM, E-only hormone therapy was strongly associated with increased odds of C4.DIF.

In this study, several epidemiologic factors were similarly associated across HGSC subtypes, including increased risk associated with family history of ovarian cancer and decreased risk associated with OC use, having three or more pregnancies, having breastfed for more than 2 years, and ever use of alcohol—suggesting common biological effects. Still, the magnitude of associations for some factors varied across HGSC subtypes, specifically highlighting the potential role of past smoking history and genital powder use in risk of C2.IMM; older age at diagnosis and history of endometriosis in increased risk and daily aspirin use in reduced risk of C5.PRO; and younger age at diagnosis, smoking history, family history of breast cancer only, and endometriosis in risk of C4.DIF. We also illustrate that risk factor profiles for C4.DIF and C2.IMM are more similar to each other than to those for C1.MES and C5.PRO. In the only other study on epidemiologic factors by HGSC subtypes, which was smaller ( n = 193) and was not restricted to HGSC, OC use and pregnancy history were similarly not strongly differentially associated with HGSC subtypes ( 39 ). In that study, Schildkraut and colleagues accounted for intratumoral heterogeneity by using the probability of subtype assignment in their modeling approach. They reported that family history of breast or ovarian cancer was more likely to be observed among C2.IMM, and we also observed that family history of ovarian cancer with or without breast cancer was most strongly associated with C2.IMM. These findings are consistent with the observation that a family history of either ovarian or breast cancer is associated with less aggressive disease (defined as surviving for five or more years after diagnosis), as survival is better for C2.IMM ( 28 ). George and colleagues ( 60 ) reported that C2.IMM is more common in tumors with germline or somatic aberrations in BRCA1 than in BRCA2 . We observed that history of breast cancer without ovarian cancer was associated with C4.DIF only ( 39 ), which aligns with the observations from the original PrOTYPE article that BRCA1 and BRCA2 mutations were most common (∼37%) in C4.DIF ( 35 ). Indeed, our reanalysis of the data in George and colleagues ( 60 ) shows that the prevalence of germline BRCA1 and BRCA2 mutations was highest in C4.DIF (25%). Chen and colleagues ( 31 ) reported that patients with HGSC classified as C4.DIF were on average approximately 4 to 9 years younger at diagnosis than the other subtypes, which is consistent with tumors diagnosed among women with BRCA mutations. In the current article, we also observed that C4.DIF was diagnosed approximately 3 years younger on average than the other HGSC subtypes. Many of the subtype-specific associations observed may be due to inflammation-related pathways. Current smoking has been reported to be associated with highly aggressive disease (death within 1 year of diagnosis; ref. 28 ), but our study suggests associations with C2.IMM and C4.DIF, which have better survival. We also observed different associations for current versus former smokers, which may relate to the recovery of the immune response among former smokers, misclassification with self-reported data, differences in smoking duration, or time since smoking cessation ( 15 , 18 , 61 ). A recent study using data from the Nurses’ Health Study reported that early-life exposure to cigarette smoke was associated with changes to the tumor immune microenvironment (including activation of cytotoxic T cells; ref. 62 ), which may explain the results observed for the association between smoking history and C2.IMM. Smoking has also been associated with certain cytokines and MUC16 expression ( 63 – 65 ), which are characteristic of C4.DIF. Recent evidence from cohort studies has indicated increased risk of HGSC among those with a history of endometriosis ( 26 , 66 , 67 ), complementing the results observed in the current study between C5.PRO and C4.DIF subtypes. We observed that ever use of alcohol was associated with decreased risk of all HGSC subtypes, but prior studies of alcohol consumption have been mixed. In the pooled analyses in Kelemen and colleagues ( 68 ), several of the larger studies, included in that article and in the current analysis, observed a decreased risk of EOC associated with ever use [e.g., Diseases of the Ovary and their Evaluation (DOV), Australian Ovarian Cancer Study (AUS), and New England Case-Control Study (NEC)]. Results for ever use of alcohol were not presented by histotype, but they reported a 12% decreased risk of HGSC associated with consumption of more than two drinks per day. An inflammation-related risk score was developed in OCAC using 12 epidemiologic factors (alcohol use, aspirin use, other NSAID use, BMI, environmental smoke exposure, history of pelvic inflammatory disease, polycystic ovarian syndrome, endometriosis, menopausal hormonal therapy, physical inactivity, smoking status, and talc use), which was shown to be associated with ovarian cancer mortality ( 69 ). This study provides further evidence that prediagnostic behavioral and lifestyle factors likely affect inflammation, the development of the tumor immune microenvironment, and the immune response, which are also reflected in the results of the current article. Intratumoral heterogeneity is a concern. The PrOTYPE algorithm assigns a probability for each of the four HGSC subtypes for every tumor (summing to 100%), and the final subtype assignment was based on the subtype with the highest probability. Nearly all samples in our study show probabilities of multiple subtypes (Supplementary Fig. S1), which has been observed previously ( 31 , 32 , 35 ). We accounted for intratumoral heterogeneity by restricting analyses to samples with a probability of a subtype assignment of >80% and by performing case-only analyses controlling for the confidence of subtype assignment. Although a substantial proportion of cases were excluded in the sensitivity analysis (41%), Supplementary Fig. S1 demonstrates that the majority of HGSC cases are classified with a >50% probability for one subtype and that the contribution of the other three subtypes is typically low. There has been some debate in the literature about the optimal number of HGSC molecular subtypes, with studies reporting two to five HGSC subtypes, using various methods and data sources ( 29 – 38 ). Konecny and colleagues compared survival patterns between classifications of three and four subtypes and observed that there were larger survival differences between the four subtypes, leading the authors to conclude that four subtypes were more clinically relevant. Indeed, most studies have reported similar differences in survival across the four subtypes, regardless of the methods used to define them. Improved precision in molecular subtyping will likely be addressed through analysis of single-cell RNA sequencing and will help clarify the relative contributions of gene expression patterns in the tumor versus the tumor microenvironment ( 29 – 38 , 70 ). This study has additional limitations. First, we analyzed 11 case–control studies, all of which had variation in data collection, question administration for exposure information, patterns of missingness, and selection of controls, which may influence the reported results. However, because ovarian cancer is rare, pooled analyses and consortium-based studies have been highly effective at understanding ovarian cancer etiology. Second, despite this being the first and largest study to date, many of the estimates were statistically imprecise because of the small sample sizes within subtypes, potentially contributing to chance findings. Still, all of the epidemiologic factors evaluated have been previously studied in relation to ovarian cancer and have been shown to be associated with an increased risk or decreased risk of HGSC. These established relationships provide evidence of an association, which supports our approach to examine the estimates and their precision, as well as the decision to not account for multiple testing. Therefore, additional studies in larger populations will be necessary to replicate the findings. Third, there could be residual confounding because we only controlled for age, parity, OC use, and study site. Fourth, the PrOTYPE assay was implemented in more than one iteration, and there may be batch effects that could explain variation across study sites; however, Ovarian Tumor Tissue Analysis has previously demonstrated that samples are classified as the same subtypes across batches ( 71 ). Finally, our study includes primarily individuals self-reporting White race, which may not generalize to all populations. Future studies are needed to further clarify the observed patterns of associations and will benefit from more diverse study populations ( 38 ). The observed patterns of similarities and differences in epidemiologic factors by biologically relevant subtypes provide information about the etiology of HGSC subtypes. Our study provides evidence that risk factor profiles could be important drivers of tumor heterogeneity that can influence survival and be used in risk modeling to identify individuals who are more likely to have aggressive tumors.

Epithelial ovarian cancer (EOC) represents a heterogeneous disease with distinct histotypes that have different cells of origin, molecular features, epidemiologic risk factors, clinical characteristics, and survival ( 1 ). The 2014 and 2020 World Health Organization guidelines ( 2 , 3 ) classify EOC into five main histotypes: high-grade serous carcinoma (HGSC), low-grade serous tumor, endometrioid tumor, clear-cell tumor, and mucinous tumor. HGSC is the most common histotype, representing approximately 70% of EOC diagnoses ( 4 , 5 ). Previous studies have identified associations between epidemiologic factors and EOC, overall and by EOC histotype ( 6 – 18 ). Factors associated with a decreased risk of EOC include oral contraceptive (OC) use, parity, having a full-term pregnancy after 35 years of age, breastfeeding, tubal ligation, and aspirin use ( 7 , 10 , 17 , 19 ). Those associated with an increased risk include age, family history of breast and ovarian cancers, polygenic risk score (PRS), lifetime ovulatory years, and estrogen (E)-only hormone replacement therapy ( 13 , 20 – 24 ). Several of these factors are associated with all EOC histotypes, including age, parity, and OC use, although the strength of the association varies across histotypes. Other factors have heterogeneous associations ( 6 , 11 , 18 , 25 ). For example, endometriosis is primarily associated with a risk of endometrioid and clear-cell tumors ( 11 , 26 ), and cigarette smoking is associated with an increased risk of mucinous tumor ( 18 ). Associations between many epidemiologic factors and HGSC are weaker than the corresponding associations with the other histotypes ( 6 , 27 ). Some epidemiologic factors have been associated with tumor aggressiveness, specifically. For example, high body mass index (BMI) and current smoking are associated with highly aggressive invasive serous EOC (death within 1 year of diagnosis), whereas family history of ovarian or breast cancer is positively associated and parity is negatively associated with less aggressive disease (lived five or more years; ref. 28 ). Tothill and colleagues ( 29 ) first reported molecular subtypes of HGSC in 2008, which were subsequently reproduced and refined across multiple studies ( 30 – 35 ). There are an estimated four gene expression–based subtypes of HGSC labeled using combined terminology from the Tothill and The Cancer Genome Atlas articles ( 29 – 36 ): mesenchymal (C1.MES), proliferative (C5.PRO), immunoreactive (C2.IMM), and differentiated (C4.DIF). These HGSC subtypes demonstrate distinct gene expression signatures and survival patterns, with C1.MES and C5.PRO having worse survival compared with C2.IMM and C4.DIF ( 29 – 38 ). To date, only one previous study has evaluated associations between epidemiologic factors and gene expression subtypes accounting for intratumoral heterogeneity ( 39 ). Clarifying risk factor associations with HGSC subtypes is necessary to identify targets for risk reduction and develop risk prediction models. In this study, we evaluated associations of reproductive and hormonal characteristics and demographic and lifestyle factors with HGSC subtypes ( 35 ).

This analysis includes data from 11 case–control studies from North America ( n = 7), Europe ( n = 3), and Australia ( n = 1; refs. 8 – 10 , 12 , 13 , 16 , 20 , 40 – 50 ). For cases, formalin-fixed, paraffin-embedded tumor samples were assayed using a 518-marker or 340-marker NanoString CodeSet developed through a collaborative effort in the Ovarian Tumor Tissue Analysis consortium to identify markers associated with survival ( 35 ). All studies include HGSC cases confirmed through pathology evaluation following the 2014 World Health Organization guidelines ( 51 ), a comparable disease-free control group, data on epidemiologic factors, and a sample size larger than 30 cases and 30 controls. Case samples were included if they were of adnexal or presumed adnexal origin, did not receive neoadjuvant treatment, and NanoString data passed quality control measures. In the 11 studies with 2,261 eligible cases and 17,278 controls, we excluded 191 cases and 645 controls because they lacked data on OC use and pregnancy information. Thus, we were able to evaluate data from 2,070 cases and 16,633 controls for analyses. Additionally, we included data from a European study with 220 well-annotated cases ( 52 ), increasing the number for the case-only analysis to 2,290 cases ( Table 1 ). Each study received institutional review board approval and obtained written consent from participants. Distributions of controls and cases by HGSC subtype for the 12 studies. All cases were assigned a gene expression subtype using the PrOTYPE assay, a validated 55-gene NanoString-based classifier described in detail elsewhere ( 35 ). PrOTYPE classifies HGSC cases into four gene expression subtypes: C1.MES, C2.IMM, C4.DIF, and C5.PRO. A probability of assignment is generated for each of the four HGSC subtypes (summing to 100%), and the final subtype assignment is the subtype with the highest probability. Each PrOTYPE prediction also generates an entropy score. Entropy is defined as the uncertainty inherent in the prediction of subtype classification and is inversely correlated with the probability of assignment. Epidemiologic factors included reproductive and hormonal characteristics and demographic and lifestyle factors that have been associated with EOC, as well as modifiable factors that have been less consistently associated with EOC, primarily based on previous Ovarian Cancer Association Consortium (OCAC) studies. We included a previously derived and validated PRS( 27 ), family history of breast or ovarian cancer defined as no known first-degree family history of breast or ovarian cancer, first-degree relative with a history of breast cancer (but not ovarian cancer), or first-degree relative with a history of ovarian cancer (with or without a first-degree relative with breast cancer). Reproductive and hormonal characteristics included: OC use (ever or never), duration of OC use (never, <6 months, 6 months to <5 years, 5 to <10 years, or ≥10 years), number of full-term pregnancies (0, 1, 2, or 3+), age at last pregnancy (6 months–2 years, or >2 years), age at menarche (≤11, 12–13, or ≥14 years), menopausal hormone therapy use (never, estrogen only for <10 years, E only for ≥10 years, E and progestin for <10 years, E and progestin for ≥10 years, use of both, and use of unknown hormone therapies; not restricted to postmenopausal women), lifetime ovulatory years ( 23 ), endometriosis (yes or no), and tubal ligation (yes or no). Demographic and lifestyle characteristics were also evaluated including age at diagnosis (<49, 50–54, 55–59, 60–64, 65–69, or ≥70 years; case-only analysis), BMI at 18 years of age and within 5 years of diagnosis/interview data (<18.5, 18.5–24.9, 25–29.9, or ≥30 kg/m 2 ), regular aspirin use (nonregular use or regular use), frequency of aspirin use (nonregular use, use <30 days/month, and use ≥30 days/month), genital powder use (never or ever), smoking status (never, former, or current), lifetime alcohol use (never, ever, or former), and recent alcohol use (no alcohol use in 5 years before reference date or any). Aspirin use, smoking status, and lifetime alcohol use were defined at the time of diagnosis for cases or interview/comparable reference date for the controls. Supplementary Table S1 illustrates variables included for each study site. If data were missing, then that study was excluded from the analyses of that variable. We used polytomous logistic regression to compute ORs and 95% confidence intervals (95% CI) associating epidemiologic factors with HGSC subtypes. Models were adjusted for age at diagnosis (cases) or at interview/reference date (controls), number of full-term pregnancies, OC use, and study site. For the two studies that did not collect information on the number of full-term pregnancies, we used number of live births as a proxy. As this study aimed to estimate heterogeneity in the associations between epidemiologic factors and HGSC subtypes based on the magnitude of the estimates and their associated precision, no hypothesis testing was performed ( 53 – 57 ). We performed two analyses to account for potential misclassification from assigning the subtype with the highest probability. First, we repeated the case–control analyses, restricting to cases with >80% probability of subtype assignment ( n = 1,228 cases; 59%). Second, we conducted a case-only analysis to compute ORs and 95% CIs associating epidemiologic risk factors with HGSC subtypes with adjustment for the entropy score using polytomous regression. This OR approximates the ratio of two subtype-specific ORs to elucidate the subtype-specific heterogeneity ( 58 ). We compared the C5.PRO, C2.IMM, and C4.DIF categories with C1.MES because C1.MES had the largest sample size and lowest entropy scores. To examine patterns of associations across HGSC subtypes, we performed unsupervised hierarchical clustering of the four subtypes with normalized β-estimates for the associations between dichotomized factors of interest and HGSC subtypes using complete linkage and uncentered correlation (Pearson coefficient; refs. 9 , 59 ). Nearly all of the cases and controls were self-reported White race or of European decent (93.3% cases and 90.4% controls). In a sensitivity analysis, we restricted the cases and controls to White individuals. Analyses were conducted using STATA (version 16.1, STATA Corporation) and R v4.0. The full individual patient data are not publicly available but can be requested through the existing data request processes of the OCAC ( https://ocac.ccge.medschl.cam.ac.uk/ ).

This table shows the availability of epidemiologic factors used in the analyses across the different studies. This table shows the associations between epidemiologic factors and HGSC subtypes restricted to HGSC cases with probability of subtype assignment >80%. This table shows the associations between epidemiologic risk factors and HGSC subtypes, restricted to White cases and controls Distribution of subtype proportions by subtype assignment