A novel serum-based steroid-protein panels for differentiating ovarian cancer from non-malignant adnexal masses.

Between December 2021 and February 2025, a total of 103 participants were recruited, and preoperative morning serum samples were collected (Fig.  1 ). Among them, 47 (45.6%) had ovarian cancer, including 42 (89.4%) with primary ovarian cancer, one (2.1%) with recurrent disease, and four (8.5%) with metastatic disease from other primary sites. Participants with metastatic disease were excluded from the primary outcome analysis. Of the remaining 43 participants with ovarian cancer, 39 (90.7%) had epithelial tumors, with 32 (74.4%) classified as HGSOC, 2 (4.7%) as clear-cell, 2 (4.7%) as endometrioid, 2 (4.7%) as carcinosarcoma, and 1 (2.3%) as low-grade serous carcinoma. The remaining 4 cases (9.3%) had sex-cord stromal malignant tumors, including 3 (75%) with adult granulosa cell tumors and 1 (25%) with high-grade Sertoli-Leydig tumor. Regarding cancer staging, 7 (16.3%) cases were diagnosed at FIGO stage I, 2 (4.7%) at FIGO stage II, 26 (60.5%) at FIGO stage III, and 6 (14.0%) at FIGO stage IV; data was missing for two patients (4.7%). In terms of treatment status, 39 patients (90.7%) were treatment-naïve, four (9.3%) had been admitted after receiving neoadjuvant chemotherapy. Fig. 1 Study flow diagram. PCOS, polycystic ovary syndrome Study flow diagram. PCOS, polycystic ovary syndrome The control group consisted of 56 participants with non-malignant adnexal masses, including 8 (14.3%) with non-neoplastic adnexal masses, 33 (58.9%) with benign ovarian tumors, and 15 (26.8%) with borderline ovarian tumors. Among borderline ovarian tumors, 10 (66.7%) were FIGO stage I, one (6.7%) was FIGO stage II, and 3 (20%) were FIGO stage III. Table  1 provides an overview of the basic demographics and clinical characteristics of the cohort. Cases and controls differed significantly in age at operation, height, and history of oral contraceptive use. No significant differences were found for oral contraceptive use in the past 3 months, weight, BMI, menopausal status, smoking status, use of hormone replacement therapy (past or in the previous 3 months), or parity. A detailed description of the clinical and histopathological characteristics of the study participants is provided in Supplementary Table 1. Table 1 Demographics and clinical characteristics of participants, by primary outcome definition of ovarian cancer Diagnosis based on reference standard All participants ( n  = 99) Primary/recurrent ovarian cancer ( n  = 43) Non-malignant adnexal masses ( n  = 56) Age (years); p = 0.006 Median (IQR) 63.00 (53.00–68.00) 54.50 (47.00–64.00) 57.00 (50.00–65.50.00.50) Height (cm); p = 0.041 Median (IQR) 164.00 (159.20–167.00) 165.00 (162.00–170.00.00.00) 165.00 (160.00–169.00.00.00) Missing 5 (11.6%) 7 (12.5%) 12 (12.1%) Weight (kg) Median (IQR) 66.00 (59.00–75.00) 65.00 (60.00–80.25.00.25) 65.00 (59.00–75.75.00.75) Missing 5 (11.6%) 6 (10.7%) 11 (11.1%) Body mass index (kg/m 2 ) Median (IQR) 23.00 (22.00–27.90.00.90) 23.80 (22.00–28.10.00.10) 23.75 (22.00–28.00) Missing 6 (14.0%) 7 (12.5%) 13 (13.1%) Menopausal status Premenopausal 9 (20.9%) 21 (37.5%) 30 (30.3%) Postmenopausal 34 (79.1%) 35 (62.5%) 69 (69.7%) Race or ethnicity White 39 (90.7%) 52 (92.9%) 91 (91.9%) Missing 4 (9.3%) 4 (7.1%) 8 (8.1%) Smoking status Never 24 (55.8%) 32 (57.1%) 56 (56.6%) Current 13 (30.2%) 12 (21.4%) 25 (25.3%) Ex-smoker 3 (7.0%) 8 (14.3%) 11 (11.1%) Missing 3 (7.0%) 4 (7.1%) 7 (7.1%) History of hormone replacement therapy use No 34 (79.1%) 48 (85.7%) 82 (82.8%) Yes 5 (8.9%) 3 (5.4%) 8 (8.1%) Missing 4 (7.1%) 5 (8.9%) 9 (9.1%) Hormonal replacement therapy in the past 3 months No 37 (86.0%) 49 (87.5%) 86 (86.9%) Yes 2 (4.7%) 1 (1.8%) 3 (3.0%) Missing 4 (9.3%) 6 (10.7%) 10 (10.1%) History of oral contraceptive use; p = 0.018 No 25 (58.1%) 21 (37.5%) 46 (46.5%) Yes 12 (27.9%) 31 (55.4%) 43 (43.4%) Missing 6 (14.0%) 4 (7.1%) 10 (10.1%) Oral contraceptive use in the past 3 months No 40 (93.0%) 51 (91.1%) 91 (91.9%) Yes 0 1 (1.8%) 1 (10%) Missing 3 (7.0%) 4 (7.1%) 7 (7.1%) Parity Median (IQR) 2.00 (1.00–2.00) 2.00 (1.50–2.00.50.00) 2.00 (1.00–2.00) Missing 21 (48.8%) 21 (37.5%) 42 (42.4%) Tumor stage according to FIGO I 7 (16.3%) 10 (BOTs, 66.7%) 17 (17.2%) II 2 (4.7%) 1 (BOTs, 6.7%) 3 (3.0%) III 26 (60.5%) 3 (BOTs, 20%) 29 (29.3%) IV 6 (14.0%) – 6 (6.1%) Unknown 2 (4.7%) – 2 (2.0%) Histopathological type Epithelial 39 (90.7%) – 39 (39.4%) High-grade serous 32 (74.4%) – 32 (32.3%) Low-grade serous 1 (2.3%) – 1 (1.0%) Clear cell 2 (4.6%) – 2 (2.0%) Endometrioid 2 (4.6%) – 2 (2.0%) Carcinosarcoma 2 (4.6%) – 2 (2.0%) Sex-cord stromal 4 (9.3%) – 4 (4.0%) Primary/recurrent disease Primary disease 42 (97.7%) – 42 (42.4%) Recurrent disease 1 (2.3%) – 1 (1.0%) Chemotherapy status Treatment-naïve 39 (90.7%) – 39 (39.4%) Post NACT 4 (9.3%) – 4 (4.0%) Non-malignant adnexal masses Non-neoplastic adnexal masses Normal adnexa with other pelvic pathology – 5 (0.9%) 5 (5.1%) Sactosalpinx – 1 (1.8%) 1 (1.0%) Paraovarian cyst – 1 (1.8%) 1 (1.0%) Ovarian torsion with necrosis – 1 (1.8%) 1 (1.0%) Benign ovarian tumors Simple cyst – 9 (16.1%) 9 (9.1%) Mature teratoma – 6 (10.7%) 6 (6.1%) Fibroma – 3 (5.4%) 3 (3.0%) Endometrioma – 4 (7.1%) 4 (4.0%) Serous cystadenoma or cystadenofibroma – 7 (12.5%) 7 (7.1%) Mucinous cystadenoma or cystadenofibroma – 4 (7.1%) 4 (4.0%) Borderline ovarian tumors Serous – 7 (12.5%) 7 (7.1%) Mucinous – 6 (10.7%) 6 (6.1%) Sero-mucinous – 2 (3.6%) 2 (2.0%) Data are n (%) unless otherwise specified. Statistical significance was determined by Mann-Whitney U test for continuous variables and Fisher exact test or Chi square test of independence for categorical variables. Two-sided p values below the significance threshold 0.05 are reported. All participants were female. BOT, borderline ovarian tumor; FIGO, the International Federation of Gynecology and Obstetrics; NACT, neoadjuvant chemotherapy. Demographics and clinical characteristics of participants, by primary outcome definition of ovarian cancer 63.00 (53.00–68.00) 54.50 (47.00–64.00) 57.00 (50.00–65.50.00.50) 164.00 (159.20–167.00) 165.00 (162.00–170.00.00.00) 165.00 (160.00–169.00.00.00) 66.00 (59.00–75.00) 65.00 (60.00–80.25.00.25) 65.00 (59.00–75.75.00.75) 23.00 (22.00–27.90.00.90) 23.80 (22.00–28.10.00.10) 23.75 (22.00–28.00) 2.00 (1.00–2.00) 2.00 (1.50–2.00.50.00) 2.00 (1.00–2.00) Data are n (%) unless otherwise specified. Statistical significance was determined by Mann-Whitney U test for continuous variables and Fisher exact test or Chi square test of independence for categorical variables. Two-sided p values below the significance threshold 0.05 are reported. All participants were female. BOT, borderline ovarian tumor; FIGO, the International Federation of Gynecology and Obstetrics; NACT, neoadjuvant chemotherapy. Preoperative serum steroid levels differed between patients with ovarian cancer and patients with non-malignant adnexal masses (Table  2 ). Univariate analysis showed significantly lower levels of testosterone (T), 11β-hydroxy-testosterone (11OHT), and 11-keto-testosterone (11KT) in cases compared to controls ( p  = 0.012, < 0.001, < 0.001, respectively). After adjusting for age and menopause, 11OHT and 11KT remained significantly lower (p adj <0.001, 0.003, respectively), while T showed borderline significance (p adj =0.08). In contrast, cortisol levels were higher in cases vs. controls, after adjusting for age and menopause (p adj : 0.027). No other significant hormone differences were found between the two groups. As expected, patients with ovarian cancer had higher CA-125 and HE4 levels as well as higher ROMA index score compared to controls. Steroid hormones showed generally poor correlations with CA-125 and HE4 levels, with 11KT exhibiting the strongest negative association with CA-125 (Spearman’s ρ = − 0.34, p  = 0.003) and DHEA showing the strongest positive association with HE4 (Spearman’s ρ = 0.33, p  = 0.001) (Supplementary Table 2). These trends of lower T, 11OHT, and 11KT and higher cortisol levels in cases compared to controls were consistent when patients with borderline ovarian tumors were excluded from the control group (Supplementary Table 3). In a separate sub-analysis comparing patients with HGSOC to controls with non-malignant adnexal masses, similar trends were observed. Patients with HGSOC had significantly lower levels of T (p adj =0.012), 11OHT (p adj =0.001), and 11KT (p adj =0.003), as well as higher cortisol levels (p adj =0.039), compared to controls (Supplementary Table 4). Cases with HGSOC had also significantly higher levels of CA-125, HE4 levels and ROMA index score compared to controls. Table 2 Preoperative serum concentrations of steroid hormones and CA-125 in patients with ovarian cancer ( n  = 43) and controls ( n  = 56), based on the primary outcome definition of ovarian cancer Analyte Primary/recurrent OC ( n  = 43) Non-malignant adnexal masses ( n  = 56) P value Median IQR Median IQR Unadjusted Adjusted Classic androgens (nM) DHEA 7.04 3.82–13.47 9.44 4.95–17.36 0.061 0.453 A4 2.21 1.51–3.63 2.82 1.89–3.93 0.151 0.467 T 0.55 0.36–1.06 0.80 0.58–1.18 0.012 0.080 11-oxyandrogens (nM) 11OHA4 5.26 3.67–7.40 4.94 3.17–7.59 0.518 0.659 11KA4 0.53 0.25–0.70 0.57 0.39–0.89 0.249 0.295 11OHT 0.34 0.16–0.56 0.68 0.41–0.97 < 0.0001 < 0.0001 11KT 0.49 0.17–0.62 0.68 0.51–0.95 < 0.0001 0.003 Glucocorticoids (nM) 17α-hydroxy-progesterone 0.81 0.56–1.63 1.13 0.65–1.87 0.105 0.286 11-deoxycortisol 0.50 0.14–0.99 0.63 0.36–1.02 0.344 0.351 Cortisol 542.10 459.30–648.60.30.60 482.20 350.80–581.00 0.075 0.027 Cortisone 43.09 31.69–58.10 51.15 35.51–63.73 0.108 0.230 Mineralocorticoids (nM) Corticosterone 14.07 9.18–20.81 12.62 7.75–23.41 0.646 0.357 Clinical biomarkers CA-125 (kU/L) 325.00 55.00–779.5.00.5 21.67 14.70–34.25.70.25 < 0.0001 0.081 HE4 (pmol/L) 245.00 99.25–848.25.25.25 57.50 51.00–70.50.00.50 < 0.0001 0.431 ROMA index score 84.39 57.57–96.02 14.24 9.40–21.41.40.41 < 0.0001 < 0.0001 Unadjusted p values were calculated using the non-parametric Mann-Whitney U test. Adjusted p value were calculated with robust ANCOVA (Analysis of Covariance) test with age at operation and menopause status as confounders. P values below 0.05 were considered statistically significant. 11OHA4, 11β-hydroxy-androstenedione; 11OHT, 11β-hydroxy-testosterone; 11KA4, 11-keto-androstenedione; 11KT, 11-keto-testosterone; A4, androstenedione; BOT, borderline ovarian tumor; CA-125, cancer antigen 125; DHEA, dehydroepiandrosterone; IQR, interquartile range; OC, ovarian cancer; T, testosterone. Preoperative serum concentrations of steroid hormones and CA-125 in patients with ovarian cancer ( n  = 43) and controls ( n  = 56), based on the primary outcome definition of ovarian cancer Unadjusted p values were calculated using the non-parametric Mann-Whitney U test. Adjusted p value were calculated with robust ANCOVA (Analysis of Covariance) test with age at operation and menopause status as confounders. P values below 0.05 were considered statistically significant. 11OHA4, 11β-hydroxy-androstenedione; 11OHT, 11β-hydroxy-testosterone; 11KA4, 11-keto-androstenedione; 11KT, 11-keto-testosterone; A4, androstenedione; BOT, borderline ovarian tumor; CA-125, cancer antigen 125; DHEA, dehydroepiandrosterone; IQR, interquartile range; OC, ovarian cancer; T, testosterone. In univariate logistic regression, HE4 was the most effective biomarker for distinguishing primary/recurrent ovarian cancer from non-malignant adnexal masses, with an AUC of 0.873. Among steroids, 11KT performed best (AUC: 0.709) (Table  3 ). Performance metrics for other steroid hormones are presented in Supplementary Table 5. For multivariate models, stepwise feature selection identified 11KT, 11β-hydroxy-androstenedione (11OHA4), 11OHT, and age at operation as the top predictors for distinguishing cases from controls. The two-variable model incorporating 11KT and 11OHA4, achieved an AUC of 0.770. Adding 11OHT and age at operation improved the AUC to 0.813. Incorporating CA-125 into this four-parameter model (11KT, 11OHA4, 11OHT, and age at operation) further increased the AUC to 0.907, significantly outperforming CA-125 alone (AUC: 0.868; p  = 0.0003) and the ROMA index (AUC: 0.884; p  = 0.039) (Table  3 ; Fig.  2 A, C). Similarly, adding HE4 instead of CA-125 to the four-parameter model improved the AUC to 0.911, significantly outperforming HE4 alone ( p  = 0.0001) and the ROMA index ( p  = 0.016) (Table  3 ; Fig.  2 B, C). In the subset analysis excluding borderline ovarian tumors from the control group, HE4 remained the best-performing single biomarker (AUC: 0.898), whereas 11OHT performed best among steroids (AUC: 0.701). The four-parameter model achieved an AUC of 0.823. The four-parameter model combined with CA-125 achieved an AUC of 0.922, significantly outperforming CA-125 alone (AUC: 0.872; p  < 0.0001) and the ROMA index (AUC: 0.901; p  = 0.03). Similarly, the four-parameter model combined with HE4 achieved an AUC of 0.932, significantly outperforming HE4 alone ( p  = 0.0003) and the ROMA index ( p  = 0.004) (Supplementary Table 6). In the second sub-analysis (postmenopausal patients only), the CA-125 was the best performing biomarker (AUC: 0.837), while 11OHT was the top steroid predictor (AUC: 0.705). The four-parameter model achieved an AUC of 0.777. The four-parameter model plus CA-124 achieved an AUC of 0.873, significantly outperforming CA-125 alone (AUC: 0.837; p  = 0.002) as well as ROMA (AUC: 0.846; p  = 0.041). The four-parameter model plus HE4 achieved an AUC of 0.872, performing better than HE4 alone (AUC: 0.832; p  = 0.005), and ROMA alone ( p  = 0.065) (Supplementary Table 7). In the third sub-analysis (early-stage ovarian cancer vs. non-malignant adnexal masses), CA-125 was the best-performing individual biomarker (AUC: 0.710), while 11KT performed best among steroids (AUC: 0.679). The four-parameter model alone performed comparably to CA-125 (AUC: 0.706). The four-parameter model plus CA-125 achieved an AUC of 0.808 and significantly outperformed CA-125 alone ( p  = 0.001) as well as the ROMA index (AUC: 0.714, p  = 0.002). The four-parameter model plus HE4 achieved an AUC of 0.770, significantly outperforming HE4 (AUC: 0.642, p  = 0.001) but was comparable to the ROMA index ( p  = 0.155) (Supplementary Table 8). In the fourth sub-analysis (epithelial ovarian cancer vs. non-malignant adnexal masses), HE4 was the best performing biomarker (AUC: 0.919), while 11KT performed best among steroids (AUC: 0.732). The four-parameter model achieved an AUC of 0.817. The four-parameter model plus CA-125 achieved an AUC of 0.923 and significantly outperformed CA-125 alone (AUC: 0.903, p  = 0.011). The four-parameter model plus HE4 achieved an AUC of 0.935 and performed better than HE4 alone ( p  = 0.054). Both steroid-protein models performed comparably to the ROMA index (AUC: 0.924, p  > 0.05 in both cases) (Supplementary Table 9). In the fifth sub-analysis (HGSOC vs. non-malignant adnexal masses), HE4 was the best-performing biomarker (AUC: 0.940), while 11KT was the top steroid predictor (AUC: 0.707). The four-parameter model achieved an AUC of 0.786. The four-parameter model plus CA-125 achieved an AUC to 0.933, significantly outperforming CA-125 alone (AUC: 0.907; p  = 0.006) but performing comparably to ROMA (AUC: 0.921; p  = 0.576). The four-parameter model plus HE4 achieved an AUC of 0.930, performing slightly worse than HE4 alone ( p  = 0.042) but comparable to ROMA ( p  = 0.262) (Supplementary Table 10). Table 3 Diagnostic performance statistics of univariate and multivariate models by primary outcome definition of ovarian cancer (cases, n  = 43; controls, n  = 56) Model AUC Sensitivity (%) Specificity (%) PPV (%) NPV (%) F1 score (%) AIC Univariate 11OHT 0.705 (0.696–0.712) 79.5 (78.6–80.7) 49.3 (48.1–50.0) 67.6 (67.0–68.2.0.2) 64.3 (63.0–65.6.0.6) 73.1 (72.4–73.8) 122.741 11KT 0.709 (0.704–0.713) 82.7 (81.4–83.9) 45.9 (44.8–47.1) 67.1 (66.5–67.7) 66.6 (64.9–68.5) 74.1 (73.3–74.9) 119.641 CA-125 0.868 (0.862–0.873) 87.5 (87.1–87.5) 71.1 (70.0–71.9.0.9) 80.2 (79.5–80.6) 81.0 (80.5–81.2) 83.7 (83.3–83.9) 91.293 HE4 0.873 (0.865–0.878) 92.7 (921.−92.9) 70.4 (69.5–71.4) 80.7 (80.1–81.3) 87.9 (87.1–88.2) 86.3 (85.9–86.7) 88.717 Multivariate Best 2 parameters 0.770 (0.764–0.774) 88.0 (86.8–88.9) 51.5 (50.5–52.4) 70.7 (70.2–71.3) 76.3 (74.6–77.9) 78.4 (77.8–79.0) 112.820 Best 3 parameters 0.793 (0.787–0.797) 86.8 (85.7–87.5) 61.8 (60.0–63.3.0.3) 75.2 (74.3–76.0) 77.9 (76.4–78.9) 80.6 (79.9–81.3) 108.552 Best 4 parameters 0.813 (0.808–0.817) 85.2 (83.9–86.4) 64.5 (63.3–65.7) 76.2 (75.5–76.9) 76.6 (74.9–78.2) 80.5 (79.5–81.3) 104.380 ROMA index 0.884 (0.875–0.891) 91.1 (91.1-91.1-1) 78.5 (78.1–78.6) 85.0 (84.7–85.0) 86.8 (86.8–86.8) 87.9 (87.8–87.9) 79.334 Best 4 parameters + CA-125 0.907 (0.904–0.910) 88.9 (88.2–89.3) 82.0 (81.4–82.8) 86.4 (86.1–87.1) 84.9 (84.2–85.6) 87.6 (87.1–88.0) 79.293 Best 4 parameters + HE4 0.911 (0.908–0.913) 94.4 (94.0–94.6.0.6) 77. 3 (76.7–77.7) 84.4 (84.1–84.7) 91.3 (90.8–91.7) 89.1 (88.9–89.3) 80.428 11OHT, 11β-hydroxy-testosterone; 11KT, 11-keto-testosterone; AIC, akaike information criteria; AUC, area under the curve; CA-125, cancer antigen 125; NPV, negative predictive value; PPV, positive predictive value; Multivariate model with best 2 parameters = 11KT + 11OHA4; model with best 3 parameters = 11KT + 11OHA4 + age; model with best 4 parameters = 11KT + 11OHA4 + age + 11OHT. Diagnostic performance statistics of univariate and multivariate models by primary outcome definition of ovarian cancer (cases, n  = 43; controls, n  = 56) 0.705 (0.696–0.712) 79.5 (78.6–80.7) 49.3 (48.1–50.0) 67.6 (67.0–68.2.0.2) 64.3 (63.0–65.6.0.6) 73.1 (72.4–73.8) 0.709 (0.704–0.713) 82.7 (81.4–83.9) 45.9 (44.8–47.1) 67.1 (66.5–67.7) 66.6 (64.9–68.5) 74.1 (73.3–74.9) 0.868 (0.862–0.873) 87.5 (87.1–87.5) 71.1 (70.0–71.9.0.9) 80.2 (79.5–80.6) 81.0 (80.5–81.2) 83.7 (83.3–83.9) 0.873 (0.865–0.878) 92.7 (921.−92.9) 70.4 (69.5–71.4) 80.7 (80.1–81.3) 87.9 (87.1–88.2) 86.3 (85.9–86.7) 0.770 (0.764–0.774) 88.0 (86.8–88.9) 51.5 (50.5–52.4) 70.7 (70.2–71.3) 76.3 (74.6–77.9) 78.4 (77.8–79.0) 0.793 (0.787–0.797) 86.8 (85.7–87.5) 61.8 (60.0–63.3.0.3) 75.2 (74.3–76.0) 77.9 (76.4–78.9) 80.6 (79.9–81.3) 0.813 (0.808–0.817) 85.2 (83.9–86.4) 64.5 (63.3–65.7) 76.2 (75.5–76.9) 76.6 (74.9–78.2) 80.5 (79.5–81.3) 0.884 (0.875–0.891) 91.1 (91.1-91.1-1) 78.5 (78.1–78.6) 85.0 (84.7–85.0) 86.8 (86.8–86.8) 87.9 (87.8–87.9) 0.907 (0.904–0.910) 88.9 (88.2–89.3) 82.0 (81.4–82.8) 86.4 (86.1–87.1) 84.9 (84.2–85.6) 87.6 (87.1–88.0) 0.911 (0.908–0.913) 94.4 (94.0–94.6.0.6) 77. 3 (76.7–77.7) 84.4 (84.1–84.7) 91.3 (90.8–91.7) 89.1 (88.9–89.3) 11OHT, 11β-hydroxy-testosterone; 11KT, 11-keto-testosterone; AIC, akaike information criteria; AUC, area under the curve; CA-125, cancer antigen 125; NPV, negative predictive value; PPV, positive predictive value; Multivariate model with best 2 parameters = 11KT + 11OHA4; model with best 3 parameters = 11KT + 11OHA4 + age; model with best 4 parameters = 11KT + 11OHA4 + age + 11OHT. Fig. 2 Performance of logistic regression models in detecting ovarian cancer. ( A ) AUC curves comparing the Best 4 + CA-125 model to CA-125 alone for distinguishing ovarian cancer ( n  = 43) from non-malignant adnexal masses ( n  = 56). ( B ) AUC curves comparing the Best 4 + HE4 model to HE4 alone for distinguishing ovarian cancer ( n  = 43) from non-malignant adnexal masses ( n  = 56). ( C ) AUC curve comparison of the Best 4 + CA-125 model, the Best 4 + HE4 model, and the ROMA index for distinguishing ovarian cancer ( n  = 43) from non-malignant adnexal masses ( n  = 56). The Best 4 model includes 11KT, 11OHA4, age, and 11OHT as predictive parameters. 11OHA4, 11β-hydroxy-androstenedione; 11OHT, 11β-hydroxy-testosterone; 11KT, 11-keto-testosterone; AUC, area under the receiver operating characteristic curve; CA-125, cancer antigen 125; HE4, human epididymis protein 4; ROMA, risk of ovarian malignancy algorithm Performance of logistic regression models in detecting ovarian cancer. ( A ) AUC curves comparing the Best 4 + CA-125 model to CA-125 alone for distinguishing ovarian cancer ( n  = 43) from non-malignant adnexal masses ( n  = 56). ( B ) AUC curves comparing the Best 4 + HE4 model to HE4 alone for distinguishing ovarian cancer ( n  = 43) from non-malignant adnexal masses ( n  = 56). ( C ) AUC curve comparison of the Best 4 + CA-125 model, the Best 4 + HE4 model, and the ROMA index for distinguishing ovarian cancer ( n  = 43) from non-malignant adnexal masses ( n  = 56). The Best 4 model includes 11KT, 11OHA4, age, and 11OHT as predictive parameters. 11OHA4, 11β-hydroxy-androstenedione; 11OHT, 11β-hydroxy-testosterone; 11KT, 11-keto-testosterone; AUC, area under the receiver operating characteristic curve; CA-125, cancer antigen 125; HE4, human epididymis protein 4; ROMA, risk of ovarian malignancy algorithm

Adult women (≥ 18 years) who were scheduled for surgery for an adnexal mass or presumed/histologically proven ovarian cancer at the Department of Gynecology, University Medical Center Ljubljana, were enrolled. Exclusion criteria were age < 18, active non-ovarian malignancy, previous ovarian malignancy, polycystic ovary syndrome (PCOS) and non-ovarian endometriosis. Blood samples were collected between December 2021 to February 2025, 1–7 days before surgery, alongside lifestyle, medication, and clinical data. Written informed consent was obtained from all participants. Morning samples were collected and processed following a standardized procedure specifically adapted for metabolomics studies [ 18 ]; a detailed description of the collection and processing steps is provided in Sect. 1.1. in Supplementary File 1. Tissue specimens were examined by a certified pathologist according to WHO Classification of tumors, 5th edition and ICCR dataset [ 19 ]. The primary outcome was diagnostic accuracy of steroids for detecting ovarian cancer (binary outcome), defined as primary or recurrent malignant ovarian neoplasms ( n  = 43) versus non-malignant adnexal masses (defined as non-neoplastic adnexal masses or benign or borderline ovarian tumors) ( n  = 56), confirmed by histological examination obtained by surgery (reference standard). Five additional sub-analyses were performed to further evaluate the models: (1) upon excluding borderline ovarian tumors from the control group; (2) upon excluding premenopausal patients from both groups; (3) upon excluding patients with advanced-stage disease (stage III-IV according to International Federation of Gynecologic Oncology (FIGO)) from the case group; (4) upon excluding patients with non-epithelial ovarian tumors from the case group; and (5) upon excluding patients with non-HGSOC from the case group. Unconjugated steroids ( n = 17) were quantified using a previously validated LC-MS/MS method described in Gjorgoska et al. [ 20 ]. Schematic representation of the position of these steroids in steroid biosynthesis is given in Supplementary Fig. 1. Briefly, 180 µL of thawed serum, calibrators, and QC samples were mixed with stable isotope-labeled internal standards (see Sect. 1.2. in Supplementary File 1) and incubated for 15 min. After protein precipitation with 180 µL 3 M Na₂SO₄, samples were extracted using methyl- tert -butyl ether (MTBE). The organic layer was evaporated and the extracted analytes were reconstituted in LC-MS grade methanol-water (50:50) (v/v). Steroids were chromatographically separated using a Phenomenex Kinetex XB-C18 column on a Shimadzu Nexera LC system, with mass spectrometry detection on a Sciex 3500 triple quadrupole mass spectrometer using electrospray ionization. Data acquisition was performed using Analyst 1.6.2. Steroid quantification was based on peak area ratios of analytes to internal standards. All samples were anonymized prior to analysis. Steroid concentrations below the lower limit of quantification (LLOQ) were replaced by 0.5 x LLOQ for statistical analysis. Aldosterone, 5α-dihydrotestosterone (DHT), and progesterone were below the lower limit of quantification (LLOQ) in 69.2%, 58.6%, and 39.4%, respectively, and were excluded from further analysis. Serum CA-125 (kU/L) and HE4 (pmol/L) levels were routinely measured at the Clinical Institute for Clinical Biochemistry, University Medical Center, Ljubljana, using clinically validated electroluminescent immunoassays (ECLIAs), for CA-125, REF: 11,776,223,190, for HE4, REF: 05950929190, on a Cobas e411 analyzer (Roche Diagnostics GmbH, Mannheim, Germany). The ROMA index score was calculated based on the formula established by Moore and collaborators [ 21 ]. Data were anonymized and analyzed using R Studio (version 4.3.0 or higher). Statistical significance was defined as p  < 0.05. Continuous variables were expressed as median values with interquartile range (IQR) and compared using the Mann-Whitney U test. Robust ANCOVA (Analysis of Covariance) was used to adjust for age and menopause status. Categorical variables were expressed as frequencies with percentages and compared using Fisher exact test or Chi square test of independence. Machine learning was performed using the caret library [ 22 ], with a 5 × 5-fold cross-validation protocol (1000 iterations). We tested 20 variables, including 12 steroid hormones, 3 steroid pools (androgen, 11-oxyandrogen and glucocorticoid pool; see Sect. 1.3. in Supplementary File 1 for their definitions), two proteins (CA-125, HE4), ROMA index and 2 clinical parameters (age, menopause status). Continuous variables were ln-transformed and standardized. Feature selection was performed via stepAIC (MASS library). Multicollinearity was assessed using the Variance Inflation Factor (VIF) in a logistic regression model, with acceptable values < 5. Diagnostic accuracy was assessed using the area under the receiver operating curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 score and Akaike information criteria (AIC). DeLong’s test was used to compare AUCs between different models.

Ovarian cancer (OC) is the eighth most diagnosed cancer in women and most common cause of gynecologic cancer death in developed countries. In 2022, 324,603 new cases and 206,956 deaths were reported globally [ 1 ]. Epithelial ovarian cancer is the most common histological type and includes four main subtypes: serous, clear cell, mucinous, and endometrioid carcinoma. The serous subtype is further classified into low-grade and high-grade, with high-grade serous ovarian carcinoma (HGSOC) being both the most prevalent and the most lethal form [ 2 ]. Prognosis varies significantly by stage, with 5-year survival rates of 93% for cancer confined to the ovaries (localized), 75% for cancer that has spread nearby (regional), and only 31% for cancer that has spread to distant organs (distant) [ 3 , 4 ]. Unfortunately, over a third of cases are diagnosed at an advanced stage, largely due to nonspecific symptoms. Most symptomatic patients or those with abnormal clinical findings, such as suspicious ultrasound results or elevated cancer antigen 125 (CA-125) levels, do not have ovarian cancer. A study found that only 3% of premenopausal and 18% of postmenopausal patients referred through the UK’s National Health Services (NHS) expedited pathway were actually diagnosed with ovarian cancer [ 5 ]. Clearly, accurate, accessible, and cost-effective diagnostic methods are needed to distinguish malignant tumors from non-malignant adnexal masses and reduce unnecessary procedures. Currently, several risk-prediction models, based on either ultrasound features or blood biomarker levels exist that can be used to triage patients with an adnexal mass. These include CA-125, HE4, the Risk of Malignancy Index (RMI1), the Risk of Ovarian Malignancy Algorithm (ROMA), the International Ovarian Tumor Analysis (IOTA) Simple Rules, the IOTA Assessment of Different Neoplasias in the Adnexa (ADNEX) model, and the Ovarian-Adnexal Reporting and Data System (ORADS) [ 6 – 10 ]. These diagnostic tools differ in performance. CA-125 at the 35 kU/L threshold has high sensitivity (>80%) but low specificity (< 80%); RMI1, ROMA and IOTA Simple Rules achieve a better balance, exceeding 80% sensitivity and specificity for diagnosing ovarian cancer in symptomatic postmenopausal women; ADNEX has very high sensitivity (>95%) but low specificity (< 60%), leading to more false positives [ 11 ]. CA-125, used alone or as part of ROMA, lacks specificity due to its elevation in benign conditions common in premenopausal patients such as menstruation, endometriosis, and peritoneal inflammation [ 12 ], as well as conditions common in postmenopausal and elderly patients such as heart failure, cirrhosis, chronic kidney disease, and intraabdominal infections [ 13 , 14 ]. Other models relying on ultrasound findings require validation in routine practice, as scans are often performed by non-specialists. Emerging molecular biomarkers, including circulating tumor DNA, tumor cells, exosomes, and metabolites, show promise for ovarian cancer diagnosis but remain unadopted due to limited validation and high costs [ 15 – 17 ]. This study investigates serum steroids as biomarkers for diagnosing ovarian cancer in patients with adnexal masses, using machine learning. We measured preoperative serum levels of classic androgens, 11-oxyandrogens, glucocorticoids, and mineralocorticoids in 43 patients with primary/recurrent ovarian cancer and 56 with non-neoplastic adnexal masses, benign or borderline ovarian tumors, using a validated liquid chromatography-tandem mass spectrometry (LC-MS/MS) method. Logistic regression was used to assess the diagnostic potential of these hormones, alone and in combination with CA-125, human epididymis protein 4 (HE4), and clinical parameters.

In this biomarker discovery study, we analyzed preoperative steroid hormone levels in 99 women with adnexal masses to evaluate their potential in distinguishing ovarian cancer from benign conditions. Incorporating 11-oxyandrogens, patient age, and either CA-125 or HE4, we developed two steroid-protein diagnostic models. Both models significantly outperformed CA-125 or HE4 alone ( p  < 0.001 for both), as well as the ROMA index ( p  < 0.05), highlighting the added diagnostic value of steroid hormone profiling. This advantage likely reflects the ability of 11-oxyandrogens to capture tumor-associated steroidogenic activity not detected by conventional markers, thereby providing a more comprehensive biochemical profile than CA-125, HE4, or the ROMA index alone. The superior performance of the models remained consistent after excluding borderline tumors from the control group and when restricted to postmenopausal patients. Both steroid-protein models also outperformed CA-125 and HE4 in detecting early-stage disease, with the steroids + CA-125 model further surpassing ROMA. However, in subgroup analyses restricted to epithelial ovarian cancer or HGSOC versus non-malignant adnexal masses, both steroid-protein models performed comparably to the ROMA index, although they were superior in the main malignant versus non-malignant comparison. This difference may reflect the smaller number of patients in the sub-analyses or suggest that further refinement of the models is needed for histology-specific diagnosis. Overall, these subgroup findings require confirmation in larger cohorts. To the best of our knowledge, this is the first study to characterize 11-oxyandrogen levels in women with adnexal masses. It is also the first study to apply machine learning to evaluate circulating steroids alone and in combination with proteins for distinguishing malignant from non-malignant cases. Given that 11-oxyandrogens are altered in hormone-sensitive cancers [ 23 , 24 ], their integration into diagnostic models represents a novel and biologically relevant improvement. Apart from hormone-sensitive cancers, 11-oxyandrogens have been linked to disorders such as congenital adrenal hyperplasia and PCOS [ 25 ]. Their elevated levels in PCOS suggest a role of adipose tissue in regulation. However, in our study, 11OHT, 11KT, and T were lower in cases than in controls, despite no significant differences in body weight. One possible explanation is intra-tumoral steroid metabolism. We recently showed that HGSOC tumors express key steroid-metabolizing enzymes [ 26 ], potentially accounting for systemic reductions in these steroids. If these metabolites are produced locally in ovarian tumors, they could impact key cellular processes, as 11KT and 11-keto-DHT activate the androgen receptor (AR) with a potency comparable to testosterone and DHT, respectively [ 27 – 29 ]. AR expression has been linked to improved survival and greater platinum sensitivity in HGSOC [ 30 ], though mechanistic studies on its signaling in ovarian tumors are lacking. In contrast to 11-oxyandrogens, classic androgens have already been investigated as risk factors for ovarian cancer, however data are conflicting. Some studies suggest higher T levels may increase the risk of endometrioid and mucinous ovarian tumors [ 31 ], while others found no link [ 32 – 34 ]. Contrariwise, Mendelian randomization studies suggest higher T might lower the risk of ovarian cancer [ 35 ], including HGSOC [ 36 ]. Glucocorticoids have also been linked to ovarian cancer progression. Schrepf et al. reported that patients with ovarian cancer exhibited disrupted cortisol rhythms before surgery [ 37 ], which may explain the higher cortisol levels we observed in cases compared to controls in our study. Disrupted cortisol rhythms have also been associated with systemic inflammation and poorer survival [ 37 ], whereas chemotherapy was shown to normalize these rhythms [ 38 ]. Our study has several limitations. First, it is based on a small, single-center cohort. Second, technical constraints of the LC-MS/MS method prevented quantification of low-abundance steroids, such as DHT. Third, the small number of patients with early-stage ovarian cancer in our study limits the generalizability of the findings for this subgroup. As earlier detection is associated with substantially improved survival outcomes [ 3 , 4 ], further validation of the models in larger early-stage cohorts is warranted. Fourth, the small number of premenopausal patients restricted assessment of model performance in this subgroup. While 11-oxyandrogens remain stable throughout life, one key steroid in our models, 11OHT, differs significantly between pre- and postmenopausal women [ 39 ]. This difference may introduce variability or limit the models’ applicability in premenopausal patients, so this should be investigated further. Finally, despite internal cross-validation, the lack of external validation is a major limitation that may affect generalizability of our findings. Validation in larger, independent, ideally multi-centric cohorts is essential to confirm robustness and support clinical translation. Nonetheless, our study has several key strengths, including the collection of serum samples following a strict SOP adapted for metabolomics studies, the use of a validated LC-MS/MS method for multi-steroid profiling, and the incorporation of tumor biomarkers routinely employed in clinical practice. Moreover, the steroid-protein models developed in this study improved the distinction between malignant and benign adnexal masses. This has the potential to improve preoperative risk assessment, leading to more accurate triage and better clinical decision-making. Such improvements could reduce unnecessary surgeries for benign conditions and ensure timely treatment for ovarian cancer, particularly in cases where conventional biomarkers are inconclusive. With validation in larger, independent cohorts, these models have the potential to become valuable diagnostic tools, especially because mass spectrometry for steroid profiling is already well established in clinical practice. In addition, integrating further preoperative clinical data, such as ultrasound features based on IOTA criteria and computed tomography characteristics of adnexal masses, could further increase model accuracy. A further important direction for future studies is longitudinal sampling and clinical follow-up, which would allow assessment of whether circulating steroid levels are influenced by chemotherapy or tumor burden, and whether they could complement established biomarkers such as CA-125 in monitoring disease progression or treatment response.

In conclusion, circulating steroid hormones, particularly 11-oxyandrogens, offer a valuable and noninvasive biomarker class for improving the preoperative differentiation of malignant and benign adnexal masses. When combined with CA-125 or HE4, these hormones significantly enhance diagnostic accuracy beyond conventional markers and the ROMA index. The steroid-protein models remain robust across clinical subgroups and provide insights into tumor-specific steroidogenic activity, supporting their potential as a clinically meaningful tool for risk stratification and improved surgical decision-making in women with adnexal masses.

Supplementary Material 1 Supplementary Material 1 Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2 Supplementary Material 3 Supplementary Material 3