Development and validation of a serum peptidomic signature for early detection of asymptomatic ovarian cancer: A multi-center prospective study

Conceptualization, Y.W., Z.L., and H.Z.; methodology, Y.W., Z.L., H.Z., Y. Zhang, W.S., and R.B.; clinical sample collection, Y. Zhang, H.Y., S.C., C.W., R.X., and Y. Zhao; cohort maintenance, Z.L., W.S., Y.L., X.L., Y. Zhang, S.C., C.W., and Y. Zhao; code development and analysis, Z.L., B.D., R.B., Y. Zhang, W.S., Y.L., X.L., and R.X.; mass spectrometry analysis, H.Z., X.C., and Y. Zhou; original draft writing, Z.L., H.Z., R.B., W.S., and Y. Zhang; writing – review and editing, Z.L., H.Z., R.B., Y. Zhang, W.S., Y.L., X.L., and Y.W.; funding acquisition, Y.W. and S.C.; supervision, Y.W., Z.L., and H.Z.

Between June 2022 and September 2024, a total of 1,141 participants were enrolled into the SOCFCP cohort. 38 After excluding 113 OC patients (due to age, metastatic OC, reported OC-related symptoms, borderline tumor, or missing data) and 185 non-OC participants (due to age, non-fasting sampling, surgery history, benign Brenner tumor, or missing data), 843 eligible participants were included in the final analysis (see STAR Methods and Figure S1 for details). The baseline characteristics of OC patients ( n = 136) and non-OC participants ( n = 707) were summarized in Table 1 . OC patients were significantly older than non-OC participants (mean 54.9 vs. 48.5 years, p < 0.001), whereas body mass index (BMI) was similar between groups (median 22.9 vs. 22.4 kg/m 2 , p = 0.23). Significant differences between groups were also observed in the menopausal status (postmenopausal: 61.0% vs. 38.0%, p < 0.001), the BRCA mutation status (positive: 16.9% vs. 0.1%, p < 0.001), prior malignancies (6.6% vs. 2.8%, p = 0.04), and reproductive hormone therapy use (3.7% vs. 0.7%, p = 0.01). Family history of OC or breast cancer did not differ significantly between groups (7.4% vs. 5.7%, p = 0.57). Serum concentrations of CA125 (median 232.0 vs. 12.0 U/mL, p < 0.001), HE4 (median 162.0 vs. 48.4 pmol/L, p < 0.001), and ROMA (median 71.9% vs. 7.4%, p < 0.001) were significantly higher in OC patients than in non-OC participants. Among OC patients, epithelial malignancies accounted for 91.2% ( n = 124), with high-grade serous ovarian cancer (HGSOC) as the predominant histological subtype (61.0%, n = 83), followed by clear cell carcinoma (11.8%, n = 16), endometrioid carcinoma (9.6%, n = 13), and other epithelial subtypes (8.8%, n = 12). Non-epithelial malignancies comprised malignant germ cell (4.4%, n = 6) and sex cord-stromal tumors (4.4%, n = 6). Early-stage (International Federation of Gynecology and Obstetrics [FIGO] I/II) disease accounted for 41.9% ( n = 57) of cases. Among patients with early-stage disease, median serum concentrations of CA125 and HE4 were 75.4 U/mL and 73.5 pmol/L, respectively. Among the non-OC participants, 59.3% ( n = 419) were healthy controls. In the remaining 288 participants with benign ovarian lesions, the most common diagnoses were ovarian endometriosis ( n = 79, 27.4%), serous cystadenoma ( n = 22, 7.6%), and teratoma ( n = 12, 4.2%). Among the 79 cases of ovarian endometriosis, disease severity was graded according to the revised American Society for Reproductive Medicine (rASRM) classification as stage I ( n = 4, 5.1%), II ( n = 5, 6.3%), III ( n = 32, 40.5%), and IV ( n = 38, 48.1%). Table 1 Baseline characteristics of the participants Characteristic Overall ( n = 843) Non-OC ( n = 707) OC ( n = 136) p value Age, years 49.5 ± 12.0 48.5 ± 11.6 54.9 ± 12.5 <0.001 a BMI, kg/m 2 22.5 (20.7–24.5) 22.4 (20.7–24.4) 22.9 (21.1–25.0) 0.234 b Menopausal status, n (%)  Postmenopausal 352 (41.8) 269 (38.0) 83 (61.0) <0.001 c  Premenopausal 479 (56.8) 427 (60.4) 52 (38.3)  Unknown 12 (1.4) 11 (1.6) 1 (0.7) Reproductive hormone therapy, n (%)  Yes 10 (1.2) 5 (0.7) 5 (3.7) 0.013 c  No/unknown 833 (98.8) 702 (99.3) 131 (96.3) Prior malignancies, n (%)  Yes 29 (3.4) 20 (2.8) 9 (6.6) 0.037 c  No 814 (96.6) 687 (97.2) 127 (93.4) Family history of OC/breast cancer, n (%)  Yes 50 (5.9) 40 (5.7) 10 (7.4) 0.570 c  No/unknown 793 (94.1) 667 (94.3) 126 (92.6) BRCA mutation status, n (%)  Positive 24 (2.8) 1 (0.1) 23 (16.9) <0.001 c  Negative 59 (7.0) 0 (0.0) 59 (43.4)  Unknown/untested 760 (90.2) 706 (99.9) 54 (39.7) CA125, U/mL 14.0 (8.9–40.0) 12.0 (8.2–20.4) 232.0 (65.5–833.6) <0.001 b HE4, pmol/L 50.8 (40.3–65.4) 48.4 (39.2–57.9) 162.0 (74.8–386.0) <0.001 b ROMA, % 8.1 (5.4–13.2) 7.4 (5.1–10.3) 71.9 (30.9–95.2) <0.001 b FIGO staging, n (%)  Early stage (I/II) – – 57 (41.9) –  Late stage (III/IV) – – 79 (58.1) – Histological subtypes of OC, n (%)  Epithelial – – 124 (91.2) –  High-grade serous (HGSOC) – – 83 (61.0) –  Clear cell carcinoma – – 16 (11.8) –  Endometrioid – – 13 (9. 6) –  Other – – 12 (8.8) –  Non-epithelial – – 12 (8.8) –  Malignant germ cell tumor – 6 (4.4) –  Sex cord-stromal tumor – – 6 (4.4) – Diagnosis of non-OC, n (%)  Healthy controls – 419 (59.3) – –  Benign ovarian lesion – 175 (24.8) – –  Ovarian endometriosis – 79 (11.2) – –  Serous cystadenoma – 22 (3.1) – –  Mature cystic teratoma – 19 (2.7) – –  Other – 55 (7. 8) – –  Other benign gynecological conditions 113 (15.9) – –  Uterine fibroids – 42 (5.9) – –  Adenomyosis – 23 (3.3) – –  Other – 48 (6.8) – – Data are shown as mean ± SD or median (IQR) for continuous variables and n (%) for categorical variables. OC, ovarian cancer; BMI, body mass index; CA125, cancer antigen 125; HE4, human epididymis protein 4; ROMA, Risk of Ovarian Malignancy Algorithm; FIGO, International Federation of Gynecology and Obstetrics; HGSOC, high-grade serous ovarian cancer. See also Table S1 . a p values for comparisons between the OC and non-OC groups were calculated using Student’s t test for means. b p values for comparisons between the OC and non-OC groups were calculated using the Mann-Whitney U test for medians. c p values for comparisons between the OC and non-OC groups were calculated using the χ 2 test for proportions. Baseline characteristics of the participants Data are shown as mean ± SD or median (IQR) for continuous variables and n (%) for categorical variables. OC, ovarian cancer; BMI, body mass index; CA125, cancer antigen 125; HE4, human epididymis protein 4; ROMA, Risk of Ovarian Malignancy Algorithm; FIGO, International Federation of Gynecology and Obstetrics; HGSOC, high-grade serous ovarian cancer. See also Table S1 . p values for comparisons between the OC and non-OC groups were calculated using Student’s t test for means. p values for comparisons between the OC and non-OC groups were calculated using the Mann-Whitney U test for medians. p values for comparisons between the OC and non-OC groups were calculated using the χ 2 test for proportions. The participants were partitioned into internal development and external validation datasets based on participating centers. The internal development datasets comprised 684 participants from three centers (Shanghai First Maternity and Infant Hospital, Jiading Community Hospital, and Jinshan Community Hospital), which were subdivided into training ( n = 547; n = 84 OC and n = 463 non-OC) and test ( n = 137; n = 19 OC and n = 118 non-OC) datasets at an 8:2 ratio, with stratification to maintain pathological subtype representation. The external validation cohort consisted of 159 participants from Fudan University Shanghai Cancer Center and Minhang Community Hospital (33 OC and 126 non-OC) ( Figure 1 ; Figure S1 ). Figure 1 Study design: a biology-informed ML-based feature selection pipeline for biomarker discovery and model development in a multi-center prospective cohort OC, ovarian cancer; BS-OC, ovarian cancer sample before surgery; AS-OC, ovarian cancer sample after surgery; non-OC, non-ovarian cancer; Asym-OC, asymptomatic ovarian cancer. See also Figure S1 . Study design: a biology-informed ML-based feature selection pipeline for biomarker discovery and model development in a multi-center prospective cohort OC, ovarian cancer; BS-OC, ovarian cancer sample before surgery; AS-OC, ovarian cancer sample after surgery; non-OC, non-ovarian cancer; Asym-OC, asymptomatic ovarian cancer. See also Figure S1 . The internal and external datasets showed comparable distributions of OC cases (15.1% vs. 20.8%, p = 0.10), HGSOC (69.7% vs. 56.3%, p = 0.17), BRCA mutation status distribution ( p = 0.12), and reproductive hormone therapy use ( p = 0.22). Although participants in the external validation dataset were younger than those in the internal dataset (mean age 46.6 vs. 50.2 years, p < 0.001), BMI (median 22.7 vs. 22.5 kg/m 2 , p = 0.82) and CA125 levels (13.8 vs. 14.0 U/mL, p = 0.80) were similar. As expected, the external dataset contained a lower proportion of early-stage disease and exhibited higher HE4 (57.2 vs. 48.9 pmol/L, p < 0.001) and ROMA values (9.7% vs. 7.7%, p < 0.001) ( Table S1 ). All modeling and feature selection procedures were conducted using preoperative fasting serum samples. Using a MALDI-TOF MS platform validated for quantitative analysis of trace serum volumes ( Figure S2 ), we obtained approximately 43,900 raw data points per sample across a mass range of 100–16,000 Da. Following baseline correction and smoothing, 1,081 high-quality m/z features with high signal-to-noise ratios were retained for subsequent analysis. While orthogonal partial least squares discriminant analysis (OPLS-DA) showed clear separation between OC and non-OC participants, suggesting distinct serum peptidomic profiles ( Figure 2 A), its moderate Q 2 value (0.434) indicated limited predictive reliability. Differential expression analysis identified 782 significantly altered peptidomic features (false discovery rate [FDR]-adjusted p < 0.05) between the OC patients and non-OC participants ( Figure 2 B). These altered features were involved in key pathways such as complement activation (e.g., C3(1,306–1,319): log 2 FC = 0.977, p < 0.001), coagulation (e.g., PLG(524–549): log 2 FC = 0.447, p < 0.001), and lipid metabolism (e.g., LysoPC(18:2): log 2 FC = −0.257, p < 0.001). Figure 2 Biomarker discovery, panel refinement, and model performance optimization (A) Orthogonal partial least squares discriminant analysis (OPLS-DA) demonstrated distinct separation between OC ( n = 136; red) and non-OC ( n = 707; gray) serum peptidomic profiles (R 2 X = 0.481, R 2 Y = 0.736, Q 2 = 0.434). (B) Volcano plot of differentially expressed serum features in OC ( n = 136) vs. non-OC ( n = 707) participants. Significantly upregulated (green) and downregulated (purple) features (–log 10 FDR-adjusted p < 0.05) are highlighted. (C) Venn diagram identifying 24 shared, differentially expressed features ( p < 0.001) across five key clinical comparisons: OC vs. non-OC ( n = 136 vs. 707), ES-OC vs. non-OC ( n = 57 vs. 707), HGSOC vs. non-OC ( n = 83 vs. 707), non-HGSOC vs. non-OC ( n = 53 vs. 707), and OC vs. ovarian endometriosis ( n = 136 vs. 79). (D) ROC curves of individual biomarkers and the integrated 7-marker ProMS panel ( n = 843). (E) ROC curve of the final ProMS+ model in the external validation dataset ( n = 159; OC, n = 33; non-OC, n = 126). (F–M) Performance evaluation of eight LightGBM model configurations in the external validation dataset ( n = 159; OC, n = 33; non-OC, n = 126). For each model, the kernel density estimate (purple) depicts the distribution of AUCs across 50 random seeds (0–49) in the external validation set, annotated with the median and interquartile range (IQR). The corresponding radar chart visualizes the model’s performance across six metrics (AUC, sensitivity, specificity, PPV, NPV, and accuracy) at seed = 1; the polygon area represents overall performance. See also Figures S2 and S3 . Biomarker discovery, panel refinement, and model performance optimization (A) Orthogonal partial least squares discriminant analysis (OPLS-DA) demonstrated distinct separation between OC ( n = 136; red) and non-OC ( n = 707; gray) serum peptidomic profiles (R 2 X = 0.481, R 2 Y = 0.736, Q 2 = 0.434). (B) Volcano plot of differentially expressed serum features in OC ( n = 136) vs. non-OC ( n = 707) participants. Significantly upregulated (green) and downregulated (purple) features (–log 10 FDR-adjusted p < 0.05) are highlighted. (C) Venn diagram identifying 24 shared, differentially expressed features ( p < 0.001) across five key clinical comparisons: OC vs. non-OC ( n = 136 vs. 707), ES-OC vs. non-OC ( n = 57 vs. 707), HGSOC vs. non-OC ( n = 83 vs. 707), non-HGSOC vs. non-OC ( n = 53 vs. 707), and OC vs. ovarian endometriosis ( n = 136 vs. 79). (D) ROC curves of individual biomarkers and the integrated 7-marker ProMS panel ( n = 843). (E) ROC curve of the final ProMS+ model in the external validation dataset ( n = 159; OC, n = 33; non-OC, n = 126). (F–M) Performance evaluation of eight LightGBM model configurations in the external validation dataset ( n = 159; OC, n = 33; non-OC, n = 126). For each model, the kernel density estimate (purple) depicts the distribution of AUCs across 50 random seeds (0–49) in the external validation set, annotated with the median and interquartile range (IQR). The corresponding radar chart visualizes the model’s performance across six metrics (AUC, sensitivity, specificity, PPV, NPV, and accuracy) at seed = 1; the polygon area represents overall performance. See also Figures S2 and S3 . We first constructed an initial peptidomic signature model (Full-ProMS) by integrating all 1,081 serum-peptidomic features using a light gradient boosting machine (LightGBM) approach. With the random seed fixed at 1, the model yielded an area under the curve (AUC) of 0.958 (95% confidence interval [CI]: 0.925–0.991) in the external validation dataset. At a high-sensitivity operating point (95.0% sensitivity), Full-ProMS demonstrated a higher specificity (73.7%) than CA125 (66.2%), HE4 (17.8%), and ROMA (25.7%). To develop a compact model with enhanced biological interpretability and generalizability, we applied a four-step selection workflow to refine the 1,081 initial features into a minimal set of key biomarkers ( Figure 1 ). First, differential analysis across five key comparisons (OC vs. non-OC, early-stage OC [ES-OC] vs. non-OC, HGSOC vs. non-OC, non-HGSOC vs. non-OC, and OC vs. ovarian endometriosis) identified 756, 561, 626, 649, and 368 significantly dysregulated features (FDR-adjusted p < 0.05), respectively. The intersection of these analyses yielded 24 shared candidate features, comprising 16 upregulated and 8 downregulated molecules. Upregulated candidates included FPA(10–16), CFH(176–183), GSN(29–41), CORO1C(407–419), RPL18(120–132), CAVIN2(2–15), FPA(1–16), ITIH4(651–666), ITIH4(630–644), FPA(1–16)+PO 3 , TF(70–97), PLG(524–549), FGA(513–558), FGA(576–629), APOC2(29–101), and APOC2(23–101). Downregulated candidates included: LysoPC(18:2), FN1(1,616–1,628), MOGS(709–720), NCAM1(362–372), F13A1(14–38), APOC3(21–99)-0a/2, APOC3(21–99)-0a, and APOC3(21–99)-1 ( Figure 2 C). Second, a literature review highlighted the importance of complement C3 in OC diagnosis and prognosis. 39 , 40 Supporting its biological significance, the concentration of fragment C3(1,306–1,319) was consistently and significantly upregulated across key comparative groups: OC vs. non-OC (log 2 FC = 0.977, p < 0.001), ES-OC vs. non-OC (log 2 FC = 0.882, p < 0.001), non-HGSOC vs. non-OC (log 2 FC = 0.944, p < 0.001), and HGSOC vs. non-OC (log 2 FC = 0.999, p < 0.001). A similar expression pattern was observed for the closely related fragment C3(1,304–1,319). Based on convergent evidence from published literature and our own data, both C3 fragments were included as candidates for further study (see full list in Table S2 ). Third, tandem MS (MS/MS) analysis confirmed high-confidence fragmentation patterns for 18 of the 26 candidate features ( Table S2 ). Finally, analysis of feature importance rankings derived from LightGBM and Shapley additive explanations (SHAP) analysis ( Figure S3 ) identified top candidate features predominantly associated with coagulation, complement, and lipid metabolism pathways. Based on a forward selection approach, we identified a minimal set of seven key biomarkers—FPA(1–16), FPA(1–16)+PO 3 , PLG(524–549), C3(1,306–1,319), C3(1,304–1,319), CFH(176–183), and LysoPC(18:2). Although LysoPC(18:2) is a lipid rather than a peptide, it was detected in the same MALDI-TOF MS workflow, emerged from the identical feature-selection pipeline, and was retained because its lipid-metabolism-related signal offered a biological dimension distinct from the peptide markers. The six peptide markers were validated by annotated MS/MS fragmentation, whereas LysoPC(18:2) was assigned by authentic-standard-matched liquid chromatography (LC)-MS/MS/multiple reaction monitoring (MRM) validation using LysoPC(18:2) as the authentic standard, including retention-time agreement and the diagnostic MRM transition ( Figure 3 ; Table S10 ). The assignment, therefore, represents an authentic-standard-matched molecular-species-level identification, rather than a full isomer-resolved structural determination. The compact model based on these markers (ProMS) demonstrated significantly higher discriminatory power in the whole dataset, achieving an AUC of 0.996, compared with individual biomarkers (AUC range: 0.661–0.747) ( Figure 2 D). Figure 3 Structural validation of the ProMS biomarker panel by tandem mass spectrometry (A–F) High-confidence MS/MS spectra for six identified peptides. Fragment ions are annotated as follows: b ions (blue), y ions (green), and ions resulting from neutral loss of water (b˚/y˚) or ammonia (b/y) from precursor b/y ions (orange). Unassigned peaks are shown in gray. (G–K) Authentic-standard-matched LC-MS/MS/MRM validation of the lipid marker LysoPC 18:2 using LysoPC(18:2/0:0) as the authentic standard. (G and I) Extracted MRM chromatograms for the transition Q1/Q3 = 520/184 in the authentic standard and serum analyte, showing retention times of 3.76 min and 3.68 min, respectively. The retention time deviation between the serum analyte and the authentic standard was 0.08 min. (H and J) Targeted MRM transition signals of the authentic standard and serum analyte. (K) Optimized MRM acquisition parameters for Q1/Q3 = 520/184, corresponding to the characteristic phosphocholine head group fragment of LysoPC/PC species. See also Tables S2 and S10 . Structural validation of the ProMS biomarker panel by tandem mass spectrometry (A–F) High-confidence MS/MS spectra for six identified peptides. Fragment ions are annotated as follows: b ions (blue), y ions (green), and ions resulting from neutral loss of water (b˚/y˚) or ammonia (b/y) from precursor b/y ions (orange). Unassigned peaks are shown in gray. (G–K) Authentic-standard-matched LC-MS/MS/MRM validation of the lipid marker LysoPC 18:2 using LysoPC(18:2/0:0) as the authentic standard. (G and I) Extracted MRM chromatograms for the transition Q1/Q3 = 520/184 in the authentic standard and serum analyte, showing retention times of 3.76 min and 3.68 min, respectively. The retention time deviation between the serum analyte and the authentic standard was 0.08 min. (H and J) Targeted MRM transition signals of the authentic standard and serum analyte. (K) Optimized MRM acquisition parameters for Q1/Q3 = 520/184, corresponding to the characteristic phosphocholine head group fragment of LysoPC/PC species. See also Tables S2 and S10 . We compared eight LightGBM configurations during optimization with a fixed random seed (seed = 1): Full-ProMS (1,081 features), Int-ProMS (18 intersected, MS/MS-validated features), ProMS (the core seven-biomarker panel), and its extensions combining age and established biomarkers: ProMS-Age (ProMS + age), ProMS-CA125 (ProMS + Age + CA125), ProMS-HE4 (ProMS + Age + HE4), ProMS-ROMA (ProMS + Age + ROMA), and ProMS+ (ProMS + Age + CA125 + HE4). Model performance was evaluated in the external validation set across 50 random seeds (0–49). Kernel density estimation of AUC distributions revealed that refining the model from Full-ProMS to the 7-marker ProMS narrowed the AUC interquartile range ([IQR]: 0.021 vs. 0.031) but significantly reduced the median AUC (0.892 vs. 0.936, p < 0.001). Subsequent integration of age (ProMS-Age) yielded a tighter (IQR: 0.016) and significantly improved AUC distribution (median: 0.900, p = 0.004). Among biomarker-augmented models, ProMS+ achieved optimal performance with the highest median AUC (0.992) and narrowest IQR (0.0026), demonstrating superior robustness. ProMS+ also produced the largest radar chart area (2.364) and an AUC of 0.995 in external validation ( Figures 2 F–2M), leading to its selection as the final model. To elucidate the biological relevance of the serum-peptidomic features, pathway enrichment analysis was performed on the 26 candidate biomarkers. Results revealed significant enrichment in three key pathways: coagulation, complement activation, and lipid metabolism ( Figure 4 C). Coagulation-related enriched terms included blood coagulation GO: 0006957) and degranulation (R-HSA-114608). Complement-related enriched terms included complement activation via the alternative pathway (GO: 0072378) and complement cascade (R-HSA-166658). Meanwhile, Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment identified the integrated term complement and coagulation cascades (KEGG: hsa04610), reinforcing the functional interconnection of these processes in OC. Lipid metabolism-related pathways were highlighted by significantly enriched terms such as triglyceride metabolic process (GO: 0006641), cholesterol metabolism (KEGG: hsa04979), and plasma lipoprotein assembly (R-HSA-76009). Figure 4 Postoperative biomarker dynamics and pathway enrichment underpin biological plausibility (A) Schematic of the postoperative follow-up study. Paired serum samples were collected from 45 ovarian cancer patients before (BS-OC) and 8–90 days after (AS-OC) tumor-resection surgery. (B) Relative serum levels of the seven panel biomarkers, CA125, and HE4 in non-OC controls ( n = 707) and preoperative (BS-OC, n = 45) and postoperative (AS-OC) samples ( n = 45). Boxplots show the median and IQR, overlaid with individual measurements. Group means are indicated by blue dots. Comparisons involving non-OC controls were analyzed using two-sided Wilcoxon rank-sum tests, whereas the paired BS-OC versus AS-OC comparison was analyzed using a two-sided paired Wilcoxon signed-rank test. Significance levels are denoted according to BH-adjusted p values. Detailed p values are provided in Table S3 . (C) Functional enrichment analysis of the 26-candidate serum-peptidomic features. Significantly enriched Gene Ontology (GO) and Reactome terms are plotted by –log 10 (FDR). Terms are color coded by primary biological pathway: coagulation (purple), complement activation (green), and lipid metabolism (brown). Postoperative biomarker dynamics and pathway enrichment underpin biological plausibility (A) Schematic of the postoperative follow-up study. Paired serum samples were collected from 45 ovarian cancer patients before (BS-OC) and 8–90 days after (AS-OC) tumor-resection surgery. (B) Relative serum levels of the seven panel biomarkers, CA125, and HE4 in non-OC controls ( n = 707) and preoperative (BS-OC, n = 45) and postoperative (AS-OC) samples ( n = 45). Boxplots show the median and IQR, overlaid with individual measurements. Group means are indicated by blue dots. Comparisons involving non-OC controls were analyzed using two-sided Wilcoxon rank-sum tests, whereas the paired BS-OC versus AS-OC comparison was analyzed using a two-sided paired Wilcoxon signed-rank test. Significance levels are denoted according to BH-adjusted p values. Detailed p values are provided in Table S3 . (C) Functional enrichment analysis of the 26-candidate serum-peptidomic features. Significantly enriched Gene Ontology (GO) and Reactome terms are plotted by –log 10 (FDR). Terms are color coded by primary biological pathway: coagulation (purple), complement activation (green), and lipid metabolism (brown). Serum samples were collected postoperatively from 45 OC patients after a median follow-up of 12 days (IQR: 11–19 days). Changes in the seven biomarkers were assessed by comparing postoperative levels with both preoperative samples ( n = 45) and non-OC participants ( Figure 4 A). In preoperative samples, LysoPC(18:2) intensity was lower in OC patients, whereas the remaining six biomarkers were elevated compared with non-OC controls (all p < 0.05, Figure 4 B). Following tumor resection, three of seven peptidomic markers decreased toward non-OC levels: FPA(1–16) (Benjamini-Hochberg (BH)-adjusted p = 2.5 × 10 -4 ), PLG(524–549) ( p = 9.6 × 10 -4 ), and C3(1306–1319) ( p = 2.6 × 10 -2 ). The remaining four markers—C3(1,304–1,319), CFH(176–183), FPA(1–16)+PO 3 , and LysoPC(18:2)—showed no significant postoperative change (all BH-adjusted p > 0.05). CA125 ( p = 4.7 × 10 -6 ) and HE4 ( p = 1.0 × 10 -6 ) also declined significantly after surgery, consistent with reduced tumor burden ( Table S3 ). Compared with CA125, HE4, and ROMA, ProMS+ achieved the highest discrimination across all subgroups, as shown in the receiver operating characteristic (ROC) curves and radar chart metrics ( Figure S4 ; Table S4 ). Moreover, under a high-sensitivity operating point (95.0% sensitivity) pertinent to asym-OC detection, ProMS+ demonstrated superior diagnostic performance compared with conventional biomarkers ( Table 2 ; Table S9 ), achieving excellent overall discrimination (OC vs. non-OC: AUC, 0.996; 95% CI, 0.993–0.999; specificity, 98.0%). Notably, for ES-OC detection—a paramount clinical challenge—ProMS+ maintained a specificity of 92.6% (AUC = 0.993). This was markedly superior not only to individual conventional markers (CA125 [44.7%], HE4 [11.2%], and ROMA [24.0%]) but also to the composite models Age + CA125 + HE4 (specificity: 18.1%, AUC = 0.821) and Age + ROMA (specificity: 24.5%, AUC = 0.827). In the context of ovarian endometriosis, a common benign condition that severely compromises the specificity of CA125 (6.3%), ProMS+ retained a high specificity of 92.4% (AUC: 0.987), effectively mitigating false-positive results. The model also demonstrated high and balanced performance across major histologic subtypes, showing exceptional specificity in HGSOC (98.0%) and non-HGSOC (98.7%). Notably, in non-HGSOC, its specificity (98.7%) and AUC (0.997) far exceeded those of ROMA (22.1% and 0.826, respectively). Table 2 Specificity at 95.0% sensitivity and AUC of ProMS+ and conventional biomarkers in OC detection ES-OC ( n = 57) vs. non-OC ( n = 707) HGSOC ( n = 83) vs. non-OC ( n = 707) Non-HGSOC ( n = 53) vs. non-OC ( n = 707) OC ( n = 136) vs. ovarian endometriosis ( n = 79) OC ( n = 136) vs. non-OC ( n = 707) Specificity (%) at 95.0% sensitivity ProMS+ 92.6 98.0 98.7 92.4 98.0 Age + CA125 + HE4 18.1 80.8 28.6 34.2 28.6 Age + ROMA 24.5 91.4 25.9 35.4 38.8 CA125 44.7 74.1 44.7 6.3 66.2 HE4 11.2 57.1 12.0 40.5 17.8 ROMA 24.0 95.2 22.1 58.2 25.7 AUC (95% CI) ProMS+ 0.993 (0.987, 0.999) 0.996 (0.992, 1.000) 0.997 (0.992, 1.000) 0.987 (0.977, 0.998) 0.996 (0.993, 0.999) Age + CA125 + HE4 0.821 (0.746, 0.895) 0.958 (0.923, 0.993) 0.855 (0.789, 0.922) 0.917 (0.880, 0.954) 0.918 (0.883, 0.953) Age + ROMA 0.827 (0.758, 0.895) 0.972 (0.947, 0.997) 0.834 (0.763, 0.904) 0.918 (0.881, 0.956) 0.918 (0.885, 0.951) CA125 0.877 (0.833, 0.922) 0.955 (0.934, 0.977) 0.899 (0.855, 0.943) 0.775 (0.713, 0.837) 0.933 (0.911, 0.956) HE4 0.752 (0.639, 0.811) 0.944 (0.904, 0.985) 0.744 (0.658, 0.831) 0.926 (0.891, 0.960) 0.866 (0.821, 0.911) ROMA 0.825 (0.752, 0.897) 0.979 (0.958, 0.999) 0.820 (0.741, 0.900) 0.954 (0.928, 0.979) 0.917 (0.881, 0.953) The cutoffs of ProMS+, CA125, HE4, and ROMA are 0.245, 35 U/mL, 70 pmol/L, and 11.4%, respectively. ES-OC, early-stage ovarian cancer; non-OC, non-ovarian cancer; HGSOC, high-grade serous ovarian cancer; non-HGSOC, ovarian cancer subtypes excluding HGSOC; OC, ovarian cancer. See also Tables S4 , S5 , S7 , and S9 and Figure S4 . Specificity at 95.0% sensitivity and AUC of ProMS+ and conventional biomarkers in OC detection The cutoffs of ProMS+, CA125, HE4, and ROMA are 0.245, 35 U/mL, 70 pmol/L, and 11.4%, respectively. ES-OC, early-stage ovarian cancer; non-OC, non-ovarian cancer; HGSOC, high-grade serous ovarian cancer; non-HGSOC, ovarian cancer subtypes excluding HGSOC; OC, ovarian cancer. See also Tables S4 , S5 , S7 , and S9 and Figure S4 . We further assessed the ability of ProMS+ to rectify cases misclassified by conventional biomarkers ( Table S5 ). Among the 21 OC patients falsely labeled negative by CA125, ProMS+ correctly identified 17 participants, reducing the false-negative rate (FNR) by 81.0%. Of the 106 non-OC participants incorrectly flagged positive by CA125, ProMS+ accurately reclassified 99 participants, lowering the false-positive rate (FPR) by 93.4%. Corresponding improvements were observed for cases misclassified by HE4 (FNR reduction: 81.2%, FPR reduction: 95.8%) and ROMA (FNR reduction: 77.8%, FPR reduction: 96.4%). When compared with the integrated Age + CA125 + HE4 model, ProMS+ further reduced the FNR among OC patients by 45.0% and the FPR among non-OC participants by 78.2%. Similar improvements were observed against the Age + ROMA baseline, with reductions of 50.0% in FNR and 71.1% in FPR, respectively. To evaluate the performance of ProMS+ in realistic early detection contexts, we conducted Monte Carlo simulations (each n = 100,000) across seven prevalence scenarios derived from published early detection and risk-reducing surgery cohorts, comprising two normal-risk settings (prevalence 0.06% 5 and 0.37% 15 ) and five high-risk settings (prevalence 1.00%–8.30% 41 , 42 , 43 , 44 , 45 ) reflecting BRCA1/2 carrier individuals with a strong family history and women undergoing risk-reducing salpingo-oophorectomy (RRSO) ( Table S6 ). Endometriosis prevalence was set at 0.36 in both cohorts. 46 In the normal-risk simulation, ProMS+ achieved an AUC of 0.980, outperforming CA125 (0.874), HE4 (0.875), and ROMA (0.942). Similarly, in the high-risk simulation (prevalence 1.00%), ProMS+ maintained superior discrimination (AUC = 0.983 vs. 0.894, 0.873, and 0.943, respectively). A minimum positive predictive value (PPV) of 0.10 is widely regarded as necessary for effective OC detection. 47 At the two normal-risk prevalence settings (0.06% and 0.37%), no marker approached this threshold, although ProMS+ outperformed CA125, HE4, and ROMA by a substantial margin at 0.37% prevalence (ProMS+: 0.082 vs. CA125: 0.010, HE4: 0.035, and ROMA: 0.016; Table S6 ). Among the high-risk scenarios, ProMS+ was the first to cross the 0.10 benchmark—achieving a PPV of 0.201 (95% CI: 0.189–0.213) at a prevalence of 1.00%, at which CA125 (0.027), HE4 (0.090), and ROMA (0.042) all remained below the threshold. Across all seven prevalence settings, ProMS+ maintained the highest PPV, increasing commensurately with prevalence from 0.015 at 0.06% to 0.689 (0.680–0.697) at 8.30%. In addition, we compared ProMS+ against the Age + CA125 + HE4 baseline to evaluate the incremental contribution of the peptidomic features in screening-relevant settings. Across all seven prevalence scenarios, ProMS+ consistently yielded higher PPVs than the baseline model. The absolute PPV gain attributable to the peptidomic features increased with prevalence, from 0.058 at 0.37% to 0.314 at 8.30% ( Table S6 ; Figure 5 ). Figure 5 Validation of ProMS+ in prevalence-specific simulated populations Seven simulated cohorts ( n = 100,000) were generated across two normal-risk and five high-risk prevalence scenarios. Normal-risk scenarios included 0.06% and 0.37% OC prevalence; high-risk scenarios included 1.00%, 2.60%, 4.10%, 5.00%, and 8.30%. Endometriosis was set at 36% within the non-OC subset. ProMS+, CA125, HE4, and ROMA were evaluated at pre-specified cutoffs. Square markers indicate PPV estimates, horizontal error bars indicate 95% bootstrap CIs, and the dashed line marks the screening-relevance threshold of PPV = 0.10. ProMS+ exceeded this threshold in all high-risk scenarios but not in normal-risk scenarios. See also Table S6 . Validation of ProMS+ in prevalence-specific simulated populations Seven simulated cohorts ( n = 100,000) were generated across two normal-risk and five high-risk prevalence scenarios. Normal-risk scenarios included 0.06% and 0.37% OC prevalence; high-risk scenarios included 1.00%, 2.60%, 4.10%, 5.00%, and 8.30%. Endometriosis was set at 36% within the non-OC subset. ProMS+, CA125, HE4, and ROMA were evaluated at pre-specified cutoffs. Square markers indicate PPV estimates, horizontal error bars indicate 95% bootstrap CIs, and the dashed line marks the screening-relevance threshold of PPV = 0.10. ProMS+ exceeded this threshold in all high-risk scenarios but not in normal-risk scenarios. See also Table S6 . To evaluate whether demographic or clinical imbalances between the OC and non-OC groups influenced model performance, we assessed ProMS+ discrimination across histological, clinical, and genetic subgroups ( Table S7 ). AUCs ranged from 0.994 to 1.000 across all strata examined. Discrimination was comparable between epithelial and non-epithelial malignancies (AUC: 0.996 [95% CI: 0.993–0.999] vs. 0.998 [0.994–1.000]) and between BRCA mutation carriers and non-carriers (0.999 vs. 0.997). Performance remained similarly unaffected by menopausal status (postmenopausal: 0.997, premenopausal: 0.994), prior malignancy history (0.997 after exclusion), or hormone therapy use (0.997 after exclusion). Stratification by family history of OC or breast cancer yielded comparable AUCs (0.996–1.000), although the limited number of family-history-positive OC cases ( n = 10) constrains interpretation in this subgroup. These findings indicate that the discriminative performance of ProMS+ is not attributable to demographic, clinical, or genetic confounding. To assess whether the imbalanced OC/non-OC ratio influenced model performance, we performed balanced-resampling sensitivity analysis by down-sampling non-OC training samples and retraining ProMS+ across 50 balanced-resampling runs. In the external validation set, ProMS+ maintained stable performance under balanced resampling, with a median AUC of 0.990 (0.986–0.991), area under the precision-recall curve (AUC-PR) of 0.962 (0.950–0.969), F1-score of 0.872 (0.852–0.896), and balanced accuracy of 0.924 (0.900–0.941). Subgroup-level metrics are provided in Table S8 ; as expected, precision-recall metrics were lower for smaller subgroups such as ES-OC and non-HGSOC, reflecting both limited sample sizes and the inherent difficulty of detecting these cases under balanced conditions. Across all subgroups, AUCs remained ≥0.959, indicating that the rank-discriminative ability of ProMS+ was not driven by class imbalance.

Further information and requests for resources should be directed to and will be fulfilled by the lead contact Prof. Yu Wang (email: [email protected] ). This study did not generate new unique reagents. • The serum peptidomic data generated in this study have been deposited in the OMIX database at the National Genomics Data Center (NGDC)/China National Center for Bioinformation (CNCB) under the accession number OMIX018710, which is associated with the BioProject PRJCA069438 ( https://ngdc.cncb.ac.cn/bioproject/browse/PRJCA069438 ). The processed feature matrix and necessary clinical metadata are also provided as Data S1 . • The analysis code and scripts used in this study are available on GitHub: https://github.com/dongbowen1996d/Early-Detection-of-Asymptomatic-Ovarian-Cancer-A-Multi-Center-Prospective-Study . • Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. The serum peptidomic data generated in this study have been deposited in the OMIX database at the National Genomics Data Center (NGDC)/China National Center for Bioinformation (CNCB) under the accession number OMIX018710, which is associated with the BioProject PRJCA069438 ( https://ngdc.cncb.ac.cn/bioproject/browse/PRJCA069438 ). The processed feature matrix and necessary clinical metadata are also provided as Data S1 . The analysis code and scripts used in this study are available on GitHub: https://github.com/dongbowen1996d/Early-Detection-of-Asymptomatic-Ovarian-Cancer-A-Multi-Center-Prospective-Study . Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Early detection of asym-OC remains a critical clinical challenge. 1 In this multicenter prospective study, we developed and validated ProMS+, a ML-based model that synergistically combines a 7-marker serum peptidomic panel with age, CA125, and HE4. ProMS+ exhibited high accuracy in discriminating asym-OC from non-OC participants, with particular superiority in detecting early-stage disease (ES-OC) compared with conventional biomarkers. Its performance was further evaluated through independent external validation, longitudinal postoperative follow-up, and Monte Carlo simulation in populations with varying OC prevalence. This study builds upon early serum peptidomic research in cancer detection. 12 , 22 , 32 , 33 , 34 , 35 The previous study of Petricoin et al. showed that serum proteomic patterns could distinguish OC from controls using surface-enhanced laser desorption/ionization (SELDI)-TOF mass spectrometry, 22 but the biological identity of the discriminatory mass-spectral features was unknown. According to the “peptidome hypothesis” proposed later, 13 LMW serum fragments serve as a rich archive of real-time biological activity within the tumor microenvironment, potentially enabling earlier detection than approaches reliant on tumor-shed macromolecules 13 , 14 —unlike conventional biomarkers that are related to tumor burden or liquid biopsy approaches that identify tumor-specific genomic alterations (cfDNA/ctDNA). 15 , 16 , 17 , 18 Villanueva et al. provided the key mechanistic underpinning for this hypothesis by showing that cancer-specific serum peptidome patterns arise primarily from differential exoprotease activities acting on abundant circulating proteins—particularly those of the coagulation and complement cascades—rather than from changes in parent protein abundance alone, 19 a finding later extended to a functional assay that directly monitors peptidase activities. 48 In the present study, we applied this framework to OC, coupling independent external validation with a biology-informed feature selection strategy. Serum peptidomic profiling by MALDI-TOF mass spectrometry was advanced by prior work on magnetic particle-assisted automated sample processing 23 and later made more accessible through open-source high-throughput magnetic bead workflows. 24 More recently, a similar nanoparticle-enhanced laser desorption/ionization (NELDI)-MS platform was applied to a large OC cohort, where a serum metabolic panel combined with ROMA achieved AUCs of 0.95–0.99 for distinguishing malignant from benign ovarian masses, 49 also supporting the clinical potential of this technological approach. In this study, ProMS+ utilizes a MALDI-TOF MS platform with automated preprocessing to leverage nanomaterials-pretreated serum peptidomic profiling, targeting LMW analytes (100–16,000 Da), 36 and provides orthogonal coverage to high-molecular-weight (HMW) glycoprotein markers used alone (e.g., CA125: 200–1,000 kDa 50 , 51 , 52 ; HE4: ∼25 kDa 53 ), and the combination of these two approaches provided a synergistic effect. The performance of ProMS+ is consistent with the above hypothesis, stemming from its distinct biological basis, particularly addressing the “black-box” limitation of prior models, where the biological relevance of discriminatory features remained unclear. Nevertheless, direct experimental validation of the underlying mechanisms is required to substantiate this interpretation. In addition to age, CA125, and HE4, the ProMS+ panel comprises six MS/MS-supported peptidomic markers and one LC-MS/MS/MRM-validated lipid marker, LysoPC(18:2), which are collectively involved in coagulation cascade activation, complement pathway dysregulation, and lipid metabolic reprogramming. In OC patients, the six protein markers were significantly upregulated compared with non-OC participants, while lysophosphatidylcholine (LysoPC(18:2)) was downregulated. The complement-derived markers showed a dual pattern: elevated C3(1,306–1,319) and C3(1,304–1,319)—whose parent protein C3 is regulated by TWIST1 in OC cells 39 and whose C3f domain is cleaved by tumor-resident MMP-9 54 —may reflect pro-tumoral complement activation, 40 , 55 , 56 while increased CFH(176–183) could indicate tumor co-option of complement regulation for immune evasion. 57 , 58 Coagulation pathway involvement was suggested by elevated FPA(1–16) and FPA(1–16)+PO 3 —thrombin-cleaved fibrinopeptide A fragments, the phosphorylated form of which has been reported in OC patient serum 59 —and by PLG(524–549), a plasminogen fragment, consistent with OC-associated hypercoagulability. 59 , 60 , 61 , 62 LysoPC(18:2) downregulation is consistent with lipid metabolic reprogramming in OC. 63 Collectively, these molecular identities align with the exoprotease-mediated peptidome hypothesis, 19 although the specific proteases operating in the OC microenvironment remain to be identified. Postoperatively, three of seven markers decreased toward non-OC levels significantly, providing preliminary support for a tumor-associated origin. As these pathways may also be activated in other cancers or inflammatory conditions, whether these associations reflect causal links to OC pathogenesis, or represent an epiphenomenon of the tumor-associated host response, remains to be determined through independent functional studies. A well-recognized limitation in high-dimensional multi-omics research is the susceptibility to overfitting and restricted generalizability, often arising from high feature-to-sample ratios and participant-specific biases. 64 , 65 The importance of rigorous pre-analytical and analytical standardization in serum peptidomics has been recognized since early efforts to identify and correct common sources of bias in peptide signature discovery. 28 To address these challenges, we crucially employed a large-scale, prospective, multi-center cohort 38 and tested the performance of the model through independent external validation. Furthermore, a rigorous biology-informed ML-based feature selection pipeline 65 , 66 , 67 was used to distill 1,081 initial features down to a 7-marker panel, substantially reducing model complexity and the risk of overfitting. The robustness of this refined feature set was further strengthened by identifying biomarkers across multiple comparisons (e.g., early/late-stage, HGSOC/non-HGSOC, and versus ovarian endometriosis), which reduces the risk of degraded performance caused by variations in pathological type distribution and enhances the model’s applicability to heterogeneous populations. Furthermore, the incorporation of longitudinal postoperative follow-up data added a valuable temporal dimension for biological interpretability beyond cross-sectional comparisons of biomarker, 68 , 69 although these dynamics require confirmation in larger cohorts with benign surgical controls. This multi-faceted validation framework collectively enhanced the methodological robustness of our findings. ProMS+ showed consistently higher specificity than traditional biomarkers at equivalent sensitivity levels. In OC detection, where disease prevalence in the general population is extremely low, maintaining high specificity is a critical prerequisite for clinical utility: at a prevalence of 0.06%, even 98% specificity would yield a PPV below 3%, and achieving a clinically acceptable PPV of ≥10% requires specificity exceeding 99.5%. Sensitivity, specificity, and PPV, therefore, need to be evaluated jointly rather than as isolated metrics. To estimate real-world performance of the model, we conducted Monte Carlo simulations (each n = 100,000) across seven prevalence settings derived from published cohorts, spanning 0.06% (average-risk general population) to 8.30% (BRCA1/2 carriers at RRSO). Two observations are noteworthy. First, across the full prevalence range, ProMS+ maintained the highest PPV among all tested markers, indicating that the peptidomic features contribute discriminative information beyond that of conventional biomarkers even after accounting for varying disease prevalence. Second, ProMS+ crossed the widely cited PPV benchmark of 0.10 at a prevalence no less than 1.00%—a prevalence reflective of high-risk groups—whereas at the normal-risk settings (0.06% and 0.37%), no marker reached this threshold. While current clinical guidelines do not endorse screening for OC in the general population, RRSO is recommended for individuals with hereditary breast and ovarian cancer (HBOC) syndrome, among whom occult cancer is found in 3.5%–4.6% of patients. 41 Collectively, these observations suggest that peptidomic-based approaches such as ProMS+ may be best suited to enriched-risk surveillance settings. This is particularly true for those who do not meet the criteria of RRSO, for whom effective early detection remains an urgent and unmet need. These results should, however, be interpreted with caution: the sensitivity and specificity inputs were derived from the present cohort and the simulation does not capture real-world demographic or referral-pattern variability; the PPV estimates, therefore, represent an extrapolation pending prospective validation. As noted above, several caveats concerning biological specificity, peptide identification confidence, and the postoperative analysis warrant consideration in interpreting our findings. Beyond these, we acknowledge the following limitations. First, this study was conducted in a cohort enriched with OC patients rather than in an asymptomatic intended-use high-risk population; prospective validation in such a population is, therefore, required. Second, although the six peptide markers are mechanistically distinct from CA125—arising from tumor-microenvironment proteolytic activity rather than peritoneal shedding—CA125 can be elevated even in early-stage patients and is a component of ProMS+. Coupled with the moderate proportion of early-stage cases in the external validation cohort, this makes disentangling tumor-burden-dependent from tumor-burden-independent contributions inherently challenging, and the early-stage subgroup estimates should, therefore, be considered hypothesis generating. Third, sample sizes for certain rare OC histotypes remain limited despite stratified sampling during data partitioning to ensure proportional representation across training and test datasets. Fourth, demographic and clinical heterogeneity across participating centers presented a challenge, potentially limiting the consistency of model performance despite standardized protocols. Fifth, the optimal operating threshold for ProMS+ may be context dependent. Finally, whether an early detection strategy incorporating ProMS+—used either alone or alongside complementary modalities (e.g., cfDNA assays or imaging)—can reduce OC mortality in a high-risk population will require evaluation in large-scale, long-term randomized controlled trials. In conclusion, ProMS+ represents a promising serum peptidomic-based model for the early detection of asym-OC. To establish its clinical utility, future prospective studies in intended-use populations—together with functional characterization of the underlying biology—are needed.

Ovarian cancer (OC) is the most lethal gynecologic malignancy and the eighth leading cause of cancer-related death among women worldwide. 1 While early-stage (I/II) disease has a 5-year overall survival rate of 70%–95%, approximately 80% of OC patients are diagnosed at advanced stages (III/IV), where the survival rate plummets to 10%–40%. 2 This dramatic contrast is largely driven by the challenge of early detection, as early-stage OC is typically asymptomatic or presents with nonspecific symptoms such as abdominal bloating or digestive issues. 3 The high mortality associated with late diagnosis underscores the urgent need for effective strategies to detect OC at its most treatable, asymptomatic stage. Current OC detection relies primarily on serum biomarkers—cancer antigen 125 (CA125) and human epididymis protein 4 (HE4) in particular—transvaginal ultrasound, and Risk of Ovarian Malignancy Algorithm (ROMA). However, CA125-based early detection strategy has hardly been shown to reduce OC mortality: a meta-analysis of four large randomized trials (293,587 women)—including the landmark UKCTOCS 4 , 5 and PLCO 6 , 7 studies—found no significant mortality benefit, while false-positive results led to major surgical complications in 3%–15% of cases and 9.7–32.5 unnecessary surgeries per 1,000 women screened. 8 For HE4 and ROMA, evidence in the early detection setting is even more limited: neither was developed for detecting asymptomatic disease—HE4 was identified as a marker for discriminating benign from malignant pelvic masses 9 and ROMA was developed and validated for triaging women with an established adnexal mass. 10 These approaches collectively miss 25%–50% of early-stage cases, 4 underscoring the need for detection strategies that capture orthogonal biological information. The underlying biology of ovarian carcinogenesis provides a compelling rationale for approaches through blood-based liquid biopsies. 11 The disease induces systemic pathophysiological alterations, including genomic instability, epigenetic dysregulation, and metabolic reprogramming, which are reflected in peripheral blood as aberrant proteomic and metabolic signatures. 12 , 13 , 14 While approaches through circulating cell-free DNA (cfDNA) 15 , 16 , 17 and circulating tumor DNA (ctDNA) 18 detect tumor-specific genomic alterations, peptidomic signatures capture a complementary dimension of tumor biology. The circulating peptidome—low-molecular-weight (LMW) peptide fragments in serum—arises from differential exopeptidase activities in the tumor microenvironment that generate cancer-specific cleavage patterns from abundant host proteins, particularly those of the coagulation and complement cascades. 19 Rather than reflecting the abundance of parent proteins, these fragments serve as functional readout of tumor-associated proteolytic activity, integrating upstream genetic and microenvironmental perturbations into a terminal disease signature. 13 , 20 , 21 , 22 Serum peptidomic profiling by MALDI-TOF mass spectrometry—established through prior work on magnetic particle-assisted automated sample processing and high-throughput peptide extraction 23 , 24 —has shown potential for cancer detection and classification across a range of solid tumors, including brain, thyroid, breast, prostate, bladder, lung, and hepatocellular cancers. 19 , 25 , 26 , 27 , 28 , 29 , 30 , 31 However, their application to OC remains constrained by methodological issues, including small sample size, cross-sectional design, symptom-driven recruitment, and inadequate external validation. 32 , 33 , 34 , 35 Moreover, a common key limitation in high-throughput omics research is biological interpretability. Critically, in studies with limited samples and potential sampling biases, purely data-driven models risk capturing statistical artifacts or cohort-specific noise rather than biologically generalizable signatures. This limitation is compounded by a scarcity of longitudinal validation, such as tracking postoperative biomarker changes, which is critical for establishing true links to the disease. We have previously developed a high-throughput, nanomaterials-assisted MALDI-TOF mass spectrometry (MS) platform that uses trace amounts of serum for the minimally invasive, rapid (less than 30 min), and simultaneous detection of six cancers (liver, lung, pancreatic, colorectal, gastric, and thyroid). 36 This approach was specifically designed to overcome key clinical barriers, including poor reproducibility in sample preparation, high cost, and lengthy testing times. 37 In this study, using samples from the prospective cohort Shanghai Ovarian Cancer and Family Care Project (SOCFCP; ClinicalTrials.gov: NCT06118307 ), 38 we aimed to develop and validate a machine-learning (ML)-based model for asymptomatic OC (asym-OC) detection, with the specific objectives of identifying a concise, biologically relevant serum peptidomic panel through rigorous feature selection and longitudinal and external validation, thereby establishing a robust, effective, and minimally invasive serum test for OC detection at its most curable stages.

H.Z. is a cofounder and scientific advisor of Gronbio. Y. Zhou, X.C., and B.D. are employees of Gronbio.

REAGENT or RESOURCE SOURCE IDENTIFIER Biological samples Human serum sample Shanghai Ovarian Cancer and Family Care Project (SOCFCP) ClinicalTrials.gov: NCT06118307 Standard human serum aliquot GronBio Co., Ltd, China Lot#20240125 Chemicals, peptides, and recombinant proteins C3 (1306–1319) peptide standard Sangon Biotech (Shanghai) Co., Ltd. Lot#P34162-2 Ferric chloride (FeCl3) Sigma-Aldrich Cat#701122; purity: ≥99.9% Ferrous chloride (FeCl2) Sigma-Aldrich Cat#450936; purity: 99.99% trace metals basis Anhydrous toluene Sigma-Aldrich Cat#244511; purity: 99.8% Ammonium hydroxide, 25% Sigma-Aldrich Cat# 30501 Acetonitrile Sigma-Aldrich Cat# 34851; purity: ≥99.9% Trifluoroacetic acid (TFA) Sigma-Aldrich Cat# 302031; purity: ≥99.0% alpha-cyano-4-hydroxycinnamic acid (CHCA) Sigma-Aldrich Cat#70990; Lot#MCKN6411; purity: ≥99.0% Cetyltrimethylammonium bromide (CTAB) Sigma-Aldrich Cat#H6269; purity: ≥99% C18-SiCl3 TCI Development Co., Ltd. Cat#T3815; exact product name: Trichlorooctadecylsilane; LysoPC(18:2/0:0) authentic standard MedChemExpress Cat#HY-N9410; purity: 99.12% Angiotensin II National Institute of Metrology, China Cat#NIM-RM 3622; Lot#240618; Insulin aladdin Cat#I189675; Lot#L2007100; purity/grade: ≥27USP units/mg Cytochrome c aladdin Cat#C301837; Lot#D2115275; purity: ≥95.0% Critical commercial assays CMIA of tumor markers (HE4 and CA125) Abbott ARCHITECT i2000 SR Deposited data Protein sequence UniProt https://www.uniprot.org/ Modification database Unimod https://www.unimod.org/ PeptideAtlas peptide database PeptideAtlas https://peptideatlas.org/ CancerPDF peptide database CancerPDF http://crdd.osdd.net/raghava/cancerpdf/ Metabolite Database Human Metabolome Database (HMDB) https://hmdb.ca/ GO functionally annotated database Gene Ontology (GO) http://geneontology.org KEGG functionally annotated database KEGG https://www.kegg.jp/ Reactome functionally annotated database Reactome https://reactome.org/ Peptidomics dataset from this study National Genomics Data Center OMIX018710; BioProject PRJCA069438; https://ngdc.cncb.ac.cn/bioproject/browse/PRJCA069438 Software and algorithms R R Foundation for Statistical Computing 4.5.2; https://www.r-project.org/ Python Python Software Foundation v3.11.3; https://www.python.org/ Risk of Ovarian Malignancy Algorithm (ROMA) Moore et al. 10 https://doi.org/10.1016/j.ygyno.2008.08.031 Peptide spectral matching PyOpenMS version 3.4.0; https://pyopenms.readthedocs.io/ Light Gradient Boosting Machine (LightGBM) LightGBM developers v4.4.0; https://github.com/Microsoft/LightGBM Scikit-learn scikit-learn developers v1.2.2; https://scikit-learn.org/ SHapley Additive exPlanations (SHAP) SHAP developers shap v 0.46.0; https://github.com/shap/shap Differential analysis ggplot2 developers v3.4.4; https://ggplot2.tidyverse.org/ Orthogonal partial least squares-discriminant analysis SIMCA v14.1 Enrichment analysis clusterProfiler v4.18.1; https://bioconductor.org/packages/clusterProfiler/ Optuna hyperparameter optimization Optuna v3.6.1; https://optuna.org/ Original analysis code GitHub https://github.com/dongbowen1996d/Early-Detection-of-Asymptomatic-Ovarian-Cancer-A-Multi-Center-Prospective-Study Other Clot activator tubes Becton Dickinson (BD) Cat#367955 Fe 3 O 4 @C 18 magnetic beads/nanoparticles GronBio Co., Ltd., China N/A Hydrophobic chip GronBio Co., Ltd., China Cat#007100 MALDI-TOF mass spectrometry Gronbio, China GR-TOF 2000 MS/MS analyses Bruker Daltonics, Germany UltrafleXtreme mass spectrometer LC MS/MS analyses SCIEX, USA AB SCIEX 5500 Qtrap-MS TRIPOD statement Collins et al. 70 https://doi.org/10.1186/s12916-014-0241-z This study was conducted using prospective data from the Shanghai Ovarian Cancer and Family Care Project (SOCFCP; ClinicalTrials.gov: NCT06118307 ) cohort. 38 This ongoing multicenter cohort was established to discover biomarkers associated with ovarian cancer (OC) onset, treatment and prognosis through a rigorous study design, and comprises three subcohorts: a bidirectional patient cohort, a prospective high-risk cohort, and a healthy control cohort. Participants of this study were from five centers: two municipal hospitals (Shanghai First Maternity and Infant Hospital and Fudan University Shanghai Cancer Center) and three community hospitals (Jiading, Minhang, and Jinshan). Initially, a total of 249 OC patients were enrolled from the SOCFCP cohort from two municipal hospitals. Inclusion criteria were as follows: (a) histopathologically confirmed malignant ovarian cancer, fallopian tube cancer, or primary peritoneal carcinoma, and (b) available staging according to the FIGO 2014 criteria. Among these patients, 113 were excluded due to the following reasons: age younger than 20 years ( n = 1), recurrent or metastatic OC ( n = 31), borderline tumors ( n = 26), presence of OC-related symptoms at enrollment ( n = 40), or missing clinical data for age, BMI, CA125, or HE4 ( n = 15). A total of 136 asymptomatic OC patients were included in the final analysis, comprising 57 patients with early-stage disease (FIGO I–II) and 79 patients with late-stage disease (FIGO III–IV). In this study, early-stage OC (ES-OC), defined pathologically as FIGO stage I/II, was used as the operational proxy for the asymptomatic OC (asym-OC) screening scenario, since FIGO stage is an objective pathological classification available for every OC case, whereas symptom status at biomarker sampling, although the intended clinical target, is harder to standardize across centers. The two concepts overlap substantially in this cohort: most asym-OC cases are expected to be early-stage, while symptomatic ES-OC and asymptomatic late-stage OC, although both possible, are less common. Participants were documented as asymptomatic through a review of individual patient case records at the time of enrollment, including incidental findings on imaging for other indications, pelvic mass on routine examination, or investigation of non-specific abdominal complaints that did not meet criteria for OC-related symptoms. Borderline ovarian tumors and metastatic tumors were excluded to maintain a homogeneous study cohort focused on primary invasive ovarian cancer, as their distinct pathology, biology, and clinical management are not aligned with the objectives of this study. A total of 892 non-OC participants were initially enrolled in the SOCFCP cohort. Inclusion criteria for the non-OC group were as follows: (a) no history of malignant tumors; (b) benign ovarian lesions diagnosed by physician assessment; or (c) healthy controls. Among these participants, 185 were excluded for the following reasons: age younger than 20 years ( n = 6), non-fasting blood sampling ( n = 6), history of uterine fibroids and ovarian surgery ( n = 81), Benign Brenner Tumors ( n = 22), or missing data for age, BMI, CA125, or HE4 ( n = 70). A total of 707 non-OC participants were included in the final analysis. This comprised 419 healthy controls and 288 individuals with benign ovarian lesions, among which all cases of ovarian endometriosis were confirmed by surgical histopathology, ensuring diagnostic rigor for this key pathological entity. The severity of ovarian endometriosis was evaluated and staged (I–IV) based on surgical records according to the revised American Society for Reproductive Medicine (rASRM) classification. Participant age was 49.5 ± 12.0 years (mean ± SD) overall, with 54.9 ± 12.5 years for OC patients and 48.5 ± 11.6 years for non-OC participants. As ovarian cancer occurs exclusively in individuals with female reproductive organs, this study was conducted solely in women (gender based on clinical records), making sex/gender comparisons not applicable. The study protocol was approved by the Ethics Committee of Shanghai First Maternity and Infant Hospital, Tongji University (Approval No. KS23167). All study procedures complied strictly with the ethical principles outlined in the Declaration of Helsinki. Venous blood samples were collected from all participants after an overnight fast (≥8 h) and prior to diagnostic confirmation. To establish peptidomic baselines, 3 mL of blood was drawn into clot activator tubes (BD 367955) using standardized protocols. Tubes were inverted 8–10 times, allowed to clot at room temperature for 30–120 min, and centrifuged at 3,000 rpm for 15 min. Serum aliquots were stored at −80°C for subsequent analysis. For OC patients undergoing surgery ( n = 45), an additional serum sample was collected ≥8 days postoperatively to minimize perioperative metabolic interference, with identical processing and storage protocols as baseline samples. Serum HE4 and CA125 levels were quantified using chemiluminescent microparticle immunoassay (ARCHITECT i2000 SR). The Risk of Ovarian Malignancy Algorithm (ROMA) was calculated from HE4, CA125, and menopausal status following the method of Moore et al. 10 The predefined postoperative follow-up window for paired BS-OC/AS-OC analyses was 8–90 days after surgery. All reagents were analytical reagent grade and were used as supplied. Ferric chloride (FeCl 3 ), ferrous chloride (FeCl 2 ), anhydrous toluene, ammonium hydroxide, acetonitrile, trifluoroacetic acid (TFA), α-cyano-4-hydroxycinnamic acid (CHCA), and cetyltrimethylammonium bromide (CTAB) were purchased from Sigma-Aldrich. C 18 -SiCl 3 was obtained from TCI Development Co., Ltd. FeCl 2 (3.17 g) and FeCl 3 (8.11 g) were dissolved in 50 mL of deionized ultrapure water. To this solution, 300 mL of toluene was added, and the mixture was subjected to ultrasonic agitation for 30 min. Subsequently, 8 drops of high-purity CTAB were added to the reaction flask, followed by rapid stirring for 10 min. A 25% ammonium hydroxide solution was then slowly injected into the flask using a micro-injection pump. The reaction was carried out in an 80 ± 2°C ultrapure water bath under an argon atmosphere for 1 h. The magnetic beads were subsequently separated using a magnet and washed three times with deionized ultrapure water. The dried Fe 3 O 4 nanoparticles (7.0 g) were dispersed in anhydrous toluene and sonicated for 1 h. After replacing the solvent with an additional 25 mL of anhydrous toluene, 2.5 mL of C 18 -SiCl 3 was added dropwise with continuous stirring. The reaction was carried out in an ultrapure water bath at 50 ± 2°C under an argon atmosphere for 5 h. The C 18 -decorated magnetic beads were then separated using a magnet and washed three times with toluene. Fe 3 O 4 @C 18 (1.5 mg) and 200 μL of pre-equilibration buffer (20% acetonitrile, 0.1% trifluoroacetic acid) were combined and subjected to ultrasonic mixing for 5 min. The Fe 3 O 4 @C 18 was then separated using a magnet, and the liquid was discarded. A serum sample (100 μL) was mixed with 100 μL of pre-equilibration buffer, and the resulting solution was combined with theFe 3 O 4 @C 18 . After magnetic separation, the mixture was washed with 200 μL of washing buffer (0.1% trifluoroacetic acid). The supernatant was collected after magnetic separation of Fe 3 O 4 @C 18 following a 3-min oscillation elution with 50 μL of desorption solution (70% acetonitrile, 0.1% trifluoroacetic acid). All mass spectrometry (MS) measurements were performed on a GR TOF-2000 MS (GronBio Co., Ltd, China) across a mass range of 100–16,000 Da. Typically, 1 μL of pretreated serum was mixed with 0.5 μL of α-cyano-4-hydroxycinnamic acid (CHCA) matrix solution (10 mg/mL in 70% acetonitrile), then spotted onto a hydrophobic chip (GronBio Co., Ltd, China). After vacuum drying, spectra were acquired where each sample was measured at 9 distinct spots with 90 laser shots per spot using a random walk pattern, yielding a total of 810 shots per sample. Only spectra containing at least one peak with a resolving power >50 and a signal-to-noise ratio >3 were retained; for m/z values above 6,000, a resolving-power threshold of 20 was applied. Instrument performance was monitored daily using a frozen aliquot of a standard human serum. Additionally, a custom calibration standard comprising angiotensin II (m/z 1047.19), insulin (m/z 5808.60), and cytochrome c (m/z 12360.97) was analyzed during each run. Results were deemed consistent if the standard spectrum intensities remained within predefined limits. If any calibration peak deviated by more than 500 ppm, m/z alignment was performed before retesting to maintain accuracy. MS/MS analyses were performed using an UltrafleXtreme mass spectrometer (Bruker Daltonics, Germany) operating in TOF/TOF LIFT mode. First, a target peptide reference repository was then constructed by integrating peptide information from PeptideAtlas and CancerPDF, and precursor ion m/z values obtained from primary mass spectrometry (MS1) were used to screen candidate peptides with a precursor mass tolerance of ±100 ppm. Next, for each candidate peptide, theoretical MS/MS fragment spectra were generated using pyOpenMS (version 3.4.0) and matched against the experimental MS/MS spectra with a fragment ion mass tolerance of 0.02 Da. Finally, to ensure high-confidence peptide identification and minimize false-positive matches, each peptide-spectrum match was evaluated using a hypergeometric distribution framework to calculate hyperscores and p -values. All p -values were adjusted using the Benjamini–Hochberg procedure to obtain q-values, and final peptide identifications were retained only if they satisfied FDR <5% (q 25. This three-step strategy enabled robust MS/MS identification by integrating precise precursor-ion matching (MS1) with high-confidence fragment-ion validation (MS/MS). For lipid verification, LysoPC(18:2) was analyzed by targeted LC-MS/MS/MRM using an authentic standard of 1-linoleoyl-2-hydroxy- sn -glycero-3-phosphocholine [LysoPC(18:2/0:0)]. The standard and serum analyte were measured under the same LC-MS/MS/MRM conditions on a Waters Acquity UPLC system coupled to an AB SCIEX 5500 QTRAP mass spectrometer, using an Acquity UPLC HSS T3 column. LysoPC(18:2) was monitored using the MRM transition Q1/Q3 = 520/184, corresponding to the characteristic phosphocholine head group fragment of LysoPC/PC species. Retention-time deviation was calculated as the absolute difference between the retention time of the serum analyte and that of the authentic standard. Lipid assignment was based on retention-time agreement with the authentic standard and the diagnostic MRM transition. The assignment was reported at the authentic-standard-matched molecular-species level rather than as a full isomer-resolved structural determination. Raw MS data contained approximately 43,900 data points per sample across the 100-16,000 Da range and underwent several preprocessing steps using Python. First, signal fidelity was optimized through Savitzky-Golay filtering, employing local polynomial regression to smooth spectral curves while preserving critical peak morphology. Subsequently, baseline distortions were systematically removed using Statistical-sensitive Non-linear Iterative Peak-clipping (SNIP) algorithm, which iteratively eliminates low-frequency drift components without compromising peak amplitude integrity. Peaks were normalized using the Total Ion Current (TIC) method to minimize experimental variations and enable comparative analysis across samples. After these preprocessing steps, a total of 1,081 reproducible m/z features were extracted for downstream analysis. Batch- and center-related analytical variation was controlled through run-level calibration, QC-based acceptance criteria, batch-level stability assessment, and independent reproducibility evaluation. For each analytical run, a custom calibration standard containing angiotensin II, insulin, and cytochrome c was analyzed. Runs were accepted only when standard spectrum intensities remained within predefined limits; if any calibration peak deviated by more than 500 ppm, m/z alignment was performed before retesting. Batch-level stability was evaluated across six independent analytical batches. To explicitly quantify and control inter-laboratory variability, three laboratory/operator/instrument/reagent-lot combinations were tested over five days with five replicates per day. Inter-laboratory reproducibility was calculated using an ANOVA-based variance-component method, and two-sided 95% confidence intervals were estimated using the corresponding effective degrees of freedom and chi-square distribution. All evaluated features should meet the predefined CV ≤ 15% acceptance criterion. This study was conducted and reported in accordance with the TRIPOD guidelines 70 and the step-by-step guide of clinical prediction model development. 71 To better externally validate the developed model, the study cohort was divided based on the clinical center of recruitment. The development dataset included patients from Shanghai First Maternity and Infant Hospital, Jiading Community Hospital, and Jinshan Community Hospital. The external validation dataset included patients from Fudan University Shanghai Cancer Center and Minhang Community Hospital. The development dataset was further split (8:2 ratio) into training and test datasets using stratified random sampling based on pathological subtype. A Light Gradient Boosting Machine (LightGBM) diagnostic model (Scikit-learn v1.2.2) was constructed and optimized through 200 Optuna hyperparameter trials, using a GBDT booster (learning rate = 0.1) with gradient-based one-sided sampling (GOSS) and exclusive feature bundling (EFB). Model robustness was evaluated via 5-fold cross-validation. To derive a compact panel with robust performance and high biological interpretability, we applied a four-step biology-informed ML-based feature selection workflow on all 1,081 serum peptidomic features: (a) differential analysis across five clinically relevant comparisons (OC vs. non-OC; early-stage OC vs. non-OC; HGSOC vs. non-OC; non-HGSOC vs. non-OC; and OC vs. ovarian endometriosis) to retain only intersecting significantly dysregulated features (FDR <0.05); (b) literature review upon OC-related biomarkers; (c) MS/MS validation to exclude features lacking clear fragmentation patterns; and (d) ML-based feature-importance ranking to derive the final panel from the 18 candidate peptidomic features generated in the previous steps. Feature contributions were quantified using SHapley Additive exPlanations (SHAP) values computed with the shap Python package (v0.46.0). The final 7-marker panel was determined through an integrated evaluation of model-derived feature importance and existing literature evidence. Three features - (LysoPC(18:2), FPA(1–16) and CFH(176–183)) - were first selected due to biological relevance reported in literature. The remaining candidates were ordered by descending SHAP importance, and were added to the base model one at a time in that order. At each step, a LightGBM classifier was trained on the training set and applied to the held-out test set to obtain predicted probabilities. To assess whether each successive addition improved performance, the test-set predictions were bootstrapped (B = 5000 resamples with replacement), yielding an empirical sampling distribution of the AUC at each step. Pairwise 95% confidence intervals were then computed for the AUC difference between successive models. Adding the fifth feature (m/z 1202.6) significantly degraded performance relative to the four-addition model (ΔAUC = +0.018 in favor of four additions; 95% CI [+0.002, +0.039], probability of superiority = 98.8%), indicating overfitting. We constructed and compared eight peptidomic signature (ProMS) models: Full ProMS (all 1,081 features), Int-ProMS (18 intersected and MS/MS-validated features), ProMS (seven-feature panel), ProMS-Age (panel + Age), ProMS-CA125 (ProMS + Age + CA125), ProMS-HE4 (ProMS + Age + HE4), ProMS-ROMA (ProMS + Age + ROMA), and ProMS+ (ProMS + Age + CA125 + HE4). Model performance was evaluated using multiple metrics, including the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy. To provide an intuitive comparison of the model performance across the six aforementioned metrics, a radar chart was employed for visualization. To compare the performance stability of the eight models, we generated kernel density estimation (KDE) plots for the distribution of AUC values obtained from 50 independent runs (each using a unique random seed from 0 to 49). Each model was evaluated in both the whole dataset and the independent external validation dataset. To estimate the potential performance of ProMS+ in populations with lower disease prevalence than the present clinical cohort, Monte Carlo simulations were conducted following the approach used by the NCI CISNET program. 72 , 73 For each simulation ( n = 100,000), synthetic individuals were generated by fitting Gaussian kernel density estimators to each input variable within each clinical subgroup and randomly sampling from the corresponding subgroup-specific distributions. The simulated values were entered into the locked ProMS+ model to generate predicted probabilities. ROMA scores were recomputed for each simulated individual using the age-stratified Moore et al. formula. AUC was estimated with 95% confidence intervals using the DeLong method, and PPVs were calculated at pre-specified clinical cutoffs using bootstrap resampling. Seven prevalence settings were derived from published cohorts: 0.06% (UKCTOCS; Menon et al., 2021), 0.37% (SEER and U.S. Census data), 1.00% (OC3 study), 2.60% (GOG-0199; Greene et al., 2011; Sherman et al., 2014), 4.10% (Fudan University Cancer Center cohort), 5.00% (meta-analysis of RRSO studies; Cheng et al., 2020), and 8.30% (Peking Union Medical College Hospital BRCA1/2 carrier cohort). The prevalence of endometriosis was uniformly set at 0.36 46 in both populations. We note that KDE sampling preserves marginal distributions but does not retain multivariable correlations present in real populations; the simulation results therefore represent an extrapolation of within-cohort performance and should be interpreted accordingly. To assess the influence of class imbalance, we performed a balanced-resampling sensitivity analysis. Non-OC samples in the internal training set were randomly down-sampled to match the number of OC samples, generating 1:1 OC/non-OC training datasets across 50 independent seeds (0–49). For each balanced subsample, ProMS+ was retrained from scratch using the same ten input variables and evaluated in the internal test set, external validation set, and clinically relevant subgroups. Imbalance-sensitive metrics, including AUC-PR, F1-score, MCC, balanced accuracy, and confusion matrices, were calculated. The statistical details of experiments, including the exact value of n and what it represents, as well as the definition of center and dispersion measures, can be found in the Results, Tables, and corresponding figure legends. All analyses were performed using R (v4.5.2) or Python (v3.11.3). Normality of continuous variables was assessed using the Shapiro-Wilk test. Group comparisons were conducted using χ 2 , Fisher’s exact, Student’s t-, one-way ANOVA test, Mann–Whitney U, or Kruskal–Wallis tests as appropriate. For the postoperative biomarker analysis, comparisons between preoperative OC samples (BS-OC) and non-OC samples were performed using two-sided Wilcoxon rank-sum tests, whereas paired preoperative/postoperative comparisons (BS-OC vs. AS-OC) were analyzed using two-sided Wilcoxon signed-rank tests. p values were adjusted using the Benjamini-Hochberg procedure to control the false discovery rate (FDR) across the biomarker panel. Multiple testing correction was applied using the Benjamini-Hochberg procedure (FDR <0.05). Differential metabolites (FDR <0.05) were visualized via volcano plots (ggplot2 v3.4.4). Orthogonal partial least squares-discriminant analysis (OPLS-DA, SIMCA v14.1) was performed to cluster peptidomic features, evaluating overall group differences. Candidate features were functionally annotated and enriched through GO, KEGG, and Reactome databases (ClusterProfiler via BiocManager v1.30.25; FDR <0.05).

We sincerely thank Dr. Yi Liu for his valuable guidance on model construction and methodological support in machine learning, which significantly contributed to this work. We also thank Ms. Mengtian Zheng for her important assistance in sample collection and processing. Besides, we are grateful to all study participants for their invaluable contributions. This work was supported by the 10.13039/501100012166 National Key Research and Development Program of China ( 2024YFA0917700 ), the Shanghai Municipal Commission of Science and Technology ( 25TS1412200 ), the 10.13039/501100001809 National Natural Science Foundation of China (grant nos. 82272888 and 82204047 ), Shanghai Hospital Development Center Foundation (grant nos. SHDC12022106 and 2022SKLY-12 ), Shanghai First Maternity and Infant Hospital (grant nos. 2023B01 and 2023A01 ), and Health Commission of Shanghai Pudong District (grant nos. PWZxq2022-03 and PW2023D-14 ).