Plasma proteome–metabolome signatures enable non-invasive early detection and lymph node risk stratification in breast cancer Authors: Wei Zhang, Yao Yao, Yimeng Wang, Lesang Shen, Jinghua Ding, Yuxuan Zhu, Heli Xu, Yinkuan Shao, Xidong Gu, Haiqi Lu, Jun Zhou, Haiming Deng, Jun Zhou, Wuzhen Chen, Wenjie Xia, Jinxing Jiang, Xiuyan Yu, Shanshan Sun, Jiaxin Chen, Jian Liu, Dezheng Wang, WenJia Liu, Ziao Lin, Kailun Xu, Qijun Wu, Jian Huang, Hengyu Li, Zhuohang Yu, Chao Ni Published: 2026 ABSTRACT AbstractDespite widespread use of ultrasound and mammography, the accuracy of early breast cancer detection remains suboptimal, particularly in Asian women with dense breast tissue, underscoring the unmet need for biologically informed, non-invasive diagnostic approaches. Notably, systematic characterization of circulating proteomic and metabolomic alterations in early-stage breast cancer remains limited, especially in large, well-validated cohorts. Here, leveraging a registered multicenter prospective study (NCT06016790) together with an independent external validation cohort, we enrolled 662 participants across nine healthcare institutions to evaluate a plasma-based multi-omics liquid-biopsy framework for non-invasive detection. Data-independent acquisition proteomics and untargeted metabolomics profiled 5,549 proteins and 630 metabolites, with targeted validation performed in independent cohorts. Integrated analyses revealed coordinated molecular remodeling characterized by enrichment of cytoskeleton- and adhesion-associated proteins (for example ACTN1, VCL, ITGA2B and MYH9) together with rewiring of lipid-metabolic pathways. Network and trajectory modeling further identified 13 malignancy-associated protein modules and two progression-linked metabolic trajectories. Based on these features, we developed ProMeta-BC, a combined plasma proteome-metabolome model incorporating 15 proteins and 5 metabolites, which achieved robust discrimination between benign and malignant lesions (AUC 0.973 and 0.951 in training and validation cohorts, respectively), and ProMeta-BC-LN for prediction of axillary lymph-node metastasis (AUC 0.841 and 0.753). Notably, in cases with discordant radiological and pathological findings, the model detected 8 of 9 imaging-negative cancers and correctly reclassified 38 of 56 patients with inconsistent lymph-node calls, indicating complementary clinical utility. Together, these findings show that coordinated molecular signatures encoded in the circulating proteome–metabolome capture disease-relevant biology and provide a scalable, interpretable framework for non-invasive detection and stratification in breast cancer.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12943-026-02713-7.Keywords: Breast cancer, Early detection, Plasma proteomics, Metabolomics, Multi-omics integration Introduction Breast cancer remains one of the most prevalent and lethal malignancies among women globally, with clinical outcomes largely contingent upon disease stage at diagnosis, particularly the extent of regional lymph node involvement [1, 2]. Although imaging-based screening, including ultrasound and mammography, plays a central role in early detection [3], its diagnostic accuracy remains suboptimal for early-stage lesions, particularly in Asian women with a high prevalence of dense breast tissue [4, 5]. Moreover, conventional imaging modalities primarily capture structural and morphological abnormalities, providing limited insight into tumor-associated biological processes and functional state [6, 7]. Consequently, there is a critical unmet need for non-invasive detection strategies that not only improve early diagnostic sensitivity but also capture systemic biological alterations associated with early tumor development. Cancer initiation and progression are not abrupt events but are accompanied by prolonged, coordinated remodeling across multiple molecular layers, including immune regulation, metabolic reprogramming, and extracellular signaling [8]. Growing evidence indicates that tumor-associated protein secretion, immune modulation, and metabolic perturbations leave detectable molecular imprints in peripheral circulation well before tumors become radiologically apparent [9, 10]. Accordingly, liquid biopsy has emerged as a promising bridge linking tumor biology to clinical detection. Recent efforts have focused primarily on circulating cell-free DNA (cfDNA)–based approaches, including methylation profiling and fragmentomic analysis, which have demonstrated encouraging performance in early cancer detection [11–13], including our recent development of a cfDNA fragmentomics-based model (TuFEst) in large prospective cohorts. Despite their diagnostic potential, cfDNA-based strategies largely reflect tumor-derived genomic and epigenetic features but offer limited biological interpretability at the functional molecular level. In particular, they provide relatively little information regarding systemic alterations in circulating proteins and metabolites, which represent downstream effectors of tumor–host interactions and are difficult to validate within in situ tumor ecosystems. Beyond genomic alterations, increasing evidence suggests that early tumor-associated molecular changes can be captured in peripheral blood at the proteomic and metabolomic levels [14–16]. At the proteomic level, Pitteri et al. reported elevated plasma EGFR levels prior to breast cancer diagnosis, conferring a 2.90-fold increased risk in the highest quartile [17]. In addition, breast cancer cells secrete multiple proteins, including IGFBP3, CD44, osteopontin, and syndecan-1, which are detectable in circulation even at early stages [18]. At the metabolomic level, early-stage breast cancer has been associated with characteristic alterations in choline-containing phospholipids, glutamine, and sphingomyelins, reflecting enhanced membrane biosynthesis and metabolic reprogramming [19, 20]. However, most existing studies have primarily reported associative relationships between individual proteins or metabolites and breast cancer, are often limited by small sample sizes, retrospective designs, or single-center settings, and rarely extend to systematic model construction and independent validation. Whether circulating proteomic and metabolomic signatures can be integrated into robust, generalizable diagnostic frameworks for early breast cancer therefore remains largely unexplored. Furthermore, comprehensive characterization of plasma proteome–metabolome landscapes in Asian populations has been particularly scarce, despite substantial differences in genetic background, lifestyle, and environmental exposure compared with Western cohorts. Here, leveraging a registered multicenter prospective study for early breast cancer detection (NCT06016790) together with an independent external validation cohort spanning nine healthcare institutions, we enrolled 662 participants, with early-stage disease accounting for 79.5% of cancer cases, to systematically evaluate a plasma-based proteomic and metabolomic liquid biopsy framework. Using data-independent acquisition (DIA) proteomics and untargeted metabolomics, followed by targeted validation with parallel reaction monitoring (PRM) and quantitative metabolomics, we comprehensively profiled circulating molecular alterations associated with early breast cancer. Through integrative network analysis and machine-learning–based modeling, we developed ProMeta-BC, a plasma proteome–metabolome diagnostic model enabling robust benign–malignant discrimination, as well as ProMeta-BC-LN for preoperative prediction of axillary lymph node metastasis. Together, our study establishes a systematically validated, biology-informed multi-omics framework for non-invasive early breast cancer detection, provides a population-specific molecular reference for Chinese women, and offers mechanistic insights into systemic tumor–host interactions with potential implications for precision diagnosis and triage of patients with breast lesions, risk stratification, and early intervention. Results Study design and multicenter cohort architecture To evaluate the clinical applicability of plasma proteomic–metabolomic signatures for non-invasive early detection of breast cancer, we established a multicenter prospective study framework encompassing discovery and independent validation cohorts (Fig. 1). A total of 662 female participants were enrolled, including 501 individuals from a registered multicenter prospective study (NCT06016790) conducted across six tertiary hospitals and one community medical center, and 161 participants from an independent external validation cohort comprising two additional institutions. The study population included 218 benign breast disease controls and 444 pathologically confirmed breast cancer patients, with early-stage disease (stage 0–IIa) accounting for 79.5% of cancer cases, consistent with an early-detection–oriented diagnostic context in patients referred for evaluation of breast lesions. All participating centers followed standardized protocols for peripheral blood collection and plasma processing prior to multi-omics profiling. Overview of the multicenter plasma proteome–metabolome framework for early detection and lymph node assessment Participants were stratified into three independent cohorts according to analytical workflow. The discovery cohort (Cohort 1, n = 150) underwent data-independent acquisition (DIA) proteomics combined with untargeted metabolomics, enabling comprehensive profiling of 5,549 proteins and 630 metabolites to delineate the circulating molecular landscape and identify candidate biomarkers. Two independent validation cohorts (Cohort 2, n = 351; Cohort 3, n = 161) were subsequently analyzed using parallel reaction monitoring (PRM)-based targeted proteomics and quantitative metabolomics to rigorously assess feature stability, reproducibility, and cross-platform robustness. Cohort 3 was independently collected from Changhai Hospital and Chaohui Community Health Service Center, with no sample-source overlap with Cohort 1 or Cohort 2, and was pre-specified as a held-out external validation set not involved in any feature selection or model training step. Baseline clinical characteristics were highly comparable across the three cohorts (Supplementary Table 1). No significant differences were observed in age, body mass index (BMI), clinical stage, molecular subtype distribution, or TNM classification (Figures S1A–D). In addition, age and BMI were well balanced between benign and malignant groups (Figure S1C), and disease stage and nodal status distributions were consistent across cohorts (Figure S1D). These results indicate that the multicenter cohorts were well matched and provided a robust foundation for downstream molecular analyses and predictive model construction. Shared and subtype-specific remodeling of the circulating proteome in early-stage breast cancer Using DIA-based plasma proteomics data from the discovery cohort (Cohort 1), we systematically compared circulating protein expression profiles between breast cancer patients and benign controls. The number of quantifiable proteins in overall malignant samples and very early stage (stage 0-Ⅰ) was significantly higher than in benign controls (Wilcoxon rank-sum test, P = 0.040 and P = 0.025, respectively), indicating that tumor-associated biological alterations are already reflected in peripheral circulation at very early disease stages (Fig. 2A, Supplementary Table 2a). Plasma proteomic differential features of malignant breast cancer and molecular subtypes. A Comparison of quantified plasma protein numbers across benign breast disease, Stage 0–I, and malignant breast cancer samples (Wilcoxon rank-sum test, P < 0.05); B PCA of the plasma proteome showing the distribution of benign and malignant samples in the global expression space; C Volcano plots of differential plasma proteins comparing malignant breast cancer overall and molecular subtypes with benign breast disease. Proteins with an adjusted P value < 0.05 and an absolute fold change exceeding the predefined threshold are considered significantly increased (red) or decreased (blue); proteins not meeting these criteria are shown in grey. Representative subtype-associated proteins are labeled; D Pathway enrichment analysis of proteins altered between malignant breast cancer and benign breast disease based on Gene Ontology (GO) and KEGG annotations; E Common and specific pathway enrichment features across molecular subtypes. Left: Pathways consistently enriched across all subtypes; Right: Pathways exhibiting subtype-specific enrichment patterns. F Subtype-specific pathway enrichment analysis based on differential plasma proteins. Dot size represents fold enrichment, and color intensity indicates statistical significance Principal component analysis (PCA) revealed substantial overlap between malignant and benign samples in the global proteomic space (Fig. 2B), suggesting that early-stage breast cancer is characterized by coordinated alterations in specific protein subsets rather than widespread disruption of the circulating proteome. Differential expression analysis identified 479 significantly altered proteins, including 308 upregulated and 171 downregulated species. Upregulated proteins were predominantly involved in actin cytoskeleton organization (TUBA4B, TPM4, MYL1) and cell–substrate adhesion (RSU1, RAP1B) (Fig. 2C, Supplementary Table 2c). Functional enrichment analysis further demonstrated that these proteins were enriched in extracellular vesicle biology, cell adhesion, and cytoskeletal regulation, together with integrin-mediated signaling and multiple metabolic pathways, including glycolysis, hexose catabolism, and purine nucleotide metabolism (Figs. 2D, E, Supplementary Table 2d, e). Notably, these pathways were consistently enriched across major molecular subtypes, indicating the presence of a shared circulating malignant program. Given the biological heterogeneity among breast cancer subtypes, we next performed subtype-resolved proteomic analyses. Despite substantial overlap in the global expression space (Figure S2A), subtype–benign comparisons revealed a core set of concordantly regulated proteins shared across HR⁺/HER2⁻, HER2⁺, and TNBC subtypes (Fig. 2C; Figure S2B), supporting the existence of common breast cancer–associated proteomic modules. In addition to this shared framework, each subtype exhibited characteristic deviations aligned with its molecular context (Fig. 2C and F; Figure S2C). HR⁺/HER2⁻ tumors were enriched for hormone-responsive pathways, including estradiol response and estrogen receptor signaling. HER2⁺ tumors preferentially exhibited activation of phosphatidylinositol metabolism and JAK–STAT signaling, consistent with aggressive biological behavior and central nervous system metastasis tropism [21, 22]. In contrast, TNBC was characterized by prominent enrichment of immune-related pathways, particularly PD-1/PD-L1 signaling, reflecting its immunogenic phenotype. Collectively, these findings indicate that the circulating proteome simultaneously encodes a conserved malignant core program and subtype-specific biological modulation, providing a molecular basis for developing broadly applicable yet biologically informed blood-based biomarkers for early breast cancer detection. Targeted plasma proteomics reveals malignancy-associated protein modules and candidate biomarkers Building upon preliminary untargeted plasma proteomics profiling, we performed targeted proteomics to achieve high-precision quantification of 312 candidate proteins. Principal component analysis (PCA) of the targeted proteome revealed clear separation between benign and malignant samples in the global expression space (Fig. 3A), supporting systemic proteomic reprogramming associated with malignant transformation. Differential analysis identified consistent alterations across overall malignancy and three major molecular subtypes (HR⁺/HER2⁻, HER2⁺, and TNBC), with 250 proteins commonly dysregulated across subtypes (Fig. 3B; Figure S3A, Supplementary Table 3b). These shared alterations were predominantly characterized by upregulated proteins, including PHB1, HK1, COTL1, and TMEM109, which have been implicated in mitochondrial metabolism, cytoskeletal dynamics, cell migration, and tumor-microenvironment remodeling in multiple malignancies, processes closely linked to tumor invasion and metastatic progression [23–26]. Together, these proteins constitute a biologically coherent candidate set for subsequent functional and network-level analyses. Modular reconstruction of targeted plasma proteome and candidate panel identification. A PCA scatter plot of the plasma protein expression matrix, displayed by benign and malignant sample grouping. B Log₂ fold change distribution from protein differential analysis across four comparisons (malignant, HR⁺/HER2⁻, HER2⁺, TNBC versus benign). C Protein co-expression network and module division (ME01-ME13) constructed based on WGCNA, with representative functional annotations labeled for each module. D Visualization of correlations between module eigengenes and continuous variables as well as categorical clinical variables. E Protein co-expression network constructed based on WGCNA in Cohort 2. Each node represents a protein, with node color depth indicating log₂ fold change in the malignant versus benign sample comparison. Labeled nodes correspond to hub proteins in their modules, with label colors denoting the module assignment. F mFuzz clustering results of protein expression trajectories based on clinical stage, showing change trends of different clustering patterns. G Workflow for integrative identification of candidate proteins To dissect the organizational architecture underlying these alterations, we constructed a weighted co-expression network using weighted gene co-expression network analysis (WGCNA). Thirteen stable and functionally distinct modules were identified (Fig. 3C; Figure S3B, C, Supplementary Table 3c), enriched in processes including actin cytoskeleton regulation (ME01), ECM–receptor interaction and integrin signaling (ME07), sphingolipid signaling (ME09), glycolysis/gluconeogenesis (ME06), and FcγR-mediated phagocytosis (ME04). Correlation analysis revealed that most modules were significantly associated with malignant status, clinical stage, and molecular subtype (Fig. 3D; Figure S3D, Supplementary Table 3d). Hub protein analysis further identified ACTN1, VCL, YWHAE, and LDHA as highly connected network nodes, reflecting their central roles in cytoskeletal remodeling, signal transduction, and metabolic reprogramming associated with malignant phenotypes (Fig. 3E; Figure S3E, Supplementary Table 3e) [27–30]. To characterize dynamic proteomic changes during disease progression, we applied mFuzz soft clustering analysis across clinical stages (benign → stage 0 → stages I–IV) (Fig. 3F; Figure S3F, Supplementary Table 3f). Multiple protein clusters exhibited progressive abundance changes with advancing stage. Proteins showing stage-dependent upregulation were enriched in cytoskeletal and adhesion-related pathways (e.g., MYH9 and ILK), consistent with their established roles in breast cancer invasion and metastasis [31, 32]. In contrast, proteins exhibiting progressive downregulation included epithelial and immune-related molecules such as CDH1 and CXCL10, indicating attenuation of epithelial integrity and immune surveillance during malignant progression [33, 34]. Although MYH9 and ILK have been extensively linked to tumor progression and poor prognosis in tissue-based studies [35, 36], their expression patterns and clinical relevance in peripheral blood remain largely unexplored. Integrating differential expression significance (P < 0.05, |log₂FC| > 0.5), network centrality, and dynamic trajectory patterns, we ultimately prioritized 27 candidate proteins displaying highly consistent expression, structural connectivity, and temporal behavior (Fig. 3G), constituting a core biomarker panel with strong translational potential. Plasma metabolomic remodeling and functional alterations in early-stage breast cancer Consistent with functional enrichment analyses revealing prominent involvement of metabolic pathways (Figs. 2D, E), and WGCNA hub protein analysis identifying metabolism-related regulators such as HK1, LDHA, and YWHAE (Fig. 3E), we next characterized the plasma metabolome to complement the proteomic findings [37]. By integrating untargeted and targeted metabolomic platforms, we reliably quantified more than 600 metabolites passing stringent quality-control criteria, covering major biochemical classes. Lipid species constituted the largest fraction (43.4%), followed by amino acids (21.3%) and xenobiotics (14.3%), indicating broad representation of systemic metabolic processes (Fig. 4A; Figure S4A, Supplementary Table 4a). Plasma metabolomic differential features of malignant breast cancer and molecular subtypes. A Compositional distribution of quantified metabolites in untargeted plasma metabolomics, classified by major biochemical categories (n = 630). B PCA of the plasma metabolome, with benign and malignant samples highly overlapping in the overall metabolic profile. C Volcano plot of differential metabolites between malignant breast cancer and benign breast disease, with significantly upregulated and downregulated metabolites screened based on FDR-corrected P values. D Metabolite interaction and pathway topology network, displaying coordinated alterations in amino acid, lipid, carbohydrate, and xenobiotic metabolism pathways. E Heatmap of differential metabolites showing consistent change direction in malignant versus benign comparison and maintained consistency across molecular subtypes (HR⁺/HER2⁻, HER2⁺, TNBC), grouped and displayed by metabolite category. F KEGG pathway enrichment results for different subgroups (overall malignant, HR⁺/HER2⁻, HER2⁺, TNBC). G Pathway enrichment features based on metabolite set enrichment analysis (MSEA), showing normalized enrichment scores (NES) for each subgroup Principal component analysis (PCA) revealed substantial overlap between malignant and benign samples in the global metabolomic space (Fig. 4B), suggesting that early-stage breast cancer is characterized by coordinated alterations in specific metabolic pathways rather than widespread metabolic disruption. Differential analysis identified 138 significantly altered metabolites, including 92 upregulated and 46 downregulated species (Fig. 4C, Supplementary Table 4c). Pathway enrichment analysis using KEGG and metabolite set enrichment analysis (MSEA) revealed prominent enrichment of lipid metabolic pathways, including unsaturated fatty acid biosynthesis, linoleic acid and arachidonic acid metabolism, and fatty acid synthesis and degradation, together with aromatic amino acid metabolism, particularly phenylalanine–tyrosine–tryptophan pathways (Fig. 4D, F and G; Figures S4C, D, Supplementary Table 4f, g), Previous studies indicate abnormal lipid and amino acid metabolism played a vital role in breast cancer initiation and progression [38, 39]. Consistent with this, our findings indicate that early breast cancer–associated metabolic remodeling could also be detected in peripheral blood. Notably, several metabolites identified in our analysis have established functional relevance in breast cancer biology. Fatty acid derivatives such as arachidonic acid and linoleic acid have been implicated in immune regulation, tumor metabolism, and therapeutic response, while L-serine, a central node in amino acid metabolism, supports nucleotide biosynthesis and tumor proliferation [40]. Metabolic network topology analysis further integrated pathway-level information and revealed that linoleic acid– and arachidonic acid–related lipid species formed highly interconnected clusters, indicating a coordinated systemic remodeling of fatty acid metabolism in breast cancer (Figs. 4F, G). Together, these results suggest that dysregulation of lipid and amino acid metabolism extends beyond the local tumor microenvironment into peripheral circulation, underscoring the potential of circulating metabolic signatures as indicators of early malignant transformation. To evaluate subtype-related metabolic heterogeneity, we performed intersubtype comparisons across HR+HER2-, HER2+, and TNBC groups. Our analyses identified concordantly regulated metabolites shared across subtypes, forming conserved breast cancer–associated metabolic clusters (Fig. 4E; Figure S5A). These subtype-consistent metabolites were predominantly involved in lipid and amino acid metabolism, with pathway enrichment patterns remaining highly consistent across subtypes (Figs. 4F, G; Figures S5, S6). Collectively, these results demonstrate that early-stage breast cancer is characterized by systemic metabolic remodeling that is largely conserved across molecular subtypes. Targeted metabolic analysis delineates malignancy-associated metabolic architecture To validate and refine candidate metabolites identified by untargeted metabolomics, we performed targeted quantitative metabolomic profiling. Orthogonal partial least squares–discriminant analysis (OPLS-DA) revealed clear separation between benign and malignant samples in the targeted metabolic feature space, indicating that the selected metabolites robustly capture systemic metabolic alterations associated with malignant transformation (Fig. 5A; Figure S7A). Differential analysis identified a coordinated pattern of metabolic reprogramming characterized by 24 upregulated and 17 downregulated metabolites (Fig. 5B, Supplementary Table 5b). Multi-layer integrated analysis dissects plasma metabolic remodeling features associated with breast cancer. A OPLS-DA-based plasma metabolome analysis, showing the distribution of benign and malignant samples in metabolic feature space. B Differential metabolite volcano plot, with x-axis showing log₂ fold change (malignant vs benign) and y-axis showing −log₁₀(P value); dot color indicates upregulation, downregulation, or non-significance, and dot size indicates VIP value. C Dynamic change patterns of key metabolite clusters identified by mFuzz soft clustering analysis. Clusters 1 and 9 are shown as representative trajectories, illustrating pronounced and contrasting abundance changes across clinical stages (Benign → stage 0 → I → II → III → IV). D Spearman correlation heatmap between proteome WGCNA modules and metabolite clusters, with dot size indicating significance level. E Multi-layer network integration schematic of protein-pathway-metabolic pathway-metabolite, with different colors and shapes representing node types and connecting lines representing cross-layer associations Among the upregulated metabolites, multiple lipid species, including phosphatidylethanolamines (PEs), together with amino acid–related metabolites such as 2-hydroxybutyric acid and orotic acid, were markedly increased, whereas purine derivatives and acylcarnitines were substantially reduced. Decreased hypoxanthine levels suggest reprogramming of purine salvage pathways involved in tumor–immune interactions [41], while reduced acylcarnitines reflect impaired mitochondrial fatty acid β-oxidation, a metabolic adaptation that promotes cellular plasticity and anabolic remodeling in cancer cells. Together, these bidirectional alterations indicate coordinated remodeling of membrane lipid metabolism and nitrogen utilization during malignant progression. To characterize dynamic metabolic changes aligned with disease development, we applied mFuzz-based soft clustering across clinical stages (benign → stage 0 → stages I–IV), identifying ten reproducible metabolic trajectories (Fig. 5C; Figures S7B, C, Supplementary Table 5c). Among these clusters, Cluster 1 exhibited a progressive decline with advancing stage, whereas Cluster 9 displayed a monotonic stage-dependent increase. Cluster 1 was predominantly composed of medium-chain acylcarnitines, including octanoylcarnitine, decanoylcarnitine, and decenoylcarnitine. Consistent with studies across multiple solid tumors, reduced circulating acylcarnitines represent a pro-tumorigenic metabolic reprogramming characterized by suppressed mitochondrial fatty acid oxidation and preferential diversion of lipid substrates toward anabolic processes supporting membrane biosynthesis and biomass accumulation [42, 43]. In contrast, Cluster 9 comprised several microbiota–derived metabolites, including phenylacetylglutamine (PAGln), 3-indoxyl sulfate, and N-acetyltryptophan, together with the pyrimidine precursor orotic acid. PAGln signals through adrenergic receptors and has been associated with lethal cancer risk in prospective cohorts, whereas orotic acid accumulation reflects enhanced de novo pyrimidine biosynthesis required for rapid tumor proliferation. These trajectories reveal that metabolic remodeling during early breast cancer progression occurs in coordinated, pathway-level patterns rather than through isolated metabolite fluctuations. Importantly, the trajectory framework was also independently validated, with Cluster 1 and Cluster 9 exhibiting consistent stage-associated dynamics in an external cohort (Figure S8B). Cross-omics integration analysis further revealed extensive functional coupling between metabolite trajectories and WGCNA-derived protein modules (Fig. 5D, Supplementary Table 5d). Multiple significant associations involved modules related to cytoskeletal remodeling, cell–adhesion signaling, and immune regulation. Specifically, Cluster 1 showed negative correlations with ME01 (actin cytoskeleton regulation) and ME02, suggesting coupling between reduced fatty acid oxidation and cytoskeletal reorganization. Conversely, Cluster 9 displayed strong positive correlations with ME09 (sphingolipid signaling), ME10, and ME01, indicating close linkage between metabolic reprogramming and structural remodeling during tumor invasion. In addition, ME06 (glycolysis/gluconeogenesis) and ME07 (ECM–receptor interaction) exhibited coordinated associations with multiple metabolite clusters, reflecting complex metabolic–signaling crosstalk. To further delineate the mechanistic architecture underlying these associations, we constructed multilayer functional networks integrating proteins, protein pathways, metabolite pathways, and metabolites (Fig. 5E, Supplementary Table 5e). Network analysis identified several cross-omics hub nodes, including actin cytoskeleton regulation linked to sphingolipid signaling, integrin–ECM signaling interfacing with glycerophospholipid metabolism, and pyrimidine metabolism serving as a central hub connecting Cluster 9 metabolites to carbon metabolism networks. These cross-layer interactions mechanistically integrate metabolic adaptation with cytoskeletal remodeling, adhesion–migration dynamics, and proliferative signaling. Notably, Cluster 9 metabolites exhibited strong enrichment in lipid- and nucleotide-related metabolic programs and tight connectivity with key protein modules (ME01 and ME09/ME10), defining a coordinated cross-omics remodeling axis. Based on its consistent stage-dependent trajectory, biological relevance, and network centrality, Cluster 9 was prioritized as the candidate metabolic biomarker panel. The seven metabolites within this cluster were subsequently validated by targeted quantification and advanced to downstream diagnostic modeling. ProMeta-BC enables robust benign–malignant discrimination across cohorts Based on 27 proteins and 7 metabolites prioritized from targeted proteomics (Fig. 3) and metabolomics (Fig. 5) in Cohort 2 (n = 351), we constructed ProMeta-BC, an integrated plasma proteome–metabolome diagnostic model for breast cancer (Fig. 6A). Feature selection was performed using recursive feature elimination combined with multivariate screening, resulting in a compact panel of 20 high-confidence features (15 proteins and 5 metabolites) representing the most informative cross-omics subset (Figure S9A). Among multiple candidate algorithms evaluated, the stacking ensemble model (SEM)–based implementation achieved the optimal balance between discriminative performance and model stability (Fig. 6B, Supplementary Table 6b). Construction and validation of the ProMeta-BC diagnostic model for benign-malignant discrimination. A Schematic diagram of model construction and validation workflow. Based on plasma data from Cohort 2, integrating 27 targeted proteins, 7 metabolites, and their combined features, diagnostic models were constructed through feature screening, recursive feature elimination, and multiple machine learning algorithms (GLM, LR, SVM, GBM, MLP, SEM), and externally validated in independent cohorts (Cohort 3) and a separate set of 9 independent patients. B Comparison of model performance of different machine learning methods in the training set, evaluating model stability and overall discriminative ability by AUC distribution. C ROC curves for benign–malignant discrimination in the external validation cohort (Cohort 3), comparing ProMeta-BC with its single-omics counterparts (ProMeta-BC-Proteome and ProMeta-BC-Metabolome). Left, all malignant cases; right, early-stage (0–I) cases. D Ranking of relative contributions of key features to benign–malignant prediction in the ProMeta-BC integrated feature panel. E Distribution of ProMeta-BC SEM-derived prediction scores in the external validation set (Cohort 3), shown as PC1 versus predicted malignancy probability. Left, color-coded by benign (NBC) versus malignant (BC); right, color-coded by clinical stage. F Confusion matrices of ProMeta-BC performance in Cohort 3 stratified by age (≤40 years; ≥55 years) and BMI (≤25; ≥28), demonstrating robust predictive performance across demographic and physiological subgroups. G Example application of ProMeta-BC in imaging-benign cases. Among patients classified as benign by ultrasound, ProMeta-BC identified pathologically confirmed malignant lesions To evaluate diagnostic performance, ProMeta-BC was systematically compared with its single-omics counterparts across internal testing and independent validation cohorts (Fig. 6C; Figure S9C). In the internal test set of Cohort 2, ProMeta-BC achieved an AUC of 0.973 (95% CI: 0.940–0.993), outperforming proteome-only (AUC = 0.930, 95% CI: 0.849–0.990) and metabolome-only models (AUC = 0.883, 95% CI: 0.809–0.942). Comparable performance was observed in the independent external validation cohort (Cohort 3, n = 161), where ProMeta-BC achieved an AUC of 0.951 (95% CI: 0.906–0.984), again exceeding proteome-only (AUC = 0.925, 95% CI: 0.863–0.969) and metabolome-only (AUC = 0.859, 95% CI: 0.790–0.918) models. This Cohort 3 estimate was tightly concordant with the training-set performance obtained by repeated stratified 5-fold cross-validation (mean AUC = 0.950 ± 0.030 across 50 folds; ΔAUC = 0.001 versus Cohort 3; Supplementary Figure S9D). The observed repeated cross-validation AUC (0.950) was 10.3 standard deviations above the null distribution mean (0.500; null SD = 0.044), with no permutation reaching it (empirical p = 0.0099; Supplementary Figure S9E), together suggesting that the reported performance was not substantially inflated by overfitting or feature-selection leakage. These results demonstrate preserved generalizability across cohorts and underscore the additive diagnostic value of integrative proteomic and metabolomic profiling. Notably, ProMeta-BC maintained high accuracy for early-stage breast cancer (stage 0–I), achieving AUCs of 0.962 (95% CI: 0.939–0.981) and 0.955 (95% CI: 0.911–0.987) in the internal and external validation sets, respectively. To elucidate the biological relevance of the selected features, we examined feature-level contributions within the integrated model. Several proteins involved in cytoskeletal remodeling and adhesion dynamics, including ITGB3, TLN1, and ACTN1, exhibited substantial predictive weights (Fig. 6D), consistent with the malignant-associated module ME01 identified in network analyses (Fig. 3D). In parallel, amino acid–related metabolites such as L-serine support anabolic growth and nucleotide biosynthesis [44], whereas alterations in tryptophan metabolism, exemplified by N-acetyltryptophan, have been linked to immune evasion and malignant evolution [45]. Visualization of ProMeta-BC–derived prediction scores demonstrated clear separation between benign and malignant samples across cohorts (Fig. 6E; Figures S9B; Figures S10–S11, Supplementary Table 6e). Given the substantial heterogeneity in systemic metabolic environments associated with age and body composition [46, 47], we next evaluated the stability of ProMeta-BC across clinically relevant demographic and physiological subgroups. Across age strata, ProMeta-BC showed stable predictive performance. In patients ≤ 40 years, the model achieved a positive predictive value (PPV) of 84.6% and a negative predictive value (NPV) of 85.7%, despite dense breast tissue and atypical imaging features. In those ≥ 55 years, PPV and NPV were 97.1% and 85.2%, supporting stable performance across age groups. Performance remained robust across BMI-defined subgroups. In patients with BMI ≤ 25, ProMeta-BC yielded a PPV of 96.2% and an NPV of 94.3%. In the BMI ≥ 28 subgroup, PPV reached 100%, while NPV was not reported due to the extremely limited number of benign cases (only two benign cases), indicating minimal confounding by body composition or systemic metabolic background (Fig. 6F). As PPV and NPV are inherently dependent on disease prevalence, we further projected these values across a range of 0.1%–66.5% via Bayes’ theorem using the sensitivity and specificity estimated in Cohort 3, with 95% uncertainty bands derived from bootstrap confidence intervals (Figure S12; Supplementary Table 6f). Across the intended-use prevalence range of 5%–50%, projected PPV ranged from 27.9% to 88.0% and NPV from 94.9% to 99.7%; Bayes-derived values at the Cohort 3 observed prevalence (66.5%) closely matched empirically measured values, validating the projection. At very low prevalences (≤ 0.5%), projected NPV exceeded 99.9% and PPV fell below 4% — mathematically expected in disease-rare settings and beyond the intended clinical scope of the assay. Together, these findings support the broad applicability of ProMeta-BC in heterogeneous real-world populations. To further assess its clinical utility in challenging diagnostic scenarios, we evaluated ProMeta-BC in 9 early-stage breast cancer patients initially classified as imaging-benign (BI-RADS 3). Using plasma samples collected at the time of initial imaging assessment, ProMeta-BC correctly identified 8 of 9 cases (88.9% sensitivity) in a single-blind analysis (Fig. 6G). All detected cases corresponded to early-stage, node-negative tumors (DCIS, n = 6; pT1, n = 3), indicating that ProMeta-BC can complement conventional imaging by improving detection of occult malignancies at an early, potentially curable stage. ProMeta-BC-LN enables prediction of axillary lymph node metastasis Although targeted profiling was designed for benign–malignant discrimination rather than lymph node stratification, multiple plasma protein modules were significantly associated with axillary lymph node status (Fig. 3D), motivating extension of the ProMeta-BC framework to develop ProMeta-BC-LN for preoperative nodal metastasis prediction (Fig. 7A). In Cohort 2, proteomic and metabolomic features were subjected to multistep feature screening and model selection, resulting in a compact panel of six highly informative features and selection of the SEM for model construction (Figure S13A). Construction and validation of the ProMeta-BC-LN model for prediction of axillary lymph node assessment. A Workflow of ProMeta-BC-LN construction. Feature screening was performed on 312 proteins and 170 metabolites in Cohort 2, with six informative features retained through LASSO and recursive feature elimination. Models were trained using multiple machine-learning algorithms and externally validated in Cohort 3. B-C ROC curves of ProMeta-BC-LN for distinguishing N0 and N+ patients in Cohort 2 (training set) and Cohort 3 (external validation set). D-E Confusion matrices illustrating classification performance of ProMeta-BC-LN in internal validation set (D) and external validation cohorts (E). F Application of ProMeta-BC-LN in cases with discordant imaging and pathological lymph-node assessments Using this feature panel, we evaluated the performance of ProMeta-BC-LN for discriminating LN negative (N−) and LN positive (N+) patients. In the training cohort, ProMeta-BC-LN achieved an AUC of 0.841 (95% CI: 0.741–0.922), outperforming proteome-only and metabolome-only models (Fig. 7B; Figures S14, S15). This performance was preserved in the independent external validation cohort (Cohort 3), where ProMeta-BC-LN achieved an AUC of 0.753 (95% CI: 0.641–0.843) and exceeded ultrasound-based assessment with an AUC of 0.585–0.719 [48], demonstrating robust cross-cohort generalizability (Fig. 7C). To elucidate the biological basis of nodal risk prediction, we examined feature-level contributions within the ProMeta-BC-LN panel (Figure S12B). Several components have established roles in lymphatic dissemination. For example, clusterin (CLU) promotes tumor invasion and metastatic competence and has been associated with lymph node involvement in breast cancer [49], whereas DYNLL1 is essential for cytoskeleton-dependent cell migration and metastatic spread [50]. These biologically relevant features support the mechanistic plausibility of the model. Importantly, ProMeta-BC-LN demonstrated a strong ability to identify patients without lymph node metastasis, as reflected by consistently high negative predictive values across cohorts. In the internal validation set, 45 of 48 patients predicted as node-negative were pathologically confirmed as non-lymph node metastasis (NPV = 93.8%). Similarly, in the external validation cohort, 69 of 80 predicted node-negative patients were pathologically confirmed (NPV = 86.3%) (Figs. 7D-E; Figure S13C-D, Supplementary Table 7d). This high NPV highlights the potential clinical utility of ProMeta-BC-LN for identifying patients who may safely avoid unnecessary axillary interventions. In clinically challenging scenarios with discordant imaging and pathological assessments, ProMeta-BC-LN also provided complementary stratification beyond conventional imaging (Fig. 7F). Notably, among patients deemed node-positive by imaging but confirmed as node-negative on pathology (Group B), the model correctly reclassified 27 of 32 cases (84.4%) as LN negative, supporting more confident de-escalation of unnecessary axillary interventions. In contrast, among imaging-suspected node-negative but pathologically positive cases (Group A), the model identified 11 of 24 patients as LN positive, providing additional support for detecting occult nodal involvement. The relatively modest positive predictive value observed is likely attributable to the low nodal tumor burden in our cohort. Nevertheless, the model’s high NPV reinforces its value in ruling out lymph node involvement when imaging results are ambiguous. Collectively, these findings demonstrate that plasma proteome–metabolome signatures captured by ProMeta-BC-LN provide independent and complementary information for axillary lymph node evaluation, with particular utility in reducing false-positive imaging assessments and supporting more individualized surgical decision-making. Discussion In this multicenter prospective study, we establish an integrated plasma proteome–metabolome framework for early-stage breast cancer detection and nodal risk assessment, providing a biologically informed complement to conventional imaging-based screening. By systematically capturing coordinated molecular remodeling in peripheral circulation, our approach extends beyond structural lesion detection to reflect systemic tumor–host interactions at early disease stages. Through rigorous external validation and targeted confirmation, we demonstrate that circulating proteomic and metabolomic signatures can be translated into robust, generalizable diagnostic models with clinically actionable performance. To our knowledge, this work addresses a long-standing gap in translating circulating proteome–metabolome signals into deployable tools for early breast cancer diagnosis. Currently, breast cancer screening still predominantly relies on mammography and ultrasound, but the accuracy substantially drops in women with dense breast tissue, a demographic especially common in Asian populations [51]. In a U.S. analysis, mammography sensitivity in Chinese women dropped from ~ 95% in non-dense breasts to ~ 87% in dense breasts (and to ~ 74% in dense breasts of Filipino women) [52]. Although ultrasound is often used as an adjunct for dense breasts and could increase the detection sensitivity, its benefits come with trade-offs: higher false-positive rates and operator dependence [53], underscoring an unmet need in Asian women for complementary detection strategies. Amid limitations of imaging, liquid biopsy strategies have emerged as promising tools for early breast cancer detection. In recent years, cell-free DNA (cfDNA) based assays – including mutation panels, targeted methylation profiling, and fragmentomics, as well as circulating tumor cell (CTC) –based approaches have been developed for early cancer detection [54]. Recent advances in cfDNA methylation have enabled multi-cancer early detection tests at high specificity. For instance, a targeted cfDNA methylation panel was shown to detect various cancers with > 99% specificity; in breast cancer, however, such DNA-based tests still exhibit limited sensitivity in early stages (reported < 10% detection in stage I and < 50% in stage II disease) [55]. Our group’s work with cfDNA fragmentomics has proved that integrating fragmentation features with machine learning can dramatically improve detection sensitivity for early BC, achieving high sensitivity (95%) and specificity (78.3%) for early cancer detection and reliably identifies malignancies missed by conventional imaging [56]. Despite these advances, DNA-centric approaches principally capture tumor cell-derived molecules (tumor DNA shed or epigenetic markers), and yield little information about tumor–host interactions or systemic biology, including the early-stage immune response and metabolic changes induced by the tumor. This motivates complementary approaches that capture the broader physiological alterations accompanying tumor development. Emerging evidence indicates that early-stage malignancies can instigate measurable changes in the circulating proteome and metabolome, reflecting the host’s reaction to the nascent tumor. Unlike directly tumor-derived signals, these proteomic/metabolic changes mainly represent downstream functional outputs of tumor-host or tumor-microenvironment interactions [57], and could also leave a “fingerprint” in circulation well before the tumor is radiologically visible. For example, a landmark prospective study found that plasma EGFR levels were elevated up to 17 months before breast cancer diagnosis; women in the highest EGFR quartile had a 2.9-fold increased risk of developing BC compared to the lowest quartile [17]. Other tumor-associated proteins, including IGFBP3, osteopontin, soluble CD44, and syndecan-1, have also been detected in peripheral blood of early-stage breast cancer patients [18]. Unlike traditional tumor biomarkers such as CA125, CA153 or CEA, most of these molecules functionally participate in extracellular matrix remodeling, cell adhesion, angiogenesis, and immune modulation [58]. Consistent with tissue-based studies, our plasma proteomic analysis revealed upregulation of cytoskeletal and adhesion-related proteins, including ACTN1 and integrins, in early-stage disease, showing their overexpression promotes invasion and stromal remodeling [59, 60].In line with these observations, our proteomic profiling and ProMeta-BC also identified and incorporated upregulated proteins involved in cytoskeletal regulation (TUBA4B, TPM4, MYL1), integrin-mediated signaling (ITGA2B, ITGB3), extracellular vesicle biology (DYNLL1), and immune-related pathways (CLU), forming conserved malignant-associated modules across multiple cohorts. Together, the concordance between prior evidence and our prospective multi-center data supports the robustness and biological relevance of circulating proteomic signatures for early breast cancer detection. In addition to the proteins, accumulating evidence also indicates that early-stage breast cancer is accompanied by characteristic systemic metabolic remodeling, particularly involving amino acid metabolism and nucleotide biosynthesis. Previous studies have consistently reported elevated serine levels in breast cancer, reflecting enhanced one-carbon metabolism that supports nucleotide synthesis, methylation reactions, and redox homeostasis during tumor proliferation [61, 62]. Similarly, increased orotic acid, an intermediate in de novo pyrimidine synthesis, has been linked to enhanced nucleotide biosynthesis required for rapid cancer cell proliferation [63]. Intriguingly, several of our top discriminative metabolites—phenylacetylglutamine and indoxyl sulfate—are well-established gut microbiota–derived tryptophan metabolites, suggesting a potential role of the gut–tumor axis in early breast carcinogenesis [64]. Consistent with these observations, our metabolomic profiling revealed coordinated alterations centered on amino acid metabolism and nucleotide biosynthesis pathways, forming conserved metabolic programs across cohorts and molecular subtypes (Fig. 4D). Moreover, dynamic trajectory analysis demonstrated gradual, stage-associated modulation of these metabolic modules (Fig. 5B), supporting their close linkage to cancer progression. Additionally, two of the top discriminative metabolites in the breast cancer cohort (Fig. 6C, cluster 9)—phenylacetylglutamine (PAGln) and indoxyl sulfate—are gut microbiota–derived metabolites with known links to diet and host physiology. PAGln is generated when gut bacteria catabolize dietary phenylalanine into phenylacetic acid, which is then conjugated with glutamine in the liver [65]. Elevated PAGln has recently been implicated in adverse outcomes in prostate cancer and in immune regulation. For example, in a prospective study from the PLCO cohort, men with the highest plasma PAGln levels had a 2–3-fold increased risk of developing lethal prostate cancer, likely reflecting diets rich in animal proteins (meat) that promote PAGln production [66] Mechanistically, high levels of PAGln could decreased infiltration of cytotoxic CD8+ T cells and attenuate anti-tumor immunity, reducing response to anti-PD-1 efficacy. On the other hand, indoxyl sulfate is a microbiota-derived tryptophan metabolite that activates aryl hydrocarbon receptor (AhR) signaling and links diet-related metabolites to immune regulation. Previous studies have reported context-dependent effects of indoxyl sulfate in cancer: chronic exposure to indoxyl sulfate has been widely reported as a pro-oxidant, pro-inflammatory uremic toxin that triggers AhR–CYP1A1–dependent oxidative stress and endothelial dysfunction, resulting in tumor initiation and progression [67, 68], while recent studies in breast cancer models suggests that indoxyl sulfate may suppress epithelial–mesenchymal transition and metastatic potential through AhR-dependent pathway [69, 70]. These controversial findings indicate that the biological impact of indoxyl sulfate is highly context-dependent, with its effects may vary across disease stages and microenvironmental conditions. Regardless, the presence of these microbiota-dependent metabolites in our model underscores the intimate connection between diet, the gut microbiome, and breast cancer biology. Beyond PAGln and indoxyl sulfate, several additional metabolites incorporated in ProMeta-BC have also been functionally implicated in tumor metabolism, immune regulation, and microenvironmental remodeling. For example, orotic acid reflects enhanced nucleotide production required for rapid tumor proliferation, including breast cancer [71]; alterations in tryptophan-related metabolites, including N-acetyltryptophan, are closely linked to immune tolerance and T-cell dysfunction [72] and elevated L-serine supports one-carbon metabolism and redox homeostasis, which facilitating anabolic growth and stress adaptation in malignant cells [73]. Collectively, these metabolites converge on core biological programs governing proliferation, immune modulation, and metabolic adaptation, suggesting that ProMeta-BC captures coordinated systemic remodeling of the tumor–host ecosystem rather than isolated tumor-derived signals. Notably, however, most existing evidence supporting the functional roles of these metabolites derives from established cancer cohorts or experimental models, and how they influence early malignant transformation remains largely unknown, warranting further mechanistic investigation. Although the present study employed high-resolution Orbitrap Astral mass spectrometry for biomarker discovery and validation, the final ProMeta-BC panel was ultimately reduced to a compact set of 15 proteins and 5 metabolites, aiming to reduce the costs and improving its translational feasibility. In future clinical implementation, these biomarkers could potentially be adapted to more accessible targeted platforms, including clinical LC-MS/MS or immunoassay-based assays for selected proteins. In addition, ProMeta-BC may be particularly valuable as an adjunctive tool in imaging-challenging scenarios, such as women with dense breasts, indeterminate BI-RADS lesions, or regions with limited radiological expertise, where it may help reduce missed diagnoses and unnecessary invasive procedures, while its value as a population-wide screening tool still requires further validation. Although our study provides a comprehensive plasma proteomic- and metabolomic-based framework for understanding EBC beyond morphology-based detection, several limitations should be acknowledged. First, although our cohorts were enriched for EBC and validated across multiple centers, the case–control design does not fully recapitulate population-based screening settings, and prospective longitudinal studies will be required to further evaluate real-world screening performance. Second, we did not perform functional experiments to determine whether and how these differential metabolites causally influence tumor initiation or dissemination, which warrants further investigation. Third, we did not perform direct comparisons with large-scale population proteomic resources such as the UK Biobank, primarily due to fundamental platform differences between affinity-based panel (Olink) assays and unbiased mass spectrometry–based proteomics, which limit cross-platform comparability (only 34.6% [108/312] protein overlap) [15, 74–76]. Fourth, although the 20-feature panel was deliberately designed for transferability to clinical triple-quadrupole platforms, formal analytical validation on sMRM-based clinical instruments and prospective health-economic evaluation have not yet been performed and are required prior to broad clinical deployment. Despite these limitations, our study establishes a population-specific molecular reference for early-stage breast cancer and provides a biologically informed foundation for advancing risk stratification, understanding early tumor–host interactions, and provide a proof-of-concept that combining plasma proteins and metabolites can yield robust, interpretable models for early-stage breast cancer detection and risk stratification. Methods Study population This multi-center cohort study enrolled 501 female participants in the primary analysis, including 337 pathologically confirmed breast cancer patients and 164 benign breast disease controls. Participants in the primary cohort were recruited as part of a registered multicenter prospective study (NCT06016790) conducted across six tertiary hospitals and one community health service center in China: the Second Affiliated Hospital of Zhejiang University School of Medicine, Hangzhou First People’s Hospital, Zhejiang Provincial Hospital of Traditional Chinese Medicine, Ningbo Medical Center Lihuili Hospital, Women’s Hospital School of Medicine Zhejiang University, Sir Run Run Shaw Hospital, and Xiangfu Community Health Service Center. In addition, an independent external validation cohort comprising 161 female participants (107 breast cancer patients and 54 benign controls) was enrolled from two additional medical institutions: Changhai Hospital and Chaohui Community Health Service Center. In addition, nine participants classified as BI-RADS 3 on imaging from the aforementioned institutions, independent of the primary and external validation cohorts, were included as an additional clinical validation subset after histopathological confirmation through biopsy or surgery. All blood samples were collected using a standardized protocol across participating centers, with collections performed in the morning after an overnight fast. Inclusion criteria were: female aged 18–70 years, histopathologically confirmed as (1) primary invasive breast cancer or carcinoma in situ, or (2) benign breast lesions, including fibroadenoma, breast adenosis, or fibrocystic hyperplasia. All enrolled individuals provided peripheral plasma samples available for analysis. Exclusion criteria included: pregnancy or lactation; history of other malignancies within the past 5 years; suspected but unconfirmed malignancy within the past year; blood transfusion within 30 days before blood collection; active mastitis or other inflammatory diseases that might affect plasma protein and metabolic profiles; and other circumstances deemed inappropriate for enrollment by the investigators. Sample collection and plasma proteomics Peripheral blood (8–10 mL) was collected in cfDNA tubes (Ardent BioMed Cat. # BY10240301) during routine blood tests or before surgery, all blood samples were collected using a standardized protocol across participating centers, with collections performed after an overnight fast. Samples were subjected to dual centrifugation (1,600×g → 16,000×g, 10 min each at 4 °C) for plasma isolation, followed by immediate storage at -80 °C. Frozen plasma was batch-shipped on dry ice to the central laboratory (OmixScience Research Institute, Hangzhou, China) for processing. Plasma samples (40 µL per sample) were processed using a Proteonano™ Plasma Proteome Enrich Kit on a Nanomation™ G1 Basic Automated Platform (Nanomics, Hangzhou, China) for protein enrichment, reduction (DTT, 5mM), alkylation (IAA,14mM), and trypsin digestion (1:50). After desalting using C18 StageTips, peptides were redissolved in water with 0.1% formic acid for LC-MS/MS analysis. Each 96 well plates contained 2–3 quality control samples and one neat plasma without nanoparticle enrichment pre-treatment. LC-MS/MS analysis was performed on a Thermo Scientific Vanquish Neo UHPLC system coupled to an Orbitrap Astral mass spectrometer. A 5.5 cm High Throughput µPAC™ Neo HPLC Column was used for peptide separation with a 24 min gradient of mobile phase B (80% acetonitrile) from 4% to 35% at 2 µL/min. Data acquisition was performed with Data Independent Acquisition (DIA) mode using an Orbitrap Astral high-resolution mass spectrometer (Thermo Scientific). The detection was in the positive mode, with a scan range of 380–980 m/z for MS1 precursor ion, and a resolution of 240,000 at 200 m/z. Normalized AGC Target was set at 500%, with a Maximum IT of 5ms. MS2 employing the DIA data acquisition mode was performed with 299 scan windows set, an Isolation Window of 2 m/z, HCD Collision Energy set at 25 eV, Normalized AGC Target at 500%, and a Maximum IT of 3ms. For targeted PRM analysis, data acquisition was performed on an Orbitrap Astral high-resolution mass spectrometer (Thermo Scientific) in positive ion mode, using the same ionization settings as the DIA workflow described above. MS1 spectra were acquired over 380–980 m/z at a resolution of 240,000 (at 200 m/z), with a normalized AGC target of 500% and a maximum injection time (IT) of 5 ms. MS2 acquisition employed a scheduled-window (2 min) inclusion list with an isolation window of 1 m/z, HCD collision energy of 25 eV, normalized AGC target of 1000%, and a maximum IT of 20 ms. A label-free targeted PRM assay, developed from the discovery proteomics, covered 479 plasma proteins (1,841 peptides). Peptides were proteotypic, 6–30 amino acids in length, free of missed cleavages, and—where alternatives existed—free of methionine; at least two peptides were monitored per protein wherever feasible. A pooled peptide QC sample (QC_mix) was injected every 20 randomized study samples, in addition to the per-plate QC controls described above. Peak picking and integration were performed in Skyline v22.2 with mProphet-based FDR control (FDR < 0.05). Peptides with > 20% missing values within a cohort were excluded, and protein-level relative intensities were computed as the median across retained peptides. Of the 479 monitored proteins, 312 met these quality thresholds across cohorts and were retained for downstream analyses. Full peptide-level transition information is provided in Supplementary Table S8. Sample collection and plasma metabolomics Plasma samples (50 µL per sample) were extracted with 200 µL cold methanol containing internal standard, and precipitated at -20 ℃ for 1 h. After centrifugation at 14,000 × g for 10 min at 4 ℃, the supernatant was transferred into a new tube, dried under nitrogen, and resuspended in 60 µL methanol/water (10/90, v/v). 5 µL of each sample was combined to make a pooled quality control (PQC) sample, which was injected every ten samples to monitor retention time and signal intensity drift. The remaining samples (30 µL) were transferred into injection vials for metabolomic analysis. Metabolites were separated using a Waters BEH C18 column (2.1 mm x 100 mm, 1.7 μm particle size) on a Vanquish Flex UHPLC (Thermo Scientific, Waltham, MA) equipped with a refrigerated autosampler (4 ℃) and column heater (40 ℃). Solvent A (0.1% formic acid in water) and Solvent B (0.1% formic acid acetonitrile/methanol 40/60) were used to elute the metabolites with a 14 min gradient. The gradient was as follows: 10% B at 0 min, 0.25 mL/min; 40% B at 3 min, 0.25 mL/min; 95% B at 5 min, 0.25 mL/min; 100% B at 8 min, 0.6 mL/min; 100% B at 10 min, 0.6 mL/min; and back to 10% B at 10.5 min, 0.25 mL/min; and equilibrate for 3 min. Samples were analyzed using a Q Exactive HF (QE-HF) (Thermo Scientific, Waltham, MA) equipped with a heated electro-spray ionization (HESI) source. All data were acquired in positive and negative mode separately using Full scan/ddMS2 (DDA) to acquire MS1 precursor for quantification and MS2 fragmentation for metabolite annotation. The Full Scan settings were as follows: 60,000 resolution, AGC target, 4e6; Maximum IT, 100 ms; scan range, 60 to 900 m/z. For Full scan/ddMS2(DDA), Top 20 MS/MS spectral (dd-MS2) @ 15,000 were generated with AGC target = 2 × 105, Maximum IT = 25 ms, and (N)CE/stepped NCE = 10, 20, 40. Metabolites detection and identification were performed using XCMS and Spectra R package by searching against online database (MoNA, GNPS, HMDB and MS-dial LipidBlast database) and in-house database. Differential analysis and functional annotation Protein and metabolite quantitative data were log-transformed and standardized prior to differential analysis between benign and malignant samples. Differential abundance was assessed using linear modeling, with P values adjusted for multiple testing using the Benjamini–Hochberg procedure. Global variation in proteomic and metabolomic profiles was first examined by principal component analysis (PCA), and supervised discrimination between groups was further evaluated by orthogonal partial least squares discriminant analysis (OPLS-DA), with model robustness assessed through permutation testing. Significantly differential proteins were subjected to functional annotation and enrichment analyses based on Gene Ontology (GO) and KEGG pathways, whereas differential metabolites were analyzed using KEGG enrichment together with metabolite set enrichment analysis (MSEA). In addition, pathway-level shifts were further assessed using gene set enrichment analysis (GSEA)-based approaches, in which proteins or metabolites were ranked by differential statistics and enrichment scores were calculated to quantify coordinated changes of predefined pathway sets, enabling systematic evaluation of pathway activity patterns and cross-omics pathway concordance. Network analysis and dynamic trajectory modeling To analyze the systematic structural features of proteome changes, WGCNA was constructed based on targeted proteome data to identify stable co-expression modules, with module eigengenes calculated. Correlations between modules and clinical variables and metabolite clusters were assessed through Spearman correlation analysis. Metabolic and proteomic dynamic changes were modeled through mFuzz fuzzy clustering analysis, using the clinical stage sequence (benign → stage 0 → stages I–IV) as a pseudo-temporal axis to identify molecular trajectories with consistent abundance trends, which were further validated in independent cohorts. Multi-layer cross-omics networks were subsequently constructed by integrating stage-dependent protein modules, metabolite trajectories, and functional pathway annotations to characterize coordinated relationships among malignancy-associated molecular programs. Machine learning model construction and validation Based on candidate proteins and metabolites identified through targeted validation, machine learning models were constructed for benign–malignant discrimination of breast cancer (ProMeta-BC) and for prediction of axillary lymph node metastasis (ProMeta-BC-LN). For ProMeta-BC, feature selection was primarily performed using support vector machine–based recursive feature elimination (SVM–RFE) to identify the most informative proteomic, metabolomic, and combined feature sets. For ProMeta-BC-LN, feature screening was conducted using least absolute shrinkage and selection operator (LASSO) regression followed by RFE, aiming to reduce redundancy, enhance model stability, and derive a compact predictive panel. In Cohort 2, samples were randomly divided into a training set (70%) and an internal test set (30%) with stratification by outcome. The analytical steps used to construct the final predictive panel — namely SVM-RFE, LASSO and SEM model training — were performed exclusively within the 70% training set, with the resulting feature panel and preprocessing parameters locked before any evaluation on the internal test set or the independent external cohort (Cohort 3). Upstream analyses used for prior biological prioritization of candidate features — WGCNA co-expression modules, mFuzz stage-trajectory clustering, and differential abundance analysis — were performed at the cohort level on Cohort 2, from which the final panel was subsequently selected within the training set. OPLS-DA was used on the full Cohort 2 only for group-level visualization (Fig. 5A) and did not contribute to the final feature panel. A stacking ensemble strategy was employed for model construction. Four base learners — generalized linear models (GLM), support vector machines (SVM), gradient boosting machines (GBM), and multilayer perceptrons (MLP) — were each trained within the training set using five-fold cross-validation. The out-of-fold predicted probabilities from each base learner were then used as meta-features for a second-level logistic regression meta-learner, which integrated the outputs of all base learners into a final prediction. We termed this approach the Stacking Ensemble Method (SEM). The SEM was implemented using StackingClassifier from scikit-learn (version 1.3.2). Final predicted probabilities were derived from the logistic link function of the meta-learner and were not subjected to additional calibration. Model performance was assessed using receiver operating characteristic (ROC) curves and the area under the curve (AUC). To quantify uncertainty in model performance, 95% confidence intervals for AUC were estimated using bootstrap resampling of predicted probabilities. Generalization ability was further evaluated in the independent external validation cohort (Cohort 3). Robustness of the training-set performance was additionally assessed by repeated stratified 5-fold cross-validation with ten independent random seeds and by a label-permutation test (n = 100), in which the full SVM-RFE, standardization, and SEM training pipeline was re-executed for each permutation. Feature contribution analysis based on model coefficients and importance measures was performed to characterize the relative contribution of individual proteins and metabolites. Supplementary Information Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Supplementary Material 4. Supplementary Material 5. Supplementary Material 6. Supplementary Material 7. Supplementary Material 8. Supplementary Material 9. Authors’ contributions CN, LH, JH, QW and ZY designed and supervised the study. WZ, YY, LS and JD drafted the manuscript. ZY, YW, HX, and XG performed bioinformatics analyses, managed data interpretation, and participated in manuscript revision. YZ, HL, JZ, JD, HD, WC, WX, JJ, XY, SS, JC, JL and DW contributed to the sample collection, data collection, and organization. ZY and CN revised the manuscript for critical content. All the authors reviewed, edited, and approved of the final manuscript, ensuring the accuracy and integrity of the work. Funding This work was supported by Noncommunicable Chronic Diseases-National Science and Technology Major Project (Grant 2025ZD0545400), the “Pioneer” and “Leading Goose” R&D Program of Zhejiang Province (2024C03050), Zhejiang Provincial Natural Science Foundation for Distinguished Young Scholars (Grant LRG25H160001), National Natural Science Foundation of China (Grants 82273275, 82473003, 82503970, 82473478, 82403888), and Medical and Health Science and Technology Project of Zhejiang Province (2025KY839). Data availability Raw and processed plasma proteomic and metabolomic data generated in this study, along with all custom analysis scripts and machine learning pipelines, and the code and outputs for the additional robustness analyses are publicly available at GitHub (https://github.com/zhuohangyu/ProMeta-BC). Declarations Ethics approval and consent to participate All procedures involving human participants were conducted in accordance with the Declaration of Helsinki and relevant ethical regulations. This multi-center study was approved by the Institutional Review Board of the Second Affiliated Hospital, Zhejiang University School of Medicine (Approval No. 2023 − 1048) as the coordinating center, as well as by the Institutional Review Boards of all participating institutions. Written informed consent was obtained from all participants prior to enrollment in the study. All participants provided voluntary consent for participation and for the use of their clinical data and biological samples for research purposes. Consent for publication All authors contributed to the writing of the manuscript. All authors approved the final version of the manuscript. Competing interests The authors declare no competing interests. Acknowledgements We gratefully acknowledge Cosmos Wisdom for providing sequencing services and technical support. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Wei Zhang, Yao Yao, Yimeng Wang, Lesang Shen and Jinghua Ding contributed equally to this work. Contributor Information Qijun Wu, Email: wuqj@sj-hospital.org. Jian Huang, Email: hjys@zju.edu.cn. Hengyu Li, Email: lihengyu@binn.cas.cn. Zhuohang Yu, Email: zhuohangyu@zju.edu.cn. Chao Ni, Email: drnichao@zju.edu.cn. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Supplementary Material 4. Supplementary Material 5. Supplementary Material 6. Supplementary Material 7. Supplementary Material 8. Supplementary Material 9. Data Availability Statement Raw and processed plasma proteomic and metabolomic data generated in this study, along with all custom analysis scripts and machine learning pipelines, and the code and outputs for the additional robustness analyses are publicly available at GitHub (https://github.com/zhuohangyu/ProMeta-BC).