Unveiling Endotypes in Systemic Lupus Erythematosus Through Multiomic Analysis: Insights Into Cardiovascular and Renal Complications

Abstract

Objective

Systemic lupus erythematosus (SLE) shows clinical and molecular heterogeneity, and cardiovascular (CV) complications and lupus nephritis (LN) remain leading causes of morbidity and mortality. This study investigated whether omic profiling can reveal molecular endotypes linked to these outcomes.

Methods

Serum from 199 patients with SLE underwent proximity extension assay–based proteomics and targeted nuclear magnetic resonance metabolomics. Unsupervised clustering was performed on proteomic data, followed by integrative multiomic factor analysis, logistic regression, and machine learning models. Cohorts with SLE from the University College London, a subset with expanded Olink Reveal panel and untargeted metabolomics, and in vitro (human umbilical vein endothelial cell and HK2) and ex vivo (rat kidneys) models exposed to patient sera were used for validation and mechanistic exploration.

Results

Proteomic clustering identified two molecular subgroups (cluster 1 and cluster 2). Compared with cluster 2, cluster 1 showed damage burden and increased rates of LN (1.8‐fold), hypertension (3‐fold), dyslipidemia (2‐fold), obesity (8‐fold), and abnormal C‐reactive protein/erythrocyte sedimentation rate (4‐fold), consistent with CV risk. A total of 47 inflammatory and organ‐damage proteins and multiple metabolites were increased in cluster 1. Neural network models based on metabolites and clinical variables discriminated clusters (area under the curve = 0.77), highlighting citrate and lipoproteins as key features. Multiomic analyses and external cohorts confirmed reproducibility and enrichment in pathways related to leukocyte trafficking, endothelial stress, and nephritis. In vitro, serum from patients with SLE induced NF‐κB activation in rat kidneys compared with controls, supporting a proinflammatory effect.

Conclusion

Multiomic profiling delineates molecular endotypes in SLE, integrating immune, vascular, and metabolic pathways associated with CV risk and LN, supporting their potential for precision risk stratification.

INTRODUCTION

Systemic lupus erythematosus (SLE) is a complex autoimmune disease with heterogeneous clinical manifestations and unpredictable progression, complicating diagnosis and treatment. ¹ It affects multiple organs, particularly the skin, joints, and kidneys, and the cardiovascular (CV) system. Lupus nephritis (LN) and CV complications are major contributors to SLE‐related morbidity and mortality.

SLE‐associated CV disease (CVD)—including early atherosclerosis, stroke, and thrombosis—arises from a multifactorial interplay of genetic predispositions (eg, STAT4, IRF5, PDGF, HAS2, ITGAM, SLC5A11), epigenetic changes (eg, DNA methylation, microRNAs, splicing), classical CV risk factors (hyperlipidemia, hypertension, obesity), autoimmunity (autoantibodies, immune cell dysregulation), complement activation, oxidative stress, endothelial dysfunction, and proinflammatory mediators. ² , ³ , ⁴ LN occurs in ~40% of patients and entails renal inflammation with potential long‐term organ damage. ⁵ Its variable clinical presentation challenges early detection and risk stratification. Traditional markers (proteinuria, serum creatinine) often fail to fully capture disease severity or dynamics. ⁶

Recent evidence supports the utility of molecular biomarkers—including anti–double‐stranded DNA (dsDNA), anti‐Smith antibodies, and others—in assessing disease activity and organ involvement. Yet, the clinical heterogeneity of SLE remains a barrier to effective management, emphasizing the need for deeper insights into its molecular basis. Identifying discrete molecular endotypes may enable personalized therapies.

Multiomic strategies offer a powerful framework for deciphering the complex biology of SLE. Unlike static genomic data, the proteome and metabolome dynamically reflect the disease state, allowing the discovery of active pathways driving pathology. Their integration enables more accurate interpretation of disease processes, uncovering phenotype‐specific molecular signatures and candidate biomarkers for stratification and prognosis. ⁷

This study aims to enhance the molecular characterization of SLE by analyzing serum proteomic and metabolomic profiles in a well‐defined patient cohort. We focus on signatures linked to inflammation, organ damage, CV risk, and LN. Complementary in vitro and ex vivo experiments—using human umbilical vein endothelial cells (HUVECs), HK2 cells, and rat kidney slices—model the endothelial and renal responses to patient serum. Advanced data analysis, including machine learning (ML), supports the classification of molecular subgroups and integrates multiomic profiles to inform patient stratification, pathophysiologic understanding, and therapeutic decision‐making.

MATERIALS AND METHODS

Patients

A total of 199 patients fulfilling the revised American College of Rheumatology criteria for SLE were included across three cohorts. ⁸ The first cohort (n = 100) belonged to the European multicenter cross‐sectional PRECISESADS Innovative Medicines Initiative study (NCT02890121; Supplementary Table 1), conducted in accordance with Good Clinical Practice guidelines and the Declaration of Helsinki, with informed consent obtained from all participants. Peripheral blood was collected in the morning after an eight‐hour fast. There were no significant differences in steroid exposure between cluster 1 (23 of 48 patients) and cluster 2 (23 of 52 patients; P = 0.866). Two additional independent cohorts comprising 99 patients with SLE (58 with CV risk and 41 with LN) were recruited from the rheumatology clinic at University College London (UCL) Hospital following approval by local ethics committees (15‐LO‐2065, 11‐LO‐0330, and 06/Q0505/79) and written informed consent. Peripheral venous blood samples were processed to obtain serum and stored at −80°C until use.

Circulating inflammatory profile

Protein levels of 184 markers—92 proinflammatory and 92 organ‐damage mediators—were measured in PRECISESADS serum using the sensitive proximity extension assay (PEA) technology (Cobiomic Bioscience; https://cobiomicbioscience.com/wp‐content/uploads/2025/07/Target‐96‐Inflammation.pdf). Additionally, proteome assays were performed in a subset of six samples (cluster 1, n = 5; cluster 2, n = 5) using PEA technology, analyzing 1,034 proteins from the Reveal panel (https://cobiomicbioscience.com/wp-content/uploads/2025/10/olink-reveal.pdf).

Serum metabolomics

Measures of 228 serum biomarkers were performed with an established nuclear magnetic resonance spectroscopy platform (Nightingale Health; https://research.nightingalehealth.com/uploads/documents/Nightingale‐Blood‐Analysis_List‐of‐Biomarkers.pdf). Moreover, an untargeted metabolomic study of the metabolite profile was conducted in a small subset of samples (cluster 1, n = 5; cluster 2, n = 5) by high‐resolution liquid chromatography mass spectrometry using both ESI‐positive and ‐negative ionization modes at the UCAI Metabolomics Facility of Instituto Maimónides de Investigación Biomédica de Córdoba, detecting 5,237 metabolites.

Treatment of HUVECs with serum from patients with SLE with high or low disease severity to identify the differential effects promoted on endothelial dysfunction

HUVECs were treated for 24 hours at 37°C with 15% serum from patients with high or low disease severity. Cells were lysed with radioimmunoprecipitation assay (RIPA) buffer, and proteome assays of 92 CV‐related proteins were performed using PEA technology (https://cobiomicbioscience.com/wp-content/uploads/2025/07/Target-96-Cardiovascular-II.pdf).

Treatment of the renal cell line HK2 with serum from patients with cluster 1 or cluster 2 lupus to identify the differential effects associated with kidney damage

HK2 proximal tubular cells were exposed to 15% serum from patients with cluster 1 and cluster 2 at 37°C for 24 hours. Cells were lysed with RIPA buffer, and proteome assays of 92 inflammation‐related proteins were performed using PEA technology (https://cobiomicbioscience.com/wp-content/uploads/2025/07/Target-96-Inflammation.pdf).

Ex vivo experiments on kidney slices from rats and p65 immunofluorescence in renal tissue

These experiments assessed the effects of serum from patients with cluster 1 and cluster 2 on renal and vascular inflammation through p65 expression. Kidneys from euthanized male Wistar rats were aseptically harvested, and renal slices were cultured for six hours with culture medium supplemented with 15% serum from healthy controls or patients with cluster 1 or cluster 2 at 37°C and 5% CO₂. Experiments were performed by authors JRM‐C and AIT under blinded conditions; they received patient sera without knowledge of their origin or cluster assignment. Only authors TC and CL‐P were aware of the identity of the sera analyzed. Finally, proteome assays in homogenized kidney slices were then performed by PEA technology, analyzing 48 proteins from the mouse cytokine panel, which is compatible with rat samples (https://cobiomicbioscience.com/wp‐content/uploads/2025/07/Target‐48‐Mouse‐Cytokine.pdf).

Renal slices were fixed in paraformaldehyde, paraffin embedded, and sectioned (3–4 μm) for p65 immunofluorescence. After deparaffinization, antigen retrieval in citrate buffer (pH 6) and permeabilization, sections were incubated overnight at 4°C with an anti‐p65 primary antibody, followed by Alexa Fluor 647–conjugated secondary antibody. Nuclei were counterstained with DAPI. Negative controls were performed by omitting the primary antibody. Images were acquired by fluorescence microscopy and analyzed using Fiji/ImageJ software.

Data analysis

Statistical analyses were performed using SPSS version 25.0 (IBM), R (with multiple analytical packages), GraphPad Prism 9, and MetaboAnalyst version 6. The significance threshold was set at P < 0.05 or false discovery rate (FDR) ≤0.05 when multiple‐testing correction was applied. Data distribution and variance homogeneity were assessed before hypothesis testing, and nonparametric methods were used when appropriate.

MetaboAnalyst was employed for unsupervised clustering analyses (including hierarchical clustering and partial least squares discriminant analysis [PLS‐DA]) and for metabolomic data exploration, whereas functional and network enrichment analyses were performed using STRING and Metascape. Multiomic data integration was conducted using the multiomic factor analysis (MOFA) framework, with log2 normalization and view scaling applied to adjust variance across omics layers before model training. Predictive modeling was performed using univariate logistic regression and six supervised ML algorithms: support vector machine, generalized linear model with lasso regularization, boosted logistic regression, XGBoost, neural network, and random forest. Models were evaluated through 10‐fold cross‐validation.

Metabolite features with >10% missing data were excluded, and the remaining missing values were imputed using k‐nearest neighbors (k = 5). Highly correlated metabolites (r > 0.95) were removed to reduce redundancy. Pearson correlation analyses were conducted to examine associations between clinical and molecular variables and visualized in R using the circlize and qgraph packages. The data sets generated and analyzed during the current study will be shared upon reasonable request from the corresponding author. PRECISESADS Clinical Consortium members are listed in Appendix A.

RESULTS

Unsupervised clustering of circulating inflammatory and organ‐damage proteomes stratifies patients with SLE into two distinctive molecular subgroups, which also exhibit relevant metabolic alterations linked to their clinical profiles

To determine the relative contribution of each omic layer, we first applied an MOFA integrating proteomic and metabolomic data. Proteomics emerged as the dominant driver of variance (44.9%) across patients, exerting a stronger discriminatory power compared to metabolomics (4.8%; Supplementary Figure 1). Based on this, we performed unsupervised clustering of circulating inflammatory and organ‐damage proteomes in patients with SLE (Supplementary Table 1), identifying two molecular subgroups with distinct clinical features (Figure 1A). Cluster 1 showed elevated disease activity (Systemic Lupus Erythematosus Disease Activity Index [SLEDAI] > 5), longer disease duration, abnormal C‐reactive protein/erythrocyte sedimentation rate, and a higher prevalence of CV risk factors (obesity, hypertension, dyslipidemia, statin use) and LN compared to cluster 2 (Figure 1B).

Figure 1.

Proteomic analysis. (A) Unsupervised heatmap of patients. High disease severity subgroup (cluster 1 [C1]) is colored in red (n = 48) and low disease severity subgroup (C2) is colored in green (n = 52). (B) Clinical and analytical profiles of patients from C1 and C2. The figure represents the most important and significant variables between both subgroups. (C) Volcano plot showing 47 significant proteins between C1 and C2 (P < 0.05, fold change [FC] ≥1.5 on the linear scale, corresponding to |log2(FC)| ≥ 0.58, false discovery rate adjusted). The x‐axis represents log2‐transformed FCs. (D) Partial least squares discriminant analysis (PLS‐DA) summarizing the differences among the protein profiles C1 (red) and C2 (green). (E) Variable importance in projection plot (VIP): proteins identified by sparse variant PLS‐DA in descending order of importance. The graph represents the relative contribution of these proteins to the variance between the two subgroups. The blue and red boxes on the right indicate whether the protein levels are increased (red) or decreased (blue) in the serum of patients across subgroups. ANA, antinuclear antibody; CNS, central nervous system; CRP, C‐reactive protein; dsDNA, double‐stranded DNA; ESR, erythrocyte sedimentation rate; SLEDAI, Systemic Lupus Erythematosus Disease Activity Index.

At the molecular level, patients with cluster 1 overexpressed 47 proteins (P < 0.05, fold change ≥ 1.5, FDR) versus cluster 2 (Figure 1C), with the top six (CD40, prostaglandin F [PGF], ENAH, NUCB2, CDCP1, interleukin‐18 [IL‐18]) linked to inflammation. PLS‐DA confirmed subgroup separation (Figure 1D), highlighting TINNI3, NCF2, CXCL10, and CXCL9 as key drivers (variable importance in projection plot > 2.2; Figure 1E). To ensure that the observed molecular differences between clusters are not biased by disease duration, we fitted linear models adjusted for age and disease duration. All proteins remained significantly associated with cluster assignment except FGF21, which was excluded (Supplementary Table 2).

Protein–protein interaction analysis via STRING revealed 46 nodes and 238 edges (P < 1.0 × 10⁻¹⁶; Supplementary Figure 2A). Enrichment analysis showed up‐regulated proteins involved in inflammatory pathways (leukocyte chemotaxis/migration, inflammatory response, myeloid leukocyte migration, and neutrophil chemotaxis), CV mechanisms (fluid shear stress, atherosclerosis, blood pressure regulation, and lipid metabolism), and nephritis‐related pathways (interferon γ and IL‐17 regulation; Supplementary Figure 2B). Comparison with healthy donors (HDs; Supplementary Figure 3) revealed that the proteomic profile of cluster 2 was more similar to HDs than to cluster 1, the latter characterized by a markedly higher number of up‐regulated inflammatory and organ damage–related proteins.

To validate the findings in an independent cohort, matched proteomic data from the UCL cohort were analyzed using the same Olink panels and analytical pipeline applied in PRECISESADS. Unsupervised clustering and principal components analysis identified two molecular subgroups (cluster 1, n = 16; cluster 2, n = 20) with clinical features closely mirroring those observed in PRECISESADS (Supplementary Figure 4A–C; Supplementary Table 3). A substantial subset of proteins distinguishing the clusters in PRECISESADS showed concordant changes in UCL, including IL‐18, CXCL9, CXCL10, CD40, CDCP1, PGF, NCF2, NUCB2, TNNI3, and FGF23 (Supplementary Figure 4D). Network and functional enrichment analyses highlighted pathways related to immune activation, endothelial/vascular stress, and nephritis (Supplementary Figure 4E–F).

These findings indicate that the proteomic endotypes defined in PRECISESADS are reproducible in an independent cohort and that the associated metabolomic and proteomic signatures represent robust, cross‐cohort molecular traits. In order to explore molecular differences between the two proteomic‐defined clusters, a metabolomic analysis was performed on matched patient samples. Significant changes in the abundance of 13 serum metabolites were identified (Figure 2A). A total of 12 metabolites were up‐regulated in cluster 1 (Figure 2A), with the top 3 being creatinine (P = 0.0034), citrate (P = 0.0037), and medium low‐density lipoprotein (LDL) triglycerides (TGs; P = 0.012). Only one metabolite was down‐regulated in cluster 1, isoleucine (P = 0.014).

Figure 2.

Metabolomic analysis using proteomic clusters. (A) Volcano plot shows 13 metabolites up‐regulated and 1 metabolite down‐regulated in cluster 1 (C1) versus C2 (fold change [FC] 1 P ≤ 0.05). (B) Forest plot showing odds ratios (ORs) and 95% confidence intervals (CIs) of serum lipids comparing all patients in the C2 to the C1 (reference group) by logistic regression analysis adjusted for disease duration, age, SLEDAI score, sex, smoking, diabetes, dyslipidemia, obesity, anti–double‐stranded DNA (dsDNA), hypertension, thrombosis, lupus nephritis, statin, steroids, prednisone, acetylsalicylic acid (ASA), anticoagulant, azathioprine, immunosuppressant, mycophenolate mofetil (MMF), antimalarials, hydroxychloroquine (HC), methotrexate, nonsteroidal anti‐inflammatory. ORs below 1 denote metabolites that are increased in C1 versus C2. (C) Significant metabolites from the Forest plot. (D) Receiver operating characteristic (ROC) curves for neural network (NN) model. (E) Variable importance plot of the NN analysis, in which the variables are ordered top to bottom as most to least important in classifying between C1 and C2. (F) Enrichment analysis using the Small Molecule Pathway Database (SMPDB). Pathways with a significant enrichment P <0.05 are shown (–log10[P value]). Enrichment ratio is indicated along the y‐axis. (G) Network of metabolic pathways derived from the SMPDB. Nodes are sized using the −log10(P value) and colored on a red‐to‐yellow gradient. The larger and redder the node, the greater the significance of the P value. Ala, alanine; AUC, area under the curve; FA, fatty acid; IDL, intermediate‐density lipoprotein; L, large; LDL, low‐density lipoprotein; M, medium; MUFA, monounsaturated FA; PL, phospholipids; S, small; SFA, saturated FA; SLEDAI, Systemic Lupus Erythematosus Disease Activity Index; TG, triglyceride; VLDL, very low–density lipoprotein; XXL, extra‐extra large. Color figure can be viewed in the online issue, which is available at http://onlinelibrary.wiley.com/doi/10.1002/art.70127/abstract.

To investigate whether metabolite levels were independently associated with subgroup classification, univariate logistic regressions were conducted for each serum metabolite. These models were adjusted for relevant clinical, demographic, and treatment covariates, including disease duration, age, and SLEDAI score. This analysis identified six metabolites—acetoacetate, citrate, creatinine, large LDL TGs, LDL TGs, and medium LDL TGs—that were significantly up‐regulated in the patients with cluster 1 compared to cluster 2 (Figure 2B and C).

Next, six supervised ML models were trained and validated to discriminate between cluster 1 and cluster 2. Given the high degree of correlation among metabolites, homology reduction was applied (see the Materials and Methods section), resulting in a data set of 111 nonredundant metabolites. The same clinical variables used in logistic regression were also included as features (Figure 2B). The best‐performing model, based on accuracy, specificity, and F1 score, was the neural network (Figure 2D and E; Supplementary Table 4). This model achieved an area under the receiver operating characteristic curve of 0.77 (Figure 2D), with a sensitivity of 70.92% and specificity of 73.08%.

Top features included citrate, mycophenolate mofetil use, very low–density lipoprotein (VLDL) particle size, pyruvate, and alanine. Citrate, acetoacetate, and creatinine—also significant in logistic regression—ranked among the top 20 features (Figure 2E), further supporting their relevance in distinguishing patients with cluster 1 from cluster 2. To determine if metabolites identified by the neural network model reflected metabolic differences between cluster 1 and cluster 2, a metabolite set enrichment analysis was performed (Figure 2E and F). Pathways related to amino acid and lipid metabolism were enriched in cluster 1, including “ketone body metabolism” (P = 0.002) and “valine, leucine, and isoleucine degradation” (P = 0.04; Figure 2F). Network analysis confirmed the connections among these pathways (Figure 2E).

To validate and extend our previous findings, we performed a massively parallel multiomic analysis in 10 patients (5 with cluster 1 and 5 with cluster 2), detecting 1,034 proteins (Olink proteomics) and 5,237 metabolites (mass spectrometry). First, MOFA (Supplementary Figure 5A) revealed that proteomics explained the majority of variance (39.25%), whereas metabolomics contributed 1.30%, consistent with the initial observations. Next, volcano plot analysis of proteomics data (fold change ≥ 2, P < 0.05; Supplementary Figure 5B) identified 340 proteins up‐regulated in clusters 1 and 2 down‐regulated in cluster 1. Metabolomics volcano analysis (fold change ≥ 2, P < 0.05; Supplementary Figure 5C) revealed 228 metabolites enriched in cluster 1 and 119 down‐regulated. Significant proteins and metabolites from these analyses were then correlated (FDR‐adjusted P < 0.05), yielding 2,467 significant correlations, predominantly positive (Supplementary Figure 5D).

Finally, biologic process enrichment analysis (STRING) of the 150 proteins involved in significant correlations highlighted relevant functional pathways (Supplementary Figure 5E), such as immune system processes, immune response and cytokine‐mediated signaling, and functional classification of the 84 metabolites involved in significant correlations (Supplementary Figure 5F) delineated inflammation‐linked metabolic signatures, including tryptophan/indole signaling, bile acid–mediated immunoregulation, lipid signaling/inflammation, and core energy metabolism.

Omics integration reveals specific biomarker signatures associated with CV risk and LN

Considering the impact of CVD risk (CVD risk⁺/CVD risk⁻) and LN (SLE‐LN⁺/SLE‐LN⁻) in SLE, we investigated molecular biomarkers to classify patients using a supervised approach.

Molecular CVD risk signature

Patients from the PRECISESADS cohort were classified based on the presence (CVD risk⁺, n = 24) or absence (CVD risk⁻, n = 76) of dyslipidemia and hypertension. The group with CVD risk was older, had longer disease duration, higher disease activity (SLEDAI > 8), and greater diabetes prevalence (Supplementary Table 5). This group showed up‐regulation of 23 circulating inflammatory and organ damage–related proteins (top 5: LHB, ADGRG1, TNFRSF9, IL‐4, and PLIN1), with only pleiotrophin (PTN) down‐regulated. Additionally, 71 significant metabolites were identified, mainly lipids and subclasses of high‐density lipoprotein, LDL, and intermediate‐density lipoprotein (Figure 3A). Multiple significant correlations (Pearson, P < 0.05) linked proteins related to chemotaxis, immune activation, and inflammation (eg, CXCL9, CD40, IL‐4) with proatherogenic metabolites (Figure 3B). These findings were validated in an independent UCL cohort (22 with CVD risk and 36 without CVD risk) with comparable clinical features (Supplementary Table 5), in which 60 metabolites were significantly up‐regulated in patients with CVD risk (Figure 4A). Comparison between PRECISESADS and UCL cohorts identified 49 overlapping significant metabolites of 71 (69% concordance; Figure 4B and C), confirming a robust metabolic signature associated with CV risk in lupus.

Figure 3.

Cardiovascular (CV) risk and lupus nephritis studies. (A) Volcano plots show 23 significant circulating inflammatory and organ‐damage proteins and 71 significantly up‐regulated metabolites in patients with CV disease (CVD). Proteomic and metabolomic differences are displayed as log2‐transformed fold changes (FCs). Thresholds correspond to an FC ≥2 on the linear scale (|log2[FC]| ≥ 1) for proteins and metabolites, with FDR‐adjusted P <0.05 for metabolites and nominal P <0.05 for proteins. Dashed vertical lines indicate the linear‐scale FC thresholds; for proteins, no significant down‐regulated proteins were detected, but the line is included for clarity. (B) Circular plots display the results of protein‐driven (black) and metabolite‐driven (gray) integration of omics in the CV risk study. (C) Volcano plots show 6 down‐regulated and 21 up‐regulated significant circulating inflammatory and organ‐damage proteins, as well as 40 significantly up‐regulated metabolites in patients with lupus nephritis risk. Thresholds and dashed vertical lines are defined as in panel A. (D) Circular plots display the results of protein‐driven (black) and metabolite‐driven (gray) integration of omics in the lupus nephritis study. In the center of each circular plot, significant correlation coefficients (Pearson, P < 0.05) are visualized through links connecting the two correlated omics. Links are colored according to the sign of the correlation coefficient: red for positive correlations and blue for negative correlations. The thickness of the links increases with the absolute value of the correlation coefficients. C, cholesterol; CE, cholesteryl ester; CRH, corticotropin‐releasing hormone; FA, fatty acid; HDL, high‐density lipoprotein; IDL, intermediate‐density lipoprotein; L, large; LDL, low‐density lipoprotein; M, medium; P, particles; PL, phospholipids; PUFA, polyunsaturated FA; S, small; uPA, urokinase plasminogen activator; VLDL, very low–density lipoprotein. Color figure can be viewed in the online issue, which is available at http://onlinelibrary.wiley.com/doi/10.1002/art.70127/abstract.

Figure 4.

Validation of metabolic signatures in the University College London (UCL) cohort. (A) Volcano plot showing the differential metabolites between the groups with and without cardiovascular disease (CVD) risk in the UCL validation cohort. Significantly regulated metabolites are identified based on –log P values and log2 fold changes (FCs). (B) Venn diagram displaying the overlap among metabolites exclusive to the PRECISESADS cohort (orange) and the UCL cohort (brown), along with the common metabolites between both cohorts. (C) Likert scale plot showing the mean values of shared metabolites between the PRECISESADS and UCL cohorts, with specific colors assigned to each study group and cohort: dark brown for CVD risk⁻ in UCL, orange for CVD risk⁻ in PRECISESADS, brown for non‐CVD risk in UCL, and yellow for non‐CVD risk in PRECISESADS. (D) Volcano plot similar to (A) but focused on the lupus nephritis signature. (E) Venn diagram comparing metabolites exclusive to the PRECISESADS cohort (blue) and UCL cohort (gray) for lupus nephritis. (F) Likert scale plot showing the mean values of shared metabolites between the PRECISESADS and UCL cohorts for lupus nephritis, with color assignments as follows: dark black for CVD risk⁺ in UCL, dark blue for CVD risk⁺ in PRECISESADS, gray for non‐CVD risk in UCL, and light blue for non‐CVD risk in PRECISESADS. C, cholesterol; CE, cholesteryl ester; DHA, docosahexaenoic acid; FA, fatty acid; HDL, high‐density lipoprotein; IDL, intermediate‐density lipoprotein; L, large; LDL, low‐density lipoprotein; M, medium; P, particle; PL, phospholipids; S, small; VLDL, very low–density lipoprotein. Color figure can be viewed in the online issue, which is available at http://onlinelibrary.wiley.com/doi/10.1002/art.70127/abstract.

Molecular LN signature

The PRECISESADS cohort included 34 patients with LN and 66 patients without LN (Supplementary Table 6). Differential analysis of serum proteome and metabolome revealed 27 up‐regulated proteins related to inflammation and organ damage (top 5: CX3CL1, IL‐2RB, CRH, and urokinase plasminogen activator [uPA]) and six down‐regulated proteins (FES, FKBP1B, WAS, BID, TNFSF24, and PRKAB1). Additionally, 40 metabolites, mainly LDL and VLDL subclasses linked to cholesterol metabolism, inflammation, and fatty acids, were significantly increased. Correlation analysis showed multiple positive associations (Pearson, P < 0.05) among proteins related to renal damage (FGF23, BID, ENTPD6, TNFSF14, and uPA) and proatherogenic/proinflammatory metabolites (LDL, VLDL subclasses, GlycA, monounsaturated fatty acid, and remnant‐C), indicating a coordinated proteomic–metabolomic profile in patients with LN (Figure 3C and D).

Validation in a UCL cohort (29 with LN, 12 without LN) with similar clinical features (Supplementary Table 6) identified 93 up‐regulated metabolites (Figure 4D) in patients with LN. Cross‐cohort comparison showed 35 overlapping metabolites (87.5% of PRECISESADS altered metabolites), confirming a robust metabolomic signature associated with LN (Figure 4E and F).

Integrated correlation networks reveal distinct protein–metabolite interactions associated with CV risk and LN in SLE

To further explore the biologic connections between omics data and disease severity in patients with SLE, we constructed integrated correlation networks based on proteins and metabolites significantly associated with CVD risk and LN. For the group with CV risk, the network was built using 23 proteins and 49 overlapping metabolites identified in the subgroup with CVD risk. Significant correlations (Pearson, P < 0.01) were found between common metabolites and five key proteins: CCL11, CXCL6, CES2, PLIN1, and RARRES1 (Supplementary Figure 6A). Expression analysis revealed that all molecules were up‐regulated in patients with cluster 1 compared to cluster 2, except for RARRES1, which showed down‐regulation (Supplementary Figure 6B).

Similarly, in the cohort with LN, a correlation network was developed using 23 proteins and 49 metabolites significantly associated with LN. Significant associations (Pearson, P < 0.01) were observed between common metabolites and five proteins: LIFR, AMN, BID, ENTPD6, and IL‐7 (Supplementary Figure 6C). Most proteins were up‐regulated in cluster 1 compared to cluster 2, whereas BID exhibited an opposite regulation pattern, being down‐regulated in the group in cluster 1 (Supplementary Figure 6D). These integrated protein–metabolite networks highlight distinct molecular signatures linked to disease severity in patients with SLE with CV risk and LN, offering insight into the underlying biologic mechanisms and potential therapeutic targets.

The direct contribution of the altered circulating molecular profile observed in patients with lupus to CV and renal phenotypes was confirmed through in vitro studies

To explore CV and renal damage mechanisms in lupus, HUVECs and renal HK2 cells were treated with serum from patients with cluster 1 and cluster 2 (Figure 5A). Cluster 1 serum induced significant increases in 29 CVD‐related proteins—including inflammatory (IL‐6, PTX3, IL‐18), growth factors (SCF, HB_EGF, PDGFb), antioxidants (HO_1, SOD2), proteolytic enzymes (MMP7, CTSL1), signaling molecules (PAPR‐1, SCR), and hemostatic regulators (TM, TF)—all linked to endothelial dysfunction and CV risk (Figure 5B and C). Similarly, renal cells treated with the serum of the cluster 1 of patients with LN displayed and altered expression of 34 inflammatory proteins compared with cells treated with the serum of the cluster 2 without LN. These altered proteins involved inflammatory cytokines (IL‐17A, IL‐13, IL‐20, and IL‐17C), growth factors (VEGFA, FGF‐21, and FGF‐19), immunomodulatory proteins (CSF‐1, IL‐15RA), chemokines (MCP‐2, CCL4, CXCL9), markers of anti‐inflammation (IL‐10RA, IL‐10RB, and LAP TGF‐beta‐1) and proteins involved in checkpoint pathways, critical in autoimmune responses (PD‐L1 and CD244; Figure 5D and E). Moreover, some of these proteins were further found altered in the serum of patients with lupus, such as MCP‐2, IL‐17C, CDCP1, CCL4, CCL25, CXCL9, and FGF‐23 (Supplementary Figure 7A). Biologic enrichment analyses revealed clear similarity, especially in inflammation and immune response pathways like “proinflammatory and profibrotic mediators” and “IL‐10 signaling” (Supplementary Figures 7B and C and 8), as well as significant enrichment of “cytokine–cytokine receptor interaction” in both data sets.

Figure 5.

In vitro cellular models reveal proteomic signatures associated with cardiovascular disease (CVD) risk and lupus nephritis (LN). (A) Schematic image of the in vitro experimental setup in which serum samples from patients with or without CVD risk (cluster 1 [C1]/C2) and nephritis (C1/C2) were incubated with human umbilical vein endothelial cells (HUVECs) and kidney tubular cells (HK2) to study cellular responses related to CVD risk and LN. (B) Heatmap of all proteins measured in the CV Olink panel, expressed in HUVECs after a 24‐hour incubation with serum from patients in C1 and C2, highlighting the proteomic profile studied for CVD risk. (C) Radar chart displaying the mean values of the significant proteins in the CV panel, scaled from 0 to 10 to facilitate visual comparison due to variability among proteins. The chart illustrates the increased expression of proteins in C1 compared to C2. (D) Heatmap of proteins measured in the inflammation‐specific Olink panel, expressed in HK2 cells incubated with serum from patients in C1 and C2, used to study LN‐related inflammation. (E) Radar chart showing the mean values of significant proteins from the inflammation‐specific panel, scaled from 0 to 10, to visualize the differences in protein expression between the C1 and C2 in the context of LN. Color figure can be viewed in the online issue, which is available at http://onlinelibrary.wiley.com/doi/10.1002/art.70127/abstract.

Ex vivo exposure of rat kidney slices to serum of patients with lupus revealed activation of inflammatory pathways, highlighting potential mechanisms underlying CV and renal damage

Lastly, to assess the ex vivo effects of altered serum proteomic and metabolomic profiles on CV and renal systems, rat kidney slices were exposed to serum from patients with SLE stratified as cluster 1, cluster 2, or HD. p65 expression was analyzed to evaluate NF‐κB activation in renal arteries and tissue (Supplementary Figure 9A). Mean fluorescence intensity showed significantly higher p65 in vascular endothelium of slices treated with cluster 1 or cluster 2 serum versus HD (Supplementary Figure 9B) and were similarly elevated in glomeruli (Supplementary Figure 9C). Both groups in cluster 1 and cluster 2 had increased p65 compared to HDs, with significant differences in vascular (cluster 1 vs HD, P = 0.05; cluster 2 vs HD, P = 0.03) and glomerular compartments (cluster 1 vs HD, P = 0.04; cluster 2 vs HD, P = 0.004). Confocal microscopy detailed these effects (Supplementary Figures 9D and E and 10).

Although ex vivo comparisons between cluster 1 and cluster 2 did not reach statistical significance, a consistent trend toward higher inflammatory activation in cluster 1 was observed across vascular and glomerular compartments. To investigate whether these subtle differences reflected underlying molecular divergence, we performed targeted proteomic profiling using the Olink Mouse Exploratory Panel on the same experimental system (Figure 6A). This analysis revealed that the average NPX values of the proteins included in the panel were significantly higher in cluster 1 than in cluster 2 (P < 0.001), indicating a globally increased inflammatory protein signature in cluster 1 (Figure 6B). Consistent with this, the heatmap (Figure 6C) demonstrated that most proteins showed higher expression levels in cluster 1 compared to cluster 2, further supporting the presence of distinct inflammatory signaling patterns among these serum‐derived subgroups.

Figure 6.

Ex vivo assessment of serum nephrotoxic potential of patients with lupus. (A) Schematic representation of the ex vivo experiment conducted to assess renal inflammation using kidney slices from male Wistar rats, incubated for six hours with 15% serum from healthy donors (HDs) and patients with lupus classified into cluster 1 (C1) and C2 subgroups. Finally, proteome assays were performed by PEA using the mouse cytokine panel compatible with rat samples. (B) Violin plots showing the distribution of mean‐centered NPX values for all proteins included in this analysis, illustrating the global variability between groups C1 and C2 and providing supportive evidence for sections B through E. (C) Heatmap depicting the average mean‐centered NPX values of the proteins shown in (B), summarizing group‐level expression patterns between C1 and C2. NPX, normalized protein expression. Color figure can be viewed in the online issue, which is available at http://onlinelibrary.wiley.com/doi/10.1002/art.70127/abstract.

DISCUSSION

SLE is a highly heterogeneous autoimmune disease that complicates diagnosis and treatment. By integrating proteomic and metabolomic analyses, this study identified two distinct proteomic endotypes using unsupervised approaches, which were further supported by metabolomic alterations and associated clinical features. The more severe endotype was characterized by increased CV risk and renal involvement alongside dysregulated inflammatory mediators and metabolic disturbances. Overall, this multiomic and ML framework improves patient stratification and provides a basis for more personalized therapeutic strategies.

Among the 46 proteins up‐regulated in cluster 1, both well‐established inflammatory mediators and less‐characterized proteins associated with NF‐κB signaling and CVD were identified, underscoring their potential roles in SLE pathophysiology. ⁹ Notably, several of these proteins (eg, IL‐6, tumor necrosis factor [TNF], IL‐8, CXCL9/10/11, IL‐18, and oncostatin M) have been linked to inflammatory processes beyond systemic inflammation, including tissue‐specific alterations such as renal or hepatic involvement, among others, which may contribute to the molecular differences observed among the identified endotypes.

Previous studies have identified biomarkers of organ involvement in SLE, including CD163 and IL‐16 for renal disease ¹⁰ and von Willebrand factor and GDF15 for CV risk. ¹¹ , ¹² Although some of these markers were not included in our analysis, our results revealed convergent biologic themes, such as enhanced myeloid/macrophage activation in LN and signatures of endothelial stress associated with CV risk. Differences among studies may reflect heterogeneity in patient populations, disease activity, and analytical platforms.

Consistent with our findings, Petrackova et al reported the up‐regulation of several overlapping proteins in patients with SLE with organ damage or biopsy‐proven active LN. ¹³ Similarly, Ruacho et al identified a subset of patients characterized by a systemic inflammatory state enriched in TNF, IL‐6, chemokines, and interferon γ, which mirrors the cytokine profile observed in our cluster 1. ¹⁴ These mediators promote the recruitment of immunocompetent cells, reinforcing a pathogenic inflammatory loop driven in part by TNFα‐induced chemokine expression. ¹⁵ The proteomic stratification was independently validated in an external cohort, in which it remained associated with key clinical features such as renal involvement and inflammation, with a high degree of overlap in dysregulated proteins across cohorts.

Metabolomic analyses revealed that cluster 1 exhibited elevated creatinine and citrate levels—markers of renal dysfunction and metabolic alterations—along with changes in LDL lipid profiles. Citrate links glycolysis and lipid metabolism and may contribute to enhanced inflammation in cluster 1, ¹⁶ whereas creatinine, beyond reflecting renal function, may indicate broader metabolic disturbances. In contrast, decreased isoleucine levels in cluster 1 are consistent with its role in regulating innate and adaptive immunity and key metabolic pathways. ¹⁷

ML models integrating clinical and metabolomic data accurately discriminated cluster 1 from cluster 2, underscoring the added value of multiomic approaches for patient stratification. Logistic regression confirmed increased levels of citrate, creatinine, and LDL subclasses in cluster 1. In addition, alterations in acetoacetate were observed, a metabolite previously linked to rheumatoid arthritis and vascular disease through its capacity to promote oxidative stress and lipid peroxidation. ¹⁸ , ¹⁹

In contrast to earlier studies that applied supervised ML to metabolomic data based on predefined clinical categories, ²⁰ , ²¹ our approach identified a metabolomic signature using ML within unsupervised proteomic‐derived clusters. Notably, the neural network highlighted citrate and isoleucine, together with clinical variables such as thrombosis and anti‐dsDNA levels, as key contributors to a disrupted metabolic network involving energy metabolism, immune activation, and oxidative stress. Altered branched‐chain amino acid metabolism, including reduced isoleucine levels, may further contribute to inflammatory and metabolic dysregulation in SLE through impaired ketogenesis and mitochondrial function. ²² , ²³ , ²⁴ , ²⁵

Lupus‐associated dyslipidemia, characterized by elevated TGs and VLDL, is driven by imbalances in fatty acid saturation, with increased saturated fatty acids promoting inflammation and disease progression. ²⁶ Although creatinine is not directly involved in energy metabolism, altered levels may reflect changes in protein metabolism or renal function, indirectly linking it to isoleucine metabolism. ²⁷ These findings, supported by metabolite set enrichment analysis, enhance our understanding of lupus‐related metabolic pathways and may aid diagnosis, endotype discrimination, and personalized therapeutic strategies.

To further extend our findings, we performed deeper integrated proteomic and untargeted metabolomic analyses in a smaller subset of patients, allowing a more detailed exploration of the connection among these omic layers. This approach recapitulated the inflammatory endotype identified in cluster 1 and uncovered additional dysregulated pathways consistent with immunometabolic reprogramming. Specifically, the convergence of inflammatory proteomic signals with metabolic alterations—particularly involving lipid metabolism, tryptophan/kynurenine pathways, and mitochondrial energy processes—supports a model in which immunometabolic dysregulation underlies a high‐risk SLE phenotype. Dysregulated lipid metabolism has been implicated in immune cell hyperactivation and organ damage in SLE, ²⁶ whereas enhanced kynurenine pathway activity has been associated with disease activity and renal involvement. ²⁸ Mitochondrial dysfunction and oxidative stress, which promote reactive oxygen species production and inflammasome activation, are also recognized contributors to SLE pathogenesis. ²⁹ Because there were no differences in steroid exposure, it is unlikely that differential protein and metabolite expression profiles identified were driven by pharmacologic effects of steroids. Together, these results reinforce the robustness of the identified endotypes and support immunometabolic coupling as a key pathophysiologic feature in SLE. Given the close link between the identified biomarkers and CV and renal involvement, we performed stratified analyses further characterize and contextualize the underlying molecular signatures.

CV risk in SLE was associated with a distinct molecular signature comprising inflammatory and immune‐related proteins together with extensive lipid dysregulation. Key proteins involved in vascular inflammation and immune activation included MCP4, PDCD1, TNF, CXCL6, and CCL11, whereas regulators of lipid and energy metabolism such as PLIN1, PTN, and 4E‐BP1 highlighted pathways linked to cholesterol handling and fat accumulation, central to atherosclerosis. ³⁰ , ³¹ , ³² Additional proteins related to hormonal signaling, vascular regulation, and drug or lipid metabolism were also identified. ³³ , ³⁴ , ³⁵ , ³⁶ Metabolically, CV involvement was characterized by alterations across multiple lipoprotein subclasses, cholesterol, fatty acids, and apolipoproteins, all closely linked to CV dysfunction. ³⁷ Integrative network analyses revealed strong connections between immune‐related proteins and lipid metabolites, reinforcing immunometabolic dysregulation as a key contributor to CVD severity in SLE. These results highlight the need to consider CV risk when interpreting systemic molecular profiles in SLE because vascular comorbidities can significantly influence circulating proteomic and metabolomic signatures.

Renal involvement was associated with a distinct immunometabolic signature, characterized by dysregulated proteins involved in immune cell recruitment and renal damage, including CX3CL1, IL‐2RB, CRH, MCP1, and TRANCE, ³⁸ , ³⁹ , ⁴⁰ , ⁴¹ together with lipid alterations dominated by VLDL and LDL subclasses. ⁴² , ⁴³ Integrative analyses revealed strong proteomic–metabolomic crosstalk, supporting coordinated immune and metabolic dysregulation in LN. These findings were validated in an independent cohort, indicating a persistent molecular footprint associated with renal involvement.

Despite the breadth of analyses performed across multiple omic layers and independent cohorts, our study converged on a robust proteomic and metabolomic signature composed of a defined set of molecules that consistently discriminated patient subgroups including the following: CCL20, IL‐10, CXCL11, CDCP1, MCP‐1, MCP‐2, CXCL10, CCL4, IL‐6, MCP‐3, IL‐8, EN‐RAGE, citrate, creatinine, acetoacetate, large and medium LDL TGs, tryptophan/indole derivatives, bile acids, and lipid‐mediated metabolites. These markers capture key biologic processes underlying disease heterogeneity and may serve as a valuable foundation for future mechanistic studies, biomarker validation efforts, and patient stratification approaches.

To investigate the functional relevance of the identified biomarkers, in vitro studies were performed in endothelial (HUVEC) and renal tubular (HK2) cells. Serum from patients with cluster 1 induced endothelial activation and dysfunction, characterized by increased inflammatory mediators, oxidative stress markers, and dysregulated angiogenic signaling, all processes implicated in CV risk in SLE. ⁴⁴ , ⁴⁵ , ⁴⁶ These findings indicate that cluster 1 serum actively promotes endothelial injury and proatherogenic responses. Similarly, exposure of HK2 cells to serum from patients with LN, including those in cluster 1, triggered the up‐regulation of multiple proteins involved in inflammation, immune cell recruitment, and fibrotic signaling. Key mediators associated with glomerular injury, Th17‐driven inflammation, and fibrogenesis were identified, supporting a direct role of patient serum in promoting tubular dysfunction and LN progression. ⁴⁷ , ⁴⁸ , ⁴⁹ , ⁵⁰ , ⁵¹

Ex vivo experiments using rat kidney tissue exposed to patient serum revealed increased expression of proteins related to inflammation, metabolic regulation, and tissue injury, accompanied by a trend toward activation of the NF‐κB pathway as a potential contributing mechanism. These findings indicate that serum from patients with cluster 1 exerts a direct molecular impact on renal tissue, consistent with enhanced inflammatory and metabolic stress responses. Together, these data further support the notion that patients in cluster 1 display a systemic molecular profile capable of driving renal damage, linking circulating immunometabolic alterations to tissue‐level pathology. ⁵² , ⁵³ Overall, the functional analyses support that the identified molecular signatures have a direct biologic impact on vascular and renal tissues, linking immunometabolic dysregulation to organ‐specific damage in SLE and reinforcing their relevance as mechanistic biomarkers and therapeutic targets.

The CV and renal molecular signatures identified in this study define a high‐risk immunometabolic endotype in SLE, characterized by coordinated immune activation and metabolic dysregulation. Patients with increased CV risk and LN showed up‐regulation of inflammatory cytokines and chemokines, Th17‐related mediators, and broad alterations in atherogenic lipoproteins and inflammatory metabolic markers, reflecting pathways shared with nonautoimmune CVD and renal disease. Several key biomarkers overlapped with those linked to adverse outcomes in general and chronic kidney disease populations, supporting the relevance of these signatures beyond lupus‐specific inflammation. Although limited by the cross‐sectional design, the concordance with established CV risk profiles and the functional impact of high‐risk SLE sera on endothelial and renal models suggest a biologically plausible link to future CV events that merits longitudinal validation.

Despite the strengths of this comprehensive multiomic study, some limitations should be acknowledged. The cohorts were of moderate size, predominantly European, and largely characterized by mild disease activity; however, the consistent replication of molecular signatures across independent cohorts supports the robustness of our findings. The cross‐sectional design limits causal inference, highlighting the need for future longitudinal studies to assess prognostic value and clinical outcomes, as well as to validate these signatures in more diverse populations. Additionally, the presence or absence of nephrotic syndrome among patients with abnormal proteinuria may represent a potential confounder for the metabolomic profiles observed in this subgroup and should therefore be taken into account when interpreting these findings. Prospective studies with systematic characterization of nephrotic status will be required to clarify its specific impact on these metabolomic signatures.

Although HUVECs differ from the vascular endothelium lining arteries, microvasculature, or glomerular endothelial cells, and although rat kidney slices suffer from species differences, these are widely accepted models for exploratory studies investigating vascular inflammation and CV risk. This study identifies a robust immunometabolic stratification of SLE, linking distinct proteomic and metabolomic signatures to CV risk and LN. Integrating multiomic data, ML, and functional validation provides mechanistic insight into organ‐specific damage and supports biomarker‐driven patient stratification and personalized therapies.

AUTHOR CONTRIBUTIONS

All authors contributed to at least one of the following manuscript preparation roles: conceptualization AND/OR methodology, software, investigation, formal analysis, data curation, visualization, and validation AND drafting or reviewing/editing the final draft. As corresponding authors, Drs Cerdó and Lopez‐Pedrera confirm that all authors have provided the final approval of the version to be published, and take responsibility for the affirmations regarding article submission (eg, not under consideration by another journal), the integrity of the data presented, and the statements regarding compliance with institutional review board/Declaration of Helsinki requirements. All authors contributed to the investigation, formal analysis of the data, and writing, reviewing, and editing of the manuscript.

Supporting information

Disclosure Form:

Supplementary Table 1. Clinical and analytical profiles of SLE patients.

Supplementary Table 2. A general linear model (GLM) was used to assess the association between study group and biomarker levels (SPSS version 25.0). For each biomarker, the dependent variable was its measured concentration. The main independent variable (factor) was the study group, defined by the variable Group (Cluster 1 vs Cluster 2). Disease duration and age were included as covariates to adjust for potential confounding effects.

Supplementary Table 3. Clinical and analytical profiles of patients from UCL patients.

Supplementary Table 4. Comparison of predictive model performance of Cluster 1 vs Cluster 2.

Supplementary Table 5. Clinical and analytical profiles of patients from PRECISASEADS cohort and UCL cohort classified according to the presence or absence of cardiovascular risk.

Supplementary Table 6. Clinical and analytical profiles of patients from PRECISASEADS cohort and UCL cohort classified according to the presence or absence of lupus nephritis.

Supplementary Figure 1. Schematic of multi‐omics factor analysis (MOFA) workflow overview, downstream analyses, and factor determination in the PRECISESADS cohort. A) Metabolomic and proteomic data were obtained from serum samples of 100 systemic lupus erythematosus (SLE) patients. The data were put through MOFA to identify latent factors and the variance decomposition by factors. B) The determined factors and the percentage of explained variance for each view per identified factor were shown. C) Bar charts depict the total variance explained for each biological data view by all the factors combined.

Supplementary Figure 2. Enrichment analysis and PPI network of significant upregulated proteins in the Cluster 1 vs Cluster 2. A) Predicted and validated protein‐protein interactions (PPIs) among the significantly upregulated proteins found in the Cluster 1 compared to the Cluster 2, analyzed using the STRING platform. The protein network illustrates the relationships between differentially expressed proteins. B) The accompanying table summarizes significant biological pathways identified through functional enrichment analysis, focusing on immune system processes, nephritis, and cardiovascular risk, using Gene Ontology (GO) and KEGG platforms.

Supplementary Figure 3. Proteomic analysis of healthy donors (HD) versus SLE patients. A) Supervised heatmap of patients and healthy donors. The high disease severity subgroup (Cluster 1, n = 48) is shown in red, the low disease severity subgroup (Cluster 2, n = 52) in green, and healthy donors (HD, n = 19) in blue. B) Partial Least Squares Discriminant Analysis (PLS‑DA) summarizing the differences among the protein profiles of the high disease severity (Cluster 1) subgroup (red), the low disease severity (Cluster 2) subgroup (green), and the healthy donor (HD) group (blue). C) Variable importance in projection plot (VIP): proteins identified by sPLS‑DA in descending order of importance. The graph represents the relative contribution of these proteins to the variance between the three groups. The colored boxes on the right indicate whether the protein levels are increased (red) or decreased (blue) in the serum of each group. D) Volcano plot showing 31 proteins upregulated in Cluster 1 patients compared with HD (FC ≥ 1.5, FDR, p < 0.05). E) Volcano plot showing 6 proteins downregulated in Cluster 2 patients compared with HD (FC ≥ 1.5, FDR, p < 0.05).

Supplementary Figure 4. Proteomic validation in the UCL cohort. A) Unsupervised heatmap of UCL patients. Cluster 1 is colored in purple (n=16) and Cluster 2 is colored in light blue (n=20). B) Principal Component Analysis (PCA) of patients. Cluster 1 is colored in light blue and Cluster 2 in purple. C) Clinical and analytical profiles of patients from Cluster 1 and Cluster 2. The figure represents the most important and significant variables between both subgroups. D) Venn diagram showing the overlapping and unique changing proteins between the unsupervised PRECISESADS cohort and the unsupervised UCL cohort. E) STRING network of significant proteins changing between Cluster 1 and Cluster 2 that are shared between the PRECISESADS and UCL cohorts. F) Functional enrichment analysis of the proteins shown in panel E.

Supplementary Figure 5. Massively parallel multi‐omics analysis in SLE patients. A) MOFA analysis showing the relative contribution of proteomics (Olink, 1,034 proteins) and metabolomics (mass spectrometry, 5,237 metabolites) to the observed variation between clusters. B) Volcano plot of proteomics data highlighting proteins significantly different between the two clusters (fold change ≥2, p < 0.05). Log₂ fold change (FC) calculated as Cluster 1 / Cluster 2; positive values indicate enrichment in Cluster 1, negative values indicate enrichment in Cluster 2. C) Volcano plot of metabolomics data showing metabolites significantly altered between clusters (fold change ≥2, p < 0.05). Log₂ fold change (FC) calculated as Cluster 1 / Cluster 2; positive values indicate enrichment in Cluster 1, negative values indicate enrichment in Cluster 2. D) Correlation analysis between proteins and metabolites identified as significant in panels B and C, with FDR‐adjusted p‐value <0.05. A total of 2467 significant correlations are visualized as a chord plot, where black bar represent proteins, gray bar represent metabolites, and edge colors indicate correlation direction (red = positive, blue = negative). E) Biological process enrichment analysis (STRING) of proteins involved in significant correlations from panel D. F) Functional grouping of metabolites involved in significant correlations from panel D.

Supplementary Figure 6. Correlation networks and expression patterns of proteins and metabolites associated with cardiovascular risk and lupus nephritis in SLE patients. A) Correlation network showing significant associations (Pearson, p < 0.01) between proteins and metabolites in SLE patients with cardiovascular risk compared to those without. Only metabolites overlapping with the UCL cohort were included in the analysis. Nodes represent molecules colored by type (proteins‐yellow‐; metabolites –grey‐), and edge thickness reflects correlation strength. B) Heatmap displaying the relative expression levels of the proteins and metabolites from panel A in a separate comparison of SLE patients from Cluster 1 (C1) versus Cluster 2 (C2). Blue indicates downregulation in Cluster 1; red indicates upregulation. Molecules highlighted with a red square show an opposite regulation pattern compared to the cardiovascular risk analysis in panel A. C) Correlation network of proteins and metabolites significantly associated with lupus nephritis in SLE patients, following the same criteria and layout as panel A. D) Heatmap of the molecules shown in panel C, illustrating their expression in SLE patients with high versus low disease severity. Blue represents downregulation in Cluster 1; red represents upregulation. Molecules marked with a red square exhibit regulation opposite to that observed in the lupus nephritis analysis (panel C).

Supplementary Figure 7. Shared and distinct inflammatory proteins profiles between those present in the serum of patients and those secreted by the renal in vitro model. A) Venn diagram showing the overlap between significantly dysregulated inflammatory proteins identified in patient serum (Cluster 1 vs. Cluster 2; red square) and those identified in HK2 cells after incubation with serum from the same patient groups (Cluster 1 vs. Cluster 2; blue square). The intersection highlights proteins commonly dysregulated in both serum and the renal cell model. B) Functional enrichment analysis (Metascape) of significantly altered inflammatory proteins in patient serum (Cluster 1 vs. Cluster 2), illustrating biological processes associated with heightened systemic disease severity. C) Functional enrichment analysis (Metascape) of significantly altered inflammatory proteins in HK2 cells after incubation with serum from Cluster 1 vs. Cluster 2 patients, illustrating pathways related to renal inflammation and immune activation.

Supplementary Figure 8. Hematoxylin and eosin staining of rat kidney slices exposed to sera from patients with lupus nephritis at different time points (3,6,9,24 and 48 h). Representative images are shown; scale bar = 50 um.

Supplementary Figure 9. Ex vivo assessment of the serum from lupus patients on renal inflammatory profiles. A) Schematic representation of the ex vivo experiment conducted to assess renal inflammation using kidney slices from male Wistar rats, incubated for 6 hours with 15% serum from healthy donors (HD) and lupus patients classified into low‐damage (Cluster 2) and high‐damage (Cluster 1) subgroups. Immunofluorescence staining was performed with P65 to assess NF‐κB activity, as indicated by nuclear translocation of the P65 subunit. The results were visualized using confocal microscopy. B) Bar graph showing the mean fluorescence intensity of vascular inflammation in kidney slices as measured by confocal microscopy. The x‐axis represents the study groups (Cluster 1 in orange, Cluster 2 in yellow, and HD in gray), and the y‐axis shows the fluorescence intensity ratio (red area/total area). Statistical significance is indicated by asterisks, with non‐significant comparisons labeled as “ns,” and p‐values are shown. C) Bar graph similar to (B) but showing the mean fluorescence intensity of glomerular inflammation. The colors for Cluster 1 and Cluster 2 are dark blue and light blue, respectively, and the statistical analysis is presented as described in (B). D) Confocal images of vascular inflammation in kidney slices. Row 1 represents the Cluster 1, row 2 the Cluster 2, and row 3 the HD group. Column 1 shows P65 staining (red), column 2 shows DAPI staining (blue), and column 3 shows the merged image, indicating the colocalization of P65 with the nucleus, indicative of NF‐κB activation. Imaging was performed using a Thunder Imager 3D Assay fluorescence microscope (Leica Microsystems) at 20× magnification. Scale bar = 50 μm. E) Confocal images of glomerular inflammation in kidney slices. Row 1 represents the Cluster 1, row 2 the Cluster 2, and row 3 the HD group. Column 1 shows P65 staining (red), column 2 shows DAPI staining (blue), and column 3 shows the merged image, indicating NF‐κB activation. Imaging was performed using a Thunder Imager 3D Assay fluorescence microscope (Leica Microsystems) at 20× magnification. Scale bar = 50 μm.

Supplementary Figure 10. Representative immunofluorescence images of whole kidney slices treated with serum from SLE subgroups and healthy donors.

Kidney slices from male Wistar rats were incubated ex vivo for 6 hours with 15% serum from healthy donors (HD), or from SLE patients classified into low‐damage (Cluster 2) and high‐damage (Cluster 1) subgroups. Immunofluorescence staining was performed to evaluate nuclear translocation of NF‐κB subunit p65 (red), and nuclei were counterstained with DAPI (blue). Merged images (right panels) show the colocalization of p65 signal with nuclear staining, indicative of NF‐κB activation. Imaging was performed using a Thunder Imager 3D Assay fluorescence microscope (Leica Microsystems) at 20× magnification. Scale bar = 1000 μm.

APPENDIX A. PRECISESADS CLINICAL CONSORTIUM

Members of the PRECISEADS Clinical Consortium are as follows: Lorenzo Beretta, Barbara Vigone, Jacques‐Olivier Pers, Alain Saraux, Valérie Devauchelle‐Pensec, Divi Cornec, Sandrine Jousse‐Joulin, Bernard Lauwerys, Julie Ducreux, Anne‐Lise Maudoux, Carlos Vasconcelos, Ana Tavares, Esmeralda Neves, Raquel Faria, Mariana Brandão, Ana Campar, António Marinho, Fátima Farinha, Isabel Almeida, Miguel Angel Gonzalez‐Gay Mantecón, Ricardo Blanco Alonso, Alfonso Corrales Martínez, Ricard Cervera, Ignasi Rodríguez‐Pintó, Gerard Espinosa, Rik Lories, Ellen De Langhe, Nicolas Hunzelmann, Doreen Belz, Torsten Witte, Niklas Baerlecken, Georg Stummvoll, Michael Zauner, Michaela Lehner, Eduardo Collantes, Ma Carmen Castro‐Villegas, Norberto Ortego, María Concepción Fernández Roldán, Enrique Raya, Inmaculada Jiménez Moleón, Enrique de Ramon, Isabel Díaz Quintero, Pier Luigi Meroni, Maria Gerosa, Tommaso Schioppo, Carolina Artusi, Carlo Chizzolini, Aleksandra Zuber, Donatienne Wynar, Laszló Kovács, Attila Balog, Magdolna Deák, Márta Bocskai, Sonja Dulic, Gabriella Kádár, Falk Hiepe, Velia Gerl, Silvia Thiel, Manuel Rodriguez Maresca, Antonio López‐Berrio, Rocío Aguilar‐Quesada, and Héctor Navarro‐Linares.