Network-based machine learning to identify biomarkers for systemic lupus erythematosus

Abstract

Background

Systemic lupus erythematosus (SLE) is a complex autoimmune disease, making accurate diagnosis and effective treatment challenging. Despite the critical need for reliable biomarkers, conventional differential gene expression (DGE) analyses often yield high false-positive rates and have limited ability to identify clinically actionable targets, leaving a significant gap in SLE precision medicine.

Results

To address this, we developed NetSLE, a network-based machine learning framework that integrates diverse SLE-related prior knowledge with comprehensive biological networks. By effectively filtering out false positives from differentially expressed genes (DEGs), NetSLE identified a robust panel of 150 key biomarkers. Clinically, the NetSLE-derived biomarkers outperformed conventional markers and full transcriptomes in predicting disease activity across independent cohorts. Furthermore, they successfully identified experimentally validated drug repurposing candidates (e.g., lipid-modifying and antithrombotic agents) and enabled precise stratification of patients into distinct immunological subtypes (AS1 and AS2).

Conclusions

NetSLE offers a translatable approach to overcome the limitations of traditional biomarker discovery. The clinically sized 150-gene panel provides a practical tool for enhancing diagnostic precision, guiding targeted treatments, and advancing personalized medicine in SLE.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12915-026-02681-w.

Background

Identifying robust biomarkers is crucial for establishing effective clinical strategies for managing systemic lupus erythematosus (SLE), a chronic autoimmune disease with complex and varied clinical symptoms [1]. The prevalence of SLE continues to rise, exceeding 100 per 100,000 people [2], yet its intricate pathogenesis and heterogeneous phenotypes make accurate diagnosis and early management challenging [3, 4]. While conventional biomarkers like anti-dsDNA and complement levels are used, their correlation with clinical symptoms remains inconsistent [5, 6]. Moreover, there is a recognized need to identify clinically useful biomarkers that can reduce disease activity and enhance quality of life through guidance on effective treatment options [7–9].

Differential gene expression (DGE) analysis is commonly used to identify clinical biomarkers [10–16]. However, DGE alone often identifies large numbers of differentially expressed genes (DEGs) that reflect transcriptomic consequences rather than causal disease mechanisms, leading to high false-positive rates and limited clinical utility [17–19]. Therefore, a major unmet need exists to improve the prioritization of DEGs to extract truly actionable biological differences.

To overcome these traditional limitations, we hypothesized that integrating diverse SLE-associated prior knowledge (e.g., GWAS [20–23] and experimental studies [24, 25]) through a network-based machine learning (ML) approach could address the limitation of DGE analysis and identify robust SLE biomarkers. Specifically, we reasoned that applying network propagation across a comprehensive protein–protein interaction network (such as the STRING database [26]) assesses functional associations more reliably than isolated network connections, thereby filtering out false positives [27–31]. By employing an ML framework to synthesize these multiple knowledge-derived network features into a single, unified scoring metric, true disease-related DEGs can be optimally prioritized [32–34].

Driven by this hypothesis, our primary objective was to develop NetSLE, a novel network-based ML framework designed to identify and prioritize clinically actionable SLE biomarkers from DEGs. To validate the translational value of these prioritized biomarkers, we pursued three specific clinical aims: (1) predicting disease activity, (2) identifying candidates for drug repurposing, and (3) enabling precise patient stratification.

Results

Overview of the network-based machine learning framework (NetSLE)

We developed a network-based machine learning (ML) framework, NetSLE, to identify robust SLE biomarkers and overcome the high false-positive rates typical of conventional DGE analysis. In this approach, DEGs in SLE active patients were assessed for their connection to SLE prior knowledge-related genes within a protein–protein association (PPA) network (Fig. 1A–B). This utilized guilt-by-association [35, 36] implemented through network propagation [28]. The prior knowledge included genes with SLE-related variants from GWAS (505 genes), cytokines of immune cells associated with SLE (45 genes), regulators of germinal center (26 genes), and BAFF, a known drug target of the SLE-approved drug “belimumab.” To further reduce false positives, the DEGs were refined using ML, resulting in the selection of NetSLE biomarkers (Fig. 1C). Finally, we utilized the NetSLE biomarkers to discover drug-repurposing candidates and classify subtypes of SLE-active patients (Fig. 1D).

Fig. 1.

Overview of the NetSLE framework for identifying SLE biomarkers. A Differentially expressed genes (DEGs) were identified by comparing SLE active patients (red) with SLE inactive patients (blue) and healthy controls (black). B Genes related to SLE prior knowledge, highlighted in heatmaps, are mapped onto a protein–protein association (PPA) network, and represented as colored nodes. The network propagation process is illustrated by colored arrows. C SLE biomarkers were prioritized from up-regulated DEGs using network propagation scores and machine learning (ML). Identified SLE biomarkers (NetSLE biomarkers) are indicated by red circles with black borders. D These NetSLE biomarkers were used to identify potential drug repurposing candidates and to subtype SLE active patients

Prioritization and identification of SLE biomarkers

We demonstrated the NetSLE approach by effectively prioritizing SLE-associated genes. From an initial set of 1119 SLE DEGs (out of 16,970 genes), we identified 150 NetSLE biomarkers (Fig. 2A). SLE DEGs were defined as genes up-regulated in SLE active patients (330 patients) compared to inactive patients (591 patients) and healthy controls (72 controls) (see the “ Methods” section) (Fig. 2B; Additional file 2: Data S1A–C). The DEGs were prioritized by integrating network propagation scores using an ML model that ranked them by their probability of association with SLE (P_SLE; see the “ Methods” section) (Additional file 2: Data S2A–B). A bimodal distribution in P_SLE revealed distinct groups of high and low relevance to SLE in the DEGs (Bimodality Index (BI) = 2.45, BI > 1.1 indicates bimodal distribution [37]) (Additional file 1: Fig. S1), exemplified by the prioritization of DDX58 (P_SLE = 0.18, rank = 130/1119) over CHMP5 (P_SLE = 0.04, rank = 833/1119) as a biomarker (Fig. 2C).

Fig. 2.

Identification of SLE biomarkers using NetSLE approach. A Workflow for prioritizing DEGs in SLE active patients using the NetSLE method. B Volcano plots showing DEGs between SLE active patients and inactive patients or healthy controls, with upregulated genes highlighted in pick and NetSLE-identified biomarkers in red. C Network representation of NetSLE biomarkers, where gray lines indicate interactions and genes are shown as circled nodes. Nodes corresponding to NetSLE biomarkers are marked with red borders. Only interactions with a confidence socre above 900 from the STRING database are shown. The color saturation of each node reflects the propagation score and P_SLE rank, with higher-ranking genes displayed in more saturated colors. D–E Venn diagram illustrating the overlap between different gene sets. F SLE-associated gene set enrichment analysis. The top bar plot shows the P_SLE of DEGs ranked in order. Heatmaps indicate the position of NetSLE biomarkers within gene sets from Chaussabel et al., Perez et al., DisGeNET, and KEGG. The stacked bar plot displays the cumulative enrichment across four SLE-associated gene sets as a function of the top-ranked gene count (increasing in increments of 5). Enrichment is evaluated using an Enrichr-derived combined score, which mathematically integrates the magnitude of gene overlap with its corresponding statistical significance. The red dashed line marks the rank threshold (top 150) for NetSLE biomarkers. G–J Comparison of enrichment in SLE-associated gene sets between SLE biomarkers identified using various methods and NetSLE biomarkers. The number of biomarkers with the highest enrichment for each method is indicated in parentheses. The gray dotted line indicates the baseline

The identification of these NetSLE biomarkers remained robust even when evaluated across varying clinical criteria for disease activity. We confirmed this across different SLEDAI criteria (thresholds of seven versus four) [38, 39]. Despite this adjustment, a significant overlap in the identified biomarkers was observed. When the SLEDAI threshold was set to four, 933 out of 1,043 identified DEGs overlapped with the original threshold of seven (odds ratio (OR) = 717.8, Fisher’s exact test P < 1.0 × 10⁻³⁰⁸). Among the 150 biomarkers selected from these 1043 DEGs, 131 genes significantly overlapped with the previously identified NetSLE biomarkers (OR = 384.7, P = 4.6 × 10⁻¹³³) (Additional file 1: Fig. S2).

Biological relevance and validation of NetSLE biomarkers

We found that NetSLE biomarkers were significantly enriched in pathways associated with inflammatory cytokines (e.g., interferon and interleukin) and immune cell functions, underscoring their roles in regulating immune responses. Analysis with hallmark gene sets from MSigDB [40] showed significant enrichment of NetSLE biomarkers in inflammatory response pathways, including interferon alpha and gamma responses, as well as associations with IL-6, TNF-alpha, and IL-2 pathways (see the “ Methods” section) (Additional file 1: Fig. S3A; Additional file 2: Data S3). This enrichment was consistent when GO slim gene sets [41] were used, highlighting immune system processes like “defense response to other organism” and “inflammatory response” (Additional file 1: Fig. S3B). NetSLE biomarkers were also related to coagulation, wound healing, and lipid metabolic processes, supported by previous reports on these pathways in SLE pathogenesis [42–46].

We further validated the NetSLE biomarkers against established SLE-associated gene sets, demonstrating that they significantly encompass known pathogenic genes while simultaneously uncovering novel candidates (see the “ Methods” section). Among NetSLE biomarkers, 33 and 36 genes were included in the signature genes from Chaussabel et al. [16] (Enrichment (E) = 2.37, Fisher’s exact test P = 1.67 × 10⁻²²) and Perez et al. [14] (E = 2.69, P = 6.06 × 10⁻³¹), respectively (Fig. 2D; Additional file 2: Data S4, Data S5A). NetSLE biomarkers significantly encompassed conventional biomarkers (E = 7.46, P = 3.02 × 10⁻⁶), DisGeNET SLE pathogenesis genes (E = 3.20, P = 2.49 × 10⁻⁶¹), and KEGG SLE pathway genes (E = 6.22, P = 6.36 × 10⁻⁵) (Fig. 2E; Additional file 2: Data S5A). Importantly, 43 genes within the NetSLE biomarkers were newly identified, absent from previously recognized transcriptome signatures or SLE-related pathways.

Through statistical evaluation, we confirmed the efficacy of the P_SLE scoring metric as a critical indicator for filtering out false positives and prioritizing true disease-associated DEGs. We revealed that SLE-associated genes exhibited significantly higher P_SLE compared to non-SLE-associated genes. The P_SLE for signature genes identified by Chaussabel et al. (Mann–Whitney U test P = 3.6 × 10⁻²⁹) and Perez et al. (P = 6.3 × 10⁻⁴²) were significantly elevated (Additional file 1: Fig. S4). Similarly, pathogenic genes from the DisGeNET (P < 1.0 × 10⁻³⁰⁸) and KEGG (P = 1.3 × 10⁻²⁴) database, as well as conventional biomarkers (P < 5.4 × 10⁻²⁴), showed significantly higher P_SLE than other genes. Recognizing P_SLE as a crucial metric for filtering out false positives, we systematically lowered its threshold (incrementally increasing the gene count by 5) and measured the enrichment of SLE-associated gene sets to determine the optimal number of biomarkers. As the threshold was progressively lowered, the enrichment decreased accordingly, peaking when the top 150 genes ranked by P_SLE were selected (see the “ Methods” section) (Fig. 2F; Additional file 2: Data S6).

Building on this effective filtering capacity, we further evaluated NetSLE’s discovery power by comparing its performance with three established network-based methods: uKIN [31], WGCNA [47], and NetBio [32] (see the “ Methods” section). NetSLE biomarkers exhibited the most significant overlap with all four SLE-associated gene sets (Fig. 2G–J; Additional file 2: Data S5B–D), demonstrating the superior capability of our approach.

Predictive SLE disease activity using NetSLE biomarkers

In clinical applications, the 150 NetSLE biomarkers demonstrated meaningful predictive performance for SLE disease activity across independent cohorts. We performed cross-cohort validation (GSE65391 as a training set; GSE49454 and GSE72798 as independent test sets) using a stochastic gradient descent (SGD) model (Fig. 3A). Evaluated on a classification score combining AUROC, AUPRC, Accuracy, MCC, and F1-score (see the “ Methods” section), the 150 NetSLE biomarkers achieved a score of 9.16, outperforming models using negative control gene sets (non-prioritized DEGs = 4.89; all DEGs = 4.89; all 16,970 genes = 4.69; conventional biomarkers = 5.33) (Fig. 3B; Additional file 1: Fig. S5A–B). Because these 150 NetSLE biomarkers consist exclusively of up-regulated genes, we further compared this up-regulated set against a combined set derived from up- and down-regulated genes to justify excluding down-regulated genes. The model utilizing only up-regulated genes yielded an outperformed disease activity prediction performance compared to the combined set (Additional file 1: Fig. S6A–B), demonstrating enhanced clinical utility and validating our exclusive focus on the up-regulation directionality.

Fig. 3.

Validation of NetSLE biomarkers through predicting SLE disease activity. A Cross-cohort validation for predicting disease activity. GSE65391 was used as a training set, and independent datasets (GSE49454 and GSE72798) were used as a test set. B Performance comparison of various gene sets in disease activity prediction. The cumulative classification scores of the prediction are depicted through stacked bars. Area under the receiver operating characteristics (AUROC); Area under the precision-recall curve (AUPRC); Matthews correlation coefficient (MCC)

We confirmed through systematic evaluation of various conditions that incorporating comprehensive prior knowledge and utilizing the top 150 genes with the SGD model yielded the optimal predictive accuracy (Additional file 1: Fig. S7A). Exploring all possible combinations, any deviation from these conditions negatively impacted performance. Omitting prior knowledge of immune cell signaling molecules reduced the score to 4.89, while increasing the number of biomarkers to 400 lowered it to 6.74 (Additional file 1: Fig. S7B). Alternative classification models, including support vector machine (SVM), random forest (RF), logistic regression (unpenalized), logistic regression (L1-penalized), logistic regression (elastic net-penalized), and gradient boosting (GB), exhibited inferior performance, with classification scores of 5.94, 5.65, 5.58, 5.45, 7.34, and 5.50, respectively.

NetSLE biomarkers enable target-based drug repurposing for SLE

NetSLE biomarkers successfully identified drug classes associated with both established SLE treatment and potential novel therapeutic candidates. We observed that drug classes related to drug targets for SLE treatment (SLE drug targets) showed greater overlap with NetSLE biomarkers than with negative control gene sets, including non-prioritized DEGs, all DEGs, and conventional biomarkers. Specifically, immunosuppressants (ATC code: L04), antibacterials for systemic use (J01) such as tetracyclines, and stomatological preparations (A01) such as prednisolone, were significantly related to both SLE drug targets and NetSLE biomarkers (Fig. 4A–B; Additional file 2: Data S7). Notably, NetSLE biomarkers showed significant relationships with antithrombotic agents (B01) and lipid-modifying agents (C10), suggesting these as potential options for SLE management supported by previous reports [44, 48, 49], whereas negative control gene sets showed minimal or no such associations.

Fig. 4.

Drug repurposing based on the relationship between NetSLE biomarkers and drug targets. A Identification of drug classes associated with NetSLE biomarkers through drug-target interactions and categorization using ATC codes. B Bubble plot illustrating the overlap between each gene set and target genes of drug class, with circle color indicating the odds ratio and circle size reflecting the significance of overlap. Statistical significance was assessed using Fisher’s exact test. C Diagram depicting the network shortest path between NetSLE biomarkers and drug targets, along with the method for calculating proximity to evaluate the therapeutic effect of drugs (see the “ Methods” section). D The SLE mouse model (MRL/MpJ-Fas^lpr/J) was used to prepare splenocyte single-cell suspensions, which were cultured with anti-CD3, anti-CD28, and anti-IgM antibodies in the presence of drugs, and the supernatants were analyzed for IgG, IFN-γ, and MPO levels using ELISA to assess immune response under specific conditions. E Target gene of prednisolone and its first neighbors among the NetSLE biomarkers. For clarity, only interactions from the STRING PPA network with a confidence score of 900 or higher are displayed. F–H Box plots display the distribution of immune component concentrations in ex vivo mouse experiments under drug treatment conditions (Additional file 2: Data S10). Asterisks represent statistically significant differences in concentration distributions between the drug treatment and control groups. Statistical significance was determined using the Mann–Whitney U test. * P < 5.0 × 10^–2, ** P < 1.0 × 10^–2, ***P < 5.0 × 10^–3, ns (P ≥ 5.0 × 10.^–2)

Motivated by these promising class-level associations, we further validated the ability of NetSLE biomarkers to identify specific therapeutic drugs for SLE. Through network proximity analysis, we quantitatively confirmed their significant proximity to the targets of both approved SLE drugs and those currently in clinical trials (see the “ Methods” section). Using a network proximity approach [50], we found that the proximity between NetSLE biomarkers and target genes of established SLE drugs was significant within the PPA network (proximity z-score < − 1.96) (Fig. 4C; Additional file 2: Data S8). For instance, BAFF (TNFSF13B) was significantly proximal to NetSLE biomarkers (z-score = − 26.4) (Additional file 1: Fig. S8A). Furthermore, anifrolumab, recently approved for SLE, was evaluated as having a strong therapeutic potential (z-score = − 18.9) (Additional file 1: Fig. S8B). Notably, its target gene IFNAR1, is included among the NetSLE biomarkers despite not being part of the prior knowledge used in their identification. Based on the assumption that drugs currently in clinical use (both on- and off-label) and those tested in clinical trials are more effective in treating SLE compared to untested drugs, we evaluated their relative network proximities. Target genes of drugs in clinical use (mean z-score = − 5.0) and those tested in clinical trials (mean z-score = − 3.7) were significantly more proximal to NetSLE biomarkers compared to untested drugs (mean z-score = 0.2; Mann–Whitney U test P = 4.4 × 10⁻³ and P = 1.4 × 10⁻⁶, respectively) (Additional file 1: Fig. S9).

Given the ability of NetSLE biomarkers to evaluate therapeutic effects, we found that antithrombotic agents (B01) and lipid-modifying agents (C10) were likely to be effective drug classes for treating SLE through individual drug-level network analysis. A significant number of drugs related to those classes were assessed to have therapeutic effects based on the proximity of NetSLE biomarkers to their target genes. Specifically, 42 out of 52 antithrombotic agents (log₂ odds ratio (OR) = 3.5, Fisher’s exact test P = 5.5 × 10⁻¹⁵) and 14 out of 29 lipid-modifying agents (log₂ OR = 1.2, P = 3.6 × 10⁻²) were assessed as having potential therapeutic effects (Additional file 2: Data S9). For classes related to both NetSLE biomarkers and established targets, 43 out of 53 immunosuppressants (log₂ OR = 3.5, P = 1.8 × 10⁻¹⁵) and 11 out of 13 stomatological preparations (log₂ OR = 3.8, P = 4.9 × 10⁻⁵) were similarly evaluated as effective.

Experimental validation using an ex vivo SLE mouse model (MRL/lpr) confirmed that top-ranked lipid-modifying and antithrombotic agents significantly reduce key immunological factors associated with disease activity, including IFN-γ (interferon gamma), IgG (immunoglobulin G), and MPO (myeloperoxidase) (see the “ Methods” section) (Fig. 4D). We tested the lipid-modifying agents pitavastatin, rosuvastatin, ezetimibe, niacin, ciprofibrate, and fenofibrate, along with the antithrombotic agents abciximab, triflusal, apixaban, and rivaroxaban (Additional file 1: Fig. S10). Prednisolone, a broad-spectrum glucocorticoid influencing multiple immune system targets [51], was used as a positive control (NR3C1 target, z-score = − 3.4) (Additional file 1: Fig. S8C) and significantly reduced IFN-γ, IgG, and MPO (Fig. 4F–H; Additional file 2: Data S10). In comparison, all tested lipid-modifying agents significantly reduced IFN-γ, IgG, and MPO, with the sole exception of rosuvastatin in the MPO assay. Among antithrombotic agents, most (except triflusal) significantly reduced MPO, and apixaban was found to markedly reduce all three immune components (IFN-γ, IgG, and MPO).

Expression levels of NetSLE biomarkers can identify subtypes of SLE active patients

Identifying patient subtypes in SLE based on immune-related pathogenesis is essential due to the disease’s complexity and heterogeneity. We demonstrated the clinical application of NetSLE biomarkers in identifying subtypes of SLE patients with high disease activity (SLE active patients) across various cohorts (Fig. 5A). By analyzing the similarity in expression profiles of these biomarkers (see the “ Methods” section), we identified two distinct subtypes, SLE active subtype 1 (AS1) and active subtype 2 (AS2), with different immunological characteristics. In the GSE49454 cohort, six patients significantly clustered into AS1 (observed/expected number of active patients = 6/3; chi-square test P = 3.4 × 10⁻²) and 31 into AS2 (31/21; P = 6.4 × 10⁻³) (Fig. 5B). Despite comparable disease activity between these active subtypes (Mann–Whitney U test P = 0.50) (Additional file 1: Fig. S11A), their immune-related pathogenesis likely diverged, as indicated by the differing expression profiles of NetSLE biomarkers (Fig. 5B). These results were consistent across diverse racial and ethnic cohorts, underscoring the robustness of our findings (Additional file 1: Fig. S11B–C, Fig. S12, and Fig. S13A–B). Additionally, significant correlations were observed in the differential gene expression between AS1 and AS2 across cohorts in pairwise comparisons (GSE49454-GSE72798: spearman ρ = 0.67, P = 2.1 × 10⁻¹⁷; GSE49454-GSE65391: ρ = 0.90, P = 7.3 × 10⁻⁵⁴; GSE72798-GSE65391: ρ = 0.75, P = 1.1 × 10⁻²⁸), highlighting the robustness of these biomarkers in characterizing SLE subtypes (Fig. 5C).

Fig. 5.

Identification of subtypes in SLE active patients using NetSLE biomarkers. A Subtyping analysis of SLE active patients was performed using NetSLE biomarkers across various cohorts. B A heatmap shows the disease activity (red and blue representing active and inactive, respectively) and expression levels (represented in red, black, and green from high to low) of SLE patients from the GSE49454 cohort, clustered based on the transcriptomic profiles of NetSLE biomarkers. Two distinct clusters, AS1 and AS2, enriched with SLE active patients, are underlined in black. In the heatmap, the expression values are log-transformed, median polished (row-wise and column-wise medians near zero), and normalized (row and column magnitudes near 1.0). C Correlation of differential gene expression between AS1 and AS2 for NetSLE biomarkers across pairwise cohorts. Correlation was analyzed using spearman’s rank correlation analysis. ρ, correlation coefficient; n, number of NetSLE biomarkers. D–E Functional enrichment analysis for AS1 and AS2 specific biomarkers. Bar plot represents the corrected significance (adjusted p-value) of enrichment using the hypergeometric test. The terms and bars of functions associated with both AS1 and AS2 are displayed transparently. To simplify data presentation, only Reactome pathways with more than 10 associated genes from the gene sets provided by MSigDB. F–H Box plots illustrating distributions of the patients’ immune cell proportions (neutrophils and lymphocytes) across gene set-identified clusters. Statistical significance of the differences in neutrophil and lymphocyte proportions between clusters were evaluated using the Mann–Whitney U test. FC, fold change; P, p-value; adjP, adjusted p-value; ND, not detected

We explored the molecular and cellular characteristics of each active subtype. AS1 was associated with the innate immune system and neutrophils, while AS2 was linked to interferon signaling and lymphocytes of the adaptive immune system. AS1-specific genes (n = 81), which were up-regulated in AS1 in at least two cohorts, showed significant enrichment in toll-like receptor signaling and neutrophil-related pathways (Fig. 5D; Additional file 2: Data S11A). In contrast, AS2-specific genes were significantly involved in interferon signaling pathways (Fig. 5E; Additional file 2: Data S11B). Using CIBERSORTx [52] to quantify immune cell infiltration (see the “ Methods” section), we found that AS1 had a significant predominance of neutrophils (GSE49454: Mann–Whitney U test P = 3.9 × 10⁻⁴), whereas AS2 exhibited a predominance of lymphocytes (P = 1.0 × 10⁻³) (Fig. 5F–G). Furthermore, employing an alternative methodology-based tool, BayesPrism, to estimate immune cell infiltration yielded consistent results (see the “ Methods” section). This concordance validates the robustness of the distinct subtype patterns derived from the NetSLE biomarkers (Additional file 1: Fig. S14).

We verified that NetSLE biomarkers exhibited superior capability in distinguishing SLE active subtypes compared to established gene sets, showing more pronounced differences in immune cell proportions. When compared with negative control gene sets and the 797-gene set from Banchereau et al., NetSLE biomarkers consistently identified AS1 and AS2 with significantly higher neutrophil (GSE72798: P = 5.3 × 10⁻⁶; GSE65391: P = 1.5 × 10⁻³¹) and lymphocyte (P = 5.3 × 10⁻⁶; P = 2.9 × 10⁻³¹) proportions (Fig. 5F–H; Additional file 1: Fig. S15). For other tested gene sets, subtype identification was either infeasible or yielded only marginal differences in cellular fractions. To further confirm the robustness of the identified active subtypes, we applied alternative biclustering methods (spectral co-clustering and spectral biclustering) and found that the characteristic immunological pattern—AS1 with higher neutrophil proportion and AS2 with higher lymphocyte proportion—was consistently reproduced across all three cohorts (Additional file 1: Fig. S16).

Furthermore, we discovered that immunological subtypes identified by NetSLE biomarkers were associated with distinct clinical features and laboratory findings. Within the GSE65391 cohort, AS1 exhibited significantly higher levels of aldolase (ALD), platelet count, and mean corpuscular volume (MCV), whereas AS2 showed significantly higher levels of albumin (ALB) and hemoglobin (HGB) (Additional file 1: Fig. S17A). Additionally, when examining clinical phenotypes closely associated with the primary disease pathology of SLE, we observed a distinct distribution between the subtypes: nephritis and renal function abnormalities were more frequent in AS1 patients, whereas leukopenia was more frequent in AS2 patients (Additional file 1: Fig. S17B and Fig. S18).

Discussion

While conventional DGE analysis has been widely used for biomarker identification, it frequently struggles with high false-positive rates and limited generalizability, often failing to identify true disease-modifying targets [17, 19, 53]. In this study, we effectively overcame these limitations through NetSLE, a methodology integrating a network-based approach, diverse SLE-related prior knowledge, and machine learning (Fig. 1A–D). Specifically, NetSLE successfully prioritized 150 robust and clinically useful biomarkers from 1119 SLE DEGs (Fig. 2A). By incorporating prior knowledge, such as SLE-associated cytokines [54–57] and multi-ethnic GWAS data, our approach minimized single-cohort biases. Consequently, our ML technique reduced false positives, and these biomarkers were validated through significant associations with known SLE pathogenesis (Fig. 2B–C, 2E; Additional file 1: Fig. S1, S3A–B). Notably, NetSLE also uncovered novel potential biomarkers absent from prior knowledge, such as TNFAIP6 and USP18, which have recently emerged in genomic and epigenetic studies of SLE [58–60].

The core innovation of NetSLE lies in its advanced integration strategy. Biological networks are effective for identifying disease-related genes via shared pathways [28, 61], but applying them to heterogeneous diseases like SLE is challenging. While existing methods, such as uKIN [31], also integrate network approaches with prior knowledge, NetSLE demonstrated superior performance by employing machine learning to merge diverse features from multiple prior knowledge sources into a single, unified score (P_SLE). This allowed NetSLE biomarkers to exhibit stronger associations with SLE than those identified by uKIN, maximizing biomarker prioritization (Fig. 2G–J).

In clinical practice, NetSLE biomarkers demonstrated significant translational value by identifying candidates for drug repurposing. Based on the functional relevance of the biomarkers, such as lipid metabolism and wound healing/coagulation (Additional file 1: Fig. S3A–B), we newly identified lipid-modifying agents (C10) and antithrombotic agents (B01) as potential SLE treatments (Fig. 4B). We experimentally validated these candidates, building on NetSLE’s success in predicting approved SLE drugs such as belimumab and anifrolumab (Additional file 1: Fig. S8A–B). Top-ranked C10 drugs (statins, fibrates, niacin, ezetimibe) and B01 drugs (apixaban) significantly reduced key SLE-related immune components (IFN-γ, IgG, and MPO) (Fig. 4F–H). Given the links between SLE, dyslipidemia [62–65], and NET-induced thrombosis [48, 66–69], these agents offer promising mechanistic benefits. Identifying such candidates through NetSLE fulfills a critical need for novel, accessible therapeutic options.

Although we utilized an ex vivo assay using splenocytes from MRL/lpr mice, which cannot fully capture the genetic and environmental heterogeneity of human SLE pathology, previous studies have established the validity of this model for evaluating drug-induced immunological responses and attenuating disease symptoms [70, 71]. Thus, our assay remains appropriate for assessing these therapeutic effects.

Beyond biomarker discovery, NetSLE biomarkers offer concrete clinical value in three areas of SLE care: diagnosis, treatment selection, and prognosis. For early diagnosis, the 150-gene panel predicted disease activity more accurately than the full transcriptome (n = 16,970) and outperformed previously reported signature gene sets across three independent ethnic cohorts (Fig. 3B; Additional file 1: Fig. S7A–B). This supports its use as a gene-expression-based complement to conventional serological markers such as anti-dsDNA, C3, and C4, which often change only after a clinical flare has begun. For individualized treatment, the robust separation of active patients into AS1 (innate/neutrophil-driven) and AS2 (interferon/adaptive-driven) subtypes (Fig. 5A–H) offers a clear basis for matching patients to mechanism-specific therapies–for example, neutrophil- or NET-targeting agents for AS1 and approved interferon-pathway blockers such as anifrolumab for AS2 [72]. For prognosis and disease monitoring, the same panel can be measured repeatedly over time to track changes in disease activity and subtype, helping clinicians decide when to escalate or de-escalate treatment. Just as importantly, the small size of the panel makes it practical to use in real clinical settings: 150 genes fit comfortably within the capacity of targeted RNA assays already used in routine pathology, such as NanoString nCounter or targeted RNA-seq panels [73]. The assay needs only standard whole-blood RNA, and avoids the cost and complex data analysis of whole-transcriptome sequencing [74]. In practice, NetSLE could be used alongside current SLE tests rather than replacing them, making it easier for both major medical centers and community clinics to adopt. Prospective multi-center studies in clinically well-characterized patient cohorts will still be needed to confirm assay performance, demonstrate cost-effectiveness, and define the decision thresholds required for routine clinical use. Taken together, these features make NetSLE a viable candidate for use in precision medicine for SLE.

Furthermore, while the AS1 and AS2 subtypes exhibited characteristics that closely align with those defined by previously known biomarkers, our results underscore the efficiency and clinical superiority of the NetSLE panel. AS1 exhibits similarities to the subtypes identified by Banchereau et al. [15] and Toro-Domínguez et al. [75], characterized by activated neutrophil-related functions. Similarly, AS2 corresponds to subtypes with activated lymphocyte (or plasma cell) functions and interferon signaling. However, a major limitation for real-world clinical implementation is that previous subtyping models rely on highly complex panels comprising approximately 800 or 1100 genes (800 for pediatric patients), respectively. In contrast, NetSLE achieves this critical stratification using a concise panel of only 150 genes. Notably, despite its substantially smaller size, the difference in neutrophil and lymphocyte proportions between subtypes classified by NetSLE biomarkers was even more pronounced than those distinguished by Banchereau et al.’s extensive gene set (Fig. 5F–G; Additional file 1: Fig. S15). This demonstrates that NetSLE effectively distills the core biological signals from the heterogeneity of complex disease, offering a precise and practical tool for patient stratification.

The effectiveness of any network-based approach depends on the completeness of the PPA network [76]. Despite current limitations regarding proteins with unknown biological relevance, NetSLE already outperforms conventional methods. Future updates to PPA networks [26], advancements in precision protein-interaction predictions such as AlphaFold2 [77], and improved network propagation algorithms [28] will further enhance our methodology. Additionally, while relying on initial DEGs makes the approach contingent on cohort size, posing challenges for rare diseases [28], growing collaborative transcriptome datasets [78] and advanced computational techniques for batch-effect correction [79, 80] will expand the applicability of NetSLE to a broader array of diseases.

Conclusions

In this study, we introduced NetSLE, a network-based tool designed to advance the understanding and clinical management of systemic lupus erythematosus (SLE). By integrating prior disease knowledge with network propagation and machine learning techniques, NetSLE enables the identification of robust disease-associated biomarkers and provides insights into precision medicine for SLE treatment. Our results demonstrate that NetSLE not only improves the precision of biomarker discovery but also offers a framework for drug repurposing and stratifying patients in therapeutic strategies. Overall, NetSLE represents a promising approach to support the development of personalized medicine in SLE, contributing to more accurate diagnosis, prognosis, and treatment planning for this complex and heterogeneous autoimmune disease.

Methods

Processing transcriptome data and differential gene expression (DGE) analysis

Three microarray-based transcriptome databases for systemic lupus erythematosus (SLE) were obtained from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/), specifically under accession numbers GSE65391, GSE49454, GSE72798 [Additional file 2: Data S12A–B]. These datasets provided information on gene expression levels from whole blood samples and the Systemic Lupus Erythematosus Disease Activity Index (SLEDAI) for the patients. Probe IDs from each series matrix file were mapped to gene symbols using a mapping file obtained from the Ensembl biomart [81] database (https://www.ensembl.org/index.html; May 2022). In cases where multiple probe IDs corresponded to a single gene symbol, the probe with the maximum expression value was selected, following the default setting used in Gene Set Enrichment Analysis (GSEA) [82].

Differential gene expression (DGE) analysis was performed on the GSE65391 dataset using the “limma” package [83] (version 3.54.2) from the R Bioconductor. A linear model in the package was fitted to the preprocessed expression data, accounting for experimental variables such as (1) disease activity (SLE active patients vs. inactive patients) and (2) disease status (SLE active patients vs. healthy controls). Differentially expressed genes in SLE (SLE DEGs) were identified as those up-regulated (log₂ fold change > 0 and adjusted p-value < 0.05) in SLE active patients in both comparisons. Considering further utilities such as molecular detection–based diagnosis, only up-regulated genes were used. This approach is supported by biomarker development principles, whereby markers elevated above a baseline of the control group are more readily and reliably detectable under constraints of analytical sensitivity and limits of detection [84]. SLE patients with a SLEDAI score greater than seven were classified as active, while those with a SLEDAI score less than seven were classified as inactive. This classification was based on the SLEDAI distribution across the cohorts (the mean SLEDAI for patients in the three datasets is seven, rounded to the nearest integer) (Additional file 1: Fig. S19A–C) and aligns with the classification criteria used in Banchereau et al. [15] To ensure a clear distinction between active and inactive patients, those with intermediate disease activity (SLEDAI = 7) were excluded from the analyses.

Network propagation score calculation using SLE-related prior knowledge

We applied a network propagation algorithm to identify potential associations between seed genes, selected based on prior knowledge related to SLE, and other genes within the protein–protein association (PPA) network. This algorithm iteratively diffuses information from the seed genes to their neighboring genes within the PPA network, continuing until convergence. The result is a set of propagation scores assigned to each gene in the network.

To refine these scores and prioritize genes relevant to SLE, we implemented the PageRank algorithm using the “pagerank_scipy” module in the NetworkX [85] Python package (version 2.4; https://networkx.org/). PageRank assigns importance scores to nodes within the network, allowing us to identify genes highly connected to the seed genes and may play a significant role in SLE pathogenesis.

For our PPA network, we used the STRING database [26] (https://version-11-0b.string-db.org/), specifically the file “9606.protein.links.v11.0.txt.gz” and considered all interactions within the network (interactions with confidence scores > 0).

Seed genes were selected based on four sources of SLE-related prior knowledge (Additional file 2: Data S13): (1) genes identified in genome-wide association studies (GWAS) catalog database [23] as associated with SLE (n = 505; trait label: “Systemic lupus erythematosus”; downloaded in May 2021); (2) cytokines of immune cells (e.g., B-cells, T-cells, and dendritic cells) associated with SLE, provided by the ImmuneXpresso database [86] (n = 45; downloaded in May 2021); (3) regulators of the germinal center [87] (n = 26); (4) B-cell activating factor (BAFF), a known drug target for SLE treatment (belimumab), sourced from the DrugBank database [88] (n = 1; version 5.1.9). Each of these four sets of seed genes was used independently in the propagation algorithm, resulting in four distinct propagation scores per gene in the network. For the GWAS seed genes, the statistical significance values (− log p-value) from GWAS were used as personalization values in the propagation. In cases where a gene had multiple mutations, the most significant value was selected.

Identification of the NetSLE biomarkers based on the probability of association with SLE (P_SLE)

To prioritize biomarkers from SLE differentially expressed genes (DEGs), we integrated network propagation scores derived from four sources of prior knowledge using machine learning (ML). An ML model was developed to predict the DEGs (from the GSE65391 database) using propagation scores of genes within the STRING PPA network, where the genes related to each prior knowledge source were used as seed nodes. We employed a random forest classifier from the scikit-learn [89] (version 0.24.2) Python package for the prediction, configured with 1000 estimators and a maximum depth of 3. The association probability of each gene with SLE was calculated based on the classifier’s predictions. Genes with higher probabilities were considered more likely to be associated with SLE. To assess the robustness and generalization of these probabilities, we conducted Monte Carlo cross-validation (1000 iterations) with random splits of the dataset into training (90%) and test (10%) sets. The final P_SLE values for each gene were obtained by averaging the probabilities across all cross-validations. The top 150 genes, ranked by P_SLE, were identified as the NetSLE biomarkers.

Functional enrichment analysis

To investigate the biological pathways and processes associated with NetSLE biomarkers, we performed a functional enrichment analysis using the gene sets from the Molecular Signatures Database [90] (MSigDB; version 2023.2). Specifically, we downloaded the “msigdb.v2023.2.Hs.symbols.gmt” file in Jun 2024. Enrichment of NetSLE biomarkers within these gene sets was assessed using the hypergeometric test, followed by correction for multiple testing using the Benjamini–Hochberg false discovery rate method. For this analysis, we utilized the g:GOSt (functional profiling) in g:Profiler database [91], applying the aforementioned custom gene set file and statistical procedures. The same approach was applied to analyze the AS1- and AS2-specific genes.

Calculation of enrichment using gene sets associated with SLE for validation of NetSLE biomarkers

To validate the NetSLE biomarkers, we used two categories of previously identified genes associated with SLE.

The first category included SLE-specific signature genes derived from transcriptome data. We obtained these signatures based on disease-specific gene expression alterations at the module level from the work of Chaussabel et al. [16] Additionally, we extracted a list of genes from single-cell RNA-seq data shared by Perez et al. [14], which demonstrated transcriptional changes between SLE patients and healthy individuals across various cell types. We selected genes from these lists for which P_SLE was calculated, resulting in the inclusion of 411 and 261 genes, respectively, in the analysis.

The second category consisted of genes implicated in the pathogenesis and genes associated with the diagnosis and prognosis of SLE. We sourced genes associated with SLE development from the DisGeNET [92] and MSigDB [90] databases. Specifically, a search in DisGeNET using the term “Lupus Erythematosus, Systemic” (disease ID: C0024141) yielded 1883 genes linked to SLE from all available sources, including “CURATED”, “INFERRED”, and “ANIMAL MODELS”. Additionally, we obtained 139 genes associated with the SLE pathway (“KEGG_SYSTEMIC_LUPUS_ERYTHEMATOSUS”) from the KEGG pathway via MSigDB. Lastly, we acquired 56 conventional biomarkers for SLE diagnosis and prognosis, as reported by Arriens et al. [93] and Capecchi et al. [94].

We assessed the enrichment (E), which reflects the degree of overrepresentation of SLE-associated genes in the rank positions of the P_SLE-based ranked list. To calculate E, we traversed the ranked list of SLE DEGs, determining the number of associated genes present among those DEGs with ranks above a specific threshold. The calculation of E was based on the following formula: m represents the number of SLE-associated genes among genes ranked higher than the threshold, M denotes the number of genes ranked higher than the threshold (150 for NetSLE biomarkers), n is the total number of SLE-associated genes among the SLE DEGs, and N is the total number of SLE DEGs.

$$Enrichment(E)=\frac{\frac{m\times 1}{M}}{\frac{n\times 1}{N}}$$

To evaluate the integrative enrichment across various established gene sets, we utilized an Enrichr-derived combined score [95]. This metric mathematically integrates the magnitude of gene overlap with its corresponding statistical significance to provide a robust assessment of functional enrichment. The raw combined score was calculated by multiplying the base-10 logarithm of the p-value (derived from Fisher’s exact test) by the z-score (representing the deviation from the expected rank of E). To ensure comparability across difference gene sets and datasets, the raw combined scores were further normalized using a min–max scaling approach.

To evaluate the ability of NetSLE compared to other methods, we applied three network-based approaches: uKIN, WGCNA, and NetBio, all utilizing the same STRING PPA network. In uKIN, the union of four SLE prior knowledge-related genes and up-regulated SLE DEGs was used as input prior knowledge and new information, respectively, with candidate biomarkers ranked by uKIN-derived scores. NetBio identified NetSLE biomarkers as genes involved in biological pathways proximal to SLE prior knowledge. A pathway was considered proximal to SLE prior knowledge if it showed significant enrichment (adjusted p-value < 0.01) for the top 200 genes with propagated scores derived from prior SLE knowledge. Genes in the proximal pathways were ranked by their highest score among the four propagation results from four SLE prior knowledge. In uKIN and NetBio, the top 1119 genes, matching the number of up-regulated SLE DEGs, were selected for enrichment analysis. WGCNA identified 468 genes in modules associated with active SLE patients (Additional file 1: Fig. S20). These genes were ranked by centrality, all of which were included in the enrichment analysis (due to fewer than the SLE DEGs). All parameters other than those mentioned were set to default for all methods.

Machine learning for predicting disease activity (SLEDAI)

We aimed to predict disease activity in SLE patients by analyzing transcript expression levels of specific gene sets, including NetSLE biomarkers and negative control gene sets. This prediction task involved classifying SLE patients into active and inactive groups based on their SLEDAI scores, with active patients defined as having SLEDAI > 7 and inactive patients as SLEDAI < 7.

For model training, we utilized data from the GSE65391 dataset, which includes 330 active and 591 inactive patients. Two additional datasets, GSE49454 (60 active and 96 inactive) and GSE72798 (23 active and 49 inactive), were used as test sets. We implemented seven machine learning algorithms–stochastic gradient descent (SGD, “modified_huber” was used as the loss function), random forest (The number of trees was 1000), support vector machine (SVM, “linear” was used as the kernel type), three logistic regressions with “liblinear” optimization (unpenalized, L1-penalized (Lasso), and elastic net-penalized), and gradient boosting (the number of boosting stages was 1000)—using the scikit-learn Python package (version 0.24.2).

The prediction performance of each model was evaluated using multiple metrics: area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), accuracy, Matthews correlation coefficient (MCC), and F1-score. To ensure comparability across models, we applied min–max normalization to these five performance-metrics using the “minmaxscaler” module from scikit-learn and calculated a classification score by summing the normalized values. The classification scores obtained from different gene sets and models were then compared to identify the most effective approach for predicting SLE disease activity.

We defined true positive (TP) as the number of correctly predicted SLE active patients, true negatives (TN) as the number of correctly predicted inactive patients, false positives (FP) as the number of inactive patients incorrectly predicted as active, and false negatives (FN) as the number of active patients incorrectly predicted as inactive.

$$\text{AUROC}={\int }_0^1\text{TPR}({\text{FPR}}^{-1}(x))dx$$

$$\text{AUPRC}={\int }_{-\infty }^{\infty }\text{Prec}\left(c\right)dP(Y\le c)$$

$$\text{Accuracy}=\frac{TP+TN}{TP+FN+TN+FP}$$

$$\text{MCC}=\frac{TP\times TN-FP\times FN}{\sqrt{(TP+FP)(TP+FN)(TN+FP)(TN+FN)}}$$

$$F1\text{score}=2\times \frac{\text{Precision}\times \text{Recall}}{\text{Precision}+\text{Recall}}$$

Drug, indications, and target-interaction data sets

To compile comprehensive data on drugs, including target genes, approval status across indications, clinical trial information, and Anatomical Therapeutic Chemical (ATC) classification codes, we utilized the DrugBank [88] (version 5.1.9) and ChEMBL [96] (version 34) databases. From DrugBank, we extracted target gene information for 5804 drugs, focusing exclusively on human target genes, along with their corresponding ATC classification codes [Additional file 2: Data S14A]. Drugs were then categorized based on these ATC codes. We also documented the approval status of each drug, classifying it as “approved” if it had been approved for at least one indication in clinical trials.

We compiled a list of drugs administered for SLE treatment and target genes (SLE drug targets) through data from public databases and clinician advice (Additional file 2: Data S14B). In the ChEMBL with the EFO term “systemic lupus erythematosus” (MONDO:0007915), we retrieved a catalog of drugs approved for SLE treatment. Additionally, we obtained a list of drugs commonly prescribed by clinicians at Ajou University Hospital to alleviate SLE symptoms. Drugs sourced from the ChEMBL databases were annotated as on-label, while those sourced from Ajou University Hospital were annotated as off-label.

Identification of drug classes (categorized by ATC codes) related to each gene set

To identify drug classes related to specific gene sets, we curated a set of NetSLE biomarkers, known SLE drug targets, and negative control genes, which included non-prioritized DEGs as well as all DEGs. We conducted an enrichment analysis to assess the target relationships between each gene set and the drug classes. This analysis involved comparing the genes within each gene set to the drug targets of the corresponding drug classes. The significance of the overlap between the target genes of drugs in each class and the genes within each gene set was evaluated using the odds ratio and Fisher’s exact test. A gene set-drug class relationship was considered significant if the odds ratio exceeded 1 and the p-value from Fisher’s exact test was less than 0.05.

Network proximity between NetSLE biomarkers and drug targets

Within the STRING PPA network, the distance between NetSLE biomarkers and drug targets was determined as network proximity to assess their relationship. According to the approach from Guney et al. [50], the significance of proximity was evaluated by generating a reference distance distribution based on expected distances between randomly selected groups of genes with sizes and degrees similar to those of the original drug targets and disease biomarkers in the network (1000 repetitions). In this study, if the z-score, which is the network proximity score between NetSLE biomarkers and drug target genes, is less than − 1.96, the drug is evaluated to have a significant effect on the treatment of SLE. The scripts for calculating the proximity were downloaded from https://github.com/emreg00/toolbox.

Ex vivo experimental validation of drugs

MRL/MpJ-Fas^lpr/J mice used as an SLE mouse model were obtained from Jackson Laboratories and housed in the animal facility at the POSTECH Biotech Center under specific pathogen-free conditions. The study was conducted in accordance with protocols approved by the POSTECH Institutional Animal Care and Use Committee. Mice had ad libitum access to drinking water and chow. Female mice aged 8–10 weeks were used for all ex vivo experiments. All mice were euthanized by exposure to carbon dioxide (CO₂) in a gradual-fill chamber, and spleens were harvested after euthanasia.

For the SLE model mouse splenocyte-based assay, a single-cell suspension was prepared from aseptically collected spleens by crushing and passing them through 70 µM cell strainers (Falcon). Following red blood cell lysis using ACK lysis buffer, 2 × 10⁵ splenocytes per well were seeded into 96-well tissue culture plates in 200 µL of complete RPMI medium supplemented with 10% fetal bovine serum (FBS). The cells were incubated with 0.2 µg/mL anti-CD3 (BioXcell), anti-CD28 (BioXcell), and 0.2 µg/mL anti-IgM (Invitrogen) in the presence of drugs for 72 h at 37 °C with 5% CO₂. All drugs were diluted in RPMI medium supplemented with 10% FBS. Detailed treatment concentrations of the drugs are provided in Additional file 2: Data S15.

Supernatants were collected and analyzed for IgG, IFN-γ (Invitrogen), and MPO (R&D Systems) using enzyme-linked immunosorbent assay (ELISA). Enzyme-bound substrates were developed using a TMB solution (Invitrogen, 7,855,927), and the reaction was terminated with 1 M H₂SO₄. Absorbance was measured at 450 nm, and immunoglobulin concentrations were calculated based on a standard curve generated from reference serum.

Patient subtyping based on gene expression profile similarity using clustering algorithm

To identify subtypes among SLE patients with high disease activity, we applied various clustering algorithms to gene expression profiles from SLE cohorts, utilizing NetSLE biomarkers and negative control gene sets. Patients were classified as “SLE active” if their SLEDAI score was greater than 7, while inactive patients had SLEDAI score below 7. The clustering algorithms applied included:

• i.Agglomerative clustering: hyperparameters included linkage methods (single, average, complete, and ward), and Euclidean affinity.
• ii.MiniBatch K-means clustering
• iii.Birch clustering: hyperparameters included a branching factor ranging from 10 to 100.
• iv.Spectral clustering: hyperparameters included affinity options of nearest neighbors and radial basis function (rbf).
• v.Gaussian mixture model: hyperparameters included covariance types (full, tied, diag, and spherical), and initialization parameters (k-means and random).

Furthermore, to evaluate whether simultaneous clustering of samples and features (biclustering) could offer advantages over standard sequential methods, we applied the Latent Block Model (LBM)—a model-based biclustering extension of GMM—across various configurations of row and column clusters. Additionally, to ensure that the clustering structures were robust and independent of specific model assumptions, two non-parametric biclustering methods—spectral co-clustering and spectral biclustering—were implemented and evaluated.

All algorithms were implemented using the “cluster” module in the scikit-learn Python package. The number of clusters or components is searched from 4 to 12 to identify consistent subtypes with similar characteristics across heterogeneous cohorts with varying sample sizes, ethnic compositions (Additional file 1: Fig. S12), and SLEDAI distributions (Additional file 1: Fig. S19).

The optimal clustering algorithm was determined by assessing the proportion of clusters significantly enriched with active SLE patients. This enrichment was evaluated using the chi-squared test, conducted via the “stats” module in the SciPy Python package. In cases where multiple algorithms produced identical proportions of enriched clusters, the best algorithm was selected based on combined p-values across clusters, calculated using the “harmonicmeanp” [97] R package.

Inference of immune cell proportions using CIBERSORTx and InstaPrism

To analyze the proportions of immune cells in individual SLE patients using their transcriptomic data, we utilized CIBERSORTx [52]. To this end, we utilized the LM22 signature matrix [98] and gene expression profiles of individuals as a mixture input for CIBERSORTx, running the analysis through the CIBERSORTx webpage [https://cibersortx.stanford.edu/]. Default settings of CIBERSORTx “Impute Cell Fractions” were applied, including disabling quantile normalization and no batch correction. To assess the robustness of immune cell proportion estimates, we additionally applied InstaPrism [99] (v0.1.6), a computationally efficient implementation of BayesPrism [100], as an independent deconvolution method. InstaPrism was run using default settings and reference for the deconvolution was generated using LM22 reference expression profile obtained from CIBERSORTx webpage.

In the GSE49454 and GSE65391 datasets, the proportions of neutrophils and lymphocytes were derived from available clinical information. For GSE49454, the neutrophil and lymphocyte proportions were calculated by dividing their respective counts by the total leukocyte count. The GSE72798 dataset did not provide clinical information on immune cells, so these values were estimated using CIBERSORTx. From the 22 immune cell types estimated by the LM22 signature matrix, the neutrophil proportion was used as provided, and the lymphocyte proportion was derived by combining the proportions of 12 immune cell types (B cell naïve, B cells memory, Plasma cells, T cells CD8, T cells CD4 naïve, T cells CD4 memory resting, T cells CD4 memory activated, T cells follicular helper, T cells regulatory (Tregs), T cells gamma delta, NK cells resting, and NK cells activated).

Statistical analysis and software

Data analysis and p-value generation were conducted using a variety of statistical tests, including Fisher’s exact test (two-sided), Mann–Whitney U test (two-sided), Student’s t-test (two-sided), chi-squared test (one-way), and hypergeometric test. Multiple testing correction was performed using the Benjamini–Hochberg method to control the false discovery rate (FDR). The Shapiro–Wilk test was applied to assess the normality of data distributions, guiding the choice between parametric and non-parametric tests. For correlation analysis, Spearman rank correlation was utilized. The “harmonicmeanp” method was employed to aggregate p-values. All statistical analyses were performed using Python version 3.7.9 and R version 4.2.2. Python packages utilized included pandas (v 1.3.5), numpy (v 1.18.5), scipy (v 1.6.0), scikit-learn (v 0.24.2), and statsmodels (v 0.10.1). The R package used was harmonicmeanp (v 3.0.1).

Abbreviations

ACK

Ammonium–chloride–potassium

ALB

Albumin

ALD

Aldolase

AS1

Active subtype 1

AS2

Active subtype 2

ATC

Anatomical Therapeutic Chemical

AUPRC

Area under the precision–recall curve

AUROC

Area under the receiver operating characteristic curve

CBC

Complete blood count

DEG

Differentially expressed gene

DGE

Differential gene expression

EFO

Experimental factor ontology

ELISA

Enzyme-linked immunosorbent assay

FN

False negative

FP

False positive

GB

Gradient boosting

GEO

Gene Expression Omnibus

GSEA

Gene set enrichment analysis

GWAS

Genome-wide association study

HGB

Hemoglobin

Logistic

Logistic regression

MCC

Matthews correlation coefficient

MCV

Mean corpuscular volume

ML

Machine learning

MRL/lpr

MRL/MpJ–Fas^lpr

NetSLE

Network-based machine learning model integrating prior knowledge of systemic lupus erythematosus

PPA

Protein-protein association

P_SLE

Probability of association with systemic lupus erythematosus

RF

Random forest

RNA-Seq

RNA sequencing

RPMI

Roswell Park Memorial Institute

SLEDAI

Systemic lupus erythematosus disease activity index

SLE

Systemic lupus erythematosus

SGD

Stochastic gradient descent

SVM

Support vector machine

TN

True negative

TP

True positive

Authors’ contributions

M.P., D.K., J.L., J.N., C.K., S.I., S.K., and I.K. conceived the study. M.P., D.K., J.L., J.N., and I.K. designed the methodology. M.P., D.K., J.L., Y.O., and I.K. performed the formal analysis. M.P., D.K., J.L., J.N., C.K., C.S., and J.K. conducted the investigation. M.P., D.K., and J.L. curated the data. M.P., D.K., and J.N. visualized the results. S.I., S.K., and I.K. supervised the project. M.P., D.K., J.L., and J.N. wrote the original draft. All authors read and approved the final manuscript.

Funding

This work was supported by the National Research Foundation of Korea (NRF) grants funded by the Ministry of Science and ICT (MSIT) (RS-2023–00260454 and RS-2024–00345575 to S.I.). S.K. was supported by Korean National Research Foundation (2020R1A6A1A03047902, RS-2026–25472972) and IITP (RS-2019-II191906, Artificial Intelligence Graduate School Program and IITP-2024-RS-2024–00441244).

Data availability

All data generated or analyzed during this study are included in this published article, its supplementary information files and publicly available repositories [Additional file 2: Data S16]. Transcriptome and clinical data for SLE patients were obtained from the GEO repository, specifically using expression profiles from microarray datasets under accession numbers GSE65391 [15], GSE49454 [101], and GSE72798 [102]. The human protein–protein association (PPA) network was downloaded from the STRING database [26] version 11.0 (https://version-11-0b.string-db.org/). Gene lists associated with SLE-prior knowledge were sourced from the GWAS catalog [23] (https://www.ebi.ac.uk/gwas/), ImmuneXpresso [86] (http://www.immunexpresso.org), DrugBank [88] (https://go.drugbank.com/), and curated manually based on Stebegg et al. [87]. Disease-related genes were retrieved from DisGeNET [92] (https://www.disgenet.com/), and signature gene sets were obtained from the supplementary materials of Chaussabel et al. [16] and Perez et al. [14]. Molecular signature gene sets, including those for Reactome and KEGG pathways, were acquired from the MSigDB(40) database (https://www.gsea-msigdb.org/gsea/msigdb/). Information on drug targets and classes was collected from the DrugBank [88] database. Information on clinical trials and indications of drugs was collected from the ChEMBL [96] database (https://www.ebi.ac.uk/chembl/). Biomarkers for previously identified SLE subtypes were sourced from supplementary materials in Banchereau et al. [15], and the leukocyte signature matrix (LM22) was obtained from those in Newman et al. [52]. Individual data values underlying the mouse ex vivo experiments reported in this study are available in Additional File 2 (Data S10). All datasets utilized in this study are publicly available. The source codes for reproduction of the results were developed in python 3.7.9. The scripts used in this study are available at our GitHub (https://github.com/IMB-AI/NetSLE) and Zenodo (https://doi.org/10.5281/zenodo.13939175) [103] repositories.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

M.P., J.L., S.I., and I.K. were affiliated with ImmunoBiome Inc. The remaining authors declare no competing interests.