Integrative proteomic and machine learning strategies identify key biomarkers associated with response to JAK inhibitor in rheumatoid arthritis

Abstract

Introduction

Janus kinase inhibitors (JAKi) have revolutionized the treatment of rheumatoid arthritis (RA), yet therapeutic responses vary widely across patients. Identifying proteomic biomarkers of response may enhance personalized treatment strategies and improve clinical outcomes.

Materials and methods

We conducted a prospective study involving 42 RA patients treated with tofacitinib, baricitinib, or upadacitinib. Longitudinal serum proteomic profiling was performed at baseline and 6 months using the Olink® Explore 384 platform. Treatment response was defined by ACR/EULAR criteria. Biomarkers associated with good responders were identified through differential expression analysis and ensemble feature selection. Prediction models were developed using selected features, and feature importance was quantified using SHapley Additive exPlanations (SHAP).

Results

Pretreatment levels of IL-12B, IL-17A, and IFN-γ were significantly elevated in JAKi non-responders. IL-17 signaling was associated with response to baricitinib and tofacitinib, while upadacitinib responders exhibited broader suppression of immune-related proteins including IL-4/IL-13 pathways. Prediction models using proteomic features outperformed those based on clinical variables alone (AUC ≥ 0.90 vs. 0.52 - 0.57), with further improvement upon integrating clinical and proteomic data (AUC = 0.95). Among all features, IL-12B emerged as the most robust and consistent predictor of treatment response, supported by both differential expression and multiple machine learning approaches.

Conclusion

Serum proteomic profiling identifies elevated pretreatment IL-12B levels as a key predictor of JAKi non-response and uncovers distinct molecular responses to different JAK inhibitors. These findings support the utility of integrating proteomics with machine learning to advance personalized treatment strategies in RA.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s13075-026-03880-0.

Introduction

Rheumatoid arthritis (RA) is a chronic, systemic autoimmune disorder characterized by persistent synovial inflammation and joint destruction, resulting in high morbidity and reduced quality of life. Despite the growing array of therapeutic options, including conventional synthetic disease-modifying antirheumatic drugs (csDMARDs) and biologic DMARDs (bDMARDs), approximately 10% of patients remain refractory to multiple biologic therapies [1]. Factors such as genetic predisposition, environmental influences, and immune system heterogeneity contribute to the variability in treatment efficacy, underscoring the need for precision medicine to optimize therapeutic outcomes.

Janus kinase inhibitors (JAKi), such as tofacitinib, baricitinib, and upadacitinib, have emerged as a new class of targeted synthetic DMARDs (tsDMARDs) for RA. Janus kinases (JAKs) are non-receptor tyrosine kinases that associate with the cytoplasmic domains of various cytokine receptors [2]. Upon cytokine binding, JAKs become activated and phosphorylate signal transducers and activators of transcription (STATs), which dimerize and translocate into the nucleus to regulate gene expression [3]. This JAK-STAT signaling pathway mediates the effects of key pro-inflammatory cytokines such as interleukin-6 (IL-6) and interferons (IFNs), which are central to the pathogenesis of RA. Clinical trials and real-world studies have demonstrated the efficacy of JAKi in patients with inadequate responses to csDMARDs [4–8]. In recognition of their efficacy, the European League Against Rheumatism (EULAR) recommendations incorporated JAK inhibitors into the treatment algorithm at a level equivalent to that of biologic DMARDs (bDMARDs) [9]. In addition, JAKi offer practical advantages, including oral administration, less immunogenicity, and simplified manufacturing compared with biologic agents [10]. However, clinical responses to JAKi vary widely among patients, and the mechanisms underlying this heterogeneity remain poorly understood [11]. Identifying reliable biomarkers predictive of JAKi response could not only optimize patient stratification and guide therapeutic decisions, but also advance our understanding of the molecular mechanisms driving treatment efficacy and resistance.

Proteomic profiling offers a powerful approach to unravel molecular signatures associated with disease activity, prognosis, and drug response [12]. The Olink^® Explore 384 platform represents a high-throughput, high-sensitivity proteomics technology that enables the simultaneous quantification of up to 3,000 proteins from serum samples [13]. This platform is based on proximity extension assay (PEA) technology, which combines antibody-based detection with nucleic acid amplification, followed by quantification using next-generation sequencing (NGS) [14]. It provides a scalable system for multiplex protein detection with good specificity, low cross-reactivity, and reproducible performance. Additionally, the platform provides hazard ratios (HRs) for disease risk derived from large-scale external cohorts, such as the UK Biobank, enhancing the interpretability of protein-level findings [15]. These features make it an ideal tool for exploring the dynamic proteomic landscape of RA patients undergoing treatment [16].

In this study, we performed comprehensive serum proteomic profiling in RA patients initiating JAKi therapy. By analyzing both baseline and six-month post-treatment protein expression levels, we identified differentially expressed proteins (DEPs) associated with clinical response across the overall cohort and within individual JAKi subtypes. We further applied machine learning-based predictive modeling and feature selection to identify biomarkers of treatment response at the individual level. By comparing features derived from statistical testing and machine learning interpretability, this integrative approach aimed to uncover candidate biomarkers of JAKi response and facilitate personalized treatment strategies in RA.

Materials & methods

Study population

The prospective cohort included patients aged 20 years or older who were diagnosed with RA according to the 2010 American College of Rheumatology (ACR) / European League Against Rheumatism (EULAR) classification criteria [17]. Eligible participants demonstrated active disease despite prior treatment with csDMARDs and initiated JAKi therapy between June 2021 and June 2022. Patients were excluded if they had prior JAKi treatment, ongoing infections, pregnancy, recent major surgery within the past eight weeks, or severe comorbid conditions. The study was approved by the Institutional Review Board of Taichung Veterans General Hospital (No. CF21176A), and written informed consent was obtained from all participants prior to enrollment.

Assessment of treatment response

Baseline serum levels of rheumatoid factor and anti-citrullinated protein antibody (ACPA) were measured in all patients prior to initiating therapy. Disease activity of RA was assessed by the 28-joint disease activity score using erythrocyte sedimentation rate (DAS28-ESR) [18]. The score was classified into four categories: remission (≤ 2.6), low disease activity (> 2.6 to ≤ 3.2), moderate disease activity (> 3.2 to ≤ 5.1), and high disease activity (> 5.1) [19]. Treatment response was evaluated at 6 months using the EULAR response criteria to assess the efficacy of JAKi therapy [20]. A good response was defined as a DAS28-ESR reduction of more than 1.2 from baseline, with the patient achieving low disease activity (DAS28-ESR ≤ 3.2). A moderate response was characterized by either a DAS28-ESR improvement of > 0.6 to ≤ 1.2 with a current score of ≤ 5.1, or an improvement of > 1.2 with a current score > 3.2. Patients who did not meet the criteria for good or moderate responses were classified as poor responses [21]. In all downstream analyses, patients with good or moderate responses were grouped together and classified as responders, whereas those with poor responses were classified as non-responders.

Serum proteome profiling

Peripheral venous blood samples were collected at baseline and after 6 months of JAKi therapy. After clotting at room temperature for 30 min, samples were centrifuged at 3,000 rpm for 10 min. Serum was aliquoted and stored at − 80 °C until analysis. Serum proteomic profiling was performed using the Olink^® Explore 384 platform (Olink Proteomics, Uppsala, Sweden), which employs proximity extension assay (PEA) technology coupled with next-generation sequencing (NGS) for high-throughput, multiplexed protein quantification [14]. This approach combines antibody-based specificity with nucleic acid-based detection to achieve high sensitivity and precision. Raw protein expression data were processed using the Olink^® NPX Manager software, which converts PCR cycle threshold (Ct) values into log2-scaled Normalized Protein eXpression (NPX) units. Higher NPX values indicate greater protein abundance. Quality control procedures included the use of internal controls and reference samples provided by the manufacturer. To account for inter-plate variation, a bridging sample design was implemented, followed by batch correction. Proteins with more than 20% of values below the limit of detection were excluded from downstream analyses.

Differential expression analysis and pathway enrichment analysis

Differential expression analysis was conducted using the “limma” R package [22]. Fold change (FC) was calculated as the ratio of NPX values between groups, and P-values were adjusted for multiple testing using the Benjamini-Hochberg false discovery rate (FDR) procedure. Proteins were considered differentially expressed if they met both criteria: adjusted p < 0.05 and absolute log₂ fold change > 1.

To explore the biological functions and signaling pathways associated with DEPs, enrichment analysis was conducted using the WEB-based GEne SeT AnaLysis Toolkit (WebGestalt; https://www.webgestalt.org/) [23]. Pathway and functional annotations were performed based on the Kyoto Encyclopedia of Genes and Genomes (KEGG) and Gene Ontology (GO) databases [24, 25]. GO analysis included three categories: biological process (BP), molecular function (MF), and cellular component (CC). A p-value < 0.05 was considered statistically significant.

Calculation of proteomic risk scores

The proteomic risk score was developed to summarize each patient’s RA-related proteomic burden using externally derived RA-associated hazard ratios, rather than relying on interpretation of individual proteins. Specifically, the scores were calculated as the weighted sum of protein expression values, with each protein weighted by its log(hazard ratio) for rheumatoid arthritis. The hazard ratios were derived from a large reference cohort (UK Biobank), which includes proteomic profiles from over 50,000 individuals, and were obtained from the Olink^® platform (Supplementary Table S1) [15]. Only proteins significantly associated with rheumatoid arthritis (adjusted P < 0.05) were included. This external reference allows protein expression changes to be evaluated relative to a population-level healthy baseline, capturing both the overall direction and magnitude of disease-associated proteomic shifts during JAKi therapy.

Machine learning–based predictive modeling

The machine learning framework was implemented as an independent approach to identify robust, patient-level predictors of treatment response. Unlike conventional differential expression analysis, which evaluates average group-level differences within a single data modality, ML models enable simultaneous assessment of heterogeneous data modalities—including clinical variables, pretreatment proteomic profiles, and post-treatment changes—while capturing nonlinear relationships and complex feature interactions.

For preprocessing, variables with more than 30% missing values were excluded, and remaining missing values were imputed using medians. Continuous variables were normalized using Z-score normalization. To reduce dimensionality and minimize bias from any single method, we implemented a consensus-based feature selection strategy. Five algorithms with distinct selection mechanisms were used: decision tree (C4.5), random forest, LASSO, LASSO with repeated random seed initialization, and support vector machine (SVM). For each method, the top 12 features were selected based on importance scores. Importance scores were normalized across methods and averaged to generate a consensus ranking, from which the top 12 features were selected for downstream model development.

Prediction models were developed using Python (version 3.13.4) with scikit-learn [26], extreme gradient boosting (XGBoost) [27], light gradient boosting machine (LightGBM) [28], and Optuna [29] libraries. Seven supervised classifiers—including decision tree, logistic regression with elastic net regularization, support vector machine (SVM), random forest (RF), XGBoost, adaptive boosting (AdaBoost), and LightGBM—were trained using five-fold cross-validation with hyperparameter tuning within each fold. Evaluation metrics included AUC, F1 score, and accuracy. SHapley Additive exPlanations (SHAP) values were used to interpret model outputs and quantify each feature’s contribution to prediction, enabling biologically meaningful interpretation of multivariable relationships [30]. The mean absolute SHAP value across all samples was used to define the global importance of each feature. Features were ranked within each model based on the mean absolute SHAP values.

Statistical analysis

Statistical analyses were performed using R software (version 4.2.1). Categorical variables were compared using the Chi-square test or Fisher’s exact test, as appropriate. For continuous variables, the Wilcoxon rank-sum test was used for comparisons between independent groups, while the Wilcoxon signed-rank test was applied for paired samples. P values < 0.05 were considered statistically significant.

Results

Patient characteristics

The workflow of this study is illustrated in Fig. 1. A total of 42 RA patients who exhibited inadequate response to csDMARDs were enrolled in this study. Patient demographics were summarized in Supplementary Table S2. The cohort was predominantly female (80%), with a median age of 57 years (IQR: 47–66) and a median disease duration of 10.5 months (IQR: 3.5–15.3). At baseline, the median DAS28-ESR was 4.87 (IQR: 3.55–5.35), indicating persistent disease activity in the study population despite prior treatment with csDMARDs. Among these patients, 12 were treated with baricitinib, 14 with tofacitinib, and 16 with upadacitinib. Baseline disease activity, serologic markers and previous csDMARDs use were comparable across the three treatment groups. However, patients receiving upadacitinib had significantly higher baseline glucocorticoid doses (p = 0.001), and the use of sulfasalazine was more frequent in the baricitinib group (p = 0.011).

Fig. 1.

Schematic workflow of this study. Abbreviations: RA, rheumatoid arthritis; JAKi, Janus kinase inhibitors; R, responders; NR, non-responders; ΔProtein, protein expression changes in paired samples; LR, logistic regression with elastic net regularization; SVM, support vector machine; RF, random forest; XGBoost, extreme gradient boosting; AdaBoost, adaptive boosting; LightGBM, light gradient boosting machine

Table 1 presented the baseline characteristics of responders and non-responders. Responders had a significantly shorter disease duration (median 9.0 vs. 11.0 months, p = 0.047) and higher ACPA levels (median 358.0 vs. 23.7, p = 0.017). The use of hydroxychloroquine was also more frequent in responders (87.0% vs. 52.6%, p = 0.035). Otherwise, no significant differences were observed in age, sex, rheumatoid factor levels, inflammatory markers, disease activity, or use of other DMARDs (p > 0.05 for all).

Table 1.

Baseline characteristics between responders and non-responders

	Responders (n = 23)	Non-responders (n = 19)	P value
Age (years), median (IQR)	57.00 (49.00–66.00)	55.00 (46.50–65.50)	0.899
Female, n (%)	16 (69.6)	17 (89.5)	0.235
Baseline disease activity
DAS28, median (IQR)	5.26 (3.72–5.43)	4.70 (3.03–5.19)	0.225
Disease duration (months), median (IQR)	9.00 (1.25–11.50)	11.00 (9.50–16.00)	0.047
RF (IU/ml), median (IQR)	62.30 (27.10-125.80)	98.70 (14.00-381.85)	0.949
ACPA (U/mL), median (IQR)	358.00 (21.75–680.00)	23.70 (1.98-163.25)	0.017
ESR (mm/h), median (IQR)	23.00 (14.50–45.50)	20.00 (9.50–51.00)	0.869
CRP (mg/dL), median (IQR)	1.52 (0.24–2.64)	0.63 (0.11–2.67)	0.527
Baseline csDMARD and glucocorticoid use
Glucocorticoid (mg/day), median (IQR)	7.50 (5.00–10.00)	5.00 (0.00–10.00)	0.236
Cyclosporine, n (%)	3 (13.0)	0 (0.0)	0.302
Hydroxychloroquine, n (%)	20 (87.0)	10 (52.6)	0.035
Leflunomide, n (%)	5 (21.7)	5 (26.3)	0.999
Methotrexate, n (%)	16 (69.6)	11 (57.9)	0.644
Sulfasalazine, n (%)	7 (30.4)	4 (21.1)	0.737

RF rheumatoid factor, ACPA anti-citrullinated protein antibody, ESR erythrocyte sedimentation rate, CRP C-reactive protein, DAS28 Disease Activity Score in 28 joints, csDMARDs conventional synthetic DMARDs and IQR interquartile range

Baseline proteomic differences between responders and non-responders

To identify proteomic biomarkers associated with treatment response, we performed comprehensive serum protein profiling in RA patients prior to JAKi therapy using the Olink Explore 384 platform. Differentially expressed proteins (DEPs) between responders and non-responders (defined by |log₂ fold change| > 1 and p < 0.05) were visualized using volcano plots. In the overall cohort, three proinflammatory cytokines, IL-12B, IL-17 A, and IFN-γ, were significantly higher in non-responders compared to responders (Fig. 2A and B). After adjustment for baseline glucocorticoid dose as a covariate, these differences remained statistically significant. In the baricitinib group, 17 proteins were elevated in non-responders, including IL-12B and IFN-γ (Fig. 2C). Functional and pathway enrichment analysis showed that these proteins were associated with leukocyte differentiation and immune effector processes, and were involved in C-type lectin receptor and IL-17 signaling pathways. In the tofacitinib group, 13 DEPs were identified between responders and non-responders, with seven upregulated in non-responders (including IL-17 A) and six upregulated in responders. These proteins were enriched in pathways relevant to rheumatoid arthritis and IL-17 signaling, showing partial overlap with those observed in the baricitinib subgroup (Fig. 2C). In comparison, the upadacitinib group exhibited a distinct proteomic pattern, with only one protein (CCL7) upregulated in non-responders and five proteins upregulated in responders (Fig. 2C). The pathway of these DEPs showed minimal overlap with those from the other two groups, suggesting that upadacitinib may engage alternative immunomodulatory mechanisms. Across analyses, pretreatment DEPs annotated as inflammation-related proteins tended to show higher levels in non-responders (Supplementary Figure S1). The complete list of DEPs and their inflammation-related annotations is provided in Supplementary Table S3, and the number of DEPs contributing to each KEGG pathway is summarized in Supplementary Table S4.

Fig. 2.

Pretreatment differentially expressed proteins (DEPs) between JAK inhibitor responders and non-responders. A Heatmap of pretreatment protein expression profiles, with red indicating high and blue indicating low expression levels. B Volcano plots of DEPs between responders and non-responders in the entire cohort. C Volcano plots of DEPs between responders and non-responders in the baricitinib, tofacitinib, and upadacitinib subgroups. DEPs were defined by |log₂ fold change| > 1 and p < 0.05. Red and blue dots represent significantly upregulated DEPs in non-responders and responders, respectively. Accompanying bar plots display enriched pathways ranked by enrichment ratio; bars are shown in solid color if statistically significant (p ≤ 0.05). The complete list of DEPs is provided in Supplementary Table S3, and the number of DEPs contributing to each KEGG pathway is summarized in Supplementary Table S4

Differential protein changes between responders and non-responders

To investigate treatment-induced proteomic changes, paired serum samples collected at baseline and 6 months post-treatment were analyzed. For each sample, paired changes in protein expression (ΔProtein; post-treatment minus pretreatment) were calculated. Group-level mean ΔProtein values were then summarized for responders and non-responders (Fig. 3A). Overall, IFN-γ and GZMB tended to increase following JAKi therapy, whereas IL-6 showed a downward trend. Notably, in patients treated with upadacitinib, non-responders showed post-treatment upregulation of proteins involved in the PI3K–Akt pathways, JAK-STAT signaling, and IL-4/IL-13 signaling. In contrast, these proteins were upregulated in responders to baricitinib and tofacitinib, suggesting different modulation of immune signaling pathways across JAK inhibitors.

Fig. 3.

Paired protein expression changes (ΔProtein) between JAK inhibitor responders and non-responders. (A) Heatmap of average paired changes in protein expression for responders and non-responders, shown for the entire cohort (left) and within each JAK inhibitor subgroup (right). Red indicates increased expression after treatment; blue indicates decreased expression. Abbreviations: Bari, baricitinib; Tofa, tofacitinib; Upa, upadacitinib; R, responders; NR, non-responders. (B–E) Bar plots of the top 40 ΔProteins that differed between responders and non-responders, ranked by fold change. Results are shown for the entire cohort (B), and the baricitinib (C), tofacitinib (D), and upadacitinib (E) subgroups. Red bars represent greater increases in ΔProtein in responders; blue bars indicate greater decreases in ΔProtein in responders. Solid bars marked with asterisks denote statistical significance (p < 0.05)

To identify proteins whose paired changes differed significantly between responders and non-responders, we next performed differential expression analysis on ΔProtein values. Figures 3B-E showed the top 40 differentially expressed ΔProteins ranked by fold change. In the overall cohort, responders exhibited decreased PTX3 and increased FGF19 and IL-17 A levels following treatment; however, the observed fold changes were modest (< 2-fold) (Fig. 3B). IL-12B levels were consistently lower in responders at both pretreatment and post-treatment time points (Supplementary Figure S2), and thus the magnitude of its paired change did not differ significantly between groups and it was not among the top-ranked ΔProteins in Fig. 3B-E. Subgroup analyses revealed distinct response-associated patterns across JAKi. In patients receiving baricitinib or tofacitinib, proteins that differed significantly between responders and non-responders were predominantly associated with the IL-17 signaling pathway (Fig. 3C and D). On the contrary, upadacitinib induced a more distinct and broader proteomic response. Responders to upadacitinib exhibited significantly greater reductions in multiple immune-related proteins, particularly those involved in neutrophil degranulation and IL‑4/IL‑13 pathways, compared to non-responders (Fig. 3E).

Dynamics of proteomic risk scores between responders and non-responders

To evaluate whether treatment-induced proteomic changes translated into shifts in overall RA-related disease risk, we incorporated disease-associated HRs provided by the Olink^® platform (Supplementary Table S1) [15]. These HRs quantify the association between protein expression levels and the risk of diseases. An HR > 1 indicates that higher expression of a given protein is associated with increased disease risk. Using these data, we calculated a proteomic risk score for each patient at baseline and after six months of JAKi treatment. The distribution of these scores before and after treatment is shown in Fig. 4.

Fig. 4.

Pretreatment and post-treatment proteomic risk scores in each JAK inhibitor group. Green dots represent responders; orange dots represent non-responders. Lines connect paired samples

Among patients treated with upadacitinib, responders exhibited a greater reduction in risk scores than non-responders (mean change: -43.1 vs. +32.3; p = 0.065). This trend is consistent with more downregulated proteins observed in upadacitinib responders (Fig. 3E). In contrast, no significant difference in risk score reduction was observed between responders and non-responders treated with baricitinib or tofacitinib (p = 0.349 and 0.147, respectively). In summary, upadacitinib showed a trend toward a clinically relevant shift in the serum proteomic profile of responders compared to other JAKi. This includes suppression of proteins linked to innate immunity and systemic inflammation, which aligns with a corresponding reduction in disease-associated risk scores.

Predictive modeling of JAKi response using clinical and proteomic data

To address the limitation of the proteomic risk score, which failed to reflect therapeutic response across different JAKi, we sought to identify predictive features independent of JAKi subtype. Predictive models were developed using five data modalities: (1) clinical variables, (2) pretreatment proteins, (3) post-treatment proteins, (4) protein paired change (ΔProtein), (5) clinical variables and DEPs selected from previous statistical analyses (Table 1; Fig. 2B, Supplementary Figure S2). For modalities (1) to (4), features were selected using a consensus-based feature selection strategy across all data modalities. (See Supplementary Table S5 for the complete list of selected features.)

As shown in Fig. 5, models incorporating serum proteomic data outperformed those using clinical variables alone across all algorithms. Among the single-modality inputs, models built on pretreatment protein profiles achieved the highest predictive performance, with random forest, XGBoost, and LightGBM models yielding AUCs ≥ 0.90. Post-treatment protein profiles also provided strong predictive value (AUC range: 0.78–0.90), whereas models based on ΔProtein showed moderate performance (AUC range: 0.60–0.83). In contrast, models using only clinical variables demonstrated poor discrimination (AUC range: 0.52–0.58). Combining clinical variables with proteomic features further improved model performance, which is better than model based on DEPs selected from prior statistical analyses. The highest predictive performance was observed in models incorporating a multimodal feature set consisting of clinical variables, pretreatment proteomic profiles, and post-treatment profiles. Using this multimodal feature set, both random forest and SVM models achieved an AUC of 0.95. These findings highlight the added value of proteomic features, particularly at baseline, for improving the prediction of treatment response.

Fig. 5.

Performance metrics of various data modalities and machine learning algorithms. (A) Area under the curve (AUC), (B) F1 score, and (C) accuracy are shown. Rows represent different data modalities: clinical variables, pretreatment proteins, post-treatment proteins, protein paired change (delta), clinical + pretreatment, clinical + post-treatment, clinical + delta, clinical + pretreatment + post-treatment, clinical + pretreatment + delta, and differentially expressed proteins (DEPs) selected from prior statistical analyses. Columns represent algorithms: decision tree, logistic regression with elastic net regularization (ElasticNet), support vector machine (SVM), random forest (RF), extreme gradient boosting (XGBoost), adaptive boosting (AdaBoost), and light gradient boosting machine (LightGBM). Performance values are visualized using a color gradient, with blue indicating higher and yellow indicating lower values. The complete list of selected features is provided in Supplementary Table S5

Identification of key predictive features across models

To identify the most influential variables within the multimodal feature set, we developed predictive models using multiple machine learning algorithms to ensure robustness and reduce algorithm-specific bias. Features were ranked by their mean absolute SHAP values within each model (Fig. 6A and G). Across all models, pretreatment IL-12B consistently emerged as the top-ranked predictor of treatment response (Fig. 6H), and the partial dependence plots indicated that higher pretreatment IL-12B levels were associated with an increased likelihood of non-response (Supplementary Figure S3). While our previous differential expression analysis has identified elevated pretreatment levels of IL-12B, IL-17 A, and IFN-γ in non-responders, machine learning models further prioritized IL-12B as the most influential predictor. This convergence of statistical and machine learning approaches reinforces the role of pretreatment IL-12B as a key biomarker of JAKi response. Additional features, such as pretreatment SIGLEC1 and post-treatment IL-1R2, were frequently ranked among top SHAP predictors, despite not reaching statistical significance in group-level analyses. These findings suggest that machine learning may uncover complex, multivariate patterns not captured by conventional statistical tests.

Fig. 6.

SHAP (SHapley Additive exPlanations) summary plots for models based on the multimodal feature set. (A–G) SHAP summary plots for individual models: (A) decision tree, (B) logistic regression with elastic net regularization (Elastic Net), (C) support vector machine (SVM), (D) random forest (RF), (E) extreme gradient boosting (XGBoost), (F) adaptive boosting (AdaBoost), and (G) light gradient boosting machine (LightGBM). Features are ranked within each model by mean absolute SHAP values. Dot color represents the SHAP value for individual samples—red indicates higher SHAP values, while blue indicates lower SHAP values. (H) Overview of feature rankings across models, with larger and darker dots indicating higher importance (i.e., closer to rank one)

In summary, pretreatment IL-12B levels appear to be a robust biomarker associated with JAKi response, supported by both group-level statistical analyses and individual-level model interpretation. These findings underscore the value of integrating traditional and machine learning approaches for biomarker discovery in RA.

Discussion

Despite the clinical success of JAKi in RA, proteomic biomarkers associated with treatment response remain largely unexplored. To address this gap, we conducted longitudinal serum proteomic profiling in RA patients undergoing JAKi therapy. By integrating differential expression analysis, pathway enrichment, and interpretable machine learning, we aimed to identify protein biomarkers linked to JAKi treatment outcomes.

A key finding of this study is that elevated pretreatment IL-12B levels were consistently associated with poorer treatment response to JAKi across both traditional statistical analyses and machine learning-based approaches. Although IL-17 A and IFN-γ were also differentially expressed at baseline and selected in pretreatment-only models (Supplementary Table S5), their relative importance decreased after incorporating clinical variables and post-treatment proteomic profiles. This likely reflects their roles as downstream effector cytokines with more dynamic expression, whereas IL-12B represents an upstream cytokine axis that remains independently informative when heterogeneous data modalities were jointly considered. Whereas differential expression analysis captures group-level differences, the machine learning framework prioritizes features that retain predictive value in a multivariable, patient-level context. The concordance of these approaches supports IL-12B as a robust biomarker of JAKi response.

IL‑12B encodes the p40 subunit shared by IL‑12 and IL‑23, which are essential for Th1 and Th17 differentiation, respectively [31]. Both cytokines signal through JAK2, providing a mechanistic rationale for linking IL‑12B expression to the efficacy of JAK-targeted therapies. Importantly, the observed association was consistent across therapies targeting JAK1, JAK2, or both, suggesting that IL‑12B may mark a broader cytokine-driven inflammatory milieu rather than a pathway-specific effect [8, 32]. Elevated IL‑12B levels have been shown to correlate with RA disease activity and frequently co-occur with other pro-inflammatory cytokines, such as IL‑6 and IFN‑γ, many of which also signal via JAK1 [33]. These findings support the interpretation that high IL‑12B expression reflects an amplified inflammatory state that may be less responsive to JAK inhibition alone, thus explaining its association with treatment non-response in our cohort. Several studies have explored biomarkers of JAKi response across different molecular layers. A recent RNA-seq study reported enrichment of IL‑12-related pathways in JAKi responders based on treatment-induced transcriptional changes in PBMCs [34]. In comparison, our study focuses on baseline circulating IL-12B protein levels, which were higher in non-responders and may reflect a pre-existing inflammatory state less amenable to JAK inhibition. Together, these observations suggest that the IL-12 axis may play important roles at different stages of JAKi response. Other biomarkers highlight complementary aspects of immune regulation. Zhao et al. identified circulating exosomal miR‑27a‑5p as a non-invasive biomarker of tofacitinib response [35]. This miRNA regulates the Th1/Th2 balance and can target IL-12 A, potentially modulating IL‑12 signaling at the post-transcriptional level. Markers of innate immune activity have also been reported: circulating levels of soluble urokinase Plasminogen Activator Receptor (suPAR), IL‑6, and calprotectin were found to differ between responders and non-responders to JAKi [36–38]; and increased synovial neovascularization has been associated with poor JAKi response [39]. While these largely reflect innate immunity or vascular mechanisms, IL‑12B provides a different insight into cytokine-regulated adaptive immunity, potentially enabling more pathway-specific interpretation.

In our prediction modeling, serum proteomic data consistently outperformed clinical variables (AUC ~ 0.8 vs. 0.5), with pretreatment data proving more predictive than post-treatment or delta measurements. The multimodal model combining clinical, pretreatment, and post-treatment data achieved the highest AUC (~ 0.9), in line with prior experience in anti-TNF studies [40, 41]. Notably, models based solely on pretreatment proteomics performed comparably, emphasizing the dominant predictive value of baseline serum profiles. To avoid multicollinearity, we did not include all three proteomic inputs in a single model, as delta values are linearly dependent on baseline and post-treatment measurements.

Our longitudinal proteomic analysis revealed both shared and drug-specific effects of JAK inhibition. A downward trend in IL-6 levels was observed across treatment groups; however, the difference between responders and non-responders did not reach statistical significance. This observation is consistent with prior reports showing that IL‑6 levels may decline following JAKi therapy irrespective of clinical response [42]. Another notable observation involves the IL-17 family. As shown in Fig. 3A, the direction of IL-17 A change was generally consistent across different subgroups, with responders exhibiting greater increases than non-responders, whereas IL-17 C showed more variable patterns. This divergence likely reflects differential regulation of immune-derived versus tissue-derived IL-17 responses under JAK inhibition. Importantly, each JAKi was associated with distinct proteomic response patterns. In patients treated with baricitinib (JAK1/2 inhibitor) or tofacitinib (JAK1/2/3 inhibitor), DEPs between responders and non-responders were predominantly involved in the IL‑17 signaling pathway. Both baricitinib and tofacitinib have been shown to inhibit IL‑23-induced STAT3 phosphorylation via JAK2 blockade, thereby suppressing Th17-driven inflammation [43, 44]. In contrast, upadacitinib (selective JAK1 inhibitor) elicited a distinct proteomic profile characterized by downregulation of proteins associated with JAK-STAT and IL‑4/IL‑13 pathways in responders, which are key mediators of Th2-driven inflammation. This is consistent with the greater inhibitory potency of upadacitinib on Th2-related pathways compared to baricitinib and tofacitinib [45]. As shown in Figs. 3E and 4, responders to upadacitinib exhibited more downregulated proteins as well as more pronounced reduction in proteomic risk scores. As the HRs used to construct the risk score were derived from external cohorts, the score should be interpreted as a summary measure of proteomic changes rather than an indicator of long-term outcomes. In this context, the observed reduction likely reflects a shift toward a lower disease-associated molecular state. This aligns with the concept of molecular remission, as proposed by Tasaki et al., which suggested that effective RA treatment can remodel patient proteomes to resemble those of healthy individuals [46]. Interestingly, this distinct molecular profile aligns with recent real-world evidence demonstrating superior drug persistence for upadacitinib compared to other JAK inhibitors [47]. We hypothesize that the deep suppression of inflammatory pathways observed in our study provides a molecular rationale for the sustained clinical durability reported in these epidemiological cohorts. These findings suggest that the depth of proteomic response may serve as a molecular surrogate for long-term drug survival and warrant further investigation.

In summary, this study presents the first comprehensive serum proteomic profiling of RA patients receiving JAKi therapy. However, several limitations should be acknowledged. First, the modest sample size limits statistical power to detect small effect sizes or support definitive subgroup analyses; accordingly, our analyses are primarily powered to identify proteins with strong associations with treatment response. The findings should be interpreted as hypothesis-generating and require validation in larger, independent cohorts. Second, as our primary aim was exploratory biomarker discovery rather than predictive model deployment, an external test set was not allocated; instead, cross-validation was used to assess the robustness of identified markers. Third, the absence of a non-JAKi control group limits our ability to attribute observed proteomic changes specifically to JAKi effects. Future studies incorporating larger cohorts, appropriate control groups, and functional validation are warranted to confirm and translate these findings.

Conclusions

This study identifies elevated IL-12B levels as a robust biomarker of JAK inhibitor non-response, supported by both statistical and machine learning analyses. Distinct pathway modulations, such as IL‑17 signaling in baricitinib/tofacitinib responders and IL‑4/IL‑13 in upadacitinib responders, offer mechanistic insights that could inform personalized therapeutic strategies. These findings warrant validation in larger cohorts and further exploration of their clinical applicability.

Authors’ contributions

Conceptualization: T-SW, Y-MC; Methodology: T-SW, K-YC, H-YC; Data Acquisition: Y-TC, C-MK, Y-JC; Investigation, Formal analysis: S-FC, K-YC; Resources: T-SW, H-YC, Y-MC; Software: K-YC, H-YC; Data Curation: S-FC, K-YC; Visualization: S-FC; Writing–original draft: S-FC, T-SW; Writing–review & editing: H-YC, Y-MC.

Funding

This study was funded by the National Science and Technology Council, Taiwan [NSTC-111-2634-F-A49-014, NSTC-111-2218-E-039-001, NSTC-111-2314-B-075 A-003-MY3, and NSTC- 113-2314-B-040-003], Taichung Veterans General Hospital, Taiwan [TCVGH-1127301 C, TCVGH-1127302D, and TCVGH-YM1120110], Chung Shan Medical University, Taiwan [RD11216]. NSTC-114-2314-B-040-025.

Data availability

The de-identified data supporting the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

The study procedures were approved by the Joint Institutional Review Board of Taichung Veterans General Hospital, Taiwan (CF21176A) and written informed consent was obtained from all participants. The study was carried out in accordance with the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.