Circulating immune cell biomarkers in Psoriatic Arthritis: associations with clinical manifestations and treatment response in the TOFA-PREDICT trial

Abstract

Background

To investigate the associations of circulating immune cell biomarkers with clinical manifestations and treatment response in Psoriatic Arthritis (PsA), and to evaluate whether these biomarkers improve prediction of individual response to tofacitinib, methotrexate, or etanercept.

Methods

Data from the TOFA-PREDICT development cohort (n = 80) were used. DMARD naïve patients were randomized to tofacitinib or methotrexate, and csDMARD inadequate responders were randomized to tofacitinib or etanercept. Treatment response (achievement of Minimal Disease Activity) was scored at week 16. Clinical data and peripheral blood mononuclear cells were analyzed at baseline and week 16. Relevant biomarkers for clinical manifestations or treatment response were selected using sparse partial least squares (sPLS) regression and extreme gradient boosting. Associations with PsA clinical manifestations were analyzed using multivariate linear regression. Biomarker profiles for treatment response were derived by sPLS regression. Lastly, a Ridge regression model was used to evaluate whether selected biomarkers improved the prediction of individual treatment response beyond a previously developed clinical model.

Results

The biomarkers explained up to 27% and 11% of observed variance for psoriasis severity and the swollen joint count, respectively. Baseline biomarker profiles classified etanercept and methotrexate responders relatively well, but did not distinguish tofacitinib responders. Adding biomarker data to clinical data did not improve the accuracy of individual response prediction (AUC-ROC 0.74).

Conclusion

Our comprehensive set of circulating immune cell biomarkers showed no additional value in precision medicine, when added to clinical predictors. However, the observed associations provide relevant insights into pathophysiological mechanisms underlying differential treatment response and clinical manifestations.

Trial registration

EudraCT Trial registration number 2017–003900-28, registration date January 25th 2018.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13075-026-03833-7.

Background

Psoriatic Arthritis (PsA) is a heterogeneous musculoskeletal inflammatory disease, with a considerable individual variation in clinical manifestations and treatment response. Clinical manifestations include peripheral arthritis, spondyloarthritis, psoriasis, enthesitis, dactylitis and nail dystrophy. Advances in the general understanding of cytokines and signalling pathways in PsA have led to several targeted therapies, such as TNF inhibitors, IL17a inhibitors and JAK inhibitors [1]. However, the underlying cellular mechanisms driving the heterogeneous clinical manifestations of PsA remain poorly understood. These mechanisms may also provide insight into the variation in response to different treatments.

Finding effective anti-inflammatory treatment in individual PsA patients can be challenging. Response rates to methotrexate range from 40–50% [2]. The efficacy of biological DMARDs (bDMARDs) is similar, and the effectiveness of second or third line bDMARDs is decreased [3, 4]. In current guidelines, treatment recommendations are partly based on PsA clinical manifestations and previously failed treatment [5–7]. However, such an approach can lead to a delay in achieving remission, risking irreversible joint damage and a reduced quality of life for patients [8, 9].

Hence, a clinical unmet need exists to predict treatment response. Building upon this, the comparison of different treatments based on individualized predicted efficacy could support the selection of personalized, effective treatment. Addressing this unmet need, the TOFA-PREDICT trial was designed to investigate whether individual treatment response in PsA patients can be predicted using clinical data, imaging data, and cellular and molecular (i.e. omics) biomarkers [10]. Within this trial, tofacitinib and methotrexate are compared in DMARD-naïve (DN) patients, and tofacitinib and etanercept are compared in patients with inadequate response to conventional synthetic DMARD (csDMARD) treatment (csDMARD failure patients; DF patients).

The present study focuses on circulating immune cell biomarkers, assessing their associations with both clinical manifestations and treatment outcomes. In a previous TOFA-PREDICT analysis, we developed a clinical prediction model based on clinical data alone (e.g. physical examination, patient history, patient-reported data) [11]. We concluded that the performance of this model needed to be enhanced in order to support real-world treatment decisions. Here, we assess whether a combination of clinical data and circulating immune cell biomarkers can improve the prediction of individual treatment response, focusing on the comparison of tofacitinib versus methotrexate or etanercept.

Methods

Patients and treatment

The TOFA-PREDICT trial (EudraCT Trial registration number 2017–003900-28, registration date January 25th 2018) is an investigator initiated, phase III, multicentre, open-label, randomized controlled trial. Patients met the ‘ClASsification for Psoriatic Arthritis’ (CASPAR) criteria and had at least two tender joints and at least two swollen joints at baseline [10]. For the present study, we used data of the development cohort (n = 80) of the trial. Half of this cohort consisted of csDMARD Naïve (DN) patients, randomized to methotrexate (n = 20) or tofacitinib (n = 20). The other half of the cohort consisted of patients with insufficient response to current csDMARD treatment (csDMARD Failure, DF patients), randomized to add-on etanercept (n = 20) or add-on tofacitinib (n = 20). Accrual of development cohort data started in April 2018 and was completed in February 2023. The full TOFA-PREDICT trial protocol, including eligibility criteria, sample size rationale, ethical approval, and trial registration has been published elsewhere [10].

Clinical data

Treatment response was defined as achievement of minimal disease activity (MDA) at week 16, using 66 Swollen Joint Count (SJC66) and 68 Tender Joint Count (TJC68), Body Surface Area (BSA) for Psoriasis involvement, the Leeds Enthesitis Index (LEI) for enthesitis, as well as the Health Assessment Questionnaire (HAQ) score, patient pain Visual Analogue Scale (VAS) score, and patient global disease activity VAS score [12]. MDA assessment was not blinded for treatment allocation. Clinical manifestation of PsA were assessed at baseline and at week 16; disease activity (represented by Disease Activity Index in Psoriatic Arthritis score; DAPSA score), joint involvement (represented by SJC66 and TJC68), skin involvement (represented by Psoriasis Area and Activity (PASI) score and BSA), dactylitis (scored as a dactylitis count), enthesitis (represented by Leeds Enthesitis Index; LEI), and nail dystrophy (scored as presence of absence) [13–15]. At week 16, physical examination parameters were missing due to COVID-19 regulations for two patients. Therefore the MDA status could not be assessed. Clinical data of the first subsequent visit were extrapolated for the missing clinical data and for calculation of MDA in these two patients. Additional potential clinical predictors of treatment response that have previously been used to develop a clinical prediction model, were added to the current analysis [11].

Circulating immune cell biomarkers

Circulating immune cell biomarkers were analyzed at baseline and week 16 in four separate panels; general immune cell subsets (panel 1), T cell cytokine production (panel 2), T cell differentiation and tissue homing marker expression (panel 3), and Regulatory T cell (Treg) phenotype (panel 4). This selection of promising phenotypical biomarkers was based on prior research, literature and expert opinion [16, 17]. More details regarding blood sample collection, sample processing, and flow cytometry analysis are provided in Supplementary information A, gating strategies are shown in Supplementary Fig. 1.

Missing data

Missing biomarker data were imputed using multiple multivariate imputation by classification and regression trees (R package mice, version 3.16.0). All clinical and biomarker variables were taken into account, as well as the predictor*treatment interactions used in development of prediction models. This created 10 imputed datasets with complete information on all variables. All presented association measures and multivariate models were pooled across the 10 imputed datasets using Rubin’s rules.

Statistical analysis

Selection of relevant variables

Both baseline and week 16 circulating immune cell biomarkers and clinical predictors were assessed. To reduce collinearity and overfitting in this relatively small cohort, the selection of relevant biomarkers followed a stepwise approach. Phenotypically closely related biomarkers (e.g. CD25hiCD127low CD4 + T cells and the Foxp3 + Treg subset) were removed based on expected clinical relevance. For remaining variables with Spearman’s ρ > 0.8, the most clinically relevant variable was retained. Subsequently, multivariate selection was performed. Sparse Partial Least Squares (sPLS) was used for continuous outcomes (most clinical manifestations), sPLS – Discriminant Analysis (sPLS-DA) was used for binary outcomes (e.g. MDA). For the development of the treatment-effect prediction model, additional eXtreme Gradient Boosting (XG-Boost) was used for selecting relevant predictors, to decrease the chance of missing relevant treatment-specific predictors. Predictors were retained if identified in ≥ 7 of 10 imputed datasets (see Supplementary Information B).

Association of circulating immune cell biomarkers with PsA manifestations

Standardized regression coefficients from univariate linear regression analysis were used to quantify the associations of the biomarkers with disease manifestations. This was done separately for cross-sectional associations at baseline and at week 16, and for longitudinal associations (changes between baseline and week 16). Subsequently, multivariate linear regression with selected biomarkers (see previous paragraph) was performed and the adjusted R-square was calculated to quantify overall associations between the biomarkers and the clinical manifestations.

Association of circulating immune cell biomarkers with treatment response

To evaluate the association between the biomarkers and treatment response, sPLS-DA analysis was used to create separate biomarker profiles for each treatment group. A profile was created with baseline biomarkers and with biomarker changes (between baseline and week 16). The Variable Importance across the Projection (VIP) score of the biomarkers were then calculated, representing the contribution of the biomarkers to the different biomarker profiles. Component scores of the sPLS-DA models were visualized and compared between responders and non-responders per treatment group using boxplots. Because the goal of this part of the analysis was to explore associations rather than to predict response, this visual evaluation of classification was deemed more appropriate than the assessment of Area Under the Curve of the Receiver Operating Characteristics (AUC-ROC) curves.

Development of a prediction model for individual treatment response

Finally we evaluated the additional predictive value of circulating immune biomarker data over a previously developed prediction model based on clinical data only [11]. First, as described above, we used sPLS-DA and XG-Boost to select relevant biomarkers and clinical predictors at baseline. The selection was done in the complete cohort, but also in the separate treatment groups to identify treatment-specific predictors of response. A Ridge logistic regression analysis utilizing nested leave-one-out cross-validation was subsequently performed with all clinical predictors and biomarkers selected by either sPLS-DA or XG-Boost in the complete cohort and in the treatment subgroups. The combination of this penalized regression method with nested cross-validation was done to reduce the risk of overfitting as much as possible. The R package glmnet (version 4.1–6) was used for Ridge regression analysis, nestedcv (version 0.7.8) was used for nested cross-validation. Three modelling strategies for individualized treatment effect prediction were evaluated; homogenous treatment effect modelling, and risk modelling and effect modelling (both assuming a heterogeneous treatment effect). More details regarding these modelling strategies are published elsewhere and are further described in Supplementary information C [11]. Penalized regression methods do not routinely provide p-values or confidence intervals. The AUC-ROC was averaged across the imputed datasets and used to evaluate the predictive accuracy. Additionally, the predicted probabilities for response were calculated for tofacitinib and for the compared treatment (methotrexate in DN patients or etanercept in DF patients) and their difference (representing the predicted treatment effect) was calculated for all patients. The range of this treatment effect indicated the variation in preferred treatment (tofacitinib or comparator treatment) and the strength of this preference. Based on AUC-ROC and the range in predicted treatment effects, the optimal prediction model was determined. Analyses were performed using SPSS version 27.0 for Windows (IBM Corporation, Armonk, NY, USA) and R version 4.2.2 (2022–10–31 ucrt) for Windows [18].

Results

Patients and treatment

Baseline clinical data are reported in Table 1, week 16 clinical data (including MDA) are reported in Supplementary Table 3. Nail dystrophy showed unexpected fluctuations: in 20% of patients either complete resolution or new onset was scored at week 16. Because there were concerns regarding the scoring reliability, nail dystrophy was excluded from further analyses.

Table 1.

Baseline characteristics

	complete cohort	methotrexate DN	tofacitinib DN	tofacitinib DF	etanercept DF
	(n = 80)	(n = 20)	(n = 20)	(n = 20)	(n = 20)
Clinical data
Demographics
age (years)	53.9 (46.3–59.2)	47.5 (44.2–57.1)	49.3 (46.5–58.1)	58.6 (50.1–63.1)	54.9 (48.4–59.0)
female	32 (40%)	12 (60%)	5 (25%)	9 (45%)	6 (30%)
PsA duration (years)	1.3 (0.1–9.9)	0.1 (0.0–0.4)	0.1 (0.1–1.1)	11.3 (4.8–18.4)	5.5 (1.3–16.7)
csDMARD at baseline
methotrexate	32 (40%)	0 (0%)	0 (0%)	17 (85%)	15 (75%)
leflunomide	5 (6.3%)	0 (0%)	0 (0%)	1 (5%)	4 (20%)
sulfasalazine	3 (3.8%)	0 (0%)	0 (0%)	2 (10%)	1 (5%)
Composite Disease Index
DAPSA score	21.2 (17.3–29)	25.3 (17.6–34.9)	20.6 (18.9–31.2)	18.4 (15.2–23.7)	22.1 (19.9–30)
Joint domain
SJC 66	4 (3–8.2)	6 (3.8–9)	3 (2.8–7)	4 (2–5)	5 (3–10)
TJC 68	6 (4–9.2)	7 (5–11)	6 (3.8–8.2)	4 (2–7.2)	6 (4.8–10.2)
Skin domain
PASI	1.2 (0.3–4.2)	1.6 (0.7–4.9)	1.2 (0.6–4.4)	0.3 (0–1.1)	2 (1.1–3.8)
BSA	1 (1–3)	2 (1–3.2)	1 (1–2.2)	1 (0–1.2)	1 (1–3.2)
Other domains
dactylitis	0 (0–0.2)	0 (0–0)	0 (0–1)	0 (0–0)	0 (0–1)
LEI	0 (0–1)	0 (0–0.2)	0 (0–0)	0 (0–1)	0 (0–0)
nail psoriasis	53 (66.2%)	13 (65%)	14 (70%)	10 (50%)	16 (80%)
Circulating immune cell biomarkers
Immune cell subsets (% of CD45 + cells)
B cells	6.4 (5.2–8.7)	5.4 (4.9–8.1)	7.2 (6.2–9.9)	7.7 (5.7–10)	5.7 (4–7.5)
T cells	61.2 (54.9–67.9)	60.5 (56.7–66)	63.8 (56.4–68)	61.8 (54.5–68.1)	59.5 (52.2–70)
classical monocytes	13 (9.6–18.1)	14.5 (11–17.7)	13 (8.7–15.9)	9.7 (8.1–17.1)	14.7 (10.8–19.2)
intermediate monocytes	1.6 (1.1–2.4)	1.6 (1.2–2.2)	1.2 (0.8–1.9)	1.7 (1.2–2.3)	2 (1.5–2.9)
non-classical monocytes	9.7 (7.2–13.7)	11.3 (9–14)	9.6 (7.4–12.9)	8.6 (7–11.6)	9.1 (5.8–14)
mDC's	0.7 (0.5–0.9)	0.6 (0.5–0.8)	0.7 (0.5–0.9)	0.6 (0.5–0.9)	0.7 (0.5–1.1)
pDC's	0.3 (0.2–0.4)	0.3 (0.2–0.4)	0.3 (0.2–0.4)	0.2 (0.2–0.3)	0.4 (0.3–0.7)

Percentages are shown for count variables, all other variables are non-normally distributed. Median and interquartile ranges reported

PsA Psoriatic Arthritis, DN DMARD naïve patient group, DF CsDMARD failure patient group, DAPSA Disease Activity in Psoriatic Arthritis, SJC Swollen joint count, TJC Tender Joint Count, PASI Psoriasis Area and Severity Index, BSA Body Surface Area, LEI Leeds Enthesitis Index, mDC's Myeloid Dendritic Cell, pDC's Plasmacytoid Dendritic Cells

Flow cytometry quality control

General immune cell subset percentages at baseline are presented in Table 1. T cell cytokine production, tissue homing marker expression by effector memory T cells, and Regulatory T cell (Treg) phenotype data at baseline and at week 16 are presented in Supplementary Table 3. Quality control confirmed stable performance of cell frequency data (see Supplementary Fig. 2 and 3). Median fluorescence intensity values were excluded from analysis because they were less stable, possibly due to increased susceptibility to technological or biological variability [19]. No batch corrections were done, to prevent loss of biologically relevant signals. Overall, there were one or more missing biomarkers in six (7.5%) patients, due to specific marker staining errors (CD16 in one patient, BDCA2 and CD123 in two patients, Foxp3 in two patients) or a general staining error (panel 2 and panel 4 in two patients).

Association between circulating immune cell biomarkers and clinical manifestations

The correlations of all circulating immune cell biomarkers and clinical manifestations are visualized in Supplementary Figs. 4 and 5. There were no strong correlations between individual biomarkers and the clinical manifestations. Highly correlated biomarkers were removed to avoid collinearity issues when selecting relevant biomarkers.

The associations between the biomarkers and the different PsA manifestations based on univariate linear regression are visualized in Fig. 1 cross-sectionally (at baseline and at week 16) and longitudinally (change between baseline and week 16). Multivariate models (Table 2) showed modest proportions of explained variance, highest for skin involvement (adjusted R² up to 0.27 for Psoriasis BSA) and joint involvement (adjusted R² up to 0.11 for SJC66). Longitudinal biomarker changes did not meaningfully explain changes in clinical manifestations.

Fig. 1.

Associations between biomarkers and PsA clinical manifestations. Heatmap of the standardized regression coefficients of univariate linear regression analyses with the biomarker as independent variable and the clinical manifestation as dependent variable. Baseline, week 16, and longitudinal data (change between baseline and week 16) are displayed separately. To improve the side-by-side comparison of the different treatment groups, all averaged VIP values were standardized by dividing them by the maximum VIP value in the model. DAPSA: Disease Activity in Psoriatic Arthritis; SJC: swollen joint count; TJC Tender Joint Count; PASI: Psoriasis Area and Severity Index; BSA: Body Surface Area; LEI: Leeds Enthesitis Index; mDC's: myeloid Dendritic Cell; pDC's: plasmacytoid Dendritic Cells; IFNγ: Interferon-gamma; TNFα: Tumour necrosis factor-alpha; Tcm cells; central memory T cells; Tem cells; effector memory T cells; Temra cells; terminally differentiated T cells; CCR: C–C Chemokine Receptor; CLA: cutaneous lymphocyte-associated antigen; Tregs: regulatory T cells; CTLA4; Cytotoxic T-Lymphocyte-Associated protein 4; ICOS: Inducible COStimulator; TIGIT: T-cell Immunoreceptor with Ig and ITIM domains

Table 2.

Selected biomarkers and adjusted R² of the PsA clinical manifestation models

	DAPSA	SJC66	TJC68	PASI	BSA	dactylitis	LEI
Baseline models
Average adjusted R2	0.05	0.04	0.04	0.08	0.13	0.01	0.06
Selected biomarkers	CD103 + CD4 + Tem cells, ICOS + Tregs, TIGIT + Tregs	TIGIT + Tregs	CD103 + CD4 + Tem cells, CCR10 + CD4 + Tem cells	CD27 + Tregs, CTLA4 + Tregs	CD27 + Tregs, CCR10 + CD4 + Tem cells	CD4 + T cells, CTLA4 + Tregs, ICOS + Tregs, TIGIT + Tregs	intermediate monocytes, IL22 + CD8 + T cells
Week 16 models
Average adjusted R2	0.07	0.11	0.06	0.12	0.27	0.03	0.07
Selected biomarkers	CLA + CD8 + Tem cells	non-classical monocytes, CLA + CD8 + Tem cells	classical monocytes, CLA + CD8 + Tem cells	IL17a + CD8 + T cells	intermediate monocytes	CCR10 + CD4 + Tem cells	mDC's, CCR10 + CD8 + Tem cells, CD27 + Tregs
Longitudinal models
Average adjusted R2	NA	NA	NA	NA	0.06	NA	0.09
Selected biomarkers	none	none	none	none	CD4 + T cells	none	CLA + CD8 + Tem cells

For each clinical phenotype a multivariate linear regression model was made for baseline data, week 16 data, and longitudinal data (changes between week 16 and baseline), incorporating the biomarkers that were selected using sPLS. The selected biomarkers are listed here, with the adjusted R² (averaged across the 10 imputed datasets) of the model

DAPSA Disease Activity in Psoriatic Arthritis, SJC Swollen joint count, TJC Tender Joint Count, PASI Psoriasis Area and Severity Index, BSA Body Surface Area, LEI Leeds Enthesitis Index, Tem cells Effector memory T cells, ICOS Inducible COStimulator, Tregs Regulatory T cells, TIGIT T-cell Immunoreceptor with Ig and ITIM domains, CCR C–C Chemokine Receptor, CTLA4 Cytotoxic T-Lymphocyte-Associated protein 4, CLA Cutaneous lymphocyte-associated antigen, mDC's Myeloid Dendritic Cell, NA Not applicable

Association of circulating immune cell biomarkers with treatment response

The biomarker differences between MDA responders versus non-responders per treatment group are shown in Fig. 2. At baseline, significant biomarker differences between responders and non-responders were identified in the etanercept group: CD4 + T cells (p < 0.01), CD8 + T cells (p < 0.01), IFNγ + CD4 + T cells (p = 0.05), IFNγ + CD8 + T cells (p < 0.01), and CD8 + terminally differentiated T cells (Temra cells, p < 0.01). For biomarker changes between week 16 and baseline, significant differences were found in the tofacitinib DF group: classical monocytes (p < 0.01) and T cells (p = 0.02).

Fig. 2.

Biomarker differences between responders and non-responders. The mean (for normally distributed biomarkers) or the median (for non-normally distributed biomarkers) of non-responders was subtracted from the mean or median of responders. Positive values (analyzed as percentages of cell populations) are displayed in green and indicate a higher mean/median in responders, negative values are displayed in orange and indicate a higher mean/median in non-responders. Paired T-tests and Mann Whitney U tests (for normally and non-normally distributed biomarkers, respectively) were performed, using FDR correction for multiple testing. Significant differences are marked with *. DN: DMARD naïve patients; DF: csDMARD failure patients; mDC's: myeloid Dendritic Cell; pDC's: plasmacytoid Dendritic Cells; IFNγ: Interferon-gamma; TNFα: Tumour necrosis factor-alpha; Tcm cells; central memory T cells; Tem cells; effector memory T cells; Temra cells; terminally differentiated T cells; CCR: C–C Chemokine Receptor; CLA: cutaneous lymphocyte-associated antigen; Tregs: regulatory T cells; CTLA4; Cytotoxic T-Lymphocyte-Associated protein 4; ICOS: Inducible COStimulator; TIGIT: T-cell Immunoreceptor with Ig and ITIM domains

To further explore differences in biomarkers between responders and non-responders, sPLS-DA analysis was used to create baseline biomarker profiles and biomarker change profiles for treatment response (Figs. 3 and 4). Using baseline biomarkers for methotrexate and etanercept, responders were clearly separated from non-responders. In contrast, tofacitinib response (both in DN and DF patients) could not be reliably classified. The biomarker change profiles showed substantial overlap between responders and non-responders in all treatment groups and were not informative.

Fig. 3.

Contribution of biomarkers to the biomarker profiles. Variable Importance in the Projection (VIP) scores for the sparse Partial Least Squares (sPLS models) of all treatment groups, indicating the contribution of each biomarker to the biomarker profiles. Distinct patterns emerge for the different treatment groups. DN: DMARD naïve patients; DF: csDMARD failure patients; mDC's: myeloid Dendritic Cell; pDC's: plasmacytoid Dendritic Cells; IFNγ: Interferon-gamma; TNFα: Tumour necrosis factor-alpha; Tcm cells; central memory T cells; Tem cells; effector memory T cells; Temra cells; terminally differentiated T cells; CCR: C–C Chemokine Receptor; CLA: cutaneous lymphocyte-associated antigen; Tregs: regulatory T cells; CTLA4; Cytotoxic T-Lymphocyte-Associated protein 4; ICOS: Inducible COStimulator; TIGIT: T-cell Immunoreceptor with Ig and ITIM domains

Fig. 4.

Biomarker profiles for the classification of response. Boxplots are shown of the component scores of the sparse Partial Least Squares (sPLS) models. The component scores are visualized for MDA responders versus non-responders per treatment group. A clear separation of responders (green boxplots) and non-responders (orange boxplots), as seen at baseline for methotrexate and etanercept, indicates a correct classification of response. All models consisted of one component in all imputed datasets, therefore the averaged score of component 1 across the 10 imputed datasets was used. DN: DMARD naïve patients; DF: csDMARD failure patients

Prediction model for individual treatment response

Six predictors were selected for the development of a prediction model: HAQ score, classical monocytes, CD4 + T cells (% of T cells), IL22 + T cells (% of CD4 + T cells), IFNγ + T cells (% of CD4 + T cells), and CLA + T cells (% of CD8 + effector memory T cells). See Supplementary Table 4 for a summary of the selected predictors per treatment group. The AUC-ROC and the range of predicted probabilities of the treatment options were optimal using a risk modelling strategy; similar to the previously developed clinical prediction model. The AUC-ROC (Supplementary Fig. 6) did not improve compared to the previously developed clinical prediction model (0.74 for current model versus 0.76 for previous clinical model). Moreover, the range of predicted differences between the compared treatments did not increase, compared to the clinical model (see Supplementary Table 5).

Discussion

To our knowledge, this is the first study to use comprehensive flow cytometry analyses to characterize circulating immune cell biomarkers and their association with clinical manifestations of PsA, as well as response to csDMARDs, bDMARDs and tsDMARDs. The analyses in this study were divided into two parts, with different aims and methods. The first part was mainly exploratory in nature, focusing on general associations between circulating immune cell biomarkers and clinical manifestations or treatment response. The methods were aimed at explaining observed clinical manifestations and treatment response within the TOFA-PREDICT development cohort itself. The second part was focused on the development of a prediction model for individual treatment response. Here, the statistical methods were aimed more strictly at the prevention of overfitting, and at developing a prediction model that can be validated and ultimately be implemented in clinical practice. Together, these analyses provided new insights into the biology of PsA. Moreover, they highlighted challenges in translating biomarker data into clinically useful prediction tools.

Overall, associations between individual biomarkers and clinical manifestations were limited. While no single biomarker correlated strongly with the clinical manifestations, multivariate regression incorporating selected biomarker subsets explained a considerable portion of the variance, particularly for psoriasis severity (PASI, BSA) and joint involvement (SJC66). At week 16 compared to baseline, the set of biomarkers changed and the explained variance was higher, possibly reflecting treatment effects and reduced disease activity over time. Interestingly, biomarkers appeared to capture skin inflammation more consistently than joint inflammation, and a distinct selection of biomarkers for the different manifestations emerged. These findings support the hypothesis that the heterogenous disease domains of PsA are driven by different drivers of inflammation and should be treated accordingly. Hence, our data provide indirect evidence for selection of treatment based on disease phenotype, as recommended by the GRAPPA [6].

With respect to treatment response, a limited set of individual circulating immune cell biomarkers were significantly different between responders and non-responders. However, the incorporation of multiple baseline biomarkers into biomarker profiles allowed classification of response to methotrexate and etanercept, but not to tofacitinib. This illustrated that the association with response was superior in (selected) biomarker profiles, compared to that of individual biomarkers. Of note, response as defined by MDA status is a complex measure that is not likely to be captured by any single biomarker. Compared to the baseline biomarkers, changes in biomarkers between week 16 and baseline did not reflect treatment response well.

To translate our findings into clinical practice, we investigated whether adding circulating immune cell biomarkers to our previously developed clinical prediction model for individual treatment response would improve accuracy of the model. However, the circulating immune cell biomarkers did not improve the predictive accuracy of the model, nor its ability to identify treatment with the highest probability of response for individual patients. Therefore, our circulating immune cell biomarkers showed no added value for personalized treatment. Despite these negative results, we are looking forward to other ongoing projects that explore the role of circulating immune cell biomarkers for precision medicine [20], because promising results have been published in the past [21, 22].

After comprehensive analysis of the circulating immune cell biomarkers, several immune cell subsets and cytokine production patterns stood out. We will focus here on pro-inflammatory cytokines (IL-17a, IL-22 and IFN-γ), Treg status, and skin homing markers. The pro-inflammatory cytokines IL-17a, IL-22 and IFN-γ were selected as relevant biomarkers for treatment response. In the baseline biomarker profiles, decreased IL-17a and IL-22 secretion were associated with methotrexate response. Previous studies have shown a decreased production of IL-17a and IL-22 after successful methotrexate treatment, but baseline associations were not reported [23, 24]. Furthermore, in the baseline biomarker profile and in the prediction model an increased secretion of IFN-y was associated with etanercept response. This aligned with previously reported correlations [24]. Overall, the exact role of pro-inflammatory cytokines as a biomarkers for methotrexate and TNFi response remains unclear.

Tregs dysregulation has been suggested in psoriatic disease, including decreased expression of inhibitory costimulatory molecules and reduced suppressive capacity [25–29]. The results from this study confirm these findings. We observed higher expression of CD27 on Tregs to be associated with reduced psoriasis severity. CD27 functions as a costimulatory receptor on Tregs and is linked to their suppressive capacity and inhibition of Th17 responses. Given that psoriatic skin disease is primarily driven by dysregulation of the IL-23/IL-17 axis, this mechanism aligns with the observed association [30, 31]. We also observed the association of increased psoriasis severity with increased CTLA-4 expression by Tregs. CTLA-4 expression by Tregs dampens the T cell response and is reported in literature to be decreased (instead of increased) in psoriasis skin lesions [32]. Our observed paradoxical decrease in CTLA-4 expression in the peripheral blood could be explained by either differences in immune cell distribution between skin and blood, an insufficient compensatory response by Tregs, or functional exhaustion. Lastly, increased T-cell Immunoreceptor with Ig and ITIM domains (TIGIT) expression by Tregs was associated with an increased arthritis joint count at baseline. A compensatory upregulation of suppressive TIGIT + Tregs has previously been reported in active PsA [17, 33].

We evaluated two skin homing biomarkers: cutaneous lymphocyte associated antigen (CLA) and CCR10. CLA⁺ effector memory T cells (Tem cells) can migrate to the skin and contribute to chronic inflammation [34, 35]. CCR10 is one of the essential chemokine receptors for CLA + skin homing T cells to enable migration to the skin [34, 35]. Regarding skin homing, we found that that methotrexate non-responders had higher proportions of CLA + CD8 + Tem cells in circulation. Additionally, CLA + CD8 + Tem cells were a predictor for non-response to csDMARD, bDMARDs and tsDMARDs. We observed significant co-expression of CLA with CCR10 expression by both CD4 + T cells (ρ = 0.674, p < 0.05) and CD8 + T cells (ρ = 0.664, p < 0.05). Aside from skin homing markers, Tem cells were not selected as predictors of response or associated with clinical manifestation of PsA, in contrast with the findings in a recent study regarding CD3 + immune cell endotypes in PsA [36].

Our results are in line with literature, which has shown impaired CCL27/CCR10 regulation and increased CLA and CCR10 expression by Tem cells in the peripheral blood in psoriasis [37–39]. In addition, peripheral blood CCR10 + Tem cells co-expressing CLA have been reported to distinguish PsA patients from those with psoriasis [16]. However, the contribution of CLA and CCR10 in arthritis development remains unclear. Overall, the presence of skin homing Tem subsets suggests inflammation driven by tissue-resident mechanisms in patients that do not respond to DMARDs. Further characterization – including pro-inflammatory cytokine production and resistance to Treg suppression – is needed to explore this CD8 + T cell subset as a future therapeutic target.

The main limitation of this study is its relatively small sample size, resulting in an increased risk of overfitting. We applied strict predictor selection and penalized regression methods with nested cross-validation to mitigate this risk, but validation in larger cohort is required. Technical variation between flow cytometry batches and different flowcytometers could have influenced the results, although consistent quality control procedures were applied. We used a uniform gating strategy to standardize the measurements and to improve reproducibility. As a result, the accuracy of naïve, central memory, effector memory and terminally differentiated T cell measurements was decreased. Finally, only a predefined set of circulating immune cell biomarkers was assessed; additional immune subsets or multi-omics integration may provide greater predictive power.

This study provides the first comprehensive flow cytometry–based characterization of immune cell profiles in PsA in relation to clinical manifestations and csDMARDs, bDMARD, and tsDMARD response. While individual biomarkers showed only modest associations with clinical manifestations and treatment response, multivariate analyses highlighted distinct immune signatures. Regarding treatment response, the role of pro-inflammatory cytokines (IL-17a, IL-22, IFN-γ), Treg dysregulation, and skin homing biomarkers stood out. Importantly, although circulating immune cell biomarkers did not improve individual prediction of treatment response, our findings add to the growing evidence that PsA is immunologically heterogeneous and that inflammatory pathways differ across disease domains.

These insights support ongoing efforts toward phenotype-driven and biomarker-informed treatment strategies in PsA. Future work within the TOFA-PREDICT trial will focus on validating these findings and expanding toward multi-omics integration (proteomics, transcriptomics, metabolomics, and imaging), with the ultimate goal of building clinically actionable prediction models for personalized treatment in PsA.

Abbreviations

PsA

Psoriatic Arthritis

DMARDs

Disease-Modifying Antirheumatic Drugs

bDMARD

Biological DMARD

csDMARD

Conventional synthetic DMARD

DN patients

CsDMARD naïve

DF patients

CsDMARD failure

CASPAR

ClASsification for Psoriatic Arthritis

MDA

Minimal Disease Activity

SJC66

66 Swollen Joint Count

TJC68

68 Tender Joint Count

BSA

Body Surface Area

LEI

Leeds Enthesitis Index

HAQ

Health Assessment Questionnaire

VAS

Visual Analogue Scale

DAPSA

Disease Activity Index in Psoriatic Arthritis score

PASI

Psoriasis Area and Activity

BSA

Body Surface Area

Treg

Regulatory T cell

sPLS

Sparse Partial Least Squares

sPLS-DA

Sparse Partial Least Squares—Discriminant Analysis

XG-Boost

EXtreme Gradient Boosting

VIP score

Variable Importance across the Projection score

AUC-ROC

Area Under the Curve of the Receiver Operating Characteristics

mDC's

Myeloid Dendritic Cells

pDC's

Plasmacytoid Dendritic Cells

Tcm cells

Central memory T cells

Tem cells

Effector memory T cells

Temra cells

Terminally differentiated T cells

CLA

Cutaneous lymphocyte-associated antigen

CTLA4

Cytotoxic T-Lymphocyte-Associated protein 4

ICOS

Inducible COStimulator

TIGIT

T-cell Immunoreceptor with Ig and ITIM domains

Authors’ contributions

FTP measured, analysed, and interpreted the data and was a major contributor in writing the manuscript. JP was a major contributor in flowcytometry methodology. DK, ANC, and CPJB helped with flowcytometry measurements. PMJW was a major contributor in statistical analysis and methodology. All authors read and approved the final manuscript.

Funding

This study was conducted as a collaboration between University Medical Center Utrecht and Pfizer. University Medical Center Utrecht is the study sponsor and the project is co-funded by the PPP Allowance made available by Health ~ Holland Top Sector Life Sciences & Health, to stimulate public–private partnerships (grant number: LSHM17074).

Data availability

Data of the TOFA-PREDICT development cohort and the analytical code are available upon reasonable request.

Declarations

Ethics approval and consent to participate

Reported in open protocol publication [10].

Consent for publication

Not applicable.

Competing interests

HEV: paid consultant for Abbvie, Novartis, Pfizer, UCB, Johnson and Johnson. Financial grants from (pharmaceutical) companies: Galapagos and Boehringer Ingelheim. AWRVK: Paid consultant for Janssen, Novartis and UCB. Financial grants from Novartis and UCB.