CSF Cytokine Network Organization Predicts Progression Independent of Relapse and MRI Activity in Multiple Sclerosis

ABSTRACT

Objective

Progression independent of relapse activity is a major determinant of long‐term disability in multiple sclerosis, but its immunopathologic basis remains incompletely understood. We investigated whether relapse‐independent progression in radiologically stable relapsing–remitting multiple sclerosis is associated with distinct cerebrospinal fluid inflammatory profiles and whether cytokine‐network features provide information beyond single‐mediator levels.

Methods

In this prospective observational cohort study, baseline cerebrospinal fluid cytokine and chemokine profiling was performed in 346 RRMS patients, 61 primary progressive multiple sclerosis (PPMS) patients, and 196 neurological controls without inflammatory or neurodegenerative central nervous system disease. After 2‐year clinical and MRI follow‐up, 37 RRMS patients with clinical relapse and/or MRI activity were excluded, leaving 309 relapse‐free and MRI‐stable RRMS patients classified as stable RRMS (n = 241) or RRMS with PIRMA (n = 68). Cytokines were analyzed using two‐part hurdle models, cytokine interactomes, and machine‐learning classifiers.

Results

Quantitative cytokine analyses distinguished multiple sclerosis from controls but showed substantial overlap across multiple sclerosis phenotypes. RRMS with PIRMA showed higher cerebrospinal fluid levels of tumor necrosis factor‐α, interleukin‐17, and RANTES than stable RRMS. Network analyses showed a more integrated cytokine interactome in RRMS with PIRMA. The network‐level model best distinguished RRMS with PIRMA from stable RRMS, with an area under the curve of 0.86, sensitivity of 77%, and specificity of 76%.

Interpretation

Relapse‐independent progression in multiple sclerosis is associated with selective cytokine changes and broader reorganization of inflammatory network architecture, supporting cerebrospinal fluid cytokine interactome profiling as a translational approach to immune stratification.

1. Introduction

Multiple sclerosis (MS) is a chronic neuroinflammatory disease of the central nervous system (CNS), characterized by focal demyelinating lesions and neurodegeneration as key pathological features [1, 2]. Over the last decade, the progressive introduction of highly effective disease‐modifying treatments (DMTs) has substantially improved the control of acute clinical and radiologic disease activity. As a result, the major unmet clinical need has increasingly shifted from relapse control to the identification and prevention of long‐term disability accrual.

In this setting, progression independent of relapse activity (PIRA) [3] is increasingly recognized as a major driver of long‐term outcomes, emerging in the early stages of relapsing–remitting MS (RRMS) and in patients who appear clinically stable [4]. In recent literature, increasing attention has been given to progression independent of relapse and MRI activity (PIRMA), a measure that identifies patients who are clinically and radiologically stable. PIRA has been linked to chronic compartmentalized inflammation, in which sustained immune and glial activation may drive inflammation‐mediated synaptic damage and excitotoxic injury [5, 6], ultimately contributing to gray matter damage and cortical atrophy [7]. These processes have been increasingly characterized by advanced magnetic resonance imaging (MRI) techniques and supported by histopathologic evidence [8].

Recent evidence also suggests that cerebrospinal fluid (CSF) analytes may serve as biomarkers of PIRA early in the disease course [9, 10]. Indeed, previous studies have linked pro‐inflammatory and anti‐inflammatory cytokines with cortical damage [5, 11] and disability progression [8], supporting their potential role as biomarkers of PIRA. However, neuroinflammation is intrinsically complex, reflecting the activation of multiple and partially overlapping immune pathways within the CNS [8]. Cytokines do not act as isolated mediators, but as components of a dynamic inflammatory milieu characterized by redundancy, pleiotropy, and context‐dependent effects [12].

Accordingly, the clinical relevance of CSF inflammatory mediators may depend not only on their absolute concentrations but also on their functional organization [12, 13]. Advanced analytical approaches, including graph theory and machine learning‐based methods, may therefore help capture these complex inflammatory patterns and assess their potential clinical utility for phenotypic discrimination and early risk stratification in MS [14].

The aims of this study were: (i) to compare CSF inflammatory mediator levels across noninflammatory controls (NIC), radiologically stable RRMS, and PPMS; (ii) to determine whether PIRMA in relapse‐free and radiologically stable RRMS is associated with distinct quantitative or network‐level inflammatory features; and (iii) to explore the potential clinical utility of cytokine‐network–derived features for phenotypic discrimination and early risk stratification.

2. Methods

2.1. Study Design and Ethics

This was a prospective observational cohort study integrating baseline CSF profiling with 2‐year longitudinal clinical and MRI follow‐up in MS patients. The study was conducted at the Neurology Unit of IRCCS Neuromed (Pozzilli, Italy). Patients were prospectively enrolled between 2018 and 2022 and followed for 2 years; the study database was closed in 2024. The study was approved by the Ethics Committee of the Neuromed Research Institute (cod. 06–17) in accordance with the Declaration of Helsinki, and all participants provided written informed consent.

2.2. Participants, Baseline Assessment, and Follow‐Up

The study included three groups: (1) relapsing–remitting MS (RRMS) patients, (2) primary progressive MS (PPMS) patients at the time of diagnosis, and (3) NIC.

At baseline (time of diagnosis), lumbar puncture (LP) was performed as part of routine diagnostic procedures, and CSF was collected for subsequent cytokine/chemokine profiling. Baseline clinical characteristics collected at diagnosis included age, sex, MS duration, Expanded Disability Status Scale (EDSS), and CSF oligoclonal band (OCB) status. MS patients underwent prospective clinical follow‐up with EDSS assessment every 6 months during the 2‐year observation period. Disability worsening was considered confirmed when it persisted at the subsequent 6‐month follow‐up visit. For descriptive purposes, only baseline and 24‐month EDSS values are reported in the present article. Clinical and radiological follow‐up data were prospectively collected over the 2‐year study period. In RRMS, follow‐up data were used to define the final analytic cohort and to classify patients according to PIRA status. In PPMS, follow‐up data were collected and reported for descriptive purposes only, whereas PPMS patients were primarily used as a baseline comparator group for cross‐sectional CSF analyses. During follow‐up, exposure to DMT was recorded and categorized as high‐ or low‐efficacy. High‐efficacy DMTs included anti‐CD20 therapies, natalizumab, sphingosine‐1‐phosphate‐receptor modulators, and cladribine; lower efficacy DMTs included teriflunomide, dimethyl fumarate, interferons, and glatiramer acetate. NIC were assessed only at baseline and did not undergo longitudinal follow‐up in this study.

2.3. Inclusion and Exclusion Criteria

For MS patients, inclusion criteria were: fulfillment of the 2017 McDonald diagnostic criteria15; age 18–55 years; absence of relevant comorbidities; no treatment with disease‐modifying therapies (DMT) before baseline assessment; and no corticosteroid treatment within 1 month prior to baseline lumbar puncture (LP). At baseline, both RRMS and PPMS patients were required to show no gadolinium‐enhancing lesions on MRI obtained at the time of LP and baseline evaluation. For RRMS patients, an additional criterion was a minimum interval of 30 days between the most recent clinical relapse and LP. A relapse was defined as the emergence of a new neurological symptom or worsening of a pre‐existing symptom lasting ≥ 24 h in the absence of fever or infection. MS duration was defined as the interval between the first MS‐related clinical symptom and baseline assessment. Patients were excluded from the analytic RRMS cohort if they showed evidence of disease activity during the 2‐year follow‐up, defined by clinical relapse and/or radiological activity.

2.4. MRI Follow‐Up and Definition of Radiological Stability

Brain and spinal cord MRI scans were obtained at baseline and at 12 and 24 months. Radiological activity was defined by the presence of gadolinium‐enhancing lesions and/or new or enlarging T2 lesions during follow‐up. Additional details are provided in Appendix S1.

2.5. PIRMA Definition

The present analysis applies a current operational definition of PIRMA to prospectively collected longitudinal data. Within this prospectively followed and radiologically stable RRMS cohort, PIRMA was defined as confirmed disability accumulation (CDA) occurring over the 2‐year interval in the absence of both clinical relapses and radiological activity [3, 4, 15, 16]. CDA was defined as an increase in EDSS score of 1.5, 1.0, or 0.5 points if baseline EDSS was, respectively, 0, 1.0–5.0, or > 5.0. Patients with RRMS underwent clinical evaluation every 6 months, and worsening of the EDSS score was considered confirmed if it persisted at the next follow‐up visit, conducted 6 months later.

2.6. CSF Collection and Cytokine/Chemokine Quantification

CSF samples were obtained by lumbar puncture at diagnosis, at least 30 days after the most recent clinical relapse, and stored at −80°C until analysis. A Bio‐Plex multiplex assay (Bio‐Rad Laboratories, Hercules, CA, USA) was used to quantify a panel of 27 CSF inflammatory mediators, including pro‐ and anti‐inflammatory cytokines, chemokines, and growth factors. Detectability was defined relative to the assay's lower limit of detection (LLOD), and values below this threshold were treated as non‐detectable. Additional methodological details and the full list of analytes are provided in the Supporting Informations.

2.7. Two‐Part (Hurdle) Models for Between‐Group Differences in CSF Inflammatory Mediators

Because CSF cytokine concentrations showed semi‐continuous distributions, including non‐detectable values, each analyte was analyzed using a two‐part (hurdle) modeling framework [17]. In the first component, detectability was modeled using multivariable logistic regression. In the second component, analyses were restricted to subjects with detectable values only. Positive cytokine concentrations were log10‐transformed and modeled using multivariable linear regression. For both components, age, sex, MS duration, and DMT (categorized as high‐ or low‐efficacy) were considered as covariates.

For each component, the overall effect of the group was assessed by likelihood ratio testing, comparing the full model including group and covariates with a reduced model including covariates only. To obtain a single global test for each cytokine, we also computed a joint likelihood‐ratio test combining the two components of the model. Moreover, significant overall group effects were further explored using pairwise post hoc contrasts between all group pairs.

All two‐part models, likelihood ratio tests, and post hoc contrasts were implemented in MATLAB 2025b. Statistical significance was set at p ≤ 0.05. p‐values were corrected for multiple testing using the false discovery rate (FDR).

An a priori power analysis was performed using G*Power 3.1 [18] to approximate the two‐part model. The continuous component (cytokine amount) was approximated using an ANCOVA with a four‐level between‐subjects factor and relevant covariates, assuming a small‐to‐moderate effect size (Cohen's f = 0.15), α = 0.05, and power = 0.80. The binary component (detectability) was approximated using a four‐group comparison of proportions, assuming a small‐to‐moderate effect size (Cohen's w = 0.15) with the same α and power. Under these assumptions, the estimated required total sample size was approximately 488 participants. The final study sample (n = 566) exceeded this threshold, supporting adequate power to detect small‐to‐moderate group effects across both components of the two‐part model.

2.8. Cytokine Interactome Analysis

To investigate the network organization of inflammatory mediators, we constructed group‐specific cytokine interactomes based on pairwise partial Spearman correlations adjusted for age, sex, MS duration, and DMT. Overall similarity between group‐specific correlation matrices was assessed using Pearson's correlation coefficient. Between‐group differences at the edge level were tested using a permutation‐based differential correlation approach with 5000 permutations and FDR correction. For graph‐theoretical analyses, correlation matrices were thresholded at |ρ| ≥ 0.30 and represented as weighted, undirected, signed networks, corresponding to moderate correlation strength, to retain robust cytokine–cytokine associations while improving network interpretability. Node degree was used to assess topological centrality.

Modular structure was identified using the spin‐glass community detection algorithm [19], and module preservation across phenotypes was assessed using the NIC network as the reference partition. Preservation was quantified using the dominant fraction, the best‐match Jaccard index, and within‐module consensus, and cross‐group changes in module membership were visualized using alluvial plots. Correlation, permutation, and network analyses were implemented in MATLAB 2025b and R. A more detailed description of the interactome analysis is provided in the Supporting Informations.

2.9. Machine‐Learning Analysis

To explore the discriminatory ability of CSF inflammatory profiles across clinical phenotypes, we performed pairwise supervised classification analyses among the four study groups using support vector machine (SVM) classifiers with a polynomial kernel [20]. Three models were tested: a cytokine‐level model based on CSF cytokine concentrations, a network‐level model based on features derived from differential cytokine networks, and a combined model integrating both feature sets.

For the network‐level model, only cytokine pairs showing significant between‐group differences in adjusted partial Spearman correlation were retained, and for each retained edge, an interaction feature was computed at the subject level as the standardized product of the two cytokine values. Binary classifiers were trained separately for each pairwise group comparison, and performance was evaluated using nested leave‐one‐out cross‐validation. Classification performance was summarized using confusion matrix‐derived metrics and the area (AUC) under the receiver operating characteristic (ROC) curve. Analyses were implemented in MATLAB 2025b. Further methodological details are provided in the Supporting Informations.

3. Results

3.1. Study Populations

A total of 346 RRMS patients were prospectively enrolled at baseline. Of these, 37 were excluded from the present restricted analysis because they showed clinical and/or radiological disease activity during the 2‐year follow‐up. The final analytic RRMS cohort, therefore, comprised 309 radiologically stable patients. Among them, 241 were classified as stable RRMS and 68 as RRMS with PIRMA. Study subgroup selection is summarized in Figure 1. Additional participants included 61 PPMS patients evaluated at diagnosis and 196 NIC. Longitudinal follow‐up was used to classify RRMS patients according to PIRA status, whereas follow‐up data in PPMS were collected solely for descriptive purposes. Follow‐up duration was 24.1 ± 0.8 months in the whole RRMS cohort, with a range of 23–26 months. In stable RRMS patients, follow‐up duration was 24.0 ± 0.8 months (range 23–26), whereas in RRMS patients with PIRMA it was 24.4 ± 0.7 months (range 23–26).

FIGURE 1.

Flowchart of RRMS patient selection and subgroup classification. A total of 346 RRMS patients were prospectively enrolled at baseline and underwent longitudinal clinical and MRI follow‐up. During follow‐up, 37 patients were excluded from the restricted analytic cohort because they showed clinical relapse and/or MRI activity. The final analytic cohort included 309 relapse‐free and MRI‐stable RRMS patients. Among them, 241 remained stable without confirmed disability accumulation and were classified as stable RRMS, whereas 68 developed confirmed disability accumulation in the absence of both clinical relapse and MRI activity and were classified as RRMS with PIRMA. MRI, magnetic resonance imaging; PIRMA, progression independent of relapse and MRI activity; RRMS, relapsing–remitting multiple sclerosis.

Demographic and clinical data of RRMS, PPMS, and NIC patients are summarized in Table 1. Among NIC (n = 196), the indication for LP was a neurological condition without inflammatory or neurodegenerative CNS disease. Diagnoses included neuropathies (n = 89, 45.4%), functional neurological disorders (n = 39, 19.9%), migraine (n = 48, 24.5%), and benign idiopathic intracranial hypertension (n = 20, 10.2%).

TABLE 1.

Demographics and clinical data at lumbar puncture (LP) by diagnostic group.

Characteristic	RRMS (n = 309)	PPMS (n = 61)	NIC (n = 196)	p‐value
				Global	RRMS vs. PPMS	RRMS vs. NIC	PPMS vs. NIC
Age, years	35.9 ± 11.2; 34.0 (26.3–45.3)	53.0 ± 11.0; 54.0 (45.8–60.3)	48.0 ± 14.9; 48.2 (39.4–58.6)	< 0.001	< 0.001	< 0.001	0.02
Female sex, n (%)	198 (64.1)	24 (39.3)	116 (59.2)	0.002	< 0.001	0.81	0.02
MS duration, months	24.3 ± 46.8; 5.2 (1.6–24.3) (n = 304)	56.7 ± 71.1; 27.6 (12.2–88.9) (n = 61)	—	< 0.001*	—	—	—
EDSS at LP	1.8 ± 1.1; 2.0 (1.0–2.5) (n = 293)	4.4 ± 1.4; 4.5 (3.5–5.0) (n = 61)	—	< 0.001*	—	—	—
EDSS 2 years after LP	1.6 ± 1.1; 1.0 (1.0–2.0) (n = 298)	5.2 ± 1.3; 5.5 (4.5–6.0) (n = 50)	—	< 0.001*	—	—	—

Note: Age and sex are reported for relapsing–remitting multiple sclerosis (RRMS), primary progressive multiple sclerosis (PPMS), and non‐inflammatory, non‐neurodegenerative controls (NIC). MS duration, Expanded Disability Status Scale (EDSS) at lumbar puncture (LP), and EDSS 2 years after LP are reported for RRMS and PPMS only. Continuous variables are presented as mean ± SD and median (IQR), and categorical variables as n (%). Between‐group comparisons for age were performed across RRMS, PPMS, and NIC using one‐way ANOVA, followed by Tukey's post hoc test. Between‐group comparisons for sex were performed using the χ ² test, followed by pairwise χ ² tests with Bonferroni correction. Comparisons for MS duration, EDSS at LP, and EDSS 2 years after LP were performed between RRMS and PPMS only. Missing data were present for MS duration in 5 RRMS patients, for EDSS at LP in 16 RRMS patients, and for EDSS 2 years after LP in 11 RRMS and 11 PPMS patients.

^*Statistically significant between‐group differences.

Within the prospectively followed relapse‐free and radiologically stable RRMS cohort, 241 patients remained stable without confirmed disability accumulation and were classified as stable RRMS, and 68 developed confirmed disability accumulation in the absence of both clinical relapses and MRI activity and were classified as having PIRMA over 2 years of follow‐up. Demographic and clinical characteristics of stable RRMS patients and patients with PIRMA status are summarized in Table 2. Longitudinal clinical and MRI data, including relapse occurrence, MRI activity, and PIRMA classification, were available for all patients in the restricted analytic RRMS cohort. Missing values for descriptive baseline/follow‐up variables are reported in Tables 1 and 2.

TABLE 2.

RRMS characteristics by PIRA status at 2 years.

Characteristic	stable RRMS (n = 241)	RRMS with PIRMA (n = 68)	p‐value
Age, years	34.6 ± 11.0; 32.3 (25.8–42.5)	40.6 ± 10.8; 42.6 (32.6–47.5)	< 0.001*
Female sex, N (%)	160 (66.4)	38 (55.9)	0.147
MS duration, months a	22.7 ± 46.6; 4.2 (1.3–20.0) (n = 238)	31.0 ± 47.3; 12.4 (3.4–34.2) (n = 65)	0.006*
Time from last relapse to LP, days b	309.1 ± 610.5; 100.0 (55.0–282.0) (n = 163)	248.9 ± 268.1; 145.0 (64.0–353.0) (n = 51)	0.331
Ocb positive, N (%)	188/217 (86.6)	52/63 (82.5)	0.417
EDSS at LP	1.9 ± 1.1; 2.0 (1.0–2.5)	1.4 ± 1.0; 1.5 (1.0–2.0)	< 0.001*
EDSS at 2 years	1.4 ± 0.9; 1.0 (1.0–2.0)	2.4 ± 1.3; 2.0 (1.5–3.0)	< 0.001*
Follow‐up DMT efficacy class, N (%)			0.501
Low‐efficacy DMT	161 (66.8)	49 (72.1)	/
High‐efficacy DMT	80 (33.2)	19 (27.9)	/

Note: RRMS patients were divided into a subgroup that remained relapse‐free and radiologically stable without progression over 2 years (stable RRMS) and a subgroup presenting progression independent of relapse activity (RRMS with PIRMA) during follow‐up. Age, sex, oligoclonal band (OCB) status, EDSS at LP and at 2 years, MS duration, time from last relapse to LP, and follow‐up DMT efficacy class are reported. Continuous variables are presented as mean ± SD and median (IQR), and categorical variables as n (%). Comparisons between stable RRMS and RRMS with PIRMA were performed using the Mann–Whitney U test for continuous variables and χ ² or Fisher's exact test for categorical variables, as appropriate. Missing data were present for MS duration in 3 stable RRMS and 3 RRMS with PIRMA patients, for time from last relapse to LP in 78 stable RRMS and 17 RRMS with PIRMA patients, and for OCB status in 24 stable RRMS and 5 RRMS with PIRMA patients.

^a MS duration was calculated as the interval between the date of lumbar puncture (LP) and the date of symptom onset.

^b Time from last relapse to LP was calculated as the interval between the date of lumbar puncture (LP) and the date of the last clinical relapse.

^*Statistically significant between‐group differences.

3.2. Two‐Part Model of Cytokine Alterations Across MS Phenotypes

To account for the zero‐inflated distribution of cytokine concentrations, we applied a two‐part model separately testing: (I) detectability (probability of observing values > 0), (II) amount among detectable observations, and (III) a joint test integrating both components.

At the detectability level, significant global group effects were observed for GM‐CSF, IL‐15, IL‐4, IL‐6, and IL‐7, with post hoc differences occurring mainly in comparisons versus NIC. No significant pairwise differences in detectability were found among MS phenotypes. For the amount component, significant global group effects were found for MIP‐1β, MIP‐1α, FGF2, TNF‐α, IL‐6, IL‐10, and IL‐17. Most post hoc differences again involved comparisons versus NIC. Among MS phenotypes, RRMS with PIRMA showed higher detectable levels of TNF‐α, RANTES, and IL‐17 than stable RRMS, whereas PPMS showed lower IP‐10 than stable RRMS; no significant differences were found between PPMS and RRMS with PIRMA. In the joint test, significant overall group effects were observed for MIP‐1β, GM‐CSF, IL‐6, MIP‐1α, TNF‐α, IL‐15, IL‐4, and FGF2. Full pairwise results are shown in Figure 2, Panels A–C.

FIGURE 2.

Between‐group differences in CSF cytokines assessed by two‐part (hurdle) models. (A) Heatmap summarizing the results of the two‐part analyses for each cytokine. For both the detectability component (y > 0) and the amount component (y∣y > 0), the plot reports the overall group effect (“Global”), all pairwise post hoc comparisons, and the joint test combining both components (“Joint”). Color intensity represents statistical significance as −log 10(p), with darker shades indicating smaller p values. Cytokines are grouped by functional class, shown in the left annotation bar: Pro‐inflammatory cytokines, anti‐inflammatory cytokines, other cytokines, chemokines, and growth factors. (B) Boxplots with overlaid individual data points showing the distribution of selected log10‐transformed detectable cytokine concentrations across groups (NIC, stable RRMS, RRMS with PIRMA, and PPMS). Only subjects with detectable values were included in these plots. Horizontal brackets indicate significant pairwise comparisons derived from the amount component of the hurdle models. (C) Volcano plots summarizing pairwise post hoc contrasts from the hurdle models. The upper row shows results for the detectability component, expressed as log₂(odds ratio), whereas the lower row shows results for the amount component, expressed as log₂(fold change) among detectable samples only. The y‐axis reports −log10(p). Positive x‐axis values indicate higher detectability or higher concentration in the first group of the comparison, whereas negative values indicate higher values in the reference group. Dashed horizontal lines denote the significance threshold, and labeled points identify cytokines contributing most strongly to between‐group differences.

3.3. Global Similarity of Cytokine Interactomes

Correlation matrices between cytokines (interactomes) were computed for each study group using Spearman partial correlation (Figure 3, Panel A).

FIGURE 3.

Global similarity and edge‐wise differences of CSF cytokine interactomes across study groups. (A) Group‐specific adjusted partial Spearman correlation matrices for NIC, stable RRMS, RRMS with PIRMA, and PPMS. Pairwise cytokine correlations were estimated after adjustment for age, sex, and DMT. Colors represent the direction and strength of correlation, with warm tones indicating positive associations and cool tones indicating negative associations. Cytokines are ordered identically across matrices and grouped by functional class, shown in the side annotation bars: Pro‐inflammatory cytokines, anti‐inflammatory cytokines, other cytokines, chemokines, and growth factors. (B) Pearson's matrix‐to‐matrix similarity between pairs of group‐specific correlation matrices. Each dot represents one cytokine pair (upper‐triangular matrix element), plotted according to its adjusted partial Spearman correlation in the two groups being compared. The solid line indicates the linear fit, and the dashed diagonal indicates the identity line. Reported values are Pearson's correlation coefficients and corresponding p values, summarizing the global similarity of cytokine interaction architecture between groups. (C) Edge‐by‐edge differential correlation maps for all pairwise group comparisons, obtained using a permutation‐based framework with 5000 permutations and FDR correction for multiple testing. Colored cells indicate cytokine pairs showing significant between‐group differences in adjusted partial Spearman correlation. Warm colors indicate stronger correlations in the first group named in the comparison, whereas cool colors indicate stronger correlations in the second group. Cytokines are displayed in the same order as in panel A, with functional classes indicated by the side annotation bars.

To assess whether the overall architecture of cytokine interactions was preserved across groups, we compared group‐specific correlation matrices using Pearson's similarity (Figure 3, Panel B). Matrix similarity was highest between stable RRMS and PPMS (r = 0.82, p < 0.001), followed by stable RRMS and RRMS with PIRMA (r = 0.80, p < 0.001), NIC and stable RRMS (r = 0.79, p < 0.001), and RRMS with PIRMA and PPMS (r = 0.76, p < 0.001). Lower, although still highly significant, similarities were observed for NIC versus RRMS with PIRMA (r = 0.69, p < 0.001) and especially NIC versus PPMS (r = 0.62, p < 0.001).

3.4. Differential Cytokine Correlations Across Groups

Then, permutation‐based edge‐wise differences between groups were performed (Figure 3, Panel C). Compared with NIC, the stable RRMS network showed a mixed pattern of differential correlations, with both strengthened and weakened edges. These changes were relatively widespread and involved not only the pro‐inflammatory compartment, but also several cross‐compartment links connecting pro‐inflammatory cytokines with anti‐inflammatory mediators and with the chemokine/growth factor domain.

By contrast, RRMS with PIRMA vs. NIC was dominated by positive differential correlations, indicating that several cytokine pairs became more tightly coordinated in the RRMS with PIRMA phenotype. These reinforced associations were especially evident within the upper inflammatory block and in links bridging inflammatory cytokines with the lower chemokine/growth factor compartment.

The pattern was markedly different between PPMS and NIC, with negative differential correlations clearly predominating. These reductions were particularly evident in edges involving the chemokine and growth factor region of the matrix, with additional losses extending to connections between this compartment and the pro−/anti‐inflammatory cytokine blocks.

Within the RRMS spectrum, the comparison of RRMS with PIRMA vs. stable RRMS showed mostly positive differential edges, indicating that progression to the RRMS with PIRMA phenotype was associated with a selective strengthening of cytokine co‐regulation. These increases were concentrated in the pro‐inflammatory sector and in cross‐talk with chemokines/growth factors, consistent with the emergence of a more integrated inflammatory network in patients with progression independent of relapse activity.

Finally, PPMS vs. stable RRMS showed relatively few differential edges, consistent with the high global similarity observed between these two groups. However, PPMS vs. RRMS with PIRMA displayed a clearer predominance of negative differential correlations, particularly in links involving pro‐inflammatory cytokines, chemokines, and growth factors, indicating that the stronger coordinated inflammatory pattern seen in RRMS with PIRMA was not preserved in PPMS.

3.5. Graph Analysis of Cytokine Interactomes and Cluster Preservation

To further characterize the topological organization of cytokine interactions across groups, we thresholded the correlation matrices and analyzed the resulting signed networks using the spin‐glass algorithm (Figure 4, Panels A and B).

FIGURE 4.

Thresholded cytokine interactomes, node centrality, and module preservation across groups. (A) Thresholded group‐specific cytokine correlation matrices for NIC, stable RRMS, RRMS with PIRMA, and PPMS, obtained from adjusted partial Spearman correlation matrices after applying a threshold of ∣ρ∣ ≥ 0.30. Only suprathreshold associations are shown. Colors represent the sign and strength of the retained correlations, with warm tones indicating positive correlations and cool tones indicating negative correlations. Cytokines are ordered identically across groups, and side annotation bars indicate cytokine functional classes. (B) Graph representations of the corresponding thresholded cytokine interactomes. Nodes represent cytokines and are colored according to the communities identified by the spin‐glass community detection algorithm. Edges represent suprathreshold correlations, with red lines indicating positive correlations and blue lines indicating negative correlations; edge thickness reflects correlation strength. These layouts illustrate the group‐specific modular organization and the relative density of cytokine interactions. (C) Node degree distributions in each group. Bar plots show the number of suprathreshold connections for each cytokine in the thresholded networks. Cytokines are ordered according to community membership, and the side annotation bars indicate cytokine functional classes. Degree was used as a measure of topological centrality to identify the most connected inflammatory mediators in each interactome. (D) Alluvial plot showing the correspondence between cytokine modules identified in NIC and those detected in stable RRMS, RRMS with PIRMA, and PPMS. NIC‐derived modules (O1–O4) are used as the reference partition and linked to the best corresponding target modules (M1–M3) in each disease group. Flow widths are proportional to the number of cytokines reassigned from each NIC module to the target module, thereby illustrating preservation or redistribution of module membership across phenotypes. (E) Summary of module preservation for NIC‐derived communities across disease groups. Bubble size represents the dominant fraction, that is, the proportion of nodes from a given NIC module that remain assigned to the same target module in the comparison group. Bubble color represents the best‐match Jaccard index, quantifying the overlap between the NIC reference module and its closest corresponding module in the target group. Together, these metrics summarize the extent to which the modular organization observed in NIC is preserved or reorganized in stable RRMS, RRMS with PIRMA, and PPMS.

In NIC, the thresholded interactome showed a relatively structured and compartmentalized organization, with four modules identified, including a smaller chemokine‐enriched community. The corresponding network layout showed a clearly segregated architecture, with distinct node clusters connected by only a few bridging edges. In stable RRMS, the thresholded network remained relatively well organized but no longer maintained the same degree of compartmentalization observed in NIC. Matrix and graph representations suggested a partial redistribution of edges across modules, with the emergence of three broader communities and loss of the autonomous chemokine‐enriched module, which became reabsorbed into a larger inflammatory cluster. The most marked reorganization was observed in RRMS with PIRMA, whose thresholded matrix appeared denser and more interconnected than those of the other groups. The corresponding graph showed a larger number of positive edges both within and across modules, accompanied by a broader distribution of high‐degree nodes. By contrast, PPMS retained a modular configuration with three main communities, but the overall topology appeared less densely interconnected and less integrated than in RRMS with PIRMA.

Node‐degree analysis showed distinct topological profiles across groups (Figure 4, Panel C). In NIC, degree values were distributed across several nodes, without clear dominance of a single hub. Stable RRMS showed a pattern similar to that of RRMS, but with some redistribution of highly connected nodes. By contrast, RRMS with PIRMA showed several cytokines with comparatively high degree values, indicating a less evenly distributed degree profile. PPMS also showed elevated degree values in selected nodes, but this pattern was less pronounced than in RRMS with PIRMA.

To assess whether the modular organization identified in NIC was preserved across disease groups, we used the NIC‐derived communities as the reference partition and tracked their distribution in stable RRMS, RRMS with PIRMA, and PPMS (Figure 4, Panels D and E). In NIC, four modules were identified, including three major communities (O1–O3) and one smaller chemokine‐enriched community (O4). Across all disease groups, the three major NIC modules remained broadly recognizable, whereas the small chemokine‐enriched module did not retain an independent configuration.

Among the major NIC‐derived communities, O3 was the most preserved across all MS groups, with dominant‐fraction values of 1.00 in stable RRMS, 0.86 in RRMS with PIRMA, and 1.00 in PPMS, and corresponding best‐match Jaccard indices of 0.70, 0.50, and 0.70. O2 also showed substantial preservation, with dominant‐fraction values of 0.75, 0.50, and 0.75, and Jaccard indices of 0.60, 0.44, and 0.60 in stable RRMS, RRMS with PIRMA, and PPMS, respectively. By contrast, O1 was the least preserved, with dominant‐fraction values of 0.60, 0.40, and 0.60 and corresponding Jaccard indices of 0.27, 0.14, and 0.27.

A distinct pattern was observed for the small chemokine‐enriched module O4, which did not maintain an autonomous identity in any disease group. Instead, its nodes were consistently reassigned to module M1, with dominant‐fraction values of approximately 0.71 and Jaccard overlap ranging from 0.38 to 0.45.

3.6. Classification Performance of Machine Learning Models

We then assessed the discriminatory performance of the three machine‐learning approaches across all pairwise group comparisons (Figure 5).

FIGURE 5.

Receiver operating characteristic (ROC) curves showing the classification performance of the three machine learning models across all pairwise group comparisons: Stable RRMS vs. NIC, RRMS with PIRMA vs. NIC, PPMS vs. NIC, RRMS with PIRMA vs. stable RRMS, PPMS vs. stable RRMS, and PPMS vs. RRMS with PIRMA. The cytokine‐level model is shown in red, the network‐level model in blue, and the combined model in green. Filled circles indicate the operating point corresponding to the classification threshold used to derive the reported performance metrics. The dashed diagonal line represents chance‐level discrimination.

In comparisons versus NIC, the cytokine‐level model performed best for stable RRMS versus NIC (AUC 0.86; accuracy 0.79) and PPMS versus NIC (AUC 0.82; accuracy 0.73), whereas the combined model performed best for RRMS with PIRMA versus NIC (AUC 0.82; accuracy 0.76). Within the MS spectrum, the network‐level model showed the best performance for distinguishing RRMS with PIRMA from stable RRMS (AUC 0.86; accuracy 0.76), outperforming both the cytokine‐level and combined models. In comparisons involving PPMS versus RRMS subgroups, the combined model yielded the highest performance (PPMS vs. stable RRMS: AUC 0.77, accuracy 0.72; PPMS vs. RRMS with PIRMA: AUC 0.79, accuracy 0.75). Across all pairwise comparisons, sensitivity ranged from 0.54 to 0.79 and specificity from 0.63 to 0.80.

Detailed performance metrics are reported in Table 3.

TABLE 3.

Performance metrics of the three machine learning models across all pairwise comparisons.

Comparison	Model	AUC	Sensitivity	Specificity	PPV	NPV	Accuracy
Stable‐RRMS vs. NIC	Cytokine‐level	0.86	0.79	0.80	0.83	0.75	0.79
	Network‐level	0.80	0.73	0.73	0.77	0.68	0.73
	Combined	0.83	0.78	0.77	0.81	0.73	0.77
RRMS with PIRMA vs. NIC	Cytokine‐level	0.81	0.73	0.72	0.45	0.89	0.72
	Network‐level	0.74	0.69	0.70	0.43	0.88	0.70
	Combined	0.82	0.76	0.76	0.50	0.91	0.76
PPMS vs. NIC	Cytokine‐level	0.82	0.74	0.72	0.44	0.90	0.73
	Network‐level	0.76	0.72	0.68	0.40	0.89	0.69
	Combined	0.77	0.70	0.69	0.40	0.89	0.69
RRMS with PIRMA vs. stable RRMS	Cytokine‐level	0.66	0.54	0.63	0.25	0.85	0.61
	Network‐level	0.86	0.77	0.76	0.43	0.93	0.76
	Combined	0.76	0.65	0.70	0.34	0.90	0.69
PPMS vs. stable RRMS	Cytokine‐level	0.75	0.69	0.68	0.35	0.90	0.68
	Network‐level	0.75	0.69	0.71	0.37	0.90	0.70
	Combined	0.77	0.70	0.72	0.39	0.91	0.72
PPMS vs. RRMS with PIRMA	Cytokine‐level	0.75	0.68	0.70	0.67	0.71	0.69
	Network‐level	0.78	0.72	0.68	0.67	0.73	0.70
	Combined	0.79	0.75	0.75	0.73	0.77	0.75

Note: The cytokine‐level model was based on cytokine concentrations, the network‐level model on edge‐based interaction features derived from differential cytokine networks, and the combined model integrated both feature sets.

4. Discussion

In this prospective study, we combined quantitative, network‐level, and machine‐learning approaches to investigate CSF inflammation across MS phenotypes. Three main findings emerged. First, quantitative CSF cytokine profiling distinguished patients with MS from NIC but had limited ability to separate MS phenotypes. Second, in radiologically stable RRMS, quantitative analysis identified selected mediators associated with PIRA, whereas network analysis better captured inflammatory organization and more clearly discriminated RRMS with PIRMA from stable RRMS patients. Third, cytokine interactome‐derived features showed clinically relevant classification performance, supporting the potential utility of network‐informed biomarkers for early risk stratification.

With regard to the first aim of the study, the two‐part model confirmed the expected finding that the inflammatory distinction between MS and NIC was driven by a broader inflammatory activation in the CSF [21, 22]. Several cytokines differed between the two groups, including GM‐CSF, IL‐15, IL‐4, IL‐6, IL‐7, MIP‐1β, MIP‐1α, FGF2, TNF‐α, IL‐10, and IL‐17. Regarding MS patients, the most clinically relevant inflammatory differences emerged not across broad MS phenotypes, but within the RRMS spectrum according to PIRA status. Accordingly, quantitative cytokine analysis was more informative for identifying radiologically stable RRMS patients with PIRA than for distinguishing RRMS from PPMS, which showed substantial overlap and no clear quantitative separation, consistent with recent evidence on individual soluble serum and CSF biomarkers [23]. In radiologically stable RRMS, PIRMA patients showed higher levels of TNF‐α, IL‐17, and RANTES (CCL5) than PIRA‐ patients. These mediators are all biologically plausible in the context of MS progression, as they are involved in immune dysregulation, neuroinflammation, leukocyte trafficking, and synaptic dysfunction [21, 24, 25, 26]. In particular, TNF‐α has been implicated in inflammatory synaptopathy [27, 28], where it enhances glutamatergic transmission and promotes excitotoxic damage, thereby providing a mechanistic link between inflammation and neurodegeneration in MS. Our findings are also in line with previous evidence showing higher CSF TNF‐α levels in RRMS patients PIRMA than in those PIRA‐ [29]. Likewise, IL‐17 has been linked to synaptic dysfunction [30] and excitotoxic injury [31], further supporting the role of inflammatory mediators in neurodegenerative mechanisms in MS [25].

The results for the second aim of this study suggest that, beyond quantitative differences, network‐based analyses provide a more meaningful representation of inflammatory organization across groups. This interpretation is biologically plausible, as cytokines are increasingly viewed as components of dysregulated signaling networks, whose organization may convey pathophysiological information that is not captured by the levels of individual mediators alone [12]. In NIC, the CSF cytokine interactome showed a structured modular architecture, with three major communities and a smaller chemokine‐enriched module (IL‐1RA, IL‐5, IL‐8, MIP‐1α, MIP‐1β, RANTES, Eotaxin) [32]. The three main communities were broadly consistent with different immune profiles, including a Th1‐like module (TNF‐α, IFN‐γ, IL‐12, IP‐10, and MCP‐1), a Th2/regulatory‐oriented module (IL‐4, IL‐5, IL‐10, IL‐13, and IL‐1RA), and a mixed pro‐inflammatory/Th17‐related module (IL‐6, IL‐17, GM‐CSF, G‐CSF, and IL‐15) [33, 34]. Although these clusters cannot be unambiguously assigned to discrete inflammatory programs, this is consistent with the current view of neuroinflammation as the simultaneous activation of multiple, partially overlapping immune pathways rather than sharply segregated responses [35, 36].

Across MS phenotypes, the main modular structure remained recognizable, suggesting that disease‐related neuroinflammation does not generate entirely new network architectures, but rather reorganizes pre‐existing ones. The most notable change involved the smaller chemokine‐enriched module, whose components were no longer retained as an autonomous cluster in MS but instead converged into the more inflammatory communities, mainly the Th1‐like and Th17‐like modules. This pattern suggests that the transition from NIC to MS reflects a redistribution of inflammatory relationships. Within MS, the interactomes were generally less segregated than in NIC, consistent with a more integrated inflammatory organization involving cytokine, chemokine, and growth factor pathways. The most relevant network change emerged within radiologically stable RRMS. Compared with stable RRMS patients, the RRMS with PIRMA group showed a denser pattern of positive differential correlations, stronger cross‐node coupling, reduced modular segregation, and a broader distribution of highly connected nodes. Thus, the main network signature of PIRA was not simply a quantitative increase in selected mediators, but a shift toward a more tightly coordinated inflammatory architecture. In this framework, the concomitant increase of TNF‐α, IL‐17, and RANTES in RRMS with PIRMA may be interpreted as part of a broader state of enhanced inflammatory coordination rather than a simple overall increase in cytokine activity. PPMS showed a different network configuration, with less cross‐module integration than RRMS with PIRMA, and a modular interactome broadly resembling that of stable RRMS patients. One possible interpretation is that progression in PPMS is associated with inflammatory processes that are less represented in the CSF compartment, either because they are more compartmentalized within the CNS [37] or because progression depends more heavily on neurodegenerative and maladaptive remodeling mechanisms [8, 38]. A more speculative interpretation is that progression may differ biologically between RRMS with PIRMA and PPMS, with the former showing a stronger inflammatory component and the latter a more compartmentalized or neurodegenerative profile. Alternative explanations include differences in age, baseline disability, MS duration, and the possibility that CSF cytokine measurements only partially capture compartmentalized inflammatory processes within the CNS. Our data cannot settle this question, but they do support the concept that progression is biologically heterogeneous.

Considering the findings from the first two aims, the third aim of the study was to assess the clinical relevance of quantitative cytokine measures and network‐derived features for group discrimination. Single‐cytokine measures clearly separated MS from NIC, indicating that quantitative inflammatory signals are sufficient for broad case–control discrimination. By contrast, the cytokine interactome provided greater resolution within the MS spectrum, particularly for distinguishing RRMS with PIRMA from stable RRMS, a subtler but more clinically relevant contrast. In this comparison, the network‐level model achieved an AUC of 0.86, with 77% sensitivity, 76% specificity, 43% positive predictive value, and 93% negative predictive value. These results outperformed the cytokine‐level model, suggesting that progression‐related biology is better captured by inflammatory network organization than by individual cytokine levels alone. A different pattern emerged in comparisons involving PPMS, where the combined model performed best, suggesting that both quantitative and relational inflammatory information may help characterize progressive phenotypes. Overall, these findings indicate that cytokine network configuration may improve identification of patients at risk of PIRA, even after accounting for broad DMT efficacy class, although the effects of individual treatment strategies cannot be fully disentangled in this study. More broadly, this framework could be extended to other biomarker families, in line with the shift in MS research from isolated analytes toward multidimensional, network‐informed biomarker profiling supported by computational and machine‐learning approaches [39, 40, 41, 42].

This study has several strengths. First, the stringent PIRMA definition allowed us to study progression‐related inflammatory features in relapse‐free, MRI‐stable RRMS patients, using stable RRMS patients as the comparator group. The two‐part model directly addressed the coexistence of non‐detectable values and highly skewed positive distributions, a major analytical issue in CSF biomarker data. Statistical analyses were adjusted for relevant covariates, reducing potential confounding in both quantitative and network‐level comparisons. The sample size was adequate for a single‐center prospective CSF‐based study, and an a priori power analysis based on the two components of the hurdle model indicated that the final sample exceeded the estimated size required to detect small‐to‐moderate group effects. Moreover, the analytical strategy combined hurdle modeling, differential‐correlation analysis, community detection, and machine‐learning classification, enabling investigation of neuroinflammation at both the single‐mediator and network levels and translating group‐level network alterations into subject‐level classification features.

This study also has some limitations. First, the control group consisted of patients with non‐inflammatory, non‐neurodegenerative neurological conditions rather than healthy subjects. Although this limits the case–control interpretation, it reflects a common, ethically pragmatic approach in CSF research. Second, the groups were not fully homogeneous: PPMS patients were older than RRMS patients, and within RRMS, PIRMA patients were older and more disabled at baseline than stable RRMS patients. Although analyses were adjusted for these factors, residual confounding remains possible. Third, the number of RRMS with PIRMA patients was limited, and follow‐up was restricted to 2 years. Moreover, the 6‐month confirmation requirement shortened the effective PIRMA detection window to approximately 18 months. This may have affected subgroup classification, as some stable RRMS patients could accumulate disability over longer follow‐up; longer studies are therefore needed. Fourth, restricting the primary analysis to relapse‐free, MRI‐stable RRMS patients limits generalizability to broader RRMS populations with ongoing inflammatory activity. Fifth, DMT exposure was classified only by efficacy category, and more detailed treatment stratification was not feasible. Network‐level findings may also be influenced by analytical choices, although these were applied consistently across groups. Sixth, network findings may depend on analytical choices; their robustness should be tested across alternative thresholds and network‐construction strategies. Finally, the findings were not externally validated, although nested out‐of‐sample cross‐validation was used to reduce overfitting. Independent replication is needed before these models can be considered for clinical risk stratification.

5. Conclusions

PIRMA in relapse‐free and radiologically stable RRMS was associated not only with quantitative increases in selected inflammatory mediators but also with a broader reorganization of the CSF cytokine network. These results suggest that progression‐related biology may be more faithfully captured by the network architecture of inflammatory signaling than by isolated cytokine levels alone. By translating interactome‐level alterations into subject‐level features, this approach also provides a framework for deriving individualized biomarkers, with potential relevance for early risk stratification and biologically informed characterization of progression across MS phenotypes.

Author Contributions

Antonio Bruno: conceptualization, investigation, data curation, and writing – original draft; Matteo Conti: formal analysis, methodology, software, visualization, and writing – original draft; Ettore Dolcetti: investigation and data curation; Federica Azzolini: investigation and data curation; Angela Borrelli: investigation and data curation; Roberta Fantozzi: investigation and data curation; Mario Stampanoni Bassi: investigation, writing – review and editing; Giovanni Galifi: investigation and data curation; Giuseppe Maccarrone: investigation and data curation; Luca Montaguti: investigation and data curation; Marco Cervigni: investigation and data curation; Maddalena Dal Pozzo: investigation and data curation; Roberto Furlan: formal analysis and writing – review and editing; Annamaria Finardi: formal analysis, writing – review and editing; Girolama Alessandra Marfia: writing – review and editing; Marco Salvetti: writing – review and editing; Antonella Conte: writing – review and editing; Fabio Buttari: conceptualization, supervision, and writing – review and editing; Diego Centonze: conceptualization, funding acquisition, supervision, and writing – review and editing; Luana Gilio: investigation, data curation, writing – review and editing.

Funding

This project was funded by the Italian Ministry of University and Research (MUR) under the complementary actions to the National Recovery and Resilience Plan (PNRR) (Grant PNC0000001—CUP B53C22005940008—“D34Health – Digital Driven Diagnostics, Prognostics and Therapeutics for Sustainable Healthcare”, to D.C.). This work wthe European Union – National Recovery and Resilience Plan (PNRR), NextGenerationEU – NRRP Mission 6, Component 2, Investment 2.1 “Enhancement and strengthening of biomedical research in the NHS” (M6/C2_CALL 2023; project PNRR‐MCNT2‐2023‐12377850, to D.C.), the Italian Ministry of Health (Ricerca Corrente, IRCCS Neuromed, to D.C.), and the Fondazione Italiana Sclerosi Multipla (FISM, Italy), financed or co‐financed with the “5 per mille” public funding (project code 2022/S/2, to D.C.).

Conflicts of Interest

A. Bruno received honoraria for participation in advisory boards from Almirall and support for travel expenses to attend congresses from Biogen. E. Dolcetti received support for travel expenses to attend congresses from Novartis, Biogen, Roche, and Janssen, and received honoraria for participation in an advisory board from Lundbeck. R. Fantozzi received consulting fees and honoraria for advisory boards from Biogen Idec, Merck‐Serono, Novartis, Roche, and Teva. R. Furlan received speaker fees for teaching and workshops from Biogen, Merck, Novartis, Roche, Teva, and Alexion, and received unrestricted grants from Biogen and EMD Serono for educational activities, courses, or research. G.A. Marfia received speaking or consultation fees from Almirall, Bayer‐Schering, Biogen, Genzyme, Merck‐Serono, Novartis, Teva, and Sanofi‐Genzyme. A. Conte served as a board member for Roche. F. Buttari acted as an advisory board member for Teva and Roche, and received honoraria for speaking or consultation fees from Merck Serono S.p.A., Rome, Italy, an affiliate of Merck KGaA, Teva, Biogen Idec, Sanofi, and Novartis. He also received nonfinancial support from Merck Serono S.p.A., Rome, Italy, an affiliate of Merck KGaA, Teva, Biogen Idec, Sanofi, and Novartis. D. Centonze acted as an advisory board member and received honoraria for speaking or consultation fees from Alexion, Almirall, Amicus, Bayer, Biogen, BMS, Celgene, Chiesi, GW Pharmaceuticals, Horizon, Janssen, Lundbeck, Merck Serono, Neuraxpharm, Novartis, Roche, Sandoz, Sanofi‐Genzyme, Viatris, and Teva. He is also the principal investigator in clinical trials sponsored by Biogen, BMS, Merck Serono, Mitsubishi, Novartis, Roche, Sanofi‐Genzyme, and Actelion. His preclinical and clinical research was supported by grants from Bayer Schering, BMS, Biogen, Celgene, Lundbeck, Merck Serono, Novartis, Roche, Sanofi‐Genzyme, and Teva. L. Gilio, M. Conti, F. Azzolini, A. Borrelli, M. Stampanoni Bassi, G. Galifi, G. Maccarrone, L. Montaguti, M. Cervigni, M. Dal Pozzo and A. Finardi report no disclosures or conflicts of interest relevant to this manuscript.

Supporting information

Appendix S1: Supporting Information.

Funding Statement

This work was funded by Ministero Della Salute grant PNRR‐MCNT2‐2023‐12377850; Ministero dell'Università e Della Ricerca grant PNC0000001—CUP B53C22005940008; “5 per mille” public funding ; Fondazione Italiana Sclerosi Multipla .

Data Availability Statement

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