Splicing-based biomarkers define a robust multigene classifier for relapsing-remitting multiple sclerosis

Abstract

Background

Alternative splicing (AS) is recognized as a key mechanism in multiple sclerosis (MS). We aimed to construct and validate a multivariate AS-based classifier (MS-Splicing Score, MS-SS) for the discrimination of relapsing-remitting MS (RRMS) patients from healthy controls.

Methods

Three AS events (IFNAR2 exon-8 skipping, NFAT5 exon-2 skipping, PRKCA exon-3∗ inclusion) were selected based on functional and literature evidence. Isoforms were quantified via fluorescent-competitive RT-PCR in peripheral blood RNA from two independent cohorts (Italy: 37 RRMS, 50 controls; USA: 29 RRMS, 20 controls). A logistic regression model was trained to derive the MS-SS, followed by ROC analysis.

Results

The MS-SS distinguished RRMS patients from controls in both cohorts (Italy: p = 0.00083, AUC = 0.71, 95 %CI = 0.59–0.82; USA: p = 0.00074, AUC = 0.77, 95 %CI = 0.63–0.90). In the pooled dataset, the score remained significantly elevated in MS (p = 5.9 × 10⁻⁶, AUC = 0.72, 95 %CI = 0.64–0.81), and a PCA-based refinement improved classification accuracy, yielding an AUC = 0.87 (95 %CI = 0.81–0.94). At the optimal cutoff (Youden's index), the score achieved a sensitivity of 80 % and specificity of 86 %. Supervised rule-based modeling using a logic-learning machine algorithm identified interpretable splicing thresholds and enabled clinical classification at the individual level.

Conclusion

Our study introduces a novel, robust AS-based classifier for RRMS and proposes a strategy for transcriptome-based biomarker development in neuroimmunology. However, the relatively small sample sizes within each cohort may limit the generalizability of these findings, warranting larger validation studies to confirm the clinical utility of this biomarker.

1. Introduction

Alternative splicing (AS) is a post-transcriptional mechanism that enhances transcriptomic and proteomic diversity by generating multiple mRNA isoforms from a single gene [1,2]. This process is highly conserved and affects >95 % of human multiexonic genes, underscoring its critical role in numerous cellular functions [3]. The regulation of AS involves cis-acting elements -such as splicing enhancers and silencers- and trans-acting factors, including serine/arginine-rich proteins and heterogeneous-nuclear ribonucleoproteins [4,5]. Given its influence on gene expression, AS plays a pivotal role in physiological processes, particularly within the immune system, where it governs T-cell activation, migration, cytokine responses, and apoptosis [6]. Dysregulation of AS has been implicated in multiple diseases, including multiple sclerosis (MS) [7].

MS, a chronic inflammatory demyelinating disease of the central nervous system (CNS), presents with heterogeneous clinical phenotypes, typically initiating as relapsing-remitting (RRMS) and potentially evolving into secondary progressive forms (SPMS), while a minority of patients exhibit a primary progressive course (PPMS) from onset. Historically, these relapsing and progressive phases were viewed as distinct clinical entities [8]. However, recent findings on progression independent of relapse activity (PIRA) have prompted a redefinition of the disease, suggesting that relapsing and progressive features coexist along a continuum from the very onset, thereby challenging the conventional classification of its course [9].

Current evidence suggests that AS dysregulation could serve as a hallmark of the disease. Genome-wide studies have indeed revealed significant alterations in splicing patterns in MS patients, including defective RNA processing and differential exon usage [[10], [11], [12], [13], [14], [15], [16], [17], [18], [19]]. Notably, these splicing abnormalities may contribute to disease pathology by generating novel autoantigenic epitopes or disrupting immune regulation [20]. In addition, studies have shown that MS-associated loci identified through genome-wide association studies (GWAS) are enriched with splicing-related elements, further supporting the hypothesis that AS dysregulation plays a central role in disease predisposition [13,15,21].

Biomarkers, defined as measurable indicators of normal or pathological processes and responses to therapeutic interventions, have gained increasing attention in MS research. Several biomarker subtypes have been proposed, including diagnostic, prognostic, predictive, and treatment-response biomarkers [22]. Among the most established biomarkers in MS clinical practice are MRI-derived features, cerebrospinal fluid (CSF) oligoclonal bands (OCB), and serum neurofilament light chain (NfL), a marker of axonal injury. However, unlike rheumatoid arthritis, for which serological markers such as rheumatoid factor (RF) and anti-citrullinated peptide/protein antibodies (ACPA) are available, no definitive blood-based diagnostic test exists for MS [22,23].

Here, we propose a score based on AS profiles of three transcripts (PRKCA, IFNAR2, and NFAT5), which have demonstrated aberrant splice processing in MS, to achieve accurate discrimination between MS patients and healthy controls.

2. Materials and methods

2.1. MS cohorts

This study received approval from local Ethical Committees. All participants provided informed consent. Patients were diagnosed with clinically definite MS based on the revised McDonald's criteria [24].

We recruited two case-control cohorts: i) an Italian cohort, consisting of 37 RRMS patients (mean age 41.6 ± 11.7 years, 69 % female) in remission, untreated for at least one month, and 50 healthy controls (mean age 39.1 ± 9.9 years, 72 % female) with no family history of autoimmune or neurodegenerative diseases; ii) a US cohort, including 29 RRMS patients (mean age 53.9 ± 13.8y, 67 % female, 92 % Caucasian) in remission and untreated for at least one month, and 20 healthy Caucasian controls (mean age 53.4 ± 15.7y, 60 % female, 85 % Caucasian) with no relevant family history for MS.

Peripheral blood mononuclear cells (PBMCs) were isolated using the Lympholyte Cell separation medium (Cederlane Laboratories, Hornby, Ontario, Canada). Total RNA was extracted with the EuroGold Trifast kit (Euroclone, Wetherby, UK) for the Italian cohort and the miRNeasy Mini Kit (Qiagen, Hilden, Germany) for the US cohort.

2.2. Quantitation of AS events

First-strand cDNA synthesis was performed using random hexamers (Promega, Madison, WI, USA) and Superscript-III Reverse Transcriptase (Thermo Fisher Scientific), following the manufacturers' protocols. One microliter of the resulting cDNA was used as template for subsequent amplifications. Fluorescent-competitive reverse transcription (RT)-PCR assays, employing 6-FAM-labeled primers, were conducted to quantify the percentage of isoforms exhibiting non-physiological exon skipping in IFNAR2 (exon 8) and NFAT5 (exon 2), and enhanced inclusion of a non-canonical exon in PRKCA (alternative exon 3∗). Reactions were performed under standard conditions using the GoTaq DNA Polymerase (Promega) and a Mastercycler EP Gradient (Eppendorf, Hamburg, Germany). Amplified fragments were separated via capillary electrophoresis on an ABI3500 Genetic Analyzer (Thermo Fisher Scientific); peak areas were measured using GeneMapper v4.0 software (Thermo Fisher Scientific). The relative exon inclusion level (percent-spliced in, psi) was estimated for each sample by calculating the ratio of the relevant peak over the sum of all the fluorescence peak areas (set as 100 %).

The identical amplification efficiency of amplicons in each assay was verified by real-time PCR. Primer sequences are available upon request.

2.3. Dataset pre-processing, construction of a discriminative score and its evaluation

Preprocessing of AS data was conducted in R (version 4.4.2) with the aid of base R functions and additional support from the tidyverse (dplyr), caret, and MASS packages. Specifically, we: i) transformed the classification variable into a binary format (cases = 1, controls = 0); ii) addressed missing data (i.e., exon-skipping percentages; accounting for <4 % of the dataset) by imputing the mean of the respective variable; and 3) standardized exon-skipping percentages using Z-scores to ensure data comparability across samples by eliminating scale differences.

To derive an optimal classification index, a multivariate logistic regression model was applied on preprocessed data using the following equation:

$$\begin{gathered} \log \left(\frac{P\left( \\ \text{case}\right)}{1-P\left( \\ \text{case}\right)}\right)=\beta 0+\beta 1IFNAR2+\beta 2NFAT5+\beta 3PRKCA \end{gathered}$$

where P(case) represents the probability of belonging to the case group. The estimated model coefficients were used to generate a predictive score, calculated as the predicted probability from the logistic regression model.

To assess the discriminatory power of the score in distinguishing cases from controls, we performed a receiver operating characteristic (ROC) curve analysis and calculated the area under the curve (AUC) for each cohort separately (Italian and US cohorts) or together. The ROC curve was generated by plotting sensitivity (true positive rate) vs. 1-specificity (false positive rate) at various score cutoffs (t), according to the following equation:

$$\text{AUC}={\int }_0^1\text{Sensitivity}\left(t\right)d\left(1-\text{Specificity}\left(t\right)\right)$$

The optimal cutoff point was determined using Youden's Index, defined as:

$$J=\max \left(\text{Sensitivity}+\text{Specificty}-1\right)$$

where the threshold maximizing both sensitivity and specificity was selected for each analysis.

Differences in score distributions between cases and controls were visualized using boxplots and statistically examined with the unpaired two-tailed t-test (normally distributed data). Differences in the distribution of the score between the two cohorts were assessed using the Kolmogorov-Smirnov (KS) test.

2.4. Comparison with dimensionality reduction techniques

To assess whether data transformations could enhance the classification performance of the score, we applied two dimensionality reduction techniques: principal component analysis (PCA) and linear discriminant analysis (LDA). PCA was employed to reduce the dimensionality of the dataset while retaining the most significant variance in an unsupervised manner. Specifically, the first principal component (PC1) was extracted and analyzed to determine whether this transformation could enhance the separation between groups. LDA was instead utilized to compute the optimal linear discriminant function, which maximizes class separability.

The results were analyzed by comparing the ROC curves obtained from different approaches (original score obtained from the multivariate logistic regression model, PCA score, LDA score). Differences in AUCs between models were assessed using DeLong's test for paired ROC curves.

2.5. Rule-based modeling via logic learning machine (LLM) and development of prototype clinical algorithm

We employed Rulex (a data analytics platform that uses machine learning and rule-based logic to extract human-interpretable models from complex datasets; https://www.rulex.ai/) to perform supervised modeling using a LLM algorithm [[25], [26], [27]]. This method identifies logical rules of the form “IF condition(s) THEN class”, where conditions refer to psi thresholds for the analyzed AS events. Each rule was evaluated in terms of support (proportion of subjects matching the rule) and confidence (proportion correctly classified).

Thresholds identified through LLM were translated into a rule-based scoring algorithm to simulate a potential diagnostic tool. Each gene-specific psi threshold was assigned a point toward MS or control classification. The resulting cumulative score was used to stratify subjects into three categories: likely MS case, likely control, or indeterminate.

3. Results

3.1. Selection of AS events for the construction of a MS splicing score

With the aim to develop a splicing-based classifier capable of distinguishing MS cases from controls (MS-splicing score, MS-SS), we selected three AS events based on the following criteria: i) events that have been experimentally confirmed through transcriptomic or molecular assays to represent genuine splicing alterations; ii) events reported in the literature to be altered in MS individuals; iii) events hypothesized to contribute to MS pathogenesis due to their roles in immune modulation, CNS integrity, or neuroinflammation; and/or iv) AS events curated in public databases with annotations indicating their functional impact or tissue-specific regulation in neural or immune contexts. The three selected events include exon skipping in IFNAR2 (Interferon Alpha and Beta Receptor Subunit 2; exon 8) and NFAT5 (Nuclear Factor of Activated T cells 5; exon 2), as well as the aberrant inclusion of an alternative exon (exon 3∗) in PRKCA (Protein Kinase C Alpha). Table 1 summarizes the selection criteria and supporting evidence for each selected AS event.

Table 1.

Summary of criteria for selecting AS events Included in the MS-SS.

Gene	AS Event	Selection Criteria	References
IFNAR2	Skipping of exon 8	i) Demonstrated through RNA-seq and RT-PCR experiments ii) n.r. iii) Functional relevance in interferon signaling and immune response iv) Skipping event annotated in VastDB and GTex (items ENST00000342101 and ENST00000413881)	Cardamone et al., 2019; Aliaga-Gaspar et al., 2021
NFAT5	Skipping of exon 2	i) Demonstrated through RT-PCR experiments ii) Reported alteration in MS patients iii) Functional impact on T-cell regulation and blood-brain barrier hyperpermeability iv) n.r.	Paraboschi et al., 2015; Lee et al., 2019; Kunze et al., 2024
PRKCA	Aberrant inclusion of alternative exon 3∗	i) Demonstrated through RT-PCR experiments ii) Reported alteration in MS patients iii) Involvement in immune responses and synaptic plasticity iv) Skipping event annotated in VastDB and GTex (item ENST00000284384)	Paraboschi et al., 2014; Lim et al., 2015; Lordén and Newton, 2021

i) Experimental demonstration through transcript analysis; ii) Event reported to be altered in MS patients in the literature; iii) Event hypothesized to affect MS pathogenesis through immune/CNS mechanisms; iv) Annotation available in the public repositories VastDB (an atlas of alternative splicing profiles in vertebrate cell and tissue types; https://vastdb.crg.eu/wiki/Main_Page) and GTEx (The Genotype-Tissue Expression; https://www.gtexportal.org/home/); n.r., not reported.

3.2. Initial evaluation of as events in the Italian cohort

To assess the discriminatory potential of AS events prior to score construction, we analyzed the expression of the three selected splicing isoforms in peripheral blood samples from an Italian case-control cohort, which included 37 RRMS patients and 50 healthy controls. The three selected AS events were quantified using a fluorescent-competitive RT-PCR assay, enabling isoform-specific detection with high sensitivity.

As shown in Fig. 1, RRMS patients displayed a consistent trend toward increased expression of the aberrant isoforms compared to controls across all three genes. The difference reached statistical significance for NFAT5 (p = 0.031), while differences for IFNAR2 and PRKCA were marginally non-significant (p = 0.077 for both). Although not all comparisons reached conventional statistical thresholds, the directionality of the effect was consistent, supporting the hypothesis of splicing dysregulation in MS and justifying the integration of these events into a composite MS-Splicing Score.

Fig. 1.

Dysregulation of selected AS events in the Italian RRMS cohort. Quantification of three selected AS events on RNA extracted from peripheral blood samples of individuals with RRMS (red, n = 37) and healthy controls (blue, n = 50) in the Italian cohort. Each plot represents the expression levels of a specific aberrant splicing isoform measured using fluorescent-competitive RT-PCR. PCR fragments were separated by capillary electrophoresis, and relative exon inclusion levels were calculated as the ratio of the relevant peak over total fluorescence peak areas. Values were rescaled for analysis (Z-score-normalized). Data shown from left to right: IFNAR2 exon 8 skipping, NFAT5 exon 2 skipping, and PRKCA alternative exon 3∗ inclusion. Each dot represents an individual sample; boxplots show the median, interquartile range, and distribution. Statistical comparisons were performed using an unpaired two-tailed t-test, using R (version 4.4.2).

3.3. MS-SS discriminates MS cases from controls in the Italian cohort

To evaluate the combined discriminative power of the selected AS events, we constructed the splicing-based classifier (MS-SS) in the Italian cohort using multivariate logistic regression. Prior to modeling, AS data were preprocessed by imputing missing values (<4 % of the dataset) with mean substitution, and standardizing aberrant splicing event values as Z-scores. The logistic regression model incorporated the three selected AS variables (IFNAR2, NFAT5, PRKCA) and produced a composite probability score representing the likelihood of MS diagnosis.

As illustrated in Fig. 2A, the MS-SS was significantly elevated in RRMS cases compared to controls (p = 0.00083), indicating a robust separation between groups. ROC curve analysis (Fig. 2B) yielded an AUC of 0.71 (95 %CI: 0.59–0.82), confirming the high discriminatory capacity of the score in the Italian cohort. The optimal cutoff value, determined using Youden's Index, achieved a sensitivity of 59 % and a specificity of 80 %, highlighting the potential utility of this AS-based index in MS classification.

Fig. 2.

Performance of the MS-SS in the Italian cohort (A) Distribution of the MS-RR in RRMS patients (red, n = 37) and healthy controls (blue, n = 50) from the Italian cohort. The score was derived from a logistic regression model incorporating aberrant splicing event data for IFNAR2, NFAT5, and PRKCA, previously quantified using fluorescent-competitive RT-PCR (see Fig. 1 legend). The statistical comparison was performed using an unpaired two-tailed t-test. (B) ROC curve showing the classification performance of the MS-SS score in distinguishing cases from controls in the Italian cohort. The model achieved an AUC of 0.71, 95 %CI = 0.59–0.82, indicating good discriminatory power. Statistical analyses and figures were done in R (version 4.4.2).

3.4. The MS-SS retains discriminatory power in an independent US cohort

The MS-SS was validated in an independent US cohort (29 RRMS, 20 controls) using the same isoform quantification and scoring procedures as in the Italian cohort.

Here, the exon-skipping levels of IFNAR2 remained significantly elevated in patients compared to controls (p = 0.00057), while NFAT5 and PRKCA showed non-significant trends, but in the same direction (p = 0.10 and p = 0.71, respectively), the latter indicating substantial variability (Fig. 3A). Nonetheless, the composite MS-SS remained significantly higher in MS cases than in controls (p = 0.00074), supporting its discriminatory utility (Fig. 3B).

Fig. 3.

Validation of the MS-SS in an independent US cohort. (A) Quantification of IFNAR2 exon 8 skipping (left), NFAT5 exon 2 skipping (center), and PRKCA alternative exon 3∗ inclusion (right) on RNA extracted from peripheral blood samples of individuals with RRMS (red, n = 29) and healthy controls (blue, n = 20) in the US cohort. Each plot represents the expression levels of a specific aberrant splicing isoform measured using fluorescent-competitive RT-PCR. PCR fragments were separated by capillary electrophoresis, and relative exon inclusion levels were calculated as the ratio of the relevant peak over total fluorescence peak areas. Values were rescaled for analysis (Z-score-normalized). Statistical comparisons were performed using an unpaired two-tailed t-test. (B) The MS-SS score was computed via logistic regression integrating the three AS events. The statistical comparison was performed using an unpaired two-tailed t-test. (C) ROC curve showing the classification performance of the MS-SS score in distinguishing cases from controls in the US cohort. The model achieved an AUC of 0.77, 95 %CI = 0.63–0.90. Statistical analyses and figures were done in R (version 4.4.2).

Importantly, the ROC analysis yielded an AUC of 0.77 (95 %CI: 0.63–0.90; Fig. 3C), confirming the score generalizability and even slightly exceeding the performance observed in the Italian cohort (AUC = 0.71), thus supporting its robustness across independent populations.

3.5. Refined MS-SS via PCA on the pooled cohort demonstrates superior classification of MS cases

Given the absence of a significant difference in the MS-SS distributions between the Italian and US cohorts (as assessed by the KS test), we constructed and evaluated the MS-SS on the combined dataset comprising 66 RRMS patients and 70 healthy controls (n = 136). As depicted in Fig. 4A, the MS-SS was significantly elevated in MS cases compared to controls in the combined cohort (p = 5.9∗10⁻⁶). ROC curve analysis of the MS-SS on the combined cohort (Fig. 4B) yielded an AUC of 0.72 (95 %CI: 0.64–0.81), indicating a good (but not improved) discriminatory capacity on the larger dataset.

Fig. 4.

Performance of the MS-SS in pooled datasets and comparison of classification methods. (A) Distribution of the MS-SS score (computed via logistic regression integrating IFNAR2, NFAT5, and PRKCA AS events) in RRMS patients (red, n = 67) and healthy controls (blue, n = 69) across combined Italian and US cohorts. The statistical comparison was performed using an unpaired two-tailed t-test. (B) ROC curves showing the classification performance for three classification models: direct logistic model using the MS-SS score computed via logistic regression (blue, AUC = 0.72, 95 %CI: 0.64–0.81), linear discriminant analysis (LDA, green, AUC = 0.72, 0.63–0.82), and principal component analysis (PCA)-refined logistic regression (red, AUC = 0.87, 95 %CI = 0.81–0.94), showing improved discrimination with the PCA-based composite. Curves were computed on the combined Italian and US cohorts. (C) Principal component 1 (PC1) scores from PCA model on splicing biomarkers significantly differentiated RRMS patients (pink) from controls (cyan) (p = 5.7 × 10⁻¹⁴). In all boxplots, each dot represents a subject; boxplots show medians, interquartile ranges, and whiskers at 1.5 × IQR. Statistical analyses and figures were done in R (version 4.4.2).

Hence, to explore potential improvements in classification performance, we applied two dimensionality reduction techniques: PCA and LDA. While the LDA approach did not yield any significant improvement in discriminatory power compared to the original MS-SS (AUC = 0.72, DeLong's test p = 0.49), the application of PCA to the combined dataset resulted in a first principal component (PC1) that demonstrated a highly significant elevation in MS cases compared to controls (p = 5.7∗10⁻¹⁴). As depicted in Fig. 4C, the boxplots show a marked separation between the groups based on the PCA score. ROC curve analysis of the PCA score (Fig. 4B) yielded an outstanding AUC of 0.87 (95 %CI = 0.81–0.94; sensitivity = 80 %, specificity = 86 %), representing a substantial and statistically significant improvement over the original MS-SS (DeLong's test p = 0.00055).

3.6. LLM-based modeling of exon splicing thresholds reveals discriminative signatures for MS

To identify the threshold levels of AS events capable of distinguishing MS patients from healthy controls, we interpreted exon inclusion percentages (psi) as “molecular logic gates”, i.e. discrete decision points in RNA processing that may delimit pathologic from physiologic states. Using this framework, we applied a LLM approach via the Rulex platform to derive threshold-based classification rules directly from fluorescent-competitive RT-PCR-quantified splicing data. Rulex LLM modeling revealed a set of rules with strong discriminatory power. The most robust rule combined elevated exon inclusion levels in PRKCA (>0.067) and IFNAR2 (>0.465), identifying MS cases with 100 % confidence and applying to 14.7 % of the pooled cohort (20 among 136 individuals; 30.3 % considering only the MS cohort). This rule was evenly distributed across Italian and US subjects. A second high-confidence rule, based solely on IFNAR2 inclusion above 0.537, identified an additional 7.4 % of cases. A third rule involving elevated NFAT5 inclusion (psi>24.799) achieved 85.7 % confidence and was observed exclusively among Italian cases. Conversely, rules predictive of control status included patterns such as concurrent low NFAT5 (≤0.232) and IFNAR2 (≤0.273) psi (100 % confidence; 4.4 % of the cohort, all US-based), as well as combined low PRKCA and moderate NFAT5 psi values.

Building on these rules, we constructed a prototype clinical algorithm that translates splicing thresholds into discrete, additive scores reflecting case or control status (Table 2). This rule-based classifier integrates psi cutoffs across IFNAR2, PRKCA, and NFAT5, yielding a cumulative score that stratifies individuals as “likely MS cases”, “controls”, or “indeterminate”.

Table 2.

Prototype rule-based scoring algorithm for MS classification based on exon psi thresholds.

Gene	Psi	Points toward MS case	Points toward control
IFNAR2	psi>0.537	+2	0
IFNAR2	0.465<psi≤0.537	+1	0
IFNAR2	psi≤0.273	0	+1
PRKCA	psi>0.067	+1	0
PRKCA	psi≤0.016	0	+1
NFAT5	psi>24.799	+1	0
NFAT5	psi≤ 0.232	0	+1
Total score interpretation
≥3 points toward MS		→ Likely MS case
≥2 points toward control		→ Likely control
<2 toward case		→ Likely control
Mixed or ambiguous scores		→ Indeterminate; further testing recommendeda

Exon inclusion level (percent-spliced in, psi) was calculated using a fluorescent-competitive RT-PCR assay. Threshold levels were determined by supervised modeling using a LLM algorithm implemented in Rulex (https://www.rulex.ai/).

^aIn our combined cohort, 28 % of cases and 20 % of controls fell within this category.

4. Discussion

Our study confirms the contribution of AS dysregulation to the molecular architecture of RRMS and proposes a novel composite biomarker (the MS-Splicing Score, MS-SS) based on quantification of specific splicing variants in peripheral blood. By integrating “aberrant” AS profiles from IFNAR2, NFAT5, and PRKCA, we developed a logistic regression-based model that demonstrated reliable discrimination between MS cases and healthy controls across two independent cohorts.

From a biological perspective, the three AS events incorporated into our MS-SS not only demonstrated strong predictive capacity but also offer mechanistic insights into the molecular underpinnings of MS. The aberrant skipping of exon 8 in IFNAR2 may lead to the generation of a dysfunctional interferon receptor isoform, potentially impairing type I interferon signaling, an axis known to be dysregulated in MS and targeted therapeutically in RR disease [28]. Similarly, the exclusion of exon 2 in NFAT5 is likely to disrupt critical regions of this transcription factor, which modulates osmotic stress responses and T-cell activation, both relevant to autoimmune neuroinflammation [29]. Notably, Nfat5 haploinsufficiency in mice is associated with more severe neurological deficits and elevated blood-brain barrier hyperpermeability, this last one being a key pathological feature in MS [30]. Lastly, the enhanced inclusion of the non-canonical exon 3∗ in PRKCA may affect kinase domain architecture, potentially altering phosphorylation dynamics in pathways governing immune responses and synaptic plasticity [31,32]. However, we acknowledge that our study lacks functional protein-level validation of these predicted effects. Future studies should address this limitation through protein expression analysis, functional assays, and investigation of the downstream effects of these splicing alterations. This is particularly complex for PRKCA, where alternative splicing of exon 3∗ (in its canonical or extended version) can generate up to six different protein isoforms [33], each requiring individual characterization to fully understand their distinct functional contributions to MS pathophysiology. Despite this limitation, the three investigated AS events collectively converge on pathways fundamental to immune regulation and neurobiology, underscoring their dual utility as biomarkers and possible modulators of disease processes in MS.

In terms of diagnostic utility, the MS- SS offers promising discriminatory performance. In the Italian cohort, the score achieved an AUC of 0.71, while in the US cohort the performance improved to an AUC of 0.77. The observed reproducibility, even in the presence of cohort-specific variability in the significance of individual markers (e.g., PRKCA in the US cohort, being p = 0.71), suggests that multivariate integration of AS events yields more robust classification than any single isoform alone. This supports the utility of AS-based metrics as potential minimally invasive biomarkers for RRMS. Interestingly, the aberrant isoform of PRKCA, which is characterized by the inclusion of the additional non-canonical exon 3∗, has been shown to be under strong genetic control via a single nucleotide polymorphism (SNP) located at the donor splice site of this exon [33]. This SNP (rs35476409/rs61762387) modulates the splice-site recognition and isoform inclusion, functioning as a splicing quantitative trait locus (sQTL). Therefore, the limited contribution of PRKCA in the US cohort might reflect population-specific differences in allele frequencies of this regulatory polymorphism, potentially influencing the penetrance of the splicing alteration. Supporting this hypothesis, population data reveal substantial variation in the minor allele frequency of rs35476409/rs61762387 across different populations, ranging from 27.85 % in the Amish population to 31.94 % in non-Finnish Europeans, 41.81 % in admixed Americans, and 42.59 % in African Americans (data from the Genome Aggregation Database, GnomAD, https://gnomad.broadinstitute.org/; accessed August 27th, 2025). Given that our US cohort likely represents a more genetically diverse population compared to the Italian cohort, these frequency differences could explain the reduced discriminatory power of the PRKCA splicing event in distinguishing MS patients from controls in the American dataset. This observation highlights the importance of integrating genetic variability when evaluating splicing-based biomarkers across diverse populations, and suggests that population-specific calibration may be necessary to optimize biomarker performance in precision medicine applications.

Extending our analysis beyond a continuous score, we employed an LLM-based approach to identify discrete splicing thresholds that could distinguish MS patients from healthy controls with high confidence. The derived rules, combining specific AS inclusion levels of PRKCA, IFNAR2, and NFAT5, as well as the clinical algorithm built upon them, offer a novel stratification of individuals into “likely MS cases”, “controls”, and “indeterminate” categories. This threshold-based approach, while requiring further validation, highlights the potential for developing clinically actionable diagnostic rules based on specific splicing patterns.

Indeed, from the clinical point of view, the MS-SS or the LLM-based approach may represent a valuable addition to the panel of diagnostic and monitoring tools currently available for MS [22,23]. Being based on the analysis of RNA extracted from peripheral blood using a solid molecular assay such as RT-PCR, our test could be amenable to translational development. Future directions should include prospective validation in larger, treatment-naïve cohorts, assessment of longitudinal dynamics during relapse and remission, and exploration of its predictive value in therapy response. Additionally, the score could be further developed and improved through the introduction of additional AS events.

Despite the promising findings, we acknowledge the limitations of our study, particularly the relatively small sample sizes within each of the two independent cohorts. While the consistency of the MS-SS performance across both cohorts strengthens our conclusions, larger validation studies are warranted to confirm these initial findings and to fully establish the clinical utility of this novel biomarker.

CRediT authorship contribution statement

Federica Airi: Writing – review & editing, Data curation. Valeria Rimoldi: Writing – review & editing, Conceptualization. Elvezia Maria Paraboschi: Writing – review & editing, Methodology, Funding acquisition, Data curation. Valentina Pellicanò: Data curation. Damiano Verda: Writing – review & editing, Formal analysis, Data curation. Giuseppe Liberatore: Writing – review & editing, Data curation. Claudia Cantoni: Writing – review & editing, Data curation. Laura Piccio: Writing – review & editing, Data curation. Alvino Bisecco: Writing – review & editing, Data curation. Anita Capalbo: Writing – review & editing, Data curation. Giulia Cardamone: Writing – review & editing, Data curation. Eduardo Nobile-Orazio: Writing – review & editing. Giulia Soldà: Writing – review & editing, Data curation. Rosanna Asselta: Writing – original draft, Supervision, Methodology, Funding acquisition, Formal analysis, Data curation, Conceptualization.

Declaration of generative ai and ai-assisted technologies in the Writing process

During the preparation of this work the authors used ChatGPT and Gemini tools to improve language and readability. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

Declaration of competing interest

All other Authors have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper:

Rosanna Asselta reports financial support was provided by the European Union - Next Generation EU - NRRP M6C2 - Investment 2.1 Enhancement and strengthening of biomedical research in the NHS project ID: PNRR-MCNT2-2023-12378017 - Cup: E23C22000970006. Elvezia Paraboschi reports financial support was provided by Fondazione Regionale per la Ricerca Biomedica, project number: 1734478. Alvino Bisecco received speaker’s honoraria and/or compensation for consulting service and/or speaking activities and/or travel grants from Biogen, Roche, Novartis, Merck, Alexion, Amgen, Argenx, UCB and Genzyme.

Handling Editor: Dr Y Renaudineau

Data availability

Data will be made available on request.