Multiple sclerosis (MS) affects >2.8 million people globally.¹ Although clinical descriptors categorize MS into relapsing–remitting, secondary progressive and primary progressive, emerging evidence reveals that these fail to capture the underlying biological continuum.^2,3 Although distinct in their disease burden, these clinical phenotypes share pathobiological mechanisms.^4,5 As a result, our capacity to predict its course and personalize treatment remains limited by our reliance on traditional clinical descriptors. A shift towards a subtyping system grounded in pathobiological underpinnings promises to facilitate earlier, more targeted therapeutic interventions, potentially even before symptom onset, improving our ability to predict individual patient outcomes and personalize treatment strategies.

Unsupervised machine learning offers an objective way to uncover MS types based on underlying disease biology (pathobiology) rather than relying on predefined diagnostic labels or observable symptoms. For example, Gross et al.⁶ identified three distinct data-derived MS types based on different immune cell markers of CD4 and CD8 T cells, natural killer cells and cytokine compartments. These data-derived subtypes showed subtle variations in disease course and treatment responses. However, their immediate clinical use is limited because specialized immune cell markers defined these subtypes, which are not collected routinely in clinical practice.

MRI provides detailed spatial information about structural damage and lesion formation, capturing the localized effects of the disease. The unsupervised learning model Subtype and Staging Inference (SuStaIn⁷) was used previously by Eshaghi et al.⁸ on brain MRI data from 9390 MS patients to identify three distinct MRI-derived subtypes. SuStaIn clusters patients with similar progression patterns, characterizing subtypes by the alternative order in which key variables reach thresholds of abnormality. They found a ‘lesion-led’ subtype that was more responsive to high-efficacy treatments in clinical trials and had faster worsening of disability than the other two subtypes.⁸ However, MRI alone does not fully capture the underlying neuro-axonal damage that drives disease progression. MRI is non-specific for underlying pathology and can miss subtle or ongoing neurodegeneration.⁹ Integrating widely available fluid biomarkers indicative of early and ongoing neuronal injury has the potential for improved separation of patient subgroups and enhanced precision in disease course prediction and personalized treatment selection.¹⁰

The serum neurofilament light chain (sNfL) level is a systemic measure of ongoing neuronal injury, reflecting disease activity and neuronal loss.^11,12 It is a component of the neuronal cytoskeleton and, therefore, a sensitive indicator of neuro-axonal damage. In MS, the sNfL level increases at disease onset,¹³ is associated with MRI changes of disease activity and neurodegeneration¹⁴ and is sensitive to treatment effects.^15,16 Serum NfL is becoming increasingly available and accessible in clinical practice.^17,18 Yet, because serum-based biomarkers alone still face variability, low sensitivity to disease activity and imperfect assay standardization, incorporating complementary data (such as MRI metrics) can strengthen their clinical utility.^19-21

We hypothesized that integrating sNfL with MRI would yield biologically distinct MS subtypes with improved prognostic accuracy in comparison to MRI alone. In this study, using two independent cohorts, we aimed to: (i) develop and validate a combined MRI–sNfL subtyping system using unsupervised machine learning (SuStaIn); (ii) assess how incorporating sNfL can simplify and improve data-derived subtypes in comparison to using MRI alone; and (iii) evaluate how these novel subtypes evolve in their brain atrophy, treatment response and new lesion development.

This study identified two data-derived MS subtypes, distinguished by the timing of sNfL elevation within an evolving landscape of MRI-derived abnormalities. Termed ‘early-sNfL’ and ‘late-sNfL’ subtypes, these categories reflect distinct biological profiles: the early-sNfL subtype displayed elevated sNfL levels, compromised corpus callosal integrity and lesion accrual early in the disease, consistent with active inflammatory and neurodegenerative processes manifesting in parallel. In contrast, the late-sNfL subtype began with tissue-specific volumetric loss (notably in the limbic cortex and deep grey matter) before sNfL levels became abnormal, suggesting a more insidious trajectory of neurodegeneration that precedes overt neuroaxonal injury. By integrating MRI and sNfL measures in a single unsupervised model, we have defined biologically grounded MS types that capture diverse disease pathways and their clinical implications.

We trained our model on a cohort that included both relapsing–remitting and secondary progressive MS (mean disease duration 7.7 years), then validated it in a larger, younger cohort of newly diagnosed MS (including clinically isolated syndrome), confirming generalizability at early stages. Our design identified subtypes spanning the entire disease course, essential for SuStaIn model training. As expected, the younger test cohort had more active disease by gadolinium-enhancing lesion count. Although these cohorts differed demographically and clinically (a typical source of bias in machine-learning models), SuStaIn differs from typical models not specifically designed for disease progression modelling and remains robust. SuStaIn assigns both a subtype and a stage to each patient, and therefore, our design and choice of cohorts with demographic differences allowed us to validate both the subtyping and staging by the model. Integration of sNfL with MRI improved correlations with disability (EDSS) and inflammation (enhancing lesions and sNfL), demonstrating that sNfL provides complementary information.

The newly identified MRI–sNfL subtypes demonstrated demographic and biological differences. The late-sNfL subtype included more women and older individuals, whereas the early-sNfL subtype comprised younger patients with more active disease. In comparison to the late-sNfL group, early-sNfL patients showed greater reductions in active lesion counts and sNfL levels over time, particularly among those receiving treatment. Although treatments such as interferons and evobrutinib have shown limited efficacy in clinical trials, the observed differences in treatment response across subtypes support the biological relevance of our subtyping approach. In both datasets, the early-sNfL subtype exhibited faster brain atrophy rates, probably reflecting more severe neurodegeneration driven by inflammation. Together, these findings suggest that the MRI–sNfL subtypes capture meaningful differences in disease activity, treatment response and progression risk.

Various pathophysiological biomarkers to classify MS disease course based on biological understanding have been proposed.^4-6,27,44,45 Recent studies investigated MRI-based MS types with distinct patterns of brain volume loss correlated with clinical outcomes and sNfL independent of traditional clinical phenotypes, but have not used sNfL in subtyping tasks.⁴⁵ Although MRI provides unique spatial information on the pathological spread, it lacks pathological specificity.⁹ Our new MRI–sNfL model represents a step change from our previous MRI-only approach, which offered valuable but incomplete insights into MS heterogeneity. MRI-based clustering revealed distinct subtypes driven by anatomical and microstructural features, but it could not fully capture the physiological processes driving disease activity and progression. By adding sNfL (an established indicator of neuroaxonal injury), we have advanced beyond the structural snapshot provided by MRI alone. Overall, our results support the added value of sNfL to MRI. When looking at the addition of sNfL to MRI in two MRI-only models (one previously published and one with the same five MRI variables), the MRI–sNfL model provided the highest correlation with active lesions, similar correlations with disability, but better generalization when comparing correlation coefficients in the training and external test set (the correlation coefficients declined for all models, but the absolute correlation coefficients remained highest for MRI–sNfL model). The decline in EDSS correlation between all models is attributable to the limited range of EDSS in the external test cohort of newly diagnosed MS and clinically isolated syndrome patients. In addition, stronger EDSS correlations in the training set compared with the testing set are observed as expected, owing to the use of EDSS-based feature selection within the training dataset.

sNfL has been linked to both active and chronic inflammation in earlier studies.^21,46-49 The most consistent finding is its association with gadolinium-enhancing lesions, although radiological activity and sNfL rises do not always coincide.²¹ Srpova et al.⁴⁸ also showed that, in some patients, sNfL elevation precedes brain atrophy. Our results corroborate the link between active lesions and high sNfL and newly show that elevated sNfL helps to stratify patients by inflammatory profile and the temporal evolution of sNfL and MRI abnormalities in subgroups of patients.

In training and external test datasets, the early-sNfL patients showed similar patterns compared with the late-sNfL (younger in early-sNfL and more women in the late sNfL) and, importantly, a similar proportion of patients with relapsing–remitting and secondary progressive MS in the training set with these two phenotypes. Our results underline that biology-grounded MS subtypes are largely independent of clinical course descriptors. These add to the evidence that biofluid biomarkers complement MRI to define the biological basis of MS disease evolution better.^12,14,27,50 Longitudinal application of the model allows us to assess how subtype and stage assignments change over time. Although the model does not explicitly allow for deterministic transitions between subtypes, we occasionally observe ‘subtype switching’, whereby the most probable subtype assignment of a patient changes between visits. We interpret this as reflecting either uncertainty in subtype classification for borderline cases or genuine overlap between trajectories. Given that 7% of patients switched from one subtype to another in the training dataset, and 23% switched in the testing dataset, these subtypes are likely to represent a continuum of underlying pathology. SuStaIn captures this with probability-based membership, evidenced by the reduction in subtype switching when membership certainty was increased. Data-derived subtyping can, therefore, impact future disease course descriptions and prognosticate MS outcomes. From a clinical perspective, we can speculate that this might indicate a shift from a dominant inflammatory profile to a more neurodegenerative phase or vice versa. Future research with longer follow-ups can clarify this.

In the training cohort, patients classified in the early-sNfL subtype displayed a more rapid decline in gadolinium-enhancing lesions when treated, whereas those in the late-sNfL subtype did not show this rapid reduction. Interestingly, this subtype-specific effect was not seen in the control (untreated) group, suggesting that early-sNfL patients respond more robustly to treatment regarding reduction in active lesions. Conversely, in the test cohort, early-sNfL patients exhibited a faster decline in brain volume compared with late-sNfL patients, a finding not observed in the training data. Across both the training and test cohorts, early-sNfL patients also showed a higher likelihood of new lesion formation relative to late-sNfL patients. The inferred ‘stages’ of the MRI–sNfL model are a proxy for pathological accumulation that showed stronger associations with both EDSS and active lesions than our MRI-only model.

Within the treated cohorts, the early-sNfL subtype had a rapid gadolinium-enhancing lesion reduction, suggesting a more pronounced and immediate therapeutic effect in this subgroup. In contrast, the late-sNfL subtype did not have such a drop, indicating potential differences in the underlying disease mechanism. Meanwhile, in patients who were not on treatment, serum NfL levels still declined in both early- and late-sNfL subtypes over the course of the study. We attribute this reduction (despite the lack of active treatment) to the eligibility criteria of the clinical trials, which required participants to have ongoing disease activity (recent relapse or enhancing lesions). Consequently, this decrease might reflect ‘regression to the mean’ once the most active phase of inflammation subsides rather than a true therapeutic effect.⁵¹

Correlations between our model stages and EDSS were weak, mirroring the many reports of only weak biomarker–EDSS associations. This is expected, because EDSS is weighted towards motor function and captures only a narrow slice of MS burden, whereas MRI and serum biomarkers typically change before clinical symptoms become evident. Hence, attenuated cross-sectional correlations do not undermine the clinical promise of the model. Instead, they highlight the need for longer prospective studies that pair the model with broader outcome measures (cognition, quality of life and other patient-reported domains) beyond EDSS alone. Nonetheless, we can speculate that our model can provide staging of MS to facilitate future early interventions (before disability emerges) and simultaneously stratifies patients by their biomarker profiles, with potential to guide personalized therapy.

We should address several hurdles to bring our research model into everyday clinical care. Clinicians first need tools that convert routine MRI scans into precise measures of the brain structures of each patient. Although new tools are becoming available for brain MRI processing of real-world data,⁵² few hospitals have the infrastructure for these tasks. Even where such tools exist, the variability of everyday scans introduces new heterogeneity that demands further study. As sNfL is becoming widely available and digital infrastructure improves, we expect our multimodal model to evolve into future decision aid systems after further rigorous research on real-world data. Because subtype assignment is possible from a single cross-sectional scan, the model remains usable even when routine follow-up is sparse; our longitudinal analyses demonstrate that these baseline labels carry prognostic weight over time. Although harmonization can reduce variability across imaging sites, it also creates practical challenges for clinical adoption. Notably, the MRI–sNfL model retained strong performance even without harmonization (see Supplementary Table 8), supporting its use in settings with limited data or infrastructure.

Although the MRI–sNfL model provides an interpretable framework for subtyping and staging MS, it operates under several key assumptions. The SuStaIn algorithm models disease progression as a monotonic sequence of biomarker changes, with each subtype following a fixed order of abnormality accumulation. This assumption allows for tractable modelling of complex data but might limit sensitivity to fluctuating trajectories. To manage uncertainty, SuStaIn uses a Bayesian framework with Markov chain Monte Carlo sampling, which enables probabilistic estimates of both subtype and stage assignments. Although recent extensions of SuStaIn have introduced methods for accommodating missing data,⁵³ all participants in our study had complete biomarker profiles, and missing data handling was not required here. These modelling assumptions should be considered when interpreting subtype assignments, particularly in broader clinical settings.

This study has several limitations. First, we drew our training and testing samples from clinical trial cohorts, which do not fully represent the broader MS population, including those with comorbidities or underrepresented ethnic groups and primary progressive MS, with strict eligibility criteria. As a result, for example, disability range was limited in the external testing dataset. Despite this, our MRI–sNfL model stages showed significant correlations with disability, showing the added value even with limited EDSS ranges. Therefore, future studies in diverse clinical settings are required for clinical translation. Second, our model was trained on both relapsing–remitting and secondary progressive cases, and is trained across the MS continuum, yet its accuracy in late-stage disease still warrants prospective validation, because our testing cohort was limited to early MS. Third, although adding sNfL enhanced the biological relevance of the model and, in some cases, clinical correlations, other fluid biomarkers or advanced imaging modalities (such as myelin-sensitive MRI sequences) will provide more comprehensive insights into disease progression. However, addition of more advanced measures reduces the accessibility and introduces obstacles to eventual clinical translation.

Our findings advance the understanding of MS heterogeneity by revealing distinct biological trajectories rooted in MRI and fluid biomarkers. sNfL, despite being a non-specific marker related to neuronal cytoskeleton, complements MRI measures of disease activity and neurodegeneration. Integrating sNfL with MRI refines subtyping and provides a foundation for earlier, more individualized prognosis. Ultimately, this approach might pave the way for more targeted therapeutic strategies and improved patient outcomes.