The combination of CSF neurofilament light chain and glial fibrillary acidic protein improves the prediction of long-term confirmed disability worsening in multiple sclerosis Authors: Simon Thebault, Seyed-Mohammad Fereshtehnejad, Hailey P Bergman, Gautier Breville, Mohammad Abdoli, Ronald A Booth, Giulia Fadda, Mark S Freedman, Gauruv Bose Published: 2024 ABSTRACT AbstractOur objective was to evaluate the individual and combined prognostic attributes of baseline serum and CSF measurements of Neurofilament light chain (sNfL, cNfL) and glial fibrillary acidic protein (sGFAP, cGFAP) on long term clinical outcomes in MS. In this retrospective single center study, patients with serum and CSF stored at first MS presentation and > 15-years of follow-up were analyzed. NfL and GFAP were quantified from cryopreserved samples using a digital immunoassay and analyzed as predictors of confirmed disability worsening (CDW). Sixty patients (70% female) underwent baseline tandem CSF and serum sampling and were followed for a mean of 17.8 years (SD 2.5). 32 developed CDW. By logistic regression, while sNfL cNfL and cGFAP showed prognostic merit (AUC 0.72, 0.70, 0.62 respectively), sGFAP did not (AUC 0.5). The combination of cNfL and cGFAP improved CDW prediction compared to either measure considered in isolation (AUC 0.72). The optimal predictive cut-off for CDW (Youden’s index) for cNfL was 596 pg/mL and for cGFAP was 8160 pg/mL. Kaplan–Meier analysis of the cutoff-defined ‘high-high’ and ‘low-low’ combined cNfL and cGFAP groupings improved prediction of CDW compared to either marker individually (Hazard ratio 4.5 (95% CI 2.7–18.3), Logrank P < 0.0001). Cox Proportional Hazards regression demonstrated that high baseline cNfL and cGFAP were independently prognostic of subsequent CDW after adjusting for baseline age, sex, EDSS score and subsequent treatment exposure. Each unit increase in Ln(cNfL) and Ln(cGFAP) was respectively associated with an additional hazard of 2.36 (95% CI 1.12–5.52) and 2.26 (95% CI 1.03–5.21). CSF NfL and GFAP are independently prognostic of long-term clinical worsening in MS, and may represent a complementary pairing.Subject terms: Prognostic markers, Multiple sclerosis, Prognostic markers Introduction Evidence supports a role for the neuroaxonal protein biomarker neurofilament light chain (NfL) as a useful marker predictor of multiple sclerosis (MS) disease activity and clinical worsening. Already adopted in clinical practice at some centres1NfL can be reliably measured in serum and CSF2 (sNfL and cNfL, respectively), where levels closely correlate. sNfL has proven to be a convenient and objective short-term predictor and early indicator of emerging relapse biology characterised by focal inflammatory lesions seen on MRI and clinical relapses2,3, which may allow for pre-emptive treatment escalation. Even when relapse biology is well controlled, recent studies have also identified that more subtle NfL elevations precede clinical worsening associated with progressive non-relapsing biologies4–6. Age has been identified as the most important confounder of sNfL, requiring either age-specific cut-offs or adjustment before clinical interpretation of one-off values7,8. NfL may also have a role in longer-term prognostication with several groups showing group-level associations between baseline levels and clinical and MRI outcomes more than a decade later, including disability worsening, conversion to progressive MS and brain atrophy9,10. A paucity of evidence for NfL as a long-term prognosticator is in part due to the lack of available long-term cohorts, particularly those that pre-date higher efficacy therapies that have the potential to alter an individual patient’s risk for clinical worsening and obscure the long-term prognostic utility of a single baseline biomarker measurement. Better long-term prognostication for patients with MS is desirable; for instance, in identifying ‘aggressive’ or ‘benign’ patients, such that initial treatment intensity can be appropriately selected. There has been growing interest in how other protein biomarkers might supplement the information provided by NfL. Glial Fibrillary Acidic Protein (GFAP) is one marker of particular interest. Akin to NfL, GFAP is an abundant intermediate protein that is measurable in both CNS and blood, and is elevated in MS patients as a non-specific marker of CNS injury11,12. GFAP is almost exclusively expressed by activated astrocytes, another cell type particularly implicated in progressive MS biology13. The association between blood and CSF GFAP levels is weaker than for NfL, with some groups identifying CSF rather than serum levels being the most closely associated with MS outcomes14. While GFAP is cross-sectionally associated with clinical severity11, progressive disease and overall MRI lesion burden11,12,15, it is much less affected by acute focal inflammation compared to NfL4. CSF (and blood) are often obtained at the time of diagnosis in MS to test for the presence of oligoclonal bands. In this study, we wanted to assess the prognostic utility of measuring baseline NfL and GFAP in tandem from both fluid compartments in a cohort followed up for more than 15 years. We hypothesized that CSF measures of NfL and GFAP may complement each other as predictors of confirmed disability worsening (CDW). Subjects/materials and methods Study population We used biological samples (serum and CSF) from the Ottawa MS biobank1 and follow-up data from the MS clinic in Ottawa, Ontario, Canada since 1994. For this study, we were interested in samples from patients who had blood and CSF sampled at first clinical presentation as part of a diagnostic workup for whom we subsequently had at least 15 years of clinical follow-up data. Written informed consent had been obtained before inclusion in the Ottawa MS biobank. The study protocol was approved by the Ottawa Hospital Regional Ethics Board (OHSN-REB-20180518-02H). At each visit to the Ottawa MS clinic (including baseline sampling and follow-up visits), clinical data was collected including MS-related features and their dates of onset, as well as Expanded Disability Status Scale (EDSS) scores, and any treatments given. Clinical disease subtype including clinically isolated syndrome (CIS), relapsing–remitting MS (RRMS), secondary progressive MS (SPMS), or primary progressive MS (PPMS) were confirmed in retrospect using the 2017 McDonald criteria2. Other eligibility criteria included the absence of clinical or MRI disease activity in the 3 months preceding sampling, the availability of tandem CSF and serum paired sample within the first 5 years of symptom onset and least 15 years of clinical follow-up from sampling. Sample collection, storage, CSF NfL and GFAP measurement Serum samples were initially collected in serum tubes, and serum was separated following coagulation by spinning at 2000 × g for 10 min. Paired CSF samples were collected in glass tubes following lumbar puncture. Serum and CSF were then aliquoted before immediate storage in cryovials at -80 degrees Celsius. Before analysis, aliquots not previously thawed were selected and coded by a laboratory technician, blinding investigators to patient-specific details; unblinding occurred only after analysis. All samples were thawed, processed, and assayed as a single batch. Serum and CSF levels of NfL and GFAP were quantified using commercially available 4-PLEX immunoassay kits (Quanterix, cat#102,153) run on the fully automated ultrasensitive Simoa HD-X Analyzer (Quanterix). Samples were run in duplicate following manufacturers’ instructions with appropriate standards and internal controls. Clinical outcomes Patients were followed every 4–12 months, depending on clinical need. Chart review was performed to ascertain the binary presence or absence of Confirmed Disability Worsening (CDW)14 over the follow-up. CDW was defined as an increase in the EDSS score on at least two sequential visits more than 6 months apart as follows: At least 1.5 points for baseline EDSS of 0 At least 1.5 points for baseline EDSS of 0 At least 1 point for baseline EDSS between 1 and 5 At least 1 point for baseline EDSS between 1 and 5 At least 0.5 points for a baseline EDSS of at least 5.5. At least 0.5 points for a baseline EDSS of at least 5.5. We also recorded demographic details, treatment exposures and the time-to-event from MS onset to when the CDW was reached. Statistical analysis All univariate and multivariate analyses and statistical modeling were implemented using IBM SPSS Statistics software (version 23.0), Microsoft Excel 365 and Prism v.8. For all analytical procedures, a two-tailed P-value < 0.05 was considered statistically significant for the corresponding comparison or association. Data description Normality of numeric data was assessed using the D’Agostino- Pearson test. We used mean and standard deviation (SD) to describe normally distributed variables, whereas median and interquartile range (IQR) were applied to describe non-normal data. Frequency percentages were used for the description of categorical features. To assess for potential bias introduced by missing samples on patients who otherwise met inclusion criteria, we compared baseline demographic and clinical features of these patients with those for whom samples were available. Univariate analysis For univariate comparisons of normally distributed variables, we used independent samples t-test and one-way ANOVA where appropriate. For non-normally distributed variables, we used Mann–Whitney test, and Kruskal–Wallis test. In addition, Chi-square (and Fisher’s exact) test was used for between-group comparisons of categorical variables. Pearson and Spearman correlation was used to assess bivariate association between various biomarker levels where appropriate, and the corresponding correlation coefficients were calculated. Logistic regression to define biomarker cutoffs and combinations To assess the discriminative accuracy of baseline serum or CSF NfL and GFAP to predict CDW, we conducted an ROC analysis by calculating the area under the curve (AUC) and its 95% confidence interval (CI) and p-values for each biomarker. Youden’s index was used to identify unadjusted biomarker cutoffs for both NfL and GFAP. These cutoffs were also used to define combined groupings of cNfL and cGFAP; namely, patients categorized as having high levels of both markers were defined as ‘high-high’, whereas patients categorized as having low levels of both markers were defined as ‘low-low’. Logistic regression was also used to combine NfL and GFAP for ROC analysis with follow-up truncated to 15 years for consistency and CDW set as the dependent variable. Kaplan Meier survival analysis cut-off-defined and combined biomarker groupings This was used to compare the average time to develop CDW based on baseline CSF biomarker levels with Log-rank(Mantel-Cox) curve comparison, reporting P-values and hazard ratios. Cox proportional hazards regression of biomarkers and baseline covariates Scaled Schoenfeld residuals vs. time plots first tested the validity of the proportional hazards assumption. Hazard ratios were calculated for CDW based on baseline NfL and GFAP levels as well as other covariates of interest (baseline age, sex, EDSS score and DMT exposure), leveraging the long-term outcome data available on our cohort by including complete time-to event information and time-to-event information. NfL and GFAP levels were assessed both as dichotomous high/low Youden’s index-defined categorizations, and as Ln-transformed continuous variables. Standard protocol approvals, registrations, and patient consents Written informed consent had been obtained before inclusion in the Ottawa MS biobank. Approval was received from the Ottawa Health Science Network Research Ethics Board for the measurement of serum NfL in this cohort of patients. All methods were performed in accordance with the relevant guidelines and regulations of our jurisdiction, ethics board and institution. Data availability statement Consideration will be made by the authors regarding the sharing of anonymized data related to this study with qualified investigators. Results Patient cohort and baseline demographics Of 3840 patients included in the Ottawa MS biobank screened for inclusion in this study (Fig. 1), 131 patients met criteria for sampling within 5 years of onset with at least 15 years of clinical follow-up. From these, 60 patients had paired serum and CSF, 70% were female, with mean age 39 years (SD 9) at time of enrollment and the median EDSS score was 2 (IQR 1.5–2.5, range 0–4 Table 1). Of the 71 patients who met inclusion criteria but without paired samples, 70% were female, the mean age of patients was 40 (standard deviation 11) years and the median EDSS score was 2 (IQR 1). None of the patients had active clinical relapses or evidence of new radiographic disease activity in the 3 months preceding sampling. After the initial workup was completed (within the first year after initial serum and CSF sampling), 42 patients (80%) met retrospectively applied 2017 criteria for RRMS, and 18 (20%) for PPMS. With a mean follow-up duration of 17.8 (range 15.5–25.7) years, 35% of the initial CIS/RRMS patients phenoconverted to SPMS, and the median EDSS progressed to 3.0 (IQR 1.4–6, range 0–10: one patient with progressive MS died as a complication of severe progressive MS after 16.4 years follow-up). Regarding disease-modifying therapy (DMT) exposure, the long period between initial sampling (1994–2007) and the most recent clinical follow-up spanned an era when disease-modifying therapies were either not available or just coming on the market in Canada: 43 patients (73%) in this cohort never received DMT. Of those that did receive DMT, three had received this at the time of sampling, and the majority received injectable therapies only (interferon b-1a/b or glatiramer acetate). One patient with RRMS underwent autologous myeloablative hematopoietic stem cell transplantation (HSCT) after having ongoing disease activity on two disease modifying therapies, and subsequently had an improvement in her EDSS score from 5 to 4, which subsequently did not progress. One patient received natalizumab and 2 patients received mitoxantrone after failing first line treatments. Patient selection from the Ottawa MS biobank for inclusion in this study. Table 1Demographics and clinical features of study participants at baseline and the end of follow-up.Demographic featuresMS Patients (n = 60)Female sex–n (%)42 (70%)OCB positivity–n (%)50 (83%)Age (years)–mean (SD)39 (9)Initial disease course–n (%) CIS17 (28%) RRMS31 (52%) PPMS12 (20%)EDSS at baseline–median (IQR)2 (1)Prior DMT exposure–n (%)3 (4.7%)Follow-up duration–mean (SD)17.8 (2.5)Disease course at last follow-up–n (%) CIS3 (5%) RRMS32 (53%) SPMS12 (20%) PPMS12 (20%)EDSS at last follow up–median (IQR)3 (4.3)DMT exposure–total number of patients (%)17 (28%) Injectable15 (25%) Oral5 (8%) Infusion1 (2%) Other3 (5%)DMT Disease modifying treatment, EDSS Expanded Disability Status Scale, IQR Interquartile range, OCB Oligoclonal bands, SD Standard deviation. Demographics and clinical features of study participants at baseline and the end of follow-up. DMT Disease modifying treatment, EDSS Expanded Disability Status Scale, IQR Interquartile range, OCB Oligoclonal bands, SD Standard deviation. Baseline biomarker and clinical correlations Figure 2 summarizes pertinent baseline biomarker and clinical covariate Spearman correlations. We replicated a strong bivariate correlation between the serum and CSF NfL levels with a coefficient of 0.69 (P < 0.0001). We did not find a correlation between sGFAP and cGFAP (r = 0.19, P = 0.15). Whilst both sNfL and cNfL moderately correlated with cGFAP (r = 0.38, P = 0.0031 and r = 0.43, P = 0.00056 respectively), sGFAP correlated weakly with sNfL (r = 0.27, P < 0.040) and not cNfL. cGFAP negatively correlated with female sex (r = − 0.33, P = 0.01). Otherwise, although there was a positive correlation between baseline age and EDSS score (r = 0.4, P < 0.0014), age was not identified as a covariate of any baseline serum or CSF biomarker level in our cohort. Spearman correlation between serum and CSF levels of NfL and GFAP measured at baseline. NfL but not GFAP correlated well between serum and CSF. Only CSF GFAP correlated with NfL levels in both compartments. In our cohort, neither NfL or GFAP correlated with age. Legend: c: CSF GFAP: glial fibrillary acidic protein, NfL: neurofilament light chain, s: serum. We further explored the parametric correlations of NfL and GFAP between serum and CSF biofluid compartments following log transformation, again demonstrating a stronger correlation for NfL (r = 0.86, P = 0.0002) than GFAP (r = 0.5, P = 0.005) (Supplementary Fig. 1). Given the weak correlations with sGFAP and other covariates in our study, and concerns regarding GFAP stability over long term storage, we assessed the effect of sample collection year on baseline biomarker levels: spanning the 13 years from 1994 to 2007 (patients had at least 15 years of subsequent follow-up), we found no association between baseline biomarker values and collection year (Supplementary Fig. 2). Logistic regression to define biomarker cutoffs and combinations Thirty two out of 60 patients developed CDW at 15-year follow-up. Unadjusted sNfL and cNfL both showed similar diagnostic accuracy to predict CDW as a binary outcome (sNfL AUC = 0.70, 95% CI 0.57–0.84, P = 0.0077; cNfL AUC 0.68, 95% CI 0.55–0.82, P = 0.015; Fig. 3A). The optimal cut-off for sNfL was of 8.4 pg/mL (Youden’s index, experienced by 37/60 patients), which was associated with an 81% sensitivity and 61% specificity of developing CDW, positive likelihood ratio of 2.1. Similarly, a baseline cNfL greater than 596 pg/mL (experienced by 35/60 patients) was associated with a sensitivity of 69% and specificity of 61% of developing CDW, likelihood ratio of 1.8. For GFAP, only CSF measures significantly predicted CDW with an AUC of 0.62 (P = 0.032, Fig. 3-B); a baseline GFAP greater than 8160 pg/mL (experienced by 19/60 patients) was associated with a sensitivity of 47% and specificity of 82% of developing CDW, positive likelihood ratio of 2.6. Baseline sGFAP did not predict CDW in our cohort. Receiver operating characteristics (ROC) curves analysis to assess the diagnostic accuracy of unadjusted serum and CSF (A) NfL, (B) GFAP and (C) the combination of cNfL and cGFAP to predict CDW at 15 years of follow-up. sNfL and cNfL conferred similar diagnostic accuracy (A), while cGFAP but not sGFAP was contributory (B). The combination of cNfL and cGFAP improved the AUC compared to either variable considered in isolation (C). Legend: AUC: area under curve (95% confidence interval), c: CSF, CDW: confirmed disability worsening, GFAP: glial fibrillary acidic protein, NfL: neurofilament light chain, s: serum. Combining the predictive models for cNfL and cGFAP by multiple logistic regression improved the AUC compared to either of these variables considered in isolation (Fig. 3.C, AUC 0.72, 95% CI 0.59–0.84). Adjusting each of the sNfL, cNfL and cGFAP models in turn for age as a covariate, the predictive model marginally improved (sNfL, cNfL and cGFAP ROC 0.75, 0.75, and 0.74, not shown). In the combined NfL and GFAP model, replacing cNfL with sNfL with did not improve accuracy (AUC 0.72, 95% CI 0.57–0.86, not shown), hence for simplicity and constancy of CSF as the biofluid source and tandem measurement, subsequent combined analyses focussed on cNfL and cGFAP. Adding age to the combined cNfL and cGFAP model further improved CDW prediction (AUC 0.77, 95% CI 0.61–0.87, P = 0.0004, not shown). Finally, in a combined model of baseline cNfL, cGFAP, age, EDSS score and DMT exposure both cNfL and cGFAP remained independent predictors of CDW (AUC 0.82, 95% CI 0.72–0.93, P < 0.0001, not shown). Kaplan Meier survival analysis cut-off-defined and combined cNfL and cGFAP groupings We used the Youden’s index cutoffs derived previously for cNfL (596 pg/mL) and cGFAP (8160 pg/mL) to define combined cNfL + cGFAP groups of patients. Patients with both cNfL and cGFAP both above or below the respective cutoffs were categorized as high + high(n = 17) and low + low (n = 24) respectively. 14/17 patients in the high + high group developed CDW by the end of the follow-up, as opposed to 9/24 in the low-low group (Fishers Exact P = 0.01, positive likelihood ratio 3.5). Kaplan Meier survival analysis was used to compare the time to develop CDW associated with ‘high’ (red line) versus ‘low’ (blue line) cNfL (Fig. 4A) and cGFAP (Fig. 4B) and combined ‘high + high’ versus ‘low + low’ cNfL + cGFAP groupings at baseline. While higher cNfL and cGFAP were separately prognostic of time-to-CDW, patients with both high cNfL and high cGFAP had the most rapid time to, and frequency of CDW (HR 4.5, 95% CI 2.7–18.3, log rank P = 0.0001). The median time-to-CDW was 6.1 years or the high-high group and 15.0 years for the low-low group. Kaplan Meier survival curves and median survival times of MS patients to develop CDW based on cutoff-defined high (red curves) versus low (blue curves) biomarker levels of (A) cNfL, (B) cGFAP and (C) the combination of cNfL + cGFAP over the follow-up period. High versus low cNfL (A) or cGFAP (B) were individually prognostic of time-to-CDW. The combination of both high cNfL and high cGFAP (as opposed to low levels of both markers) showed a more marked separation in the survival curves (C). P-values and hazard ratios represent the logrank comparison between high and low groupings. The black dotted line denotes the survival of the total cohort. Legend: CDW: confirmed disability worsening, cGFAP: CSF glial fibrillary acidic protein, cNfL: CSF neurofilament light chain, HR: logrank hazard ratio. Cox proportional hazards regression of cNfL, cGFAP and baseline covariates Cox proportional hazards regression was used to further evaluate the long-term prognostic value of cNfL and cGFAP to predict CDW in our cohort while considering baseline covariates and complete time-to-event information. Both cNfL and cGFAP met the proportional hazard assumption. High baseline cNfL and cGFAP were independently associated with increased hazard of developing CDW, expressed either as continuous ln-transformed concentrations (Fig. 5A) or dichotomous Youden’s index-defined ‘high’/’low’ (1/0) categorizations (Fig. 5B), in both models remining significant after adjustment for sex, baseline age and EDSS score as well as subsequent disease-modifying treatment (DMT) exposure. In the continuous model, each unit increase in Ln(cNfL) and Ln(cGFAP) was associated with 1.81 (95% CI 1.19–2.72) and 1.52 (95%CI 1.08–2.78) hazard ratio of CDW. In the dichotomous model, high cNfL and GFAP were associated with a 2.36 (95% CI 1.12–5.52) and 2.26 (95% CI 1.03–5.21) hazard ratio of CDW. In both the dichotomous and continuous models, while baseline EDSS score was associated with increased risk of subsequent CDW, female sex and DMT exposure were protective. While there was a trend towards increased age decade at sampling and worse CDW outcomes, this did not reach significance in either model. Cox proportional hazard models of CDW prognostication based on baseline biomarker measurements, age, sex, EDSS score and subsequent DMT exposure with cNfL and cGFAP evaluated as (A) continuous Ln-transformed biomarker concentrations, (B) dichotomous variables (1/0) based of Youden’s index-defined high/low cutoffs, and (C) combined dichotomous model based on ‘high-high’ versus ‘low-low’ biomarker groupings. High concentrations of either cNfL or cGFAP was independently prognostic of CDW in both the continuous (A) and dichotomous (B) models after adjusting for age, sex, EDSS score and DMT exposure. In the combined dichotomous model, patients with high (as opposed to low) levels of both biomarkers had a greater hazard of developing CDW (C). Legend: CDW: confirmed disability worsening, cNfL: CSF neurofilament light chain, cGFAP: CSF glial fibrillary acidic protein, EDSS: expanded disability status scale; DMT: disease-modifying treatment. In a final Cox proportional hazards model, we compared hazard of CDW based on the combined groupings of ‘high’high’ (n = 17) vs ‘low-low’ (n = 34) cNfL and cGFAP defined previously; patients with high as opposed to low levels of both biomarkers had a 4.02 (95% CI 1.56–10.69) hazard ratio of CDW, approximately twice that of either marker considered in isolation (Fig. 5C). Discussion In this single-center, long-term cohort of MS patients with paired serum and CSF sampled as part of their initial diagnostic evaluation, we show that the combination of baseline cNfL with cGFAP aids in predicting prognosis 15 years or more later. We focused on the combination of baseline CSF measures rather than serum because serum GFAP was not contributory in our analyses. Highlighting the prognostic complementarity of these two proteins, both markers independently retained significance in multivariate models including after adjustment of confounders. While cNfL was more sensitive, cGFAP was more specific for predicting CDW, with their combination improving prognostication compared to either marker in isolation. These associations remained significant after adjustment for baseline age, sex, EDSS score and DMT exposure. While patients with high levels of both fared particularly poorly, patients with low levels of both markers had a more favorable disease trajectory. Our results add to a growing body of literature that highlights the combination of these protein markers may help identify patients at high and low risks of clinical worsening, which in turn could assist in earlier and more precise treatment optimization for patients with MS. Since newly presenting patients often undergo lumbar puncture for diagnostic purposes, our findings suggest that biomarker analysis from this opportune CSF sampling might have utility in long-term prediction and possible treatment decision making. Several groups are now investigating the possible role of blood-derived (serum) GFAP as a marker in MS, with promising emerging data15,16. However, our study found that serum and CSF GFAP levels did not correlate well in paired samples from 60 MS patients and that only cGFAP showed prognostic merit. Several groups have reported the correlations of GFAP levels using the same commercially available assay in discrete cohorts, with some representative reported r values including 0.3217, 0.4518, 0.5111, and 0.619. The correlation in our analysis was in the middle of this range of reported values, which are consistently lower than correlations reported for serum and CSF NfL. One possibility is that there is a pre-analytical or analytical issue measuring low levels of sGFAP levels in cryopreserved samples with the commercially available assay that our group and others used. To this end, we assessed biomarker levels by year of sampling in our cohort, and did not find an effect of sample storage duration on NfL or GFAP in the serum or CSF spanning a 13-year period from 1994 to 2007. Alternatively, there could be biological explanations for a looser association of serum and CSF GFAP concentrations. If consistently demonstrated in larger cohorts, this observation would suggest that GFAP is most useful when measured in the CSF as opposed to blood. An unusual feature of our study was the availability of patient samples at early disease timepoints followed by such a long follow-up. Indeed, the longevity of the follow-up spanned an era when therapies were only starting to become available in clinical practice. Less than a third of the patients were exposed to disease-modifying treatments, and the majority of those were with first generation platform therapies. Even so, we did observe a beneficial effect of treatment exposure on long term clinical outcomes. In an era where high efficacy therapies are used liberally for most patients to blunt the natural history of the disease, it may be difficult to recapitulate the heterogeneity in outcomes that we observed or correct for the effects of treatment exposure in modern cohorts of patients with MS. The significance of baseline biomarker measurements on patients exposed to these treatments warrants further study. Our study was limited by small sample size and restricted outcome information, constrained by the retrospective identification of real-world patients based on sample and follow-up data availability from a single center. To address the possibility of a selection bias, we compared baseline characteristics of the patients who were lost to follow-up and those who were not: fortunately, baseline characteristics were very similar, suggesting that there was no major bias. While age did not emerge as a significant baseline biomarker correlate or CDW covariate in our study, age is a known and important confounder of both NfL and GFAP. Our analyses relied on post hoc age adjustment; at least in the case of sNfL, age adjustment is now possible using Z-scores or percentiles from normative datasets of many thousands of putatively healthy controls using their online app7,8. We could not incorporate this method, as similar age adjustment is currently not available for CSF NfL or GFAP, which were the focus of this study. Similarly, our analysis did not consider the confounding effects of baseline comorbidities and BMI20. We also did not include subsequently ascertained clinical MS subtype as a covariate; this was not evident at the point where CSF was sampled, and our relatively small sample size limits the number of variables that we could adjust for. Finally, while none of the studied patients had clinical relapses in the 3 months preceding sampling, we did not consistently have access to confirmatory MRI data to rule out the possibility of subclinical MRI lesions in this timeframe, which may have skewed biomarker results. In conclusion, the complementary combination of cNfL and cGFAP demonstrate promising prognostic utility in MS. As baseline CSF testing is often undertaken for diagnostic purposes, we propose that the additional measurement of these two proteins could be incorporated to inform prognosis at baseline and may allow a more personalized selection of treatment regimen intensity. Supplementary Information Supplementary Information. Author contributions S.T. was responsible for the study conception and design, acquisition, analysis, interpretation and drafting of the manuscript. S.M.F. was provided expert input on analysis interpretation, as well as contributed to the drafting of the manuscript. H.B., M.A. and G.B. contributed to clinical data acquisition and analysis. R.A.B. provided expert input on biomarker data acquisition and analysis. G.F., M.S.F. and G.B. were involved in study conception, design, interpretation and manuscript drafting. All authors agreed to the submitted version of the manuscript and are accountable for their contributions. Data availability statement The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Simon Thebault, Mark S. Freedman and Gauruv Bose Co-senior authors. Supplementary Information The online version contains supplementary material available at 10.1038/s41598-024-75290-8. References Associated Data Supplementary Materials Supplementary Information. Data Availability Statement Consideration will be made by the authors regarding the sharing of anonymized data related to this study with qualified investigators. The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.