Tom Chebassier 1,2, Hoai-Thu Thai 2, Jeremy Seurat 2, Vincent Thuillier 3, Sophie Fliscounakis-Huynh 4, Julie Bertrand 1
1 Université Paris Cité and Université Paris Sorbonne Paris Nord, Inserm, IAME (Paris, France), 2 Sanofi, Translational Medicine Unit, Quantitative Pharmacology, Development Pharmacometrics, Sanofi (Gentilly, France), 3 Clinical modeling & Evidence Integration, Sanofi (Gentilly, France ), 4 Sanofi, Translational Medicine Unit, Quantitative Pharmacology, Disease modeling, Sanofi, Gentilly, France on behalf of IT&M Stats (Gentilly, France)
Introduction: Relapsing‑Remitting Multiple Sclerosis (RRMS) is an autoimmune disease characterized by demyelinating lesions and progressive neurological impairment(1). While reducing relapse frequency remains a major therapeutic objective, slowing long‑term disability progression continues to represent a critical unmet need(2). Disability in RRMS is commonly measured using the Expanded Disability Status Scale (EDSS), an ordinal composite scale built from several functional system (FS) subscores(3). Despite its wide use, EDSS remains difficult to model due to its non‑linearity(4), bounded and discretized structure, inter‑rater variability, and limited sensitivity to short‑term changes(5). These limitations motivate increasing interest in blood biomarkers that could help anticipate disability evolution(6). Serum neurofilament light chain (sNfL), reflecting neuroaxonal injury(7), has emerged as a promising biomarker of disease activity, but its quantitative relationship with disability progression remains insufficiently defined in RRMS(8).
In previous work, we developed a semi-mechanistic model for sNfL(9) capturing alemtuzumab effect and within-subject variability (WSV) was used to capture sNfL peaks due to acute inflammation. Using sNfL predictions from this model not accounting for WSV (underlying dynamics), we quantified sNfL link with time to first 6-month confirmed disability progression (6MCDP). However, focusing on 6MCDP induced a loss of information since such endpoint is derived from EDSS trajectories.
Objectives: This work aimed (i) to model disability captured by EDSS trajectories in alemtuzumab‑treated RRMS patients, and (ii) to characterize their link with sNfL.
Methods: Data were collected from the alemtuzumab arm of CARE‑MS I phase III trial and its 6‑year extension (CARE‑MS I‑EXT)(10). EDSS and FS subscores were recorded every 3 months; lymphocytes were measured every month; sNfL was measured every 6 months during the core study and every 3 months thereafter.
A nonlinear mixed‑effects modeling framework (Monolix2024R1) was used to characterize EDSS trajectories. Three approaches were evaluated:
• A Continuous‑variable (CV) model treating EDSS as a transformed continuous outcome(11),
• A Bounded‑integer (BI) model preserving the ordinal, bounded nature of the scale(12),
• An Item response theory (IRT) handling all the FS subscores(13).
Model selection relied on corrected Bayesian Information Criterion (BICc), diagnostic plots, and predictive performance(14).
The link between sNfL and disability was examined on the best EDSS model; (i) on top of the alemtuzumab effect or (ii) alone (alemtuzumab effect being completely captured by sNfL)(15). Individual sNfL predictions were incorporated into the EDSS model following a two‑stage approach. We evaluated whether incorporating sNfL peaks or excluding them improved predictive accuracy.
External evaluation was performed using the TOPAZ 5.5‑year extension study(16).
Results: Data for a total of 351 alemtuzumab‑treated RRMS patients from CARE‑MS I and CARE‑MS I‑EXT were analyzed. Alemtuzumab effect was quantified by cumulative area under the concentration curve (AUC(t)). For the CV approach, the best structure was a four‑parameter logistic model with alemtuzumab impacting progression slope. For BI and IRT approaches, optimal model was a linear latent disability trajectory with an alemtuzumab effect on progression slope.
The IRT model provided the best predictive performance and was therefore chosen to explore the association with sNfL.
In the final model, latent disability progressed linearly at 12.5% per year in absence of treatment, and in most patients at the end of the first extension (AUC(t=8 years)=50 mg.L-1.days) we predicted a 60% decrease in latent disability, reflecting alemtuzumab effect on long‑term disability accumulation, including progression independent of inflammation. Current sNfL levels added an independent contribution e.g. at 50 pg/mL, they increased the latent disability by 30%, improving model fit substantially (ΔBICc = –117), consistent with the presumed impact of acute axonal injury on disability. This combined structure best described TOPAZ EDSS trajectories. Without an alemtuzumab effect on the latent variable, the optimal alternative was a model linking sNfL underlying dynamic change from baseline to disability progression slope, which however performed worse than a model including only the alemtuzumab‑mediated progression slope modification.
Conclusion: CV, BI, and IRT approaches were evaluated to model EDSS. IRT framework demonstrated clear superiority. Accounting for sNfL peaks improved fitness, while also acknowledging that sNfL may reflect broader neurodegenerative processes that extend beyond overt clinical relapse activity(17). Nevertheless, further work is needed to develop a covariate model(18), and determine if the model can capture EDSS individual trajectories and time to disability progression. Moreover, future work will investigate whether incorporating imaging biomarkers can further enhance disability‑prediction.
References:
1. C. A. Dendrou, L. Fugger, M. A. Friese, Nat. Rev. Immunol. 15, 545–558 (2015).
2. D. S. Reich, C. F. Lucchinetti, P. A. Calabresi, N. Engl. J. Med. 378, 169–180 (2018).
3. J. F. Kurtzke, Mult. Scler. Relat. Disord. 4, 95–103 (2015).
4. K. Tilling et al., Heal. Technol. Assess. 20, 1–48 (2016).
5. A. M. Novakovic et al., J. Clin. Pharmacol. 58, 1284–1294 (2018).
6. S. Meier et al., JAMA Neurol. 80, 287–297 (2023).
7. M. Khalil et al., Nat. Rev. Neurol. 20, 269–287 (2024).
8. A. Bar-Or et al., eBioMedicine. 93, 104662 (2023).
9. T. Chebassier et al., PAGE2025 (2025)
10. J. A. Cohen et al., Lancet. 380, 1819–1828 (2012).
11. G. J. Wellhagen, M. O. Karlsson, M. C. Kjellsson, AAPS J. 23, 9 (2020).
12. G. J. Wellhagen, M. C. Kjellsson, M. O. Karlsson, AAPS J. 21, 74 (2019).
13. S. Ueckert, CPT: Pharmacomet. Syst. Pharmacol. 7, 205–218 (2018).
14. T. H. T. Nguyen et al., CPT: Pharmacomet. Syst. Pharmacol. 6, 87–109 (2017).
15. Q. L. Coënt, C. Legrand, V. Rondeau, Biostatistics. 25, 98–116 (2022).
16. A. J. Coles et al., Ther. Adv. Neurol. Disord. 16, 17562864231194823 (2023).
17. A. Abdelhak et al., JAMA Neurol. 80, 1317–1325 (2023).
18. G. J. Wellhagen et al., CPT: Pharmacomet. Syst. Pharmacol. 13, 812–822 (2024).
Reference: PAGE 34 (2026) Abstr 12147 [www.page-meeting.org/?abstract=12147]
Poster: Oral: Drug/Disease Modelling - Other Topics