Immunogenicity dynamics and covariate effects after satralizumab administration predicted with a Hidden Markov Model
Rory F. Leisegang (1), Hanna E. Silber Baumann (2), Siân Lennon-Chrimes (3), Hajime Ito (4), Elodie L. Plan (1)
(1) Department of Pharmacy, Uppsala University; (2) Roche Pharma Research and Early development, Pharmaceutical Sciences, Roche Innovation Center Basel; (3) Roche Products Ltd, Welwyn, UK; (4) Chugai Pharmaceutical Co., Ltd., Tokyo, Japan
Objectives:
Developing a deeper understanding of immunogenicity is important for biologics drug development [1,2]. Immunogenicity is the propensity of a therapeutic protein to generate an immune response to itself. While reporting of anti-drug antibodies (ADAs) increased, modeling it is seldom performed. Model-informed characterization of factors affecting treatment-induced ADAs may enable drug development decisions aiming to decreasing immunogenicity risk. The objective of this analysis was to predict ADA dynamics, including the potential influence of individual covariates, following subcutaneous satralizumab administration.
Methods:
Hidden Markov Models (HMMs) is a class of probabilistic models connecting predictions of observable variables to hidden underlying states, between which unobserved transitions may occur. Their implementation in NONMEM have been described [3,4], as well as their application to ADAs [3]. In this work, the number of Markov states were evaluated, allowing for permanent and transient ADA status, in addition to positive and negative.
Satralizumab is a humanized IgG2 monoclonal recycling IL-6 receptor antagonist antibody and is approved for treatment of neuromyelitis optica spectrum disorder. Data consisted of two pivotal phase III studies (NCT02028884, NCT02073279), in which 120 mg of satralizumab was given every 4 weeks as monotherapy or add-on to immunosuppressive therapy (IST); a phase I study in healthy volunteers (SA-001JP) [RL1] was used to inform the PK model.
Individual average steady-state concentrations (Cav,ss) were generated based on a previously developed PK model [6] fitted to PK observations with negative ADA.
Covariate relationships were evaluated structurally and through a stepwise covariate analysis. Covariates considered were demographic and laboratory findings, as well as ISTs. Model evaluation, alongside numerical considerations, was performed graphically. The chain of states presented by each subject was predicted utilizing the Viterbi algorithm. Simulations were generated as 5000 copies of profiles of interest to illustrate the impact of covariates on ADA incidence.
Results:
226 patients were included in the PK analysis and 154 in the HMM analysis. ADA observations were collected at trough every 4 weeks up to clinical cut-off (max. 4.6 years). A tiered approach was used for ADA titers.
The final model supported the quantification of emission probabilities for subjects to start either as permanently ADA-negative (34%) or transiently ADA-negative (66%). Transition probabilities governed the flux from the latter state towards (21%) a third hidden state, transiently ADA-positive, and back (16%); a permanently ADA-positive separate state was not identifiable. ADA titers were associated with truncated Poisson distributions, averaging ~9 when presented by patients in the ADA-positive state, and ~0 when from one of the ADA-negative states. These structural parameters were estimated with RSEs ranging 10-31%.
Several structural relationships were identified; ADA titer, time since starting satralizumab therapy, and time since last observation were found to impact transition probabilities. Discrete-time predicted states tended to be more like the previous state when closer in time (autocorrelation half-life of 10 days) [7]. The likelihood of ADA emergence declined exponentially (half-life of 270 days).
BMI and drug exposure were subsequently identified as significant covariates, accounting for the varying prevalence of ADAs between the studies. Higher BMIs led to greater risk to be allocated to transiently ADA-negative, to a longer time window to subsequently transit to ADA-positive, and to a lower chance to transit back. Patients with higher exposures were less likely to transition to ADA-positive, while also being linked to higher titers, and a greater risk to remain ADA-positive. Other tested covariates, including ISTs, were not retained.
Simulations, besides confirming model adequacy, supported understanding of the prevalence and clinical relevance of the findings, demonstrating the transient appearance of ADAs for a large patient population (BMI 18.5-31.5kg/m2, satralizumab Cav,ss 30-80μg/mL).[RL2]
Conclusions:
ADA dynamics may be further elucidated with modeling. HMMs constitute a framework to distinguish patients who may, and patients who may not, develop ADAs. The link between BMI and ADA emergence is likely relevant beyond this analysis and consistent with emerging evidence.
References:
[1] https://www.fda.gov/media/155871/download
[2] https://www.ema.europa.eu/en/documents/scientific-guideline/guideline-immunogenicity-assessment-monoclonal-antibodies-intended-vivo-clinical-use_en.pdf
[3] PAGE 24 (2015) Abstr 3625 [www.page-meeting.org/?abstract=3625]
[4] Brekkan A, Jönsson S, Karlsson MO, Plan EL. Handling underlying discrete variables with bivariate mixed hidden Markov models in NONMEM. J Pharmacokinet Pharmacodyn. 2019 Dec;46(6):591-604
[5] PAGE 27 (2018) Abstr 8664 [www.page-meeting.org/?abstract=8664]
[6] PAGE 29 (2021) Abstr 9665 [www.page-meeting.org/?abstract=9665]
[7] Silber HE, Kjellsson MC, Karlsson MO, The impact of misspecification of residual error or correlation structure on the type I error rate J Pharmacokinet Pharmacodyn 2009 Mar; 36(1):81-99
Disclaimer:
This work was performed by Uppsala University and was funded by F. Hoffmann-La Roche Ltd.