Characterization of anti-drug antibodies using a bivariate mixed hidden Markov model
A. Brekkan (a), B. Lacroix (b), R. Lledo-Garcia (c), S. Jönsson (a), M. O. Karlsson (a), E. L. Plan (a)
(a) Department of Pharmaceutical Biosciences (b) UCB Pharma, Braine l'Alleud, Belgium (c) UCB Pharma, Slough, UK
Objectives: Monoclonal antibodies are a well-established therapy for several chronic inflammatory diseases. A phenomenon observed with administration of this type of drugs, is the ability of the immune system to produce specific anti-drug antibodies (ADA), which may influence the pharmacokinetics (PK) of the drug and possibly affect efficacy and safety. Certolizumab Pegol (Cimzia®) is a PEGylated Fc-free anti-tumor necrosis factor α (anti-TNF-α) antibody used in the treatment of several inflammatory diseases including rheumatoid arthritis (RA). The ADA against Cimzia have been characterized with an ELISA technique in clinical settings and the incidence of immunogenicity in RA has been reported to be 9.6%, with transient and persistent PK effects1. There is an interest in characterizing the transient/persistent ADA and to investigate potential covariates and trial characteristics that may influence their occurrence. Due to drug interference with its measurement assay, false negative data may occasionally arise. Thus, in some patients, ADA may not be measureable despite the disposition of the drug being altered. In this work, a novel model-based method for ADA characterization is presented using mixed hidden Markov models (MHMM), allowing for inferences about ADA formation given a set of ADA and drug PK observations2,3.
Methods: Phase II data from a clinical trial aiming to assess the efficacy and safety of 6 doses (50-800 mg) of Cimzia versus placebo administered Q4W in patients with RA was used in this work. The total number of evaluated patients was 239 with an average of ~9 observations over a maximum of 13 weeks. A previously developed PK model not including ADA as a covariate was fit to the first dosing occasion in the data with the subsequent dosing occasions being predicted from the resulting fit.The obtained PK individual weighted residuals (PKRES) were used, in addition to ADA measurements (ADAMEAS), as continuous observed variables to inform about the two states in a bivariate-MHMM (BV-MHMM). The hidden states in the model were no ADA (SNOADA) and ADA production (SADA). The parameter estimates in the model were compared to expectations for the distributions of the observed variables and the model was used to calculate the most probable state sequence in each individual using the Viterbi algorithm. Estimation was done in NONMEM with IMPMAP.
Results: A BV-MHMM was established that included two states influencing PKRES and ADAMEAS, correlated through a bivariate normal distribution. Correlations between PKRES and ADAMEAS in SNOADA (ρNOADA) and SADA (ρADA) were estimated to be -0.12 and -0.15, respectively. Modes of the distributions were estimated as 0.36 and -1.4 for the residuals in SNOADA and SADA and fixed as, 0.6 and 2.4 U/mL for the ADAMEAS in SNOADA and SADA , respectively. These estimates were in agreement with the expectation where i) the PKRES when ADA are present would be negative (associated with model over-predictions in the presence of ADA); ii) the ADAMEAS should be positive when ADA are present. Transition probabilities were low (0.08 and ~0) for the transition from SNOADA to SADA, and SADA to SNOADA, respectively). Standard errors (SE) of the estimated parameters were <25% RSE with the exception of the transition probabilities (>100%). The mean time to transit to SADA was 47.7 days, while the observed mean time to clinical positivity was 60.9 days.
Conclusions: A BV-MHMM utilizing PKRES and ADAMEAS in characterizing ADA formation against Cimzia was developed. The model was able to characterize the transient/persistent ADA profiles and suggested ADA positivity earlier than conventional ELISA ADA measurements assays. The results suggest that the BV-MHMM may be able to identify PK altering ADA and as such the model can be considered a relevant complement for ADA characterization in addition to assay results. This model may prove to be more promising to characterize the onset of ADA formation and to explore the impact of covariates in driving the transient/persistent ADA formation and effects on Cimzia PK. Further investigations are warranted.
Acknowledgements: This study was supported by UCB Pharma.
Disclosures: RLG & BL are employees of b,c, and AB SJ, MOK, ELP of a.
References:
[1]. European Medicines Agency. Cimzia Product Information. Annex I: Summary of product characteristics. 2017. Accessed on 21/02/2017. URL = http://www.ema.europa.eu/docs/en_GB/document_library/EPAR_-_Product_Information/human/001037/WC500069763.pdf
[2]. Plan, EL., Nyberg, J., Bauer, RJ., Karlsson, MO. Handling Underlying Discrete Variables with Mixed Hidden Markov Models in NONMEM. Presented at PAGE 24. 2015. Abstr 3625. www.page-meeting.org/?abstract=3625
[3]. Brekkan, A., Karlsson, MO., Jönsson S., Plan EL. Parameter Estimation in Bivariate Mixed Hidden Markov Models. Presented at PAGE 26. 2017. Abstr 7379. www.page-meeting.org/?abstract=7379