Kamal Srinivasan 1, Samit Ganguly 1, John D. Davis 1, Lutz Harnisch 1, Jason Chittenden 1
1 Regeneron Pharmaceuticals,Inc. (Tarrytown, USA)
Introduction:
Monoclonal antibodies (mABs) typically show linear PK which are consistent for the class, under target-absent conditions. It would be useful to have a strong prior to support model‑informed drug development in applications such as study design or augmenting observed data. For a mAb without observations in clinical studies, the prior should include PK parameter estimates, their variability, and the uncertainty arising from between‑drug differences.
When co-administered as a cocktail, the physiological drivers of mAb distribution and elimination induce a strong between-analyte correlation in the interindividual variability (IIV) structure. Across studies of multiple mAb cocktails, it is possible to specify the IIV as a combination of random effects at the compound and subject level, thus separating the variability induced by physiological differences between subjects from the variability due to differences in the compounds’ properties.
Such a model is not feasible in most available nonlinear mixed effects software implementations due to the crossed subject level and compound level random effects (compounds appear with multiple subjects and subjects appear with multiple compounds). Markov Chain Monte Carlo method does provide a potential solution, and the resulting Bayesian prior yields posterior distributions that can be reused as informative priors for typical linear mAb PK estimation, and multiple downstream applications.
Objectives:
• Estimate population PK parameters (CL, Vc, Q, Vp, ka) and mAb-specific parameters across a wide range of compounds in the Regeneron clinical database;
• Quantify subject and drug IIV components within the crossed random effects framework;
• Generate posteriors that can be reused as informative priors for linear mAb PK simulation, parallel linear–TMDD models, pediatric scaling under sparse sampling, and clinical trial design.
Methods:
Data included PK for 11 monoclonal antibodies measured under target-absent conditions in healthy volunteers, administered either alone or in 2 or 3 antibody cocktail combinations and individual mABs serum concentration was measured. A linear two- compartment model with IV/SC dosing and transit absorption was fitted in Stan[1]. The variance structure used a crossed random effects design: each subject–drug parameter was represented as the sum of a subject effect, a drug effect, and a within subject drug specific deviation. The within subject component allows correlation of random effects across antibodies within the same individual.
Results:
Population PK parameters were well estimated with tight posteriors, all showing narrow 90% credible intervals and Rhat≈1 with high Effective Sample Size (ESS). Interindividual variability well partitioned across the crossed structure: subject level captured most variability, while drug level contributed smaller but analyte specific differences. The within‑subject shared–vs–independent deviation was dominated by the shared component of variability.
Individual parameter summaries across the 11 mABs showed class-consistent ranges and low relative standard errors (RSE~1.6–2.5%). Within a subject, correlations between parameters across different antibodies were uniformly high for all parameters, reflecting a large shared component and strong cross‑antibody similarity within the same individual. Posterior distributions for CL, Vc, Q, Vp and ka were substantially narrower than their corresponding priors, demonstrating strong information gain from the data and precise estimation relative to the broad prior assumptions.
Conclusions:
In summary, we developed a reusable Bayesian prior for linear mAb PK. The crossed random effect omega structure in Stan[1] partitions IIV, enabling within subject cross drug correlation and stable parameter estimation even under sparse sampling. This outperforms single analyte or shared ETA/block OMEGA approaches, which cannot reliably separate subject shared from analyte specific variability. The posteriors that can be reused as informative priors for pediatric scaling with limited PK, linear mAB PK simulation, and as linear priors in parallel linear–nonlinear (TMDD) models. The resulting posteriors enable realistic exposure predictions for future mAb candidates and support more efficient clinical trial design.
References:
[1] Carpenter, B., Gelman, A., Hoffman, M. D., Lee, D., Goodrich, B., Betancourt, M., et al. (2017). Stan: A probabilistic programming language. Journal of Statistical Software, 76(1), 1–32; Stan version 2.26.1.
Reference: PAGE 34 (2026) Abstr 12230 [www.page-meeting.org/?abstract=12230]
Poster: Methodology - Covariate/Variability Models