Learn-and-Confirm in Quantitative Systems Pharmacology: Evaluation of an Immunogenicity Platform
Hans Peter Grimm (1), Linnea C. Franssen (1), Maciej Swat (2), Andrzej M. Kierzek (2), Rachel Rose (2), Piet H. van der Graaf (2)
(1) Roche Pharma Research & Early Development, Switzerland, (2) Certara, United Kingdom
Introduction:
Immunogenicity (IG) against therapeutic proteins often impacts their pharmacokinetics (PK), pharmacodynamics, efficacy and safety. Predicting IG is complex and requires integrating compound properties, patient characteristics, and treatment parameters [1].
Building on the model by Chen et al. [2, 3], the IG Quantitative Systems Pharmacology (QSP) Consortium has been developing the IG Simulator for this purpose [4, 5]. This model combines bioinformatics predictions of antigen presentation, a PK model, and a QSP model of lymphocyte activation and anti-drug antibody (ADA) production.
Over the course of 5 years, the IG Simulator has been evaluated on a set of 15 compounds with published study data predominantly from late-stage trials. To further assess model credibility [6], we conducted this single-blinded evaluation with data from 10 compounds representative of the context of use in the drug development process.
Objectives:
- Expand the evaluation of the IG Simulator’s predictivity for ADA incidence and exposure impact
- Unbiased and stepwise assessment by keeping the modelling operator blinded to experimental outcomes
- Explore whether and how additional data from in vitro assays can be used to improve predictions
Methods:
This evaluation is based on 10 monoclonal antibodies, 4 of which are considered immuno-stimulating. Data are organized in 11 datasets with 3 from Phase I single-ascending dose trials, 7 from combined Phase I/II multiple dose trials, and 1 from a multiple dose Phase III trial. The number of subjects per dataset ranges from 12 to 182, totalling 408.
The evaluation was performed in a stepwise manner. At each step, the analyst was blinded to information from subsequent steps:
- Step 1: “Ad Hoc Prediction” based on the amino-acid sequence of the compounds, the non-linear mixed-effects model of the PK in absence of IG, and individual dosing, sampling history, and patient covariates in the trial.
- Step 2: “T-Cell Refined Prediction” additionally using T-cell activation data, and “MAPPs Refined Prediction” using MHC-associated peptide proteomics (MAPPs), cf. [1]. No other data indicative of IG were provided at this step.
- Step 3: Exploration of further model adjustments. Clinical observations of PK and ADA were provided for this step only.
The analysis workflow involved the prediction of peptide binding to MHCII using NetMHCIIpan 4.0 [7], the reduction of this set to non-self peptides using blastp with the Non-redundant UniProtKB database [8], and the generation of a virtual population with the Allele Frequency Net Database [9]. The IG Simulator Version 4.1b was used for simulations. Population PK models were translated into equivalent minimal physiologically based pharmacokinetic (mPBPK) models [10] within the IG Simulator.
Results:
Ad hoc predictions of IG incidence were within a predefined range of the respective observations for >50% of the study duration for 3 of the 11 trials, for 10-50% of the duration for 5 trials, and <10% for 3 trials. Across datasets, the predictions of incidence were within the defined range for 42% of the assessed time intervals. A trend for over-prediction was noticed – predictions were above observations for 41% of the time (vs. 17% below). This trend for over-prediction also applied to the case of immuno-stimulatory compounds.
Neither refinement with MAPPs nor with T-cell activation data (“Step 2”) improved predictions significantly. However, adjusting the threshold value for ADA positivity had a significant impact in some of the projects and further improvements could be obtained by adjusting system parameters (“Step 3”), which could reflect the impact of different mechanisms of action on the immune system. Generally, only low or moderate sensitivity was seen with respect to the parameters describing the antigen presentation or the naive antigen-specific T-cell numbers.
Conclusions:
The immense need for tools allowing to project the IG of therapeutic proteins requires to be paralleled by rigorous model qualification. In this study, the data from 10 projects between early and late clinical development were used for an unbiased evaluation of the IG Simulator. The results have provided invaluable insights into the model framework and the associated workflows. We believe this is a unique example of a development of a rigorous QSP platform through a sustained, collaborative “learn-and-confirm” approach.
References:
[1] Bray-French K et al. J Pharm Sci (2021) 110(7), 2575-84.
[2] Chen X et al. CPT:PSP (2014) 3(9), 1-9.
[3] Chen X et al. CPT:PSP (2014) 3(9), 1-10.
[4] Kierzek AM et al. CPT:PSP (2019) 8, 773-6.
[5] Shakhnovich V, et al., Clin Transl Sci. (2020) 13(2): 219-223.
[6] Musuamba FT et al. CPT:PSP (2021) 10(8), 804-825.
[7] Reynisson B et al. J Proteome Res (2020) 19(6), 2304-15.
[8] The UniProt Consortium, Nucleic Acids Res (2021) 49(D1), D480-D489.
[9] Gonzalez-Galarza FF et al. Nucleic Acids Res (2020) 48, D783-8.
[10] Li et al., AAPS J (2014) 16(5):1097-109