Predicting aberrant pharmacokinetics of bispecific antibody-derived molecules from in vitro gradient chromatographic assays and physiologically based pharmacokinetic (PBPK) modeling
Danilo Tomasoni1, Alessio Paris1, Tim Acker2, Kevin D. Cook2, Marcus Soto3, Manuel Ponce3, Esperanza Ojeda3, Diana Wong2, Shweta Mandavalli2, Veena A. Thomas2, Luca Marchetti1,4
1Fondazione The Microsoft Research - University of Trento Centre for Computational and Systems Biology (COSBI), 2Amgen Inc, Department of Pharmacokinetics and Drug Metabolism, 3Amgen Inc, Department of Pharmacokinetics and Drug Metabolism, 4University of Trento, Department of Cellular, Computational and Integrative Biology (CIBIO)
Objectives: Bispecific antibody-derived molecules (bsAbs) and half-life extended bsAbs (with YTE mutation) are an evolution of monoclonal antibodies (mAbs). Post-endocytosis, mAbs interact with neonatal Fc receptor (FcRn) in acidified endosomes and subsequently dissociate in neutral pH serum, enabling recycling of mAbs, leading to prolonged serum persistence. Half-life extension strategies exploit the intracellular trafficking and recycling mechanism by engineering the active Fc region and optimizing the pH-dependent IgG Fc-FcRn interaction. YTE mutation can extend half-life and minimize the difference between peak and trough concentration, potentially decreasing the dose amount or frequency. Yet, the introduction of the same YTE mutation in a panel of mAbs or bsAbs does not lead, in general, to the same degree of half-life benefits. Variability in FcRn-mediated recycling efficiency and non-specific off-target binding can be significant sources of inter-antibody variability in PK across YTE bsAb formats [1]. In particular, the source of non-specific clearance is often difficult to predict and can delay drug development. We propose the use of in vitro gradient chromatographic assays [2] to identify bsAb biologics with aberrant PK. Our objective is to incorporate the in vitro metrics into a physiologically based pharmacokinetic (PBPK) model to make a priori predictions of plasma pharmacokinetics and help de-risk aberrant PK behaviors. Methods: A panel of non-mouse cross-reactive bsAbs with (n=30) and without (n=41) YTE mutation has been used to extend a literature PBPK model by Chen and Balthasar for non-specific monoclonal antibodies [3]. Product-specific scaling factors (F1 and F2) have been added to scale the pinocytosis rate (CLUP) and the dissociation rate of bsAbs from FcRn (kOFF) at pH 6. The product-specific values for F1 and F2 have been initially estimated by fitting pharmacokinetic concentration time series measured in transgenic mice expressing huFcRN (Tg32). We then determined ad hoc formulas for F1 and F2 as a function of in vitro metrics by leveraging an in-house symbolic regression method that employed the product-specific values computed before to identify parsimonious formulas that maximize R² and minimize the fit error according to a k-fold cross-validation scheme (k=15 for YTE bsAbs, k=20 for bsAbs without YTE mutation). The final PBPK model extended with the identified formulas has been finally validated by considering the data of an additional pool of bsAbs with (n=5) and without (n=6) YTE mutations that were kept aside for this purpose. Results: The original PBPK model by Chen and Balthasar has been extended with four formulas, allowing the effective scale of the pinocytosis rate (CLUP) and the dissociation rate of bsAbs from FcRn (kOFF) at pH 6 in the two cases of bsAb with and without YTE mutation. Most notably, the functional form of log10(F2) has been expressed in both cases as a sigmoid directly dependent on the value of F1 and not explicitly on chromatographic measures. Log10(F1) instead, has been expressed as linear combinations of several experimental metrics, including FcRn and heparin chromatography and bsAb molecular weight. The extended model has been then validated by comparing its prediction performances with respect to the original PBPK model in an independent pool of bsAb. In the case of bsAb without YTE mutation, the absolute average fold error (AAFE) computed from the Chen and Balthasar model was 273.5 (min=1.55, max=1496.18), while the error of our model decreased to 28.45 (min=1.52, max=153.52). In the case of YTE bsAb, the AAFE computed by the Chen and Balthasar model was 10.57 (min=1.21, max=46.24), while our model obtained 2.62 (min=1.07, max=6.93). Conclusions: We propose a computational approach to use in vitro chromatographic assays for extending the Chen and Balthasar PBPK model [3] and predict aberrant PKs of therapeutic bispecific antibody-derived molecules with and without YTE mutation. Scaling factors for two model parameters have been expressed using ad hoc formulas dependent on a set of in vitro assays, allowing for an accurate product-specific prediction of plasma PK concentrations. The prediction performance in the validation set of our extended model outperforms the state-of-the-art and indicates that machine learning techniques can be effectively employed to integrate chromatographic measures into PBPK modeling and help de-risk aberrant PK behaviors.