Principal Scientist Preclinical and Translational Pharmacokinetics/Pharmacodynamics, Genentech San Francisco, California, United States
Objectives: To develop and apply a neural ordinary differential equation (Neural ODE) framework for prediction of preclinical pharmacokinetics of T-cell dependent bispecific molecules.
Methods: Deep learning has the potential to transform many aspects of drug development, as evidenced by the increasing number of AI/ML related FDA submissions year over year [1]. In particular, neural ODE frameworks have shown success in capturing and predicting pharmacokinetics (PK) and pharmacodynamics in clinical settings where dozens of observations from hundreds of patients were used for model training and validation [2,3]. However, the utility of these methods in preclinical settings with fewer observations remains an open question. In this work, we trained a previously published neural ODE [2] with cynomolgus monkey PK data for the T cell engaging antibody mosunetuzumab to benchmark the predictive capabilities of a neural ODE framework in preclinical settings. We then developed a pharmacologically-informed neural ODE framework and demonstrated its improved ability to make PK predictions without providing initial PK data points. Model training, testing, and visualization were conducted in Python.
Results: We present a previously published encoder-decoder Neural ODE framework that, with hypertuning, can predict preclinical PK across several dose levels spanning two orders of magnitude. We trained the Neural ODE on mosunetuzumab PK data [4,5] in 83 cynomolgus monkeys (~80% of total dataset) and tested individualized model predictions against PK from 20 monkeys (~20% of total dataset). However, this Neural ODE framework requires the first 7 days of PK concentrations as input for the encoder. We next developed a novel pharmacologically-informed Neural ODE (PINODE) that does not require the incorporation of early PK time-points to make individualized (weight-based) PK predictions and demonstrated its success in predicted PK across the same dose ranging study. Finally, we evaluated the ability of the PINODE to extrapolate and make PK predictions for ‘unseen’ dose regimens ranging from 0.001 mg/kg to 1 mg/kg. This framework has the potential to generalize across dosing strategies in preclinical settings and potentially be extended to project clinical PK profiles.
Conclusions: The PINODE framework accurately reproduces the PK of mosunetuzumab in preclinical settings. The framework represents the first Neural ODE that can predict individualized PK without providing initial PK data points to the encoder, significantly improving its relevance for preclinical applications. This work demonstrates an important first step toward the development of a generalized Neural ODE framework for a priori prediction of individualized PK.
Citations: Liu, Q., Huang, R., Hsieh, J., Zhu, H., Tiwari, M., Liu, G., ... & Huang, S. M. (2023). Landscape analysis of the application of artificial intelligence and machine learning in regulatory submissions for drug development from 2016 to 2021. Clinical pharmacology and therapeutics, 113(4), 771-774.
Lu, J., Deng, K., Zhang, X., Liu, G., & Guan, Y. (2021). Neural-ODE for pharmacokinetics modeling and its advantage to alternative machine learning models in predicting new dosing regimens. Iscience, 24(7).
Lu, J., Bender, B., Jin, J. Y., & Guan, Y. (2021). Deep learning prediction of patient response time course from early data via neural-pharmacokinetic/pharmacodynamic modelling. Nature machine intelligence, 3(8), 696-704.
Hosseini, I., Gadkar, K., Stefanich, E., Li, C. C., Sun, L. L., Chu, Y. W., & Ramanujan, S. (2020). Mitigating the risk of cytokine release syndrome in a Phase I trial of CD20/CD3 bispecific antibody mosunetuzumab in NHL: impact of translational system modeling. NPJ systems biology and applications, 6(1), 28.
Ferl, G. Z., Reyes, A., Sun, L. L., Cheu, M., Oldendorp, A., Ramanujan, S., & Stefanich, E. G. (2018). A preclinical population pharmacokinetic model for anti‐CD20/CD3 T‐cell‐dependent bispecific antibodies. Clinical and Translational Science, 11(3), 296-304.
