PhD Candidate University at Buffalo Buffalo, New York, United States
Background: Our central hypothesis is that antibody PK is determined by the interplay between molecular physicochemical properties of compounds and species physiology and that these relationships can be integrated into predictive mathematical models using contemporary modeling and computational systems. This work aims to develop a hybrid model for a series of mAbs in humans and expand it to preclinical species. Methods: The data comprise all available PK profiles, parameters, and structure properties of approved antibody drugs. The data were then randomly divided into a training and test dataset (70% and 30%). A neural network (NN) QSPKR model was developed in MATLAB to predict nonspecific clearances of approved antibody-based therapeutics. The inputs to the neural network were Patches of Surface Hydrophobicity Metric, Patches of Positive and Negative Charge Metrics, and Isotropic Surface Area. The NN was linked to a human minimal physiologically based pharmacokinetic model and simultaneously optimized to characterize the PK mechanistically. The unified mPBPK model has target-mediated disposition in both the tissue and the plasma and accommodates SC and IV dosing. The PK profiles of newly approved drugs were simulated as validation. The model was then simultaneously fit to mice, rats, rabbits, monkeys, and human PK data via adapted allometric scaling of network outputs. Results: A hybrid model has been developed for mAbs. The neural network portion was optimized to 80% accuracy for nonspecific and later linear antibody clearance. The sieving coefficients for tight and leaky tissue were 0.58 (%CV 34.12) and 0.85 (%CV 37.02), respectively. All estimated parameters had RSE less than 30%. The estimated allometric exponent was 0.837 (%CV 45.82), which reasonably agreed with the literature. The PK profiles of 60 drugs were adequately captured. The hybrid approach did not overtly outperform pure NLME. Conclusion: A QSPKR model of a noncongeneric series of monoclonal antibodies revealed four properties that were adequate predictors of clearance. The use of mPBPK allows the model to predict both parameters and time course data in multiple species, all while leveraging property information utilizing ML, which strengthens the predictive performance of this model and comes closer to creating aQSPKR model that aids in early-stage PK prediction in drug development.
