Prediction of plasma concentration-time profiles in mice using deep neural network integrated with pharmacokinetic models

Quantitative structure-activity relationship (QSAR) methods have emerged as powerful tools to streamline non-clinical pharmacokinetic (PK) studies, with extensive evidence demonstrating their potential to predict key in vivo PK parameters such as clearance and volume of distribution. However, despite their potential, the reliability and robustness of these techniques are often compromised by estimation errors from curve fitting, a crucial limitation that has been largely overlooked. To address this, we propose a novel QSAR method that directly leverages in vivo PK data from intravenous and oral administrations by implicitly integrating a deep neural network with a two-compartmental model. The model uses the chemical structure of the compounds along with in vitro/in silico absorption, distribution, metabolism, and excretion (ADME) features as input data and was trained on in vivo mouse PK data encompassing 1,162 compounds across 30 in-house projects. The prediction accuracy for the plasma concentration-time profiles was evaluated using 5-fold cross-validation (CV). For the test sets within the CV folds, the median R² values ranged from 0.530 to 0.673 for intravenous administration and 0.119 to 0.432 for oral administration, demonstrating a consistently improved performance over the method that explicitly used PK parameters. The impact of the input features on the model outputs was evaluated using Integrated Gradients, revealing attributes consistent with established PK principles. These findings indicate that our novel method not only exhibits enhanced performance in predicting PK profiles but also provides valuable insights into the relationship between input features and model outputs.