CAN A NEURAL NETWORK TRAINED ON IN VITRO MICROSOMAL DATA ACCURATELY PREDICT IN VIVO HEPATIC CLEARANCE OF 50 FDA‑APPROVED DRUGS

Abstract

Predicting human hepatic clearance accurately during drug development is really important since hepatic metabolism largely impacts systemic drug exposure, efficacy, and safety. Traditional in vitroin vivo extrapolation (IVIVE) methods based on microsomal intrinsic clearance typically show low predictive consistency because of factors such as protein binding, transporter-mediated processes, and complicated metabolic interactions. This study looked into whether an artificial neural network (ANN) fed with in vitro microsomal data could precisely estimate the in vivo hepatic clearance of a sample of 50 FDA-approved drugs. A retrospective computational modeling approach was taken by reusing published data of microsomal intrinsic clearance, physicochemical descriptors, and BDDCS-related parameters. The data set was multi-therapeutic class and metabolically diverse. The model incorporated variables such as the fraction unbound in plasma lipophilicity molecular properties, and BDDCS classification with the help of ANN to enhance the predictive performance. The dataset was split into training, validation, and testing sets, and the performance of the model was assessed using the correlation coefficient (R), the coefficient of determination (R), the mean absolute error (MAE), and the root mean square error (RMSE). According to the ANN model, the training, validation, and testing datasets yielded high correlations of 0.93, 0.89, and 0.87, respectively. The prediction hit rate was over 85% for the overall dataset, with BDDCS Class 1 and 2 compounds showing the most accuracy. Besides, the model was capable of accurately representing nonlinear relationships between the microsomal intrinsic clearance and the hepatic clearance values observed, leading to a better performance than using the simple assumption IVIVE linear models. On the downside, this work considered only that the microsomal system reflects the in vivo situation, IVIVE modeling only made use of the microsomal system, and these are less predictive for characterizing transporter-dependent BDDCS Classes 3 and 4 compounds, suggesting other mechanisms at play. The results show that artificial neural networks (ANNs) trained on the in vitro microsomal data can be a consistent and reliable way of predicting the in vivo hepatic clearance of a wide variety of FDA-approved drugs. The combination of physicochemical and BDDCS-informed descriptors much boosted the models' performance and backed the use of machine learning (ML)-based pharmacokinetic (PK) prediction. These findings will help to better understand the mechanism of action of drugs. This ANN-based approach could immediately lead to better drug candidate selection, reduction in experimental burden, and more accurate prediction of human pharmacokinetic behavior at the early stage of drug development.