While machine learning (ML)-based predictions for in vitro ADME assays (absorption, distribution, metabolism, excretion) and noncompartmental analysis (NCA) based pharmacokinetics (PK) parameters are well established to prioritize compounds, the prediction of full plasma concentration time profiles only recently gained attention. In drug discovery, optimizing the PK along with potency is important to ensure high quality drug candidates for patients.

In this work, the authors compared four distinct strategies to predict full concentration time profiles in a ML framework. As a proof of concept, PK predictions after intravenous administration in rats were performed using only information on chemical structure as input. The reason is that these prediction methods are supposed to be applied for compound prioritization even before synthesis (i.e., when no measured data is available). Finally, predictions were compared to corresponding in vivo data to substantiate the model's predictive ability.

The four different methods benchmarked in this study were an NCA based prediction as baseline method (Figure, left) a Pure ML approach without mechanistic PK constraints (Figure, middle-left), Compartmental PK modelsCompartmental PK models (Figure middle-right), as well as physiologically based pharmacokinetic (PBPK) models (Figure, right). Overall, the authors trained ML models on preclinical in-house i.v. PK studies for around 8,000 compounds and thoroughly evaluated the different approaches considering precision and bias of the predictions.

The authors concluded that three methods (Compartmental ML, PBPK ML and Pure ML) performed comparably well, and these three clearly outperformed the NCA-based baseline method. Overall, when combined with potency information, these ML PK modelling approaches provide the potential to improve drug discovery to prioritize candidates early based on
human dose scores.

Sort by: Newest First Oldest First Most Popular Most unliked