The study by Walter et al. (2025) evaluates and compares four machine learning (ML) approaches to predict intravenous pharmacokinetic (PK) profiles in rats, with the goal of improving early-stage drug discovery by enabling compound ranking based on predicted PK behavior prior to synthesis. The four methods include: (1) Baseline-ML, which predicts clearance and volume of distribution for a one-compartment model; (2) Pure-ML, which directly predicts drug concentrations over time; (3) Compartmental-ML, which uses ML to estimate parameters for 1- or 2-compartment models; and (4) PBPK-ML, which predicts inputs for physiologically based pharmacokinetic models. The evaluation, using a temporal data split and a dataset of ~8,000 compounds, showed that Pure-ML, Compartmental-ML, and PBPK-ML all achieved similar and significantly better accuracy than Baseline-ML, with median geometric mean fold errors (GMFE) under 3-fold compared to Baseline-ML’s 4.39-fold. While Pure-ML offered robust performance, especially at later time points, the compartmental and PBPK approaches provided interpretability and mechanistic insights, respectively. Overall, the study demonstrates that ML-driven full-profile PK prediction is feasible and practical for aiding in early drug development decisions.

Five Key Findings:

1. Superior Accuracy of Advanced Models: Pure-ML, Compartmental-ML, and PBPK-ML all significantly outperformed the simpler Baseline-ML approach, achieving median GMFEs below 3-fold.

2. Value of Multiphasic Modeling: The Baseline-ML's poor performance stemmed from its inability to model multiphasic PK behavior, reinforcing the importance of mechanistic flexibility in predictions.

3. Temporal Validation Strategy: By employing a forward-looking, temporally split dataset for training and testing, the study realistically mimicked prospective drug discovery workflows and avoided overfitting.

4. Interpretability and Trade-Offs: While Pure-ML offered excellent performance, it lacked interpretability; in contrast, Compartmental-ML and PBPK-ML delivered meaningful parameters but were more complex.

5. Utility for Drug Prioritization: These ML approaches enable in silico evaluation of PK properties, facilitating early and cost-effective prioritization of compounds when paired with potency data.

Reference