Abstract

Accurately predicting the pharmacokinetics (PK) of small-molecule candidates early in discovery can accelerate optimization cycles, reduce animal testing, and improve the quality of compounds advancing toward the clinic. However, conventional physiologically based pharmacokinetic (PBPK) modeling is often limited by throughput and by the need for extensive in vitro inputs. In this webinar, Dr. Davide Bassani, Computational DMPK Leader at Roche, presents an evaluation of SwiftPK, a corporate high-throughput PBPK (HT-PBPK) application that enables rapid PBPK simulations at scale using machine learning (ML)–predicted ADME inputs derived from chemical structure alone. Using a large in vivo rodent PK dataset (9,000 compounds), his team at Roche assessed SwiftPK performance across ten PK endpoints. Overall, most endpoints were predicted within a three- to four-fold error range, with absolute average fold errors (AAFEs) spanning 2.90–4.15 across the full dataset. Their research further demonstrated that predictive performance improves when (i) filtering for compounds predicted to be primarily cleared by hepatic metabolism (Extended Clearance Classification System, ECCS class 2) and (ii) restricting to cases where ML input predictions carry high confidence. Dr. Bassani walks through these results, and how they highlight the successful applicability of HT-PBPK in early-phase projects, especially for ECCS2-predicted compounds and with reliable input-property projections, and illustrate how HT-PBPK can support compound ranking and decision-making when experimental data are limited or unavailable.