DeepCt: Predicting Pharmacokinetic Concentration–Time Curves and Compartmental Models from Chemical Structure Using Deep Learning

Abstract

After initial triaging using in vitro absorption, distribution, metabolism, and excretion (ADME) assays, pharmacokinetic (PK) studies are the first application of promising drug candidates in living mammals. Preclinical PK studies characterize the evolution of the compound’s concentration over time, typically in rodents’ blood or plasma. From this concentration–time (C–t) profiles, PK parameters such as total exposure or maximum concentration can be subsequently derived. An early estimation of compounds’ PK offers the promise of reducing animal studies and cycle times by selecting and designing molecules with increased chances of success at the PK stage. Even though C–t curves are the major readout from a PK study, most machine learning-based prediction efforts have focused on the derived PK parameters instead of C–t profiles, likely due to the lack of approaches to model the underlying ADME mechanisms. Herein, a novel deep learning approach termed DeepCt is proposed for the prediction of C–t curves from the compound structure. Our methodology is based on the prediction of an underlying mechanistic compartmental PK model, which enables further simulations, and predictions of single- and multiple-dose C–t profiles.