DeepCt:\nPredicting Pharmacokinetic Concentration–Time\nCurves and Compartmental Models from Chemical Structure Using Deep\nLearning

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