Metabolic stability is a key property of drugs that represents the susceptibility of compounds to biotransformation. It strongly impacts drug efficacy and safety and is carefully optimized during early drug discovery. National Center for Advancing Translational Sciences (NCATS), in collaboration with IQC DMLG members, has endeavored to develop quantitative structure-activity relationship (QSAR) models that predict the P450-mediated metabolic stability. There is a paucity of publicly available data on P450-mediated metabolic stability. We have generated a comprehensive dataset of the metabolic clearance (half-life) values for more than 4000 compounds in <i>in vitro</i> assays of recombinant P450 3A4, 2C9, and 2D6, using a high-throughput, semi-automated assay. In addition to clearance, we have determined whether these same compounds inhibit the metabolism of other marker substrates by recombinant P450 CYP1A2, 3A4, 2C9, 2D6, 2C19 and CYP3A7. Finally, we have generated human liver microsomal stability data for subset of 1000 compounds. Thus, we characterized the compounds by 10 different metabolic related end-points. Since these end-points are biologically related to each other, we have implemented a novel multitask deep learning (MLT) architecture which incorporates together different types of chemical descriptors/fingerprints. In this study, Morgan, Avalon, and AtomPair fingerprints available from RDkit package were used as chemical descriptors. The MTL architecture was implemented in Keras using the Tensorflow backend. In addition to MTL model, as the baseline comparison, we constructed a consensus RandomForest model and a consensus single-task deep learning (STL) model based on three types of fingerprints for each end-point. The developed models were evaluated by fivefold cross-validation procedure and the balanced accuracy (BA) was calculated for each end-point. All models showed reasonable prediction accuracy for all activities, with BA ranging between 0.65 and 0.75. However, MTL based models showed the better performance compared to the both consensus RandomForest and consensus STL models. The obtained results suggest that the developed MTL model can be used by the scientific community engaged in drug development for the design and optimization of potential drug compounds, thereby helping accelerate the drug discovery process. The finalized model will be incorporated in the publicly available NCATS web services (https://predictor.ncats.io).

AI-driven QSAR modeling of P450-mediated drug metabolism

2019 Spring National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.