Prediction of In Vivo Pharmacokinetic Parameters and Time-Exposure Curves in Rats Using Machine Learning from the Chemical Structure.

Animal pharmacokinetic (PK) data as well as human and animal in vitro systems are utilized in drug discovery to define the rate and route of drug elimination. Accurate prediction and mechanistic understanding of drug clearance and disposition in animals provide a degree of confidence for extrapolation to humans. In addition, prediction of in vivo properties can be used to improve design during drug discovery, help select compounds with better properties, and reduce the number of in vivo experiments. In this study, we generated machine learning models able to predict rat in vivo PK parameters and concentration-time PK profiles based on the molecular chemical structure and either measured or predicted in vitro parameters. The models were trained on internal in vivo rat PK data for over 3000 diverse compounds from multiple projects and therapeutic areas, and the predicted endpoints include clearance and oral bioavailability. We compared the performance of various traditional machine learning algorithms and deep learning approaches, including graph convolutional neural networks. The best models for PK parameters achieved R2 = 0.63 [root mean squared error (RMSE) = 0.26] for clearance and R2 = 0.55 (RMSE = 0.46) for bioavailability. The models provide a fast and cost-efficient way to guide the design of molecules with optimal PK profiles, to enable the prediction of virtual compounds at the point of design, and to drive prioritization of compounds for in vivo assays.

Findings of the Second Challenge to Predict Aqueous Solubility.

Ten years ago, we issued an open prediction challenge to the cheminformatics community: would participants be able to predict the equilibrium intrinsic solubilities of 32 druglike molecules using only a high-precision (CheqSol instrument, performed in one laboratory) set of 100 compounds as a training set? The "solubility challenge" was a widely recognized success and spurred many discussions about the prediction methods and quality of data. We revisited the competition a second time recently and challenged the community to a different challenge, not a blind test this time but using a larger test set of molecules, gathered and curated from published sources (mostly "gold standard" saturation shake-flask measurements), where the average interlaboratory reproducibility for the molecules was estimated to be ∼0.17 log unit. Also, a second test set was included, comprising "contentious" molecules, the reported (mostly shake-flask) solubility of which had higher average uncertainty, ∼0.62 log unit. In the second competition, the participants were invited to use their own training sets, provided that the training sets did not contain any of the test set molecules. We were motivated to revisit the competition to (1) examine to what extent computational methods had improved in 10 years, (2) verify that data quality may not be the main limiting factor in the accuracy of the prediction method, and (3) attempt to seek a relationship between the makeup of the training set data and the prediction outcome.

Modeling of non-additive mixture properties using the Online CHEmical database and Modeling environment (OCHEM).

The Online Chemical Modeling Environment (OCHEM, http://ochem.eu) is a web-based platform that provides tools for automation of typical steps necessary to create a predictive QSAR/QSPR model. The platform consists of two major subsystems: a database of experimental measurements and a modeling framework. So far, OCHEM has been limited to the processing of individual compounds. In this work, we extended OCHEM with a new ability to store and model properties of binary non-additive mixtures. The developed system is publicly accessible, meaning that any user on the Web can store new data for binary mixtures and develop models to predict their non-additive properties.The database already contains almost 10,000 data points for the density, bubble point, and azeotropic behavior of binary mixtures. For these data, we developed models for both qualitative (azeotrope/zeotrope) and quantitative endpoints (density and bubble points) using different learning methods and specially developed descriptors for mixtures. The prediction performance of the models was similar to or more accurate than results reported in previous studies. Thus, we have developed and made publicly available a powerful system for modeling mixtures of chemical compounds on the Web.