Hamzic, Seid, Lewis, Richard, Desrayaud, Sandrine, Soylu, Cihan, Fortunato, Mike, Gerebtzoff, Gregori and Rodriguez Perez, Raquel (2022) Predicting in vivo brain penetration using multitask graph neural networks. Journal of chemical information and modeling.

Abstract

The blood-brain-barrier (BBB) is a semi-permeable interface, separating the central nervous system (CNS) from the blood stream. The physiological role of the BBB is to create a stable microenvironment for the CNS by tightly regulating the transport of molecules from the blood to the brain and vice versa. For early drug discovery teams, it can be critical to know if com-pounds are able to penetrate into the brain compartment. Generally, pre-clinical in vivo studies measuring the ratio of total and free brain and blood concentrations (Kp and Kpuu, respectively) are required to estimate the brain penetration potential of a new drug entitiy. In this work, we evaluated the performance of different machine learning approaches to predict Kp, using Novar-tis internal and publicly available experimental data. We investigated the benefit of including in vitro experimental data as auxiliary tasks in multitask graph neural network (MT-GNN) mod-els. We observed that MT-GNN models generally outperformed single-task (ST) learning ap-proaches, which were only trained on in vivo brain penetration data. The best performing MT-GNN regression model achieved a coefficient of determination (R2) of 0.42 on a prospective validation set and outperformed all tested ST models. Overall, models solely based on public data achieved lower performance on the prospective validation set compared to models built with internal data. However, the MT-GNN based upon literature data outperformed all litera-ture-based ST models, with a R2 of 0.31. Lastly, we observed that post hoc classification using a MT-GNN regression model outperformed a MT-GNN classification model, with Matthew’s correlation coefficient values of 0.66 and 0.44, respectively. Taken together, we show that the inclusion of the right auxiliary tasks improves the prediction of in vivo brain penetration using MT-GNNs.

Item Type:	Article
Keywords:	Brain penetration, multitask learning, compound property, ADME, machine learning, deep learning
Date Deposited:	12 Jul 2022 00:45
Last Modified:	12 Jul 2022 00:45
URI:	https://oak.novartis.com/id/eprint/47645