Pharmaceutical companies often invest in drug candidates that fail late in the process of drug discovery because of their toxicity. Cell-based assays, such as Alamar Blue, have successfully moved cytotoxicity testing to earlier phases of drug development. However, they are still costly and time-consuming to screen large databases of compounds. Therefore, much effort has been put to make predictive models for drug cytotoxicity. Despite the ability of Artificial Intelligence (AI) techniques, such as Deep Learning (DL), to determine molecular features related to a property, like other data-driven models, they are limited by uncertainties in the underlying data, which are usually collected from multiple diverse sources. A dataset without systematic error between different experiments is thus required to develop a reliable DL model for predicting cytotoxicity. In this work, we first present a database containing values of 50% cytotoxic concentration (CC50) for several hundred inhibitors of the HIV Reverse Transcriptase designed in our lab over the past two decades. The CC50 values have been measured on MT2 cells under consistent experimental conditions. We then employ Graph Convolutional Neural Networks (GCNN) to learn structural patterns and atomic features explaining the variance in our cytotoxicity data. In our GCNN models, molecules are represented as undirected graphs with attributed nodes. Two different convolutional operators are applied, namely, spectral- and attentional-based convolutions. Other prediction models are also generated based on physically relevant molecular descriptors and Morgan fingerprints, which are used to conduct the control study. Different neural network architectures are analyzed for GCNN models. Our results show that the two convolutional operators studied perform comparably, while both significantly outperform the control models in reproducing the corresponding experimental CC50 values. The trained GCNN models and the database will be freely accessible to the community.

Graph convolutional neural networks for high-throughput cytotoxicity prediction of drug-like compounds

ACS Spring 2022

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