Critical Assessment of Artificial Intelligence Methods for Prediction of hERG Channel Inhibition in the “Big Data” Era

• DOI: 10.1021/acs.jcim.0c00884
• Publication Date: 2020-12-01T22:30:12Z
• Authors: Vishal B. Siramsh...; Dac-Trung Nguyen; Natalia J. Martinez; Noel T. Southall; Anton Simeonov; Alexey V. Zakharov
• Keywords: hERG; chEMBL; Chemical space

The rise of novel artificial intelligence (AI) methods necessitates their benchmarking against classical machine learning for a typical drug-discovery project. Inhibition the potassium ion channel, whose alpha subunit is encoded by human ether-à-go-go-related gene (hERG), leads to prolonged QT interval cardiac action potential and significant safety pharmacology target development new medicines. Several computational approaches have been employed develop prediction models assessment hERG liabilities small molecules including recent work using deep methods. Here, we perform comprehensive comparison effect based on (random forests gradient boosting) modern AI [deep neural networks (DNNs) recurrent (RNNs)]. training set (∼9000 compounds) was compiled integrating bioactivity data from ChEMBL database with experimental generated an in-house, high-throughput thallium flux assay. We utilized different molecular descriptors latent descriptors, which are real-value continuous vectors derived chemical autoencoders trained large space (>1.5 million compounds). were prospectively validated ∼840 in-house compounds screened in same best results obtained XGBoost method RDKit descriptors. only revealed that DNNs performed significantly better than RNNs operate SMILES provided highest model sensitivity. merged into consensus offered superior performance compared reference academic commercial domains. Furthermore, shed light exploit big chemistry generate representations useful predictive modeling tailoring space.