Prospective Validation of a Comprehensive In silico hERG Model and its Applications to Commercial Compound and Drug Databases

Ligand‐based in silico hERG models were generated for 2 644 compounds using linear discriminant analysis (LDA) and support vector machines (SVM). As a result, the dataset used for the model generation is the largest publicly available (see Supporting Information). Extended connectivity fingerprints (ECFPs) and functional class fingerprints (FCFPs) were used to describe chemical space. All models showed area under curve (AUC) values ranging from 0.89 to 0.94 in a fivefold cross‐validation, indicating high model consistency. Models correctly predicted 80 % of an additional, external test set; Y‐scrambling was also performed to rule out chance correlation. Additionally models based on patch clamp data and radioligand binding data were generated separately to analyze their predictive ability when compared to combined models. To experimentally validate the models, 50 of the predicted hERG blockers from the Chembridge database and ten of the predicted non‐hERG blockers from an in‐house compound library were selected for biological evaluation. Out of those 50 predicted hERG blockers, tested at a concentration of 10 μM , 18 compounds showed more than 50 % displacement of ³H]astemizole binding to cell membranes expressing the hERG channel. K_i values of four of the selected binders were determined to be in the micromolar and high nanomolar range (K_i (VH 01 )=2.0 μM , K_i (VH 06 )=0.15 μM , K_i (VH 19 )=1.1 μM and K_i (VH 47 )=18 μM ). Of these four compounds, VH 01 and VH 47 showed also a second, even higher affinity binding site with K_i values of 7.4 nM and 36 nM , respectively. In the case of non‐hERG blockers, all ten compounds tested were found to be inactive, showing less than 50 % displacement of ³H]astemizole binding at 10 μM . These experimentally validated models were then used to virtually screen commercial compound databases to evaluate whether they contain hERG blockers. 109 784 (23 %) of Chembridge, 133 175 (38 %) of Chemdiv, 111 737 (31 %) of Asinex and 11 116 (18 %) of the Maybridge database were predicted to be hERG blockers by at least two of the models, a prediction which could, for example, be used as a pre‐filtering tool for compounds with potential hERG liabilities.