Cardio-vascular safety beyond hERG: in silico modelling of a guinea pig right atrium assay

As chemists can easily produce large numbers of new potential drug candidates, there is growing demand for high capacity models that can help in driving the chemistry towards efficacious and safe candidates before progressing towards more complex models. Traditionally, the cardiovascular (CV) safety domain plays an important role in this process, as many preclinical CV biomarkers seem to have high prognostic value for the clinical outcome. Throughout the industry, traditional ion channel binding data are generated to drive the early selection process. Although this assay can generate data at high capacity, it has the disadvantage of producing high numbers of false negatives. Therefore, our company applies the isolated guinea pig right atrium (GPRA) assay early-on in discovery. This functional multi-channel/multi-receptor model seems much more predictive in identifying potential CV liabilities. Unfortunately however, its capacity is limited, and there is no room for full automation. We assessed the correlation between ion channel binding and the GPRA's Rate of Contraction (RC), Contractile Force (CF), and effective refractory frequency (ERF) measures assay using over six thousand different data points. Furthermore, the existing experimental knowledge base was used to develop a set of in silico classification models attempting to mimic the GPRA inhibitory activity. The Naïve Bayesian classifier was used to built several models, using the ion channel binding data or in silico computed properties and structural fingerprints as descriptors. The models were validated on an independent and diverse test set of 200 reference compounds. Performances were assessed on the bases of their overall accuracy, sensitivity and specificity in detecting both active and inactive molecules. Our data show that all in silico models are highly predictive of actual GPRA data, at a level equivalent or superior to the ion channel binding assays. Furthermore, the models were interpreted in terms of the descriptors used to highlight the undesirable areas in the explored chemical space, specifically regions of low polarity, high lipophilicity and high molecular weight. In conclusion, we developed a predictive in silico model of a complex physiological assay based on a large and high quality set of experimental data. This model allows high throughput in silico safety screening based on chemical structure within a given chemical space.