Abstract
The optimal descriptors generated by the CORAL software are studied as potential models of cardiotoxicity. Two significantly different cardiotoxicity databases are studied here. Database 1 contains 394 hERG inhibitors (pIC50) and external 200 substances that are potential drugs, which were used to confirm the predictive potential of the approach for Database 1. Database 2 contains cardiotoxicity data for 13864 different compounds in a format where active is denoted as 1 and inactive is denoted as 0. The same model-building algorithms were applied to all three databases using the Monte Carlo method and Las Vegas algorithm. The latter was used to rationally distribute the available data into training and validation sets. The Monte Carlo optimization for the correlation weights of different molecular features extracted from SMILES was improved by including the conformity coefficient of the correlation prediction (CCCP). This improvement provided greater predictive potential in the considered models.