Fig. 6
From: CardioGenAI: a machine learning-based framework for re-engineering drugs for reduced hERG liability
Visualization of the CardioGenAI framework applied to nintedanib (A, B), pimozide (C, D), ibutilide (E, F), halofantrine (G, H), and astemizole (I, J). In each application, the specified maximum predicted hERG pIC50 value of any of the generated compounds was set to 6.00. For each optimization, the input molecule, the 100 generated refined molecules, and the molecules in the training set for the transformer-based models (approximately 5 million datapoints), are projected into a principal component analysis (PCA)-reduced physicochemical-based space. The input compound is colored yellow, the generated refined compounds are colored purple, and the compounds in the training set of the transformer-based models are colored red. The first two principal components explain 45.07% and 17.61% of the total variance, respectively. In each case, the CardioGenAI framework is able to identify the region of physicochemical space corresponding to compounds that are similar to the input compound, yet exhibit significantly reduced activity against the hERG channel. The densities of predicted pIC50 values against the hERG channel of the generated refined compounds as compared to that of the respective input compound are shown in [B]. Relevant metrics are shown on each plot