Utilization of Supervised Machine Learning to Understand Kinase Inhibitor Toxophore Profiles

Abstract

There have been more than 70 FDA-approved drugs to target the ATP binding site of kinases, mainly in the field of oncology. These compounds are usually developed to target specific kinases, but in practice, most of these drugs are multi-kinase inhibitors that leverage the conserved nature of the ATP pocket across multiple kinases to increase their clinical efficacy. To utilize kinase inhibitors in targeted therapy and outside of oncology, a narrower kinome profile and an understanding of the toxicity profile is imperative. This is essential when considering treating chronic diseases with kinase targets, including neurodegeneration and inflammation. This will require the exploration of inhibitor chemical space and an in-depth understanding of off-target interactions. We have developed an early pipeline toxicity screening platform that uses supervised machine learning (ML) to classify test compounds’ cell stress phenotypes relative to a training set of on-market and withdrawn drugs. Here, we apply it to better understand the toxophores of some literature kinase inhibitor scaffolds, looking specifically at a series of 4-anilinoquinoline and 4-anilinoquinazoline model libraries.