Abstract

The bromodomain family presents a druggable family of epigenetic proteins, for which multiple recent drug discovery efforts have yielded selective small molecule probes, enabling elucidation of therapeutic roles in oncology, inflammation and cardiovascular disease. Although progress has been made, the design of selective bromodomain inhibitors is still a challenge, particularly due to the lack of understanding of the drivers of selectivity. This study integrates bioactivity data from the public domain, as well as internal data from AstraZeneca to build the first in silico proteochemometric models to predict compound selectivity profiles for the bromodomain target family. The random forest algorithm was employed to build highly predictive classification models, utilising well-validated Morgan fingerprints and physicochemical property descriptors to describe the compounds and alignment dependent amino acid property descriptors to describe the targets. The models outperformed global quantitative structure activity relationship (QSAR) predictions and also quantitative sequence activity models (QSAM). Interpretation of the models led to the identification of key chemotypes and protein hotspots that contribute to the binding interaction between small molecules and bromodomains.