Harnessing Deep Belief Networks for Selective HDAC6 Inhibitors Identification

Histone Deacetylases (HDACs) play crucial roles in maintaining normal body functions and are considered in targeted drug therapy, particularly for cancer treatment. While promising inhibitors for these enzymes have been developed and several drugs targeting this enzyme class are approved, their lack of selectivity across HDAC isoforms may result in dose-dependent side effects. Hence, the design and discovery of selective inhibitors, especially for the HDAC6 isoform, is desired and well explored, though without any approved selective drug as of yet. In this study, we utilized molecular modeling tools to discover selective HDAC6 inhibitors. We developed and evaluated multiple machine learning models to predict compound selectivity for HDAC6 over other common isoforms such as HDAC1 and HDAC4. Notably, deep learning-based Deep Belief Network (DBN) models, which demonstrated superior ranking performance, were employed in a virtual screening of a molecular library for potential HDAC6-selective compounds. Additionally, physics-based molecular simulation tools such as molecular dynamics studies were conducted to assess the interactions between selected hits from the library and the HDAC6 enzyme. Consequently, we identified three compounds predicted by DBN models to be selective, which showed favorable interactions with HDAC6, comparable to those of a known HDAC6 inhibitor.