Exploring the Chemical Space of HDAC6 Inhibitors: A deep generative study using a Gradient Ascent Algorithm

This study aimed to implement the Junction Tree Variational autoencoder (JT-VAE) in conjunction with a gradient ascent algorithm to explore the chemical space of potential inhibitors of Histone deacetylase 6 (HDAC6). In the generation stage, a diverse subset of active compounds was identified using the Butina algorithm. These compounds were then subjected to chemical space exploration employing JT-VAE and a gradient ascent algorithm. The generated substances were subsequently reassessed using an Artificial neural networks model and molecular docking (PDB ID: 6CE6) studies. Thirty-one active compounds with a Tanimoto coefficient under 0.35 were identified from 5225 compounds collected from the ChEMBL database. These compounds underwent a chemical exploration stage, resulting in the generation of 303 novel substances. An artificial neural network-based quantitative structure-activity relationship (ANN-QSAR) model was constructed to predict the inhibitory values of these generated compounds, with external validation yielding an R2 value of 0.595 and an RMSE value of 0.643. Subsequently, a retrospective control protocol was used to determine the scoring function and the cutoff of binding affinity energy, using Deepcoys to generate decoys without bias at a ratio of 1:50 (active:decoys) through deep learning algorithm. Finally, Vinardo was chosen as the preferred scoring function in the GNINA software due to its superior ROC-AUC value of 0.715 when compared to two other scoring functions, and 13 distinguished compounds were identified with a pChEMBL Value threshold above 7 and binding affinity below -7.69 kcal/mol, representing a significant advancement in the field of HDAC6 inhibitor discovery. This multi-pronged approach efficiently identified potential inhibitors of Histone deacetylase 6 (HDAC6), suggesting the following stages involve synthesis and biological testing.