HVLGAN: hybrid hierarchical scaled attention-enabled latent model for structure-based drug discovery

Structure-based drug design involves utilizing the three-dimensional structure of a biological target to guide the design and development of new therapeutic compounds. Traditionally, a huge number of structure-based drug discovery methods have been adopted, but their time-consuming, erroneous molecule formation, and highly complex characteristics prevent their extensive application in drug discovery. Therefore, to mitigate such intricacies, an effective Hybrid Hierarchical Scaled attention-enabled Variational Autoencoder-based Latent Generative Adversarial Network (HVLGAN) is proposed. The inclusion of the Graph-based pocket encoding (GPE) aided in the effective generation of the Simplified Molecular Input Line Entry System (SMILES) strings to stipulate the drug discovery process with reduced computational complexity. Further, the Hybrid Hierarchical Scaled (H2S) attention strategy generates additional significant details for the effective generation of new drug molecules. In addition, the incorporation of the latent encoder and decoder enhanced the drug discovery performance by effectively processing the high-dimensional features. Nevertheless, the Variational Autoencoder (VAE) alleviated the long-term dependency problems, thereby resulting in a faster drug discovery process. Moreover, the performance validation performed in terms of performance metrics showed efficacy by attaining 0.96 validity, 0.96 novelty, and 0.96 unique scores for 90 training percentages using the MOSES package.