(152b) Evaluation of Sampling Algorithms to Explore the Latent Space Created By Deep Generative Models

In generative chemistry, artificial intelligence based design of small molecules employs a solution to discover new chemical compounds with target properties by systematically exploring a large design space. This design space is normally explored with stochastic sampling. Thus, new candidate discovery is not only dependent on initial data but also on the sampling algorithms. In this study, different sampling methods such as metropolis-hastings, ensemble slice sampling, and multi-grid search are investigated to understand the latent space exploration performance for new molecules. Specifically, we utilize syntax-directed variational autoencoders (SD-VAE) to create a latent space for designing new drug-like molecules with strong binding energy (< -9.0 kcal/mol) to SARS-Cov-2 Spike protein. The sampling algorithms were integrated with a stacked ensemble model (SEM) for efficient exploration of the latent space. Then, the drug-like molecule candidates from different sampling algorithms are validated through docking using Autodock Vina as well by performing additional structural and pharmacophore models. The top candidates were later functionalized with functional groups based on glycomaterials to design biocompatible drug-like molecules with better binding affinities to the Spike protein. This machine learning approach may help us find good synthesizable carbohydrate-based drug-like molecules for Covid-19 treatment and might be extended to discover micromolecules for different applications.