De Novo Molecular Generation Augmentation for Drug Discovery Using Deep Learning Approaches: A Comparative Study of Variational Autoencoders

Background: Drugs are defined as chemicals that induce physiological effects when ingested, and their development involves multiple stages, including discovery, design, and development, which are often complex and resource-intensive. To address these challenges, machine learning (ML) and deep learning (DL) techniques have emerged as powerful tools to optimize the drug development pipeline. Methods: This study utilized two distinct variational autoencoders: a convolutional encoder-decoder model and a convolutional-GRU-based encoder-decoder model. Employing a reparameterization technique, we aimed to improve the efficiency of de novo molecular generation. Both models were trained and evaluated on the ZINC dataset, assessing their capability to generate chemically valid and syntactically accurate molecules.Results: The convolution-GRU model demonstrated a synthesis accuracy of 96.79%, matching the performance of the convolutional encoder-decoder model. Additionally, the chemical validity of the generated compounds was notable, with unique chemical validity scores of 90.71% for the convolutional encoder-decoder model and 90.42% for the convolution-GRU model. Conclusion: The results indicate that deep molecular generative models, especially the convolution-GRU approach, significantly advance de novo molecular design. By achieving high levels of accuracy and chemical validity, these models hold promise for enhancing drug discovery processes and expediting the introduction of new therapeutics to the market.