Abstract

Computer-Aided Drug Design (CADD) is indispensable in modern drug development, enabling researchers to optimize potential therapeutic compounds with greater speed and accuracy. The integration of Artificial Intelligence (AI) has transformed CADD by allowing for rapid exploration of chemical space and targeted optimization of molecular properties. MetaDrug is an AI-based platform developed to accelerate molecular optimization during the Hit-to-Lead phase in drug development, therefore reducing costs. Using a Junctional Tree Variational Autoencoder (JT-VAE) framework that was trained on a 4.5 million compound diverse dataset from the ZINC database, a unique graph-based molecular representation of the target molecule is transformed into an embedding in the latent space. The embedding then undergoes iterative modifications via a genetic algorithm whose role is the optimization of specific properties that are predicted using fitness functions. Currently, our fitness functions cover synthetic accessibility, solubility, half-life and toxicity. All of them are derived either from experimental data obtained from public databases or from calculated properties. MetaDrug's genetic algorithm operates directly on the latent space, bypassing repeated molecular reconversion and significantly reducing processing time. Once a convergence criteria has been met, the embedding is decoded into a molecule once again. Initial validation of generated molecules has shown promising in silico results, affirming MetaDrug's potential to minimize costly and time-consuming traditional optimization steps that require chemist-led modifications. We’re currently capable of generating families of molecules with enough structural similarity to be part of an analog program. With its capacity for rapid, robust optimization, MetaDrug is designed to complement existing in silico tools and holds particular appeal for academic research and small pharmaceutical companies seeking scalable, AI-driven solutions to advance drug development pipelines.