AI-driven molecular modeling and design: from property prediction to drug generation

Integrating the techniques of deep learning, particularly graph neural network models, has made a significant advancement in drug discovery by facilitating effective exploration of chemical spaces and precise prediction of molecular properties. This paper presents a deep learning approach: molecular encoder, property predictor, molecular generator, and realness classifier (ME&PP-MG&RC). which integrates the molecular encoding and property prediction (ME&PP) component and the molecular generation and realness classification (MG&RC) component to help the prediction and discovery of new molecular compounds. Encoding molecular structures in latent space using graph neural network models, our approach allows for accurate prediction of properties such as quantum mechanical properties (HOMO, LUMO, and gap) for QM9 and pharmaceutical properties (QED, LogP, and SAS) for ZINC datasets. ME&PP component achieves the highest performance in property prediction with lower average absolute errors compared to existing methods after evaluation. Additionally, the approach produces molecules that are chemically diverse and valid, and they are checked by a realness classifier to make sure that they are practical. Results demonstrate the potential of ME&PP-MG&RC to revolutionize computational molecular design and drug discovery, providing a unified pipeline for precise property predictions, new molecular generation, and reality validation. This integrated approach represents a major step towards the automation of drug discovery and paves the way for the accelerated identification of innovative therapeutic candidates.