Graph Neural Networks for Drug Discovery: An Integrated Decision Support Pipeline

Drug Discovery is a fundamental discipline that is needed to produce new pharmacological solutions to the ever-evolving challenges of healthcare systems. Obtaining a new drug is a long and costly process, often involving many institutions and companies. In recent years, many proposals of support systems based on deep learning have been devised, in order to integrate the traditional methods and cut down the costs, in terms of time and money, to obtain a new approved drug. These deep learning based methods can help in different stages of the process, from the proposal of drug candidates to the evaluation of how to deploy them on the market. In the first stages of the process, where the molecule gets designed and undergoes several chemical evaluations, many support tools based on Graph Neural Networks have been proposed. Graph Neural Networks are deep learning models and have been demonstrated to be universal approximators on graphs, a characteristic which represents a great advantage with respect to other deep learning methods when dealing with molecular graphs. Graph Neural Networks have been successfully employed in molecular graph generation, drug side-effect and polypharmacy prediction, and in many chemical classification and regression tasks. This makes them the ideal model to build a drug-discovery support pipeline which can assist in the first stages of the drug discovery process: designing the molecular structure of candidate drugs, evaluating their chemical characteristics, predicting their probable side-effects. This paper is intended as a proof of concept of this pipeline, providing its outline and highlighting the future challenges in this direction.