Pharmaceutical giant AstraZeneca in collaboration with the University of Sheffield, has developed an artificial intelligence (AI) that could cut the cost and expedite the development of new medicines. This latest technology created by Prof Haiping Lu and his Ph.D. student Peizhen Bai in the computer science department at the University of Sheffield with AstraZeneca’s Dr. Bino John and Dr. Filip Miljkovic is described in a Nature Machine Intelligence publication.
Scientists develop AI that predicts interactions of drugs
The study shows that the AI dubbed DrugBAN is capable of predicting if a drug candidate will interact with a target protein molecule in the human body. Currently, there is an AI that can predict if drugs reach their expected targets, but the latest advancement by AstraZeneca and Sheffield University can accurately predict the interaction.
AI has the potential to offer insight into how a drug will interact with cancer-related proteins or if the candidate drug can bind to targets in the human body and result in unexpected side effects for patients. The AI has been trained to understand the protein substructures in the human body and those of drug compounds. Additionally, the tech will learn how the substructures interact with one another and then predicts how new drugs will behave.
AI is designed to learn how drugs interact with proteins
Machine Learning Professor at the University of Sheffield Haiping Lu said that they designed the AI with two goals. The first objective is to capture how drugs interact with targets at a precise scale which could offer important insights for scientists to understand the interactions at the molecular level. The other aim was to predict how the interactions would happen to accelerate the prediction process.
Important to the AI design is how the model draws insight into pairwise substructure interactions, which are the interactions that take place between proteins in the body and substructures of drug compounds. On the other hand, the majority of drug predictions AI available on the market get their learning from complete models of medicines and proteins that don’t account for their substructures and hence offer less insightful data.
