DrugBAN AI could cut costs and accelerate drug discovery

The new technology, developed by Professor Haiping Lu and his PhD student Peizhen Bai from Sheffield’s Department of Computer Science, with Dr Filip Miljković and Dr Bino John from AstraZeneca, is described in a new study published in Nature Machine Intelligence.

The study demonstrates that DrugBAN can predict whether a candidate drug will interact with its intended target protein molecules inside the human body. 

AI that can predict whether drugs will reach their intended targets already exists, but the technology developed by the researchers does this with greater accuracy and also provides insights to help scientists understand how drugs engage with their protein partners at a molecular level.

AI has the potential to inform whether a drug will successfully engage an intended cancer-related protein, or whether a candidate drug will bind to unintended targets in the body and lead to undesirable side effects for patients.

MORE FROM MEDICAL & HEALTHCARE

The AI is trained to learn the substructures of proteins in the human body as well as those of drug compounds. The technology then learns how these substructures can interact with each other, which it draws on to make predictions on how new medicines are likely to behave. 

In a statement, Haiping Lu, Professor of machine learning at Sheffield University, said: “We designed the AI with two primary objectives. Firstly, we want the AI to capture how drugs interact with their targets at a finer scale, as this could provide useful biological insights to help researchers understand these interactions on a molecular level. Secondly, we want the tool to be able to predict what these interactions will be with new drugs or targets to help accelerate the overall prediction process. The study we've published…shows our AI model does both of these.”

According to the team, the key to the AI’s design is how the model learns pairwise substructure interactions, which are the multiple interactions that can take place between substructures of drug compounds and proteins in the body. Most existing drug prediction AI learns from whole representations of drugs and proteins, which don’t capture their substructures and provide less useful insights.

In the next stage of the AI’s development, the team plans to use more in-depth data on the structure of compounds and proteins to make the AI even more accurate. 

Dr Bino John, director of Data Science, Clinical Pharmacology and Safety Sciences (CPSS), at AstraZeneca, said: “A key novelty of DrugBAN is its reliance on a bilinear attention network that allows it to learn interactions from substructures of both drugs and their targets simultaneously. We have also made the source code freely available to the public, which hopefully will support more AI approaches that will continue to accelerate drug discovery.”