New AI tool developed at the University of Oregon could speed discovery of life-changing new drugs

EUGENE, Ore. (KTVZ) — For every life-changing new drug that comes to market, many candidates fail along the way. An artificial intelligence-based tool developed at the University of Oregon could help scientists better predict how hypothetical new drugs might act in the body before running expensive tests.

Their algorithm efficiently simulates how never-seen-before molecules will move and behave, based on their chemical structure, U of O said in a news release Tuesday that continues in full below.

The research was led by doctoral student Revanth Elangovan and postdoctoral researcher Sompriya Chatterjee, in the lab of biophysicist Dhiman Ray. The team recently published the findings in Proceedings of the National Academy of Sciences.

“If my research is successful, in 10 or 20 years, we’ll be able to do simulations in real time and see how different types of drugs bind to a protein,” Ray said. “And from there, we can pick which ones we should carry forward to human trials.”

Scientists already use advanced computational tools to predict the structure of new molecules. Programs like Google’s AlphaFold can help researchers rapidly screen candidates for vaccines and other drugs based on their shape. These programs have changed the field of drug development — the scientists who built them even won the 2024 Nobel Prize in Chemistry.

While extremely powerful, structure-prediction tools are like taking a still frame from a movie, Ray said. They capture a moment in time, but don’t tell you much about the action.

To really predict how molecules will behave in the human body, scientists need to understand how they move and interact. Will the molecule stick to the desired target in the body, or will it more likely stick to something off-target?

Current simulation approaches are computationally intense, Ray said. Like a full-length feature film, they’re rich in detail but expensive to produce, and therefore out of reach for many labs.

His team’s new approach is something like a Wikipedia plot summary, a happy medium between a photo and a feature film.

The output gives scientists the basic information they need to predict how molecules will move — the end result, but also the most efficient route to get there.

Ray’s team created the model by blending artificial intelligence with physics data. Scientists used measurements of how known molecules behave in different situations and the amount of energy it takes to get them to change shapes to put parameters on the AI model, preventing it from wandering off track and wasting energy exploring unlikely scenarios.

And mathematical techniques increase the likelihood that researchers will capture the specific moment they seek: the fraction of a second when a drug binds to its target.

“That’s the power we have, being computational scientists: We can do these tricks to sample things that we cannot see in a regular simulation,” Ray said.

The code they’ve developed is freely available for others to use.

Ray and his team are particularly interested in applications in drug development. But the approach could be useful across biology and chemistry, helping researchers understand the properties of many new materials.

A next step is to make the data from the model easier to interpret, translating what comes out into a user-friendly short movie.

“We keep some of the physics insights and some of the advantages of machine learning, and combine them in a way that benefit both fields,” Ray said. “That’s the uniqueness of the projects that we are doing - not just this particular paper, but also the lab overall.”

— By Laurel Hamers, University Communications