Researcher’s AI-powered drug discovery tool supported by Empire AI

UB researcher Thomas Grant has been awarded a $2.18 million federal grant to create a new artificial intelligence-powered tool that improves scientists’ understanding of how proteins move and change shape within the human body.

The tool, called SWAXSFold, ultimately aims to speed up drug development and help scientists design better, more precise medicine for people suffering everything from cancer to Alzheimer's disease.

“The basic idea is that we’re revolutionizing how we study proteins in their natural environment,” says Grant, assistant professor of structural biology, Jacobs School of Medicine and Biomedical Sciences. “Proteins are constantly moving and changing shape. Current methods often give us static snapshots, but we need to see them in action to really understand how they work and how to design drugs that target them effectively.”

The work is supported by the computing power of Empire AI, the $500 million New York State-based research consortium advancing AI for the public good. Grant is utilizing Empire AI’s supercomputing center, located at UB, to train and validate SWAXSFold.

“Empire AI is a big part of this, as we wouldn’t be able to do this level of computation without it” he says.

SWAXS is short for small- and wide-angle X-ray scattering. It is akin to taking a very sophisticated X-ray of proteins in solution, Grant says.

“When you shine X-rays on a sample of proteins dissolved in water, the X-rays scatter off the proteins in specific patterns,” he says. “By measuring how the X-rays scatter at different angles, both small angles and wide angles, we can extract information about the protein’s size, shape and internal structure.”

SWAXS works with proteins in their natural environment, dissolved in water at room temperature, just like they are in human cells. It eliminates the need to crystallize or freeze them. It’s also fast and can handle proteins of almost any size.

The downside, Grant says, is that the data are more challenging to interpret than current methods. That’s especially true in the wide-angle region, which contains information about the protein’s internal structure. Researchers often rely on crude measures to say whether a protein model is good or bad.

“It’s kind of like having a thermometer that only tells you ‘hot’ or ‘cold’ instead of giving you an actual temperature,” he says. “Our method is like giving you the actual temperature — not only how precisely ‘good’ or ‘bad’ the model is, but also which parts of the model are better resolved than others.”

The “Fold” part of SWAXSFold is a nod to AlphaFold, an AI program that won the 2024 Nobel Prize in Chemistry for using deep learning to predict protein structures.

“We’re basically taking the AlphaFold approach and integrating it with experimental data,” Grant says.

AlphaFold predicts what a protein structure might look like based on its amino acid sequence, but proteins are dynamic and can adopt many different shapes.

“AlphaFold doesn’t know which shape the protein actually has under your specific experimental conditions,” he adds. “SWAXSFold solves this by directly incorporating SWAXS experimental data into the AI prediction process.”

So instead of just giving the AI a protein sequence and asking, “what might this look like?” — the researchers are giving it the sequence plus experimental data and asking, “what does this actually look like under these specific conditions?”

“Nobody’s done this before — integrating experimental structural data directly into the AI training and prediction process,” Grant says.

The research is important for drug discovery, Grant says, because drugs need to bind to the actual shape the protein has in a person’s body, not just any possible shape it could have.

“Right now, a lot of drug development fails because we don’t understand the true, dynamic shapes of proteins well enough,” Grant says. “By giving researchers tools to see how proteins actually move and change shape in conditions similar to what they experience in the human body, we can design much better, more targeted drugs.”

This is especially important for proteins that are involved in diseases but have been hard to develop drugs against because they’re so dynamic or because they don’t have obvious binding pockets for drugs — such as cancer-causing proteins that change shape, or proteins involved in neurological diseases.

“We’re also developing tools that will help researchers understand how disease-causing mutations change protein structure,” he says. “If we can see exactly how a mutation alters a protein’s shape and function, we can design personalized therapies targeted to that specific change.”

The five-year grant is from the National Institute of General Medical Sciences, part of the National Institutes of Health.