AI and Advanced Computing Accelerate Scientific Research at Duke

AI fuels Duke research with the potential to improve human lives, from modeling patient hearts with supercomputers and mining health data to predict ADHD risk in children to predicting how molecules will bind to proteins to develop more effective drugs for a range of diseases. 

Duke researchers who leverage AI have access to the Duke Computer Cluster, a high-performance computing resource for large-scale scientific applications. The university is also expanding its AI infrastructure with a small GPU center expected to open in 2027, which will be designed to minimize power and water consumption and carbon emissions through energy-efficient practices. Singh said that access to more GPUs will help accelerate his computational biology research.

“The investments Duke has been making have gone a long way in allowing us to address some of these questions, and from a competitive perspective, be one of the top places in the world that’s doing this kind of work,” he said. 

Accelerating Biological Discovery

Singh uses advanced computing to improve understanding of cell biology, analyzing millions of gene expression profiles and biological sequences with foundation models, which are machine learning models trained on vast datasets. Machine learning allows computer systems to learn patterns from data and make predictions, helping researchers work with biological data at a large scale while reducing the need for time-consuming manual annotation.

“In many ways I almost think of foundation models as microscopes,” Singh said. “They help us see biology with a new perspective. Each of them gives you a new aspect of life that you can study.”

Training models on “hundreds of millions of data points” requires a “massive computational effort to ingest it all and build a model that can understand it,” Singh said. Access to GPUs accelerates this process and makes it easier to build additional models adapted to specific types of research. 

Insights from Singh’s research can be applied to a wide range of projects, from developing gene therapies to identifying new drugs to treat a range of diseases. The machine learning models he trains can create abstract representations of proteins, genes and cells, allowing him to better understand the ways mutations or diseases may alter their functions. 

“If I can learn a good abstract representation of, for example, both healthy cells and cancer cells, then I can try to determine the difference between a cancer cell and try to figure out how I can zero out that difference,” Singh said. “Is there a drug that zeros out that difference?” 

Building Trustworthy AI

For Duke computer scientist Cynthia Rudin, the impact of advanced computing goes beyond faster hardware and encompasses new algorithms and machine learning methods that have changed the way AI research is done. 

“The field and what we can do keeps surprising me,” Rudin said. “In my lab, we’ve designed some algorithms that I didn’t think were possible at all.” 

For years, the standard for AI models has been “black boxes,” with internal workings that are hidden from the user. Rudin’s work focuses on building interpretable models that allow users to see how they make decisions for use in high-stakes domains such as healthcare and criminal justice. Interpretable models make it possible for experts to troubleshoot and evaluate the data and reasoning behind their predictions.  

Rudin’s algorithms have been used across a range of healthcare settings, from predicting seizures to analyzing medical images. One interpretable AI model, developed in collaboration with Duke radiologist Joseph Lo and other researchers, can predict a patient’s risk of developing breast cancer over the next 1 to 5 years by analyzing subtle imaging patterns in mammograms. 

“You should be able to see into models for high-stakes decisions,” Rudin explained. “With medical decisions, for example, there’s no reason that the algorithm can’t explain itself to you, or work with you rather than by itself.” 

Rudin and her team developed algorithms that allow physicians and practitioners to compare multiple highly accurate AI models and choose the one that best fits their needs rather than being given a single black-box prediction and expected to trust it.

“This completely changed the way we designed interpretable models, because now humans can look through the set of good models rather than just being handed a model that is reasonably good,” Rudin said. “That’s a problem I did not think would be solved in my lifetime. I thought it was just too computationally difficult, and now we can just do it in a fairly short amount of time for a reasonably sized data set.”

Turning Clinical Notes into Data 

Duke computer scientist Monica Agrawal has leveraged advanced computing to analyze electronic health records, which can help researchers better understand disease and identify opportunities for earlier intervention.

“How can we use clinical note data to automatically identify areas we might be able to intervene to improve patient outcomes?” Agrawal asked. “Patients generate a huge amount of data every time they interact with the health system. A lot of the real richness of what happened to a patient lives in their clinical notes because you can be expressive with language in a way you can’t with checkboxes and forms.” 

With advanced computing resources, researchers can analyze larger datasets than would be practical manually and evaluate a more diverse patient population. In a recent study, Agrawal and her collaborators used AI and large language models to review de-identified records from more than 40,000 patients to better understand menopause symptoms and disease risk. 

“An individual clinician might only encounter certain clinical scenarios a few times, but if you pull data from across these really large datasets, researchers can elucidate patterns beyond what any individual doctor would ever see,” Agrawal said. 

According to Agrawal, the availability of high-powered computing infrastructure like GPUs has significantly reduced the amount of time required for data analysis, enabling researchers to explore their hypotheses quickly rather than spending time manually “poring through thousands of charts.” It also enables them to more easily train and fine-tune AI models for their own datasets. 

“This speeds up a process that might have been months to a process that might be more on the order of days,” Agrawal said. “I think that’s just really exciting that we no longer have this huge bottleneck in how we try to analyze electronic health records.” 

To learn more about Duke’s diverse AI research initiatives, visit ai.duke.edu.