Duke University researchers are applying advances in computing and artificial intelligence to expand knowledge of biology and human health, while also building AI systems that people can understand and trust. The work spans computational biology, interpretable machine learning for healthcare, and large-scale analysis of electronic health records, supported by high-performance computing resources including GPUs.
"AI has been a massive multiplier in terms of answering existing questions, but the path which is super interesting to me is around the new questions we'll be able to ask," said Rohit Singh, a computational biologist at Duke University School of Medicine who uses AI to find patterns across millions of cells.
Decades of progress in computing have produced more sophisticated machine learning algorithms and better hardware, including graphics processing units that enable researchers to train large AI models and analyze large-scale datasets rapidly. Tasks that once took weeks can now be completed in days.
Accelerating biological discovery
Singh uses foundation models - machine learning models trained on vast datasets - to analyze millions of gene expression profiles and biological sequences. The models create abstract representations of proteins, genes, and cells, helping researchers understand how mutations or diseases alter their functions.
"In many ways, I almost think of foundation models as microscopes," Singh said. "They help us see biology from a new perspective-each one giving you a new aspect of life that you can study."
Training models on hundreds of millions of data points requires massive computational effort, Singh said. Access to GPUs accelerates the process and makes it easier to build additional models adapted to specific research questions.
The insights from Singh's research can apply to projects ranging from developing targeted interventions to identifying new drugs. If he can learn a good abstract representation of both healthy cells and cancer cells, he can try to determine the difference between them and ask whether a drug exists that zeros out that difference.
Duke is expanding its AI infrastructure with a small GPU center expected to open in 2027, designed to minimize power and water consumption and carbon emissions. Singh said access to more GPUs will accelerate his computational biology research and help Duke remain competitive globally. For research scientists looking to build similar skills, an AI Learning Path for Research Scientists covers the core techniques used in this kind of work.
Building trustworthy AI
For Duke computer scientist Cynthia Rudin, the impact of advanced computing goes beyond faster hardware. New algorithms and machine learning methods have changed how AI research is done.
"The field and what we can do keeps surprising me," Rudin said. "In my lab, we've designed algorithms that I didn't think were possible at all."
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. This contrasts with "black box" AI models whose internal workings stay hidden from the user.
Her algorithms have been used in healthcare settings including interpreting medical images to predict seizures and analyzing mammograms. One interpretable AI model, developed with Duke radiologists, 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 a tool that lets physicians comparing AI models understand their reasoning. Clinicians can choose from a range of accurate models that make decisions they can see, rather than being handed a single black-box prediction. The team found a way to quickly identify a reason for a single recommendation or revisit the set of good models for a more informed decision.
"This changed the way we design interpretable models, because now humans can look through the set of good models rather than just being handed a single model that is reasonably good," Rudin said. "That's a problem I did not think would be solved in my lifetime, because it could be very computationally difficult, but now we can just do it in a short amount of time for a reasonable size of data set."
Turning clinical notes into data
Duke computer scientist Monica Agrawal uses advanced computing to analyze electronic health records, helping researchers better understand disease and identify opportunities for earlier intervention.
"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 lies in clinical notes, given the flexibility of language in a way you can't capture with checkboxes and forms," Agrawal said.
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. Advanced computing resources let researchers analyze larger datasets than would be practical manually, and evaluate a more diverse patient population.
"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.
GPUs have significantly reduced the time required for data analysis, Agrawal said. Researchers can explore hypotheses quickly rather than manually poring through thousands of charts, and can 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 process in how we try to analyze electronic health records."
Why this matters for research scientists
For research scientists, the Duke work shows how AI is shifting from a tool for answering existing questions to a method for asking new ones. Foundation models trained on massive datasets can reveal patterns in biology, medicine, and clinical data that would be impossible to find manually. The practical takeaway: researchers who build skills in training and fine-tuning these models - and who understand how to make them interpretable - will be positioned to work at a scale that was previously unavailable. That includes access to GPU infrastructure, which is becoming a competitive differentiator for institutions and individual researchers alike. Those looking to develop these capabilities can explore AI for Science & Research resources covering the relevant methods and applications.
Your membership also unlocks:
