Yang Lu likes to say that searching the human genotype for a biomarker of a given disease is akin to trying to find a needle in a haystack. Or, more accurately, needles—the set of biomarker interactions out of millions of possible combinations that drive that disease’s progression.
It’s a monumental challenge beyond the analytical limits of the human brain. Artificial intelligence can help: Machine learning models can identify solutions from mountains of data, but they also look for shortcuts that can lead to false positives. And it’s important that biomedical researchers can understand why the model generated its results when they decide which candidates to experimentally study.
With all that in mind, Lu and his collaborators have created a new method, called Diamond, for interaction discovery with rigorous error control. Detailed in a September 2025 paper in the journal Nature Machine Intelligence, the researchers’ system works with a wide range of machine learning models to map genetic makeup (genotype) to genetic expression (phenotype). Diamond generates disease-specific hypotheses for researchers to further investigate.
“Biologists usually cannot afford to do experimental evaluation for 100 or 1,000 of these gene interactions,” says Lu, an assistant professor of biomedical engineering at the University of Wisconsin-Madison. “Due to budget limitations, they can only afford, say, 10. How do we make sure the 10 we provide them with are guaranteed to be the 10 most likely to trigger the disease?”
The project spans Lu’s time as a postdoctoral researcher at the University of Washington, two-and-a-half years on the faculty at the University of Waterloo and his arrival at UW-Madison in fall 2025. He sees it as a significant step toward the “holy grail” of computational biology: Harvesting scientific discoveries directly from data, without requiring repeated rounds of experiments.
“AI models are powerful in building this genotype-to-phenotype mapping, by capturing subtle patterns,” says Lu. “Once we have this model, it’s not the end of the story. It’s just the beginning of the story. We want to interrogate this model.”
The interactions between potential biomarkers are important, because they can produce outsized effects—by enhancing or repressing each other—on genetic expression compared to the sum total of their individual roles. Diamond scores each interaction’s synergistic effect and delivers a false discovery rate, a rigorous estimate of the odds that it’s wrong.
Lu and collaborators at Washington and Waterloo validated their system on datasets for disease progression in diabetes, DNA enhancers in fruit fly embryos, and mortality risk factors and outcomes.
Next, Lu’s lab plans to develop tools to sift through data to identify causal relationships and genetic pathways to disease.
“Essentially, this will be a more mechanistic understanding of what will happen in causing disease,” he says. “And we want to all of this being interrogated from this AI model. This is the big picture of what our lab is pursuing.”
Other authors on the paper include Winston Chen, a former undergraduate student at the University of Washington and current PhD student at the University of Michigan; Yifan Jiang, a PhD student at the University of Waterloo; and William Stafford Noble, a professor in the Department of Genome Sciences and in the Paul G. Allen School of Computer Science and Engineering at the University of Washington.
Top image via iStock.