A new study has developed a powerful computational method that can detect how genes interact with each other to influence complex traits in humans at a scale previously impossible. The new method was applied to massive datasets that pair individual genomes and traits to find evidence for such interactions. The findings, published in Nature Genetics, show that a person's genetic background can substantially modify how individual genetic variants affect their traits.
Why it matters
Understanding how genes interact to influence complex traits, such as Body Mass Index or total cholesterol levels, in humans is important for basic biology and precision medicine. Such interactions could explain why two individuals that carry the same genetic risk factor can have different health outcomes or why attempts to use genetic information to predict disease risk from genetics can have poorer accuracy than expected. Evidence for such interactions has been limited partly because of the lack of tools that could deal with the complexity and scale of large genomic datasets.
What the study did
This study develops a powerful computational method, FAME, that can detect and quantify the impact of genetic interactions. Previous approaches, typically attempting to find pairs of genetic variants that interact to influence a trait, struggle to detect interactions that have weak effects. FAME, instead, seeks to find genetic variants whose effect on a trait is modulated by all other genetic variants in a person's genome. By aggregating these weak effects, FAME is able to detect interaction signals that would have been missed by previous approaches. However, aggregating the interaction effects arising from hundreds of thousands of variants across the genome is a computationally daunting challenge. FAME uses sophisticated mathematical techniques that reduce the computational requirements while retaining accuracy.
What they found
Applying FAME to the UK Biobank, a massive dataset containing genetic and trait information from nearly 300,000 European ancestry individuals, the research team found 16 instances of interactions on traits that include cholesterol, liver enzymes, and testosterone. While the interaction effects tend to be small, they are often larger than the effect sizes identified by examining the variant by itself (as is common in Genome-Wide Association Studies of complex traits). Additionally, several of these interaction signals were found to replicate in an independent dataset strengthening the robustness of the findings.
What's next
The team plans on extending FAME to inspect rare variants and disease traits, and to localize the interactions in the genome. While the current study used a dataset of almost exclusively white individuals living in the UK, the team also plans on investigating how interactions vary across diverse populations.
From the experts
"This gives us a new window into how genes work together to influence traits," said senior author, Sriram Sankararaman, who is a professor of Computer Science, Computational Medicine, and Human Genetics at UCLA.
"For years, we've known that genes don't act in isolation, but we lacked the computational tools to detect these interactions at scale," said lead author, Boyang Fu, previously a Ph.D. student at UCLA and now a Postdoctoral Fellow at Harvard Medical School. "FAME allows us to test for genetic interactions across hundreds of thousands of people and millions of genetic variants in a matter of hours—something that was previously impossible. It opens a new exciting perspective to look at sequence to function relationships."
About the study
"A biobank-scale test of marginal epistasis reveals genome-wide signals of polygenic interaction effects." Nature Genetics 2025 DOI: https://www.nature.com/articles/s41588-025-02411-y
Research Team
Boyang Fu, Harvard University; Ali Pazokitoroudi, Asha Kar, Albert Xue, Aakarsh Anand, Prateek Anand, Zhengtong Liu, Päivi Pajukanta, Noah Zaitlen, and Sriram Sankararaman, UCLA; Zhuozheng Shi, University of Pennsylvania; Richard Border, Carnegie Mellon University.
Funding and Disclosures
The study was supported by grants from the NIH (GM125055, GM153406, HG006399, R01MH130581, U01MH126798, R01MH122688, R01GM142112, R01HL170604, and R01DK132775), the NSF (CAREER-1943497) and the UCLA Dissertation Year Fellowship. The authors declare no competing interests.
