UC Berkeley scientists have made a breakthrough in more efficiently linking genetic variants to unique individual traits such as inheritable disease.
The human genome has been mapped in its entirety, but unraveling the DNA to determine its association with human traits is another realm in which UC Berkeley researchers have now broken ground.
An AI model called GPN-Star is a new advancement in genomic language models, which researchers train with DNA sequences the same way they would train large language models to understand human languages. It’s allegedly outcompeting other large models such as Evo 2 and Nucleotide Transformer, which, unlike GPN-Star, aim to solve a wider range of tasks in the field.
“This is just the beginning,” said UC Berkeley professor Yun Song. “I hope that this is going to open up many doors for us and for others, and I’m excited for the opportunities.”
The team published the results in the journal Nature on Sept. 9 and has made all of the model’s predictions publicly available. Chengzhong Ye and Gonzalo Benegas, two recent UC Berkeley Ph.D. graduates, developed the model alongside Song, who is the corresponding author.
GPN-Star can predict the genetic variants that are most likely to be linked to diseases, predictions that could improve preventative measures for inheritable diseases. This data could also refine disease screening processes and open doors for more therapeutic targets for inheritable diseases, according to Brian Clarke, who is a co-author of the paper.
The model uses data from whole-genome alignment — a process that unveils similarities between different species’ DNA — to parse genomic differences. GPN-Star compares the human genome to those of primates, mammals and other vertebrates. Ye said this data gives the model more “explicit information about evolution.”
Arc Institute’s Evo 1 and Evo 2, both of which are generative AI models, developed genome sequences from scratch. Evo’s results sparked concerns over how AI might be misused or could pose a danger to public health.
Researchers stressed that GPN-Star is not a generative model but rather has one main function: to make predictions about genetic variants’ effects.
“We are using a more efficient — while more specialized — design. What we achieved is a better performance on this task of variant effect prediction with a much smaller model and much less computing,” Ye said.
Ye said the team hopes publicly accessible data will benefit the scientific community and make advancements in AI modeling easier.
Making GPN-Star’s predictions public sets the stage for future studies, according to Song.
“We really hope that people, not just at Berkeley but worldwide, will use our predictions to enable creative investigations in genomics,” Song said. “It could be (related to) human health, evolutionary biology — I think there are many applications that we envision this model to enable.”