New AI model for DNA learns from evolution to unlock secrets of the human genome
UC Berkeley researchers unveiled GPN-Star, an AI model that decodes non-coding DNA variants linked to disease using evolutionary alignments, offering faster, more efficient predictions than existing tools.
UC Berkeley scientists introduced GPN-Star, a genomic language model trained on whole-genome alignments to identify genetic variants influencing inherited traits and diseases. Unlike prior models, GPN-Star focuses on conserved and functional DNA regions by comparing genomes across species, reducing computational demands. The approach leverages evolutionary biology to prioritize data likely containing functional elements, improving accuracy in predicting variant impacts. Researchers trained the model on human-anchored alignments with primates, mammals, and vertebrates, as well as genomes of mice, fruit flies, and other species.
The model outperforms competitors in identifying pathogenic variants and regulatory elements, requiring minimal computing resources. GPN-Star was trained on data from three human-anchored alignments and five other species, using specialized algorithms to highlight conserved and divergent DNA sequences. By focusing on evolutionary conservation, the model avoids the noise of non-functional "junk DNA" that complicates other genomic AI systems. Its efficiency allows training in days or hours with few processors, compared to months for larger models.
GPN-Star’s predictions highlight genetic variants most likely to affect inherited traits, including those tied to cancer, heart disease, and autism. The team published genome-wide annotations to guide biologists in selecting targets for experimental validation. Study senior author Yun Song emphasized the model’s role in prioritizing high-impact experiments amid the vast number of untested variants in the human genome. The work was funded by NIH grants and published in Nature on September 9.
Researchers found that models trained on different evolutionary timescales excel at predicting distinct types of variants. For example, primate-focused training improved predictions for complex traits like schizophrenia risk, linked to thousands of mutations in non-coding regions. The team plans to release the model for broader use, aiming to accelerate genetic research worldwide. Co-authors included scientists from UC Berkeley, Jackson Laboratory, and the German Cancer Research Center.