Scientists have finally cracked a major puzzle in genomics: understanding what most of our DNA actually does. A new artificial intelligence model developed at UC Berkeley, called GPN-Star, can identify which genetic variants contribute to inherited traits and diseases with unprecedented accuracy, while requiring only days or hours to train instead of months.

Why Does Most of Our DNA Remain a Mystery?

When researchers first sequenced the complete human genome in 2003, they decoded all 3 billion letters of our genetic code. But here's the catch: only about 1 to 2 percent of that code actually instructs cells to build proteins. The remaining 98 percent was long dismissed as "junk DNA," but scientists now understand it contains crucial regulatory elements that control when, where, and how strongly genes are expressed. These non-coding regions could hold the key to understanding why some people develop cancer, heart disease, autism, and countless other conditions, but first researchers need to understand how variants in this DNA contribute to disease.

How Does GPN-Star Outperform Other AI Models?

Most existing genomic AI models are trained on raw, unaligned sequences from hundreds of thousands of species, which demands enormous computing resources. The massive Evo 2 model, published earlier this year, required 2,000 powerful NVIDIA graphics processors and months of training to learn patterns across more than 100,000 species. GPN-Star takes a different approach. Instead of processing raw sequences, it learns from whole-genome alignments, which are pre-processed datasets that highlight similarities and differences between species across evolutionary time. This clever shortcut means the model can be trained in just days or even hours using only a handful of processors.

"Our model excels in making predictions about the pathogenicity of genetic variants, and identifying functional versus non-functional elements in the genome," said Yun Song, professor of computer science and statistics at UC Berkeley and an investigator at the Innovative Genomics Institute.

Yun Song, Professor of Computer Science and Statistics at UC Berkeley

The team trained GPN-Star on three different human-anchored whole-genome alignments, plus alignments for mice, fruit flies, chickens, roundworms, and plants. Each alignment represented different evolutionary timescales, from recent primate evolution to ancient vertebrate divergence. This multi-scale approach revealed something surprising: models trained at different evolutionary timescales were actually optimized for interpreting different kinds of genetic variants.

What Makes This Discovery Practically Useful?

The real power of GPN-Star lies in its ability to help researchers prioritize which genetic variants to study experimentally. Scientists cannot test every single variant in the human genome, but GPN-Star's predictions can highlight the variants most likely to influence inherited traits and disease risk. The researchers have already published genome-wide predictions from their model, making these annotations freely available to the broader scientific community.

"We hope our work will help drive biological discovery. People have developed really creative tools for assaying the impact of genetic variants, but they cannot experimentally test every single variant in the genome. We believe our predictions will help to prioritize the experiments that could have the greatest impact on human health," explained Yun Song.

Yun Song, Professor of Computer Science and Statistics at UC Berkeley

The model's efficiency also matters for global health. Because GPN-Star requires minimal computational resources to train and adapt, researchers around the world can modify and improve upon the work without needing access to massive computing clusters. This democratization of genomic AI could accelerate genetic discovery in countries and institutions that lack billion-dollar budgets.

How Are Other Organizations Scaling Genomic AI?

Beyond academic research, the genomics industry is rapidly adopting AI to handle massive sequencing datasets. Ultima Genomics, a leader in ultra-high-throughput DNA sequencing, recently announced a collaboration with NVIDIA and Google to enable pangenome-aware whole genome sequencing analysis at unprecedented scale. A pangenome is a reference that captures genetic diversity across many human populations, rather than relying on a single reference genome that can introduce bias toward European ancestry.

The collaboration combines Ultima's cost-efficient sequencing platform with NVIDIA's GPU-accelerated computing and Google's open-source DeepVariant software for variant calling, which identifies genetic differences in sequenced DNA. The teams are benchmarking these pangenome-aware approaches using genetically diverse human genomes and evaluating workflows based on community resources like the Human Pangenome Reference Consortium.

"Pangenomes are a critical next step for capturing global human genetic diversity and improving the accuracy of genomic analysis. With large-scale WGS initiatives accelerating globally, this work with NVIDIA and Google allows us to pair ultra high-scale sequencing with GPU-accelerated pangenome graph analysis to move these methods from research concepts into capabilities that can be operationally deployed to support large-scale efforts," said Gilad Almogy, CEO of Ultima Genomics.

Gilad Almogy, CEO of Ultima Genomics

How to Understand the Practical Impact of Genomic AI

• Rare Disease Diagnosis: AI systems like DeepRare can interpret complex genetic variants and propose diagnoses for rare diseases by combining active information retrieval, multi-step reasoning, and traceable evidence chains, transforming diagnosis from a guessing game into a systematic process.
• Population-Scale Sequencing: As health systems and national precision medicine programs adopt whole genome sequencing, pangenome-aware analysis reduces reference bias and improves diagnostic accuracy by capturing genetic diversity from people of all ancestries, not just European populations.
• Accelerated Research Cycles: By training AI models efficiently on curated evolutionary data, researchers can generate genome-wide predictions in days rather than months, enabling faster hypothesis generation and experimental prioritization across universities and biobanks worldwide.

These advances represent a fundamental shift in how scientists approach genomics. Rather than viewing AI as a tool that will magically solve biology on its own, researchers are now using AI to augment human expertise, prioritize experiments, and handle the computational complexity that would otherwise be impossible to manage. The convergence of efficient AI models, scalable sequencing platforms, and open-source software is creating a flywheel where new datasets improve diagnostics and drug discovery, which in turn create better tools and analysis standards for the entire scientific community.

The International Conference on Genomics held in Shenzhen in September 2026 highlighted this momentum, bringing together experts from around the world to explore how artificial intelligence is reshaping life science research under the theme "Bio-Generative Intelligence for Health". As these technologies mature and become more accessible, the promise of personalized medicine based on individual genetic profiles moves closer to reality.