UC Berkeley researchers release AI model that flags harmful DNA variants with evolutionary history

Overview

Updated 51 minutes ago

UC Berkeley researchers released GPN-Star, an AI model that flags disease-causing DNA variants by learning from millions of years of evolution. Trained in days on eight processors, it outperformed rivals hundreds of times larger at interpreting human DNA.

Only 1-2% of human DNA codes for proteins. Most disease risk sits in the noncoding majority, which geneticists have struggled to interpret. GPN-Star's genome-wide scores are free to download, giving researchers a way to rank which of the billions of possible variants to test first.

Why it matters

Clinicians can't test every genetic variant in a lab. GPN-Star's free genome-wide scores help them find the variants that cause disease.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

~200M

Model parameters

About 200 million settings, far smaller than rival genomic language models.

Training processors

Trained in several days on eight processors; Evo 2 used over 2,000 for months.

106

Human traits with top variant enrichment

Variants GPN-Star ranked highest explained more heritability across 106 traits than any earlier score.

Species with published genome-wide scores

Human plus mouse, chicken, fruit fly, roundworm, and Arabidopsis thaliana.

Voices

Curated perspectives — historical figures and your fellow readers.

Ever wondered what historical figures would say about today's headlines?

People Involved

Yun Song

Senior author; professor of computer science and statistics, UC Berkeley

Led the GPN-Star project; director of the Berkeley Center for Computational Biology

Gonzalo Benegas

Co-first author of the GPN-Star study

Research scientist in Yun Song's group at UC Berkeley

Chengzhong Ye

Co-lead author; graduate student in statistics, UC Berkeley

Studying how evolutionary timescale affects variant interpretation accuracy

Organizations Involved

University of California, Berkeley

Public research university

Institution where GPN-Star was developed

UC Berkeley's computational biology and statistics groups built GPN-Star.

Nature

Scientific journal

Journal that published the GPN-Star study

Nature published the GPN-Star study on September 9, 2026.

Evo 2

Genomic language AI model

Largest rival model; outperformed by GPN-Star on variant interpretation benchmarks

The largest genomic language model, trained on more than 100,000 species using over 2,000 NVIDIA processors.

Timeline

September 2025 September 2026

2 events Latest: September 9th, 2026 · 1 week ago

September 2026
GPN-Star study published in Nature
Latest Publication

Researchers published the study and released free genome-wide variant predictions for humans and five model species.
September 9th, 2026
September 2025
GPN-Star preprint posted to bioRxiv
Publication

The team posted a preprint describing phylogeny-informed genomic language models for predicting functional constraints.
September 21st, 2025

Scenarios

Clinical genetics lab adopts GPN-Star scores for variant triage

Possible Resolves by End of 2027

Discussed by: Yun Song and the UC Berkeley research team

Song designed GPN-Star to rank which variants deserve expensive, time-consuming experimental tests. Diagnostic labs and hospital genetics departments that interpret patient genomes could formally integrate its scores into their pipelines, cutting the number of variants that need lab validation. Adoption would likely start with labs that already use computational scores for triage.

Rival genomic language models adopt evolution-based training

Likely Resolves by End of 2027

Discussed by: Genomics AI researchers building large genomic language models, including the Evo 2 team

GPN-Star trained in several days on eight processors; Evo 2 needed over 2,000 NVIDIA processors running for months. That efficiency gap gives competing labs a strong reason to shift from raw-sequence-only training to whole-genome alignment training. If the field converges, GPN-Star sets a new architectural standard even where other labs use their own models.

GPN-Star stays a research tool as clinical limits surface

Possible Resolves by End of 2027

Discussed by: The UC Berkeley team's own method caveats

GPN-Star cannot assess insertions, deletions, or large rearrangements, because those don't fit the alignment grid. It also can't resolve selection within human populations, which the researchers say would require training on modern human and ancient hominin genomes. If clinical variant interpretation needs those capabilities, adoption stays limited to research annotation.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

October 1990 – April 2003

Human Genome Project (1990–2003)

The international Human Genome Project sequenced all 3 billion base pairs of human DNA over 13 years at a cost of roughly $3 billion. When it finished, scientists had the letters but not the meaning: most variants discovered in patient genomes could not be classified as harmful or harmless.

Then

The project triggered a wave of research into genome interpretation and the discovery of disease-associated variants.

Now

Interpretation remains the bottleneck; clinical genetics still struggles to classify the majority of variants found in sequencing.

Why this matters now

GPN-Star attacks exactly the bottleneck the Human Genome Project left behind: interpreting what the letters do.

September 2012

ENCODE project (2012)

The international ENCODE consortium assigned biochemical function to 80% of the human genome, prompting debate about what 'functional' means. The claims drew criticism that biochemical activity is not the same as evolutionary function.

Then

ENCODE's data became a research staple, and the debate sharpened definitions of genome function.

Now

The field moved toward combining biochemical data with evolutionary conservation to identify meaningful functional elements.

Why this matters now

GPN-Star takes the evolutionary route, using conservation across species rather than biochemistry to mark which DNA positions matter.

November 2020

AlphaFold 2 (2020)

DeepMind's AlphaFold 2 used deep learning to predict protein 3D structure from amino acid sequence, solving a problem experimentalists had worked on for 50 years. Its predictions matched experimental accuracy for a large share of proteins.

Then

Structural biologists adopted AlphaFold predictions as a starting point for experiments, accelerating fields from drug design to enzyme engineering.

Now

AI prediction became a standard first step, with experiments used to validate the most promising hits.

Why this matters now

GPN-Star aims to do for genetic variant interpretation what AlphaFold did for protein structure: produce high-quality predictions that let researchers prioritize experiments.