UC Berkeley's GPN-Star predicts harmful DNA variants using a fraction of rival compute

Genomic language models have become one of the loudest promises in computational biology. The premise is simple: treat DNA, a string of roughly 3 billion base pairs, as a language, and let a neural network learn its grammar well enough to separate harmless differences between people from variants that cause disease. The stakes are high because only 1 to 2 percent of the human genome codes for proteins, which means the overwhelming majority of the sequence has been hard to interpret. Many variants linked to common conditions such as schizophrenia sit in that poorly understood non-coding territory, and some of those conditions are associated with as many as 10,000 mutations.

Scale has been the field's main strategy so far. The most ambitious models are trained on unaligned genomes gathered from tens of thousands of species, a method that demands extraordinary amounts of computing. Evo 2, published earlier in 2026, covered more than 100,000 species and needed over 2,000 NVIDIA H100 processors running for several months. Nucleotide Transformer, a single-sequence genomic language model, consumed 128 NVIDIA A100 GPUs for a month. Both approaches ask a network to rediscover evolutionary constraint from raw sequence, one genome at a time.

On 9 September 2026, researchers at the University of California, Berkeley published a different answer in Nature. The paper, titled Predicting genome-wide functional constraints with GPN-Star, introduces a genomic pretrained network with species tree and alignment representations, known as GPN-Star. Instead of unaligned genomes, it consumes whole-genome alignments, which relate the genomes of hundreds of species to a single reference species and expose conserved versus changed code. The largest version of the model has 200 million parameters and was trained for several days on 8 NVIDIA A100 GPUs, a fraction of what the biggest rival models required.

The authors are Ye, Benegas, Albors, Li, Prillo, Fields, Clarke and Song. The senior author is Yun Song, a professor of computer science and statistics at Berkeley and an investigator at the Innovative Genomics Institute, who also directs the Berkeley Center for Computational Biology. The work was supported in part by the National Institutes of Health.

Key Facts

Nature reported on September 9, 2026 that GPN-Star is a biologically grounded genomic language model whose phylogeny-aware architecture uses whole-genome alignments and species trees to model evolutionary relationships explicitly. Training spanned alignments across vertebrate, mammal and primate evolutionary timescales, and the model set state-of-the-art results across a wide range of variant-effect prediction tasks in both coding and non-coding regions of the human genome, according to the paper published with DOI 10.1038/s41586-026-11005-5.

The comparison with larger systems is stark. According to the study, GPN-Star at 200 million parameters outperformed classical methods such as PhyloP and PhastCons, the ensemble CADD, the single-sequence genomic language models Nucleotide Transformer 2.5B and Evo 2 40B, and an earlier model called GPN-MSA on the ClinVar, COSMIC and OMIM/HGMD benchmarks. Berkeley News reported on September 9, 2026 that the model far outpaces its competitors at identifying the most important genetic variants for inherited traits, including those that lead to disease.

The Nature paper also reports that GPN-Star beats the supervised predictors AlphaMissense and PrimateAI-3D on fine-mapped GWAS variants across 65 coding and 83 non-coding UK Biobank traits. A separate analysis described in the study covers rare-variant association testing over 34 UK Biobank quantitative traits in 161,822 unrelated individuals, where the approach surfaced on average 402 genes compared with 383 for the base DeepRVAT method, and 353 compared with 338 replicating genes.

Timescale matters. Co-first author Chengzhong Ye, a graduate student in statistics at UC Berkeley, said models trained at different evolutionary timescales were optimized for different kinds of variants, with longer-timescale models better at rare protein-coding variants and the primate model better at complex traits such as schizophrenia risk. GPN-Star (P), the primate-timescale version, improved heritability enrichment across 106 independent traits, with the most enriched variants being missense variants at 33 percent and 170-fold enrichment, and distal enhancers at 26 percent.

The Innovative Genomics Institute said on September 9, 2026 that the team published genome-wide predictions alongside the study, highlighting variants likely to have the greatest influence on inherited traits so that other biologists can identify relevant genes and regulatory elements for further study. Beyond humans, the researchers trained GPN-Star for five model organisms: Mus musculus, Gallus gallus, Drosophila melanogaster, Caenorhabditis elegans and Arabidopsis thaliana.

Analysis

What this really means is that the assumption driving the current race in biological AI, that more species and more processors inevitably produce better biology, is being challenged from an unexpected direction. GPN-Star does not win by being larger. Nature reported on September 9, 2026 that the team focused on the 200 million parameter model, trained for several days on 8 NVIDIA A100 GPUs, and that result outperformed systems built on vastly bigger budgets, including Evo 2 at 40 billion parameters and Nucleotide Transformer 2.5B. The model wins because whole-genome alignments already encode the evolutionary comparison that unaligned models must infer from scratch.

That is a meaningful shift in how the field might allocate resources. The figures tell the story: 8 A100 GPUs against more than 2,000 H100 GPUs, days against several months, 200 million parameters against 40 billion. Computation is not free, and if a fraction of the budget buys better variant rankings, the constraint on progress stops being hardware and starts being the quality of the evolutionary signal fed into the network.

The bigger picture here is that architecture and biological prior knowledge are becoming the differentiators, not raw scale alone. It also matters that GPN-Star does not only beat other language models. It beats AlphaMissense and PrimateAI-3D, supervised predictors built with labeled disease data, on fine-mapped GWAS variants across 65 coding and 83 non-coding UK Biobank traits. A model that learns from evolutionary constraint, with no direct supervision from disease labels, is competing directly with tools designed for that job. The nuance is that timescales are not interchangeable, since the primate model does better on complex traits while longer timescales do better on rare coding variants.

Why It Matters

Geneticists are flooded with candidate variants and starved for ways to rank them. Only 1 to 2 percent of the 3 billion base pair human genome codes for proteins, and the non-coding remainder is where much of the difficulty lies. A tool that reliably prioritizes which variants and regulatory elements deserve experimental follow-up can redirect scarce laboratory time toward the changes most likely to affect human health. Song said the team believes its predictions will help prioritize the experiments that could have the greatest impact on human health.

The practical dividend is the release of genome-wide predictions with the study. Rather than leaving other groups to retrain a model, the Berkeley team has published annotations that biologists can use to find relevant genes and regulatory elements for further study. That lowers the barrier for labs without large GPU clusters and turns a paper result into reusable infrastructure.

Efficiency also changes who can do this research. Because whole-genome alignments do the work of identifying conserved regions, the model can be trained in days, or even hours, using only a handful of processors, the team says. The Innovative Genomics Institute said on September 9, 2026 that the study was supported in part by NIH grants and the UC AI Science at Scale program, with GPU resources from the Chan Zuckerberg Initiative.

Next Up

The next step is already visible in the paper. The team trained GPN-Star for five model organisms beyond humans, Mus musculus, Gallus gallus, Drosophila melanogaster, Caenorhabditis elegans and Arabidopsis thaliana, which signals that the phylogeny-aware recipe is meant to travel across species rather than stay human-specific. Genome-wide predictions from the model are also available for other biologists to reuse in their own work.

Broader adoption will depend on how well the rankings hold up as more groups test them against experiments, and on whether the timescale-specific models continue to separate cleanly on rare coding variants and complex traits. For now, the Berkeley result stands as a demonstration that a 200 million parameter model trained on 8 NVIDIA A100 GPUs can set state-of-the-art results across coding and non-coding variant-effect prediction.