Genomic foundation models fall short in common variant discovery
Genomic foundation models are being promoted as a new route to interpret human DNA, but this preprint suggests their current strength lies more in rare variant modeling than in routine common variant discovery. Using the surrogate functional false discovery rate framework for functionally informed GWAS discovery, investigators compared standard functional annotations with features derived from three leading models: the sequence-to-function systems AlphaGenome and Borzoi, and the genomic language model GPN-Star. The question was simple and clinically relevant for statistical genetics: can these models help uncover more common variant associations than the annotations already used in standard practice?
The answer, at least in this analysis, was mostly no. Across 27 UK Biobank traits, GPN-Star reduced locus discovery relative to a standard GWAS, while AlphaGenome and Borzoi increased discovery, but not enough to beat the benchmark set by baseline functional annotations. Median changes in discovery were 1.8% for GPN-Star, 3.8% for AlphaGenome, and 5.1% for Borzoi. The gains were real, but they were modest, and they did not translate into a clear advantage over the established annotation set.
BaselineLD still carries most of the signal
The most consequential result came when the model-derived annotations were combined with baselineLD, a standard functional annotation framework already widely used in genetic analyses. In that setting, there was no increase over baselineLD alone. Median changes ranged from −0.1% to 0.2%, a narrow band that leaves little room for claims of additive value. In practical terms, the foundation models did not appear to contribute new information beyond what baselineLD had already encoded.
That pattern suggests the models may be learning much of the same biology captured by conventional functional annotations. Sequence context, conservation signals, and other regulatory features already represented in baselineLD may be sufficient for many common variant analyses, at least with the current model architectures and training corpora. The failure to gain additional discovery after combination is important because it argues against a simple “more AI equals more loci” assumption. Instead, the incremental value of these models seems constrained by overlap with existing annotation systems.
One sentence captures the central result plainly: the new models did not displace the old ones.
Why rare-variant success does not automatically translate
The divergence between rare and common variant utility is biologically plausible. Rare variant interpretation often depends on identifying strong local functional effects, where sequence-based models can help estimate whether a change is likely to disrupt regulation or protein function. Common variant discovery is a different problem. Signals are typically subtler, more polygenic, and more entangled with linkage disequilibrium, tissue-specific regulation, and weak effects distributed across large genomic regions.
In that setting, predictive models must do more than estimate local sequence function. They need to separate causal regulatory information from the dense background of correlated annotation and linkage structure that already influences GWAS prioritization. If a foundation model largely recapitulates what baselineLD already captures, then the added complexity may not buy much extra discovery power. The study’s results fit that logic. AlphaGenome and Borzoi improved discovery relative to a standard GWAS, but the improvement was still smaller than what baselineLD delivered on its own. GPN-Star performed worse than standard GWAS for locus discovery, reinforcing the idea that not all large-scale genomic models are equally suited to the common-variant problem.
Trait-level gains were real but limited
The analysis covered 27 UK Biobank traits, which gives the findings breadth across multiple human phenotypes. Even so, the reported gains were not uniformly dramatic enough to suggest a broad shift in discovery strategy. The median results point to a consistent hierarchy: baseline functional annotations performed best, AlphaGenome and Borzoi followed behind, and GPN-Star lagged further still. That ordering matters because it implies current model choice is not a trivial implementation detail; the underlying architecture and training objective appear to influence whether a model can support variant prioritization at all.
For researchers trying to increase locus discovery, this matters at the workflow level. A model that performs well in one genomic task may still be poorly matched to another, especially when the endpoint is not prediction of a molecular phenotype but enrichment of statistically robust GWAS signals. Discovery pipelines can amplify small differences in annotation quality, yet this preprint suggests the ceiling remains set by the overlap between learned representations and established functional annotations.
What the results imply for genomic AI development
The most defensible interpretation is not that genomic foundation models have failed, but that they have not yet surpassed the information already embedded in standard annotation resources. Larger corpora may help. So may trait-specific feature engineering. Both are explicitly raised as possible paths forward by the investigators, and both are sensible next steps because common variant discovery likely demands narrower tuning than broad-purpose sequence modeling currently provides.
That future work will need to confront a difficult question: how much genuinely new biology can these models learn, rather than simply repackage existing regulatory annotation? If the training data are not diverse enough, or if the model objective is too generic, the output may remain highly correlated with baseline annotations. In that case, gains in biological insight will be incremental rather than disruptive.
This preprint is also a reminder that model evaluation must match the clinical-genomic task. A model that performs well in rare variant settings, or in molecular prediction benchmarks, cannot be assumed to improve common variant discovery. The endpoint here was not abstract predictive accuracy; it was actual locus yield across human traits. On that measure, the current generation of genomic foundation models has not yet outpaced the standard annotation framework they were meant to extend.
A narrow path to clearer value
The study leaves a practical agenda rather than a mystery. Future iterations will need larger and more representative training corpora, stronger trait-specific calibration, and a clearer strategy for extracting information not already present in baselineLD. Until then, the most robust conclusion is straightforward: current genomic foundation models can support common variant discovery, but they do not yet improve it beyond standard functional annotations, and combining them with baselineLD does not add measurable gain.