A DNA variant is easier to interpret when the model can see what evolution did with the same genomic position across related species. GPN-Star combines whole-genome alignments with the species tree, allowing it to weigh conservation and change across vertebrate, mammalian and primate timescales rather than treating one human sequence as isolated text.[1][2]
The approach is biologically specific. Evolution has already run an immense natural experiment: variants that disrupt important functions tend to be removed or constrained, while other positions change more freely. A phylogeny-aware model tries to learn that signal directly while accounting for which species are closely or distantly related.
The GPN-Star programme
| Measure | Value | Context |
|---|---|---|
| Human models | Three | Separate vertebrate, mammal and primate whole-genome-alignment models.Source 1 |
| Largest model | 200 million parameters | The paper focuses on this scale after experiments with smaller variants.Source 1 |
| Training resource | Eight NVIDIA A100 GPUs for several days | Authors’ reported configuration; not a complete energy or cost comparison.Source 1 |
| Human prediction scope | Coding and non-coding variants | Includes pathogenic variants, fine-mapped GWAS variants and complex-trait heritability analyses.Source 1 |
| Notable regulatory result | Strong performance on distal enhancer variants | A historically difficult non-coding category.Source 1 |
| Model organisms | Five | Mouse, chicken, fruit fly, C. elegans and Arabidopsis.Source 1Source 2 |
| Required scientific input | Multispecies whole-genome alignment plus species tree | The model’s efficiency comes partly from explicit evolutionary data rather than sequence alone.Source 1 |
| Code availability | Public research code and model collection | Implementation resources are provided by the Song Lab.Source 3 |
The family tree prevents every species from counting equally
If ten closely related species share a letter, that evidence is not equivalent to ten independent evolutionary experiments. GPN-Star compresses aligned source sequences into clade-level representations and uses evolutionary distances in cross-attention. The tree helps the model interpret agreement and disagreement according to ancestry rather than raw count.[1]
Different timescales matter for different questions. Deep vertebrate conservation can reveal ancient essential functions. More recent primate or mammalian change can be more informative for regulatory elements that evolve faster. The paper reports task-dependent advantages rather than one universal best alignment.
Which evolutionary window may help which task
| Window | Signal emphasised | Potential strength | Potential blind spot |
|---|---|---|---|
| Vertebrate | Deep conservation across long evolutionary time | Ancient functional constraints and broadly conserved coding effects | Recent lineage-specific regulation may be dilutedSource 1 |
| Mammal | Intermediate evolutionary distance | Human-relevant constraint with more comparative changes than primates alone | Some very ancient or primate-specific signals may be weakerSource 1 |
| Primate | Recent evolutionary neighbourhood | Fine distinctions in human-adjacent regulatory sequence | Fewer independent substitutions and less depth for ancient functionsSource 1 |
| Single human sequence | Local sequence patterns without explicit aligned context | Works where alignments are absent or broad generative flexibility is needed | Must infer evolutionary constraint indirectly |
Variant ranking is not a diagnosis
A strong score suggests that a change occurs at a position evolution treated as functionally constrained or alters a pattern associated with deleterious variants. Clinical interpretation still depends on inheritance, population frequency, phenotype, experimental evidence and established classification standards. The model can prioritise variants for review; it does not determine a patient’s diagnosis by itself.
The distinction is especially important outside coding regions. A non-coding variant may affect an enhancer in one tissue, developmental stage or cellular state. Evolutionary constraint provides a valuable prior, but a mechanistic claim requires functional evidence.
Biological structure can substitute for brute-force scale
The authors contrast several days on eight A100 GPUs with single-sequence genomic models trained on far larger resources. The careful conclusion is not that compute no longer matters. GPN-Star receives information those systems must infer: aligned sequences and an explicit phylogeny. Better problem representation can reduce the amount of scale needed for a specialised prediction task.[1]
GPN-Star is a strong example of scientific AI becoming more domain-native. Natural language architectures inspired the transformer, but the useful advance comes from adding the object biology already uses—the species tree. The result is a smaller model with a clearer source of signal and a practical route for prioritising variants across humans and model organisms.
Sources & further reading
- Predicting genome-wide functional constraints with GPN-Star Nature · 9 Sep 2026 · Accessed 11 Oct 2026
- Phylogeny-informed genomic prediction with GPN-Star Nature Genetics · 8 Oct 2026 · Accessed 11 Oct 2026
- songlab/gpn: Genomic Pre-trained Network Song Lab on GitHub · Accessed 11 Oct 2026