Summary
A bioRxiv preprint presents PHACTn, a training-free method that uses mammalian evolutionary history to infer how tolerable single-nucleotide variants may be across the human genome. The authors report strong performance on benchmark sets of non-coding variants using four interpretable parameters and no GPU.
Researchers have presented PHACTn, a computational method for estimating the tolerance of single-nucleotide variants across the human genome. The method uses evolutionary information from the mammalian phylogenetic tree rather than a trained machine-learning model. Its results are described in a bioRxiv preprint posted on September 14, 2026.
A single-nucleotide variant, or SNV, is a change involving one DNA base. Estimating whether a particular change is likely to be tolerated can help prioritise variants for further biological or clinical investigation. This is especially difficult in non-coding DNA, which contains regulatory sequences but does not directly encode proteins and has a large, incompletely characterised functional landscape.
How PHACTn makes its predictions
PHACTn stands for Phylogeny-Aware Computing of Tolerance for nucleotide variants. The authors describe it as training-free and parameter-minimal: it uses four interpretable parameters, requires no model training and does not need a graphics processing unit.
The method examines substitutions observed across mammalian species and incorporates two features of that evolutionary history. First, it models the evolutionary independence of the substitutions, accounting for the fact that observations in related species may not represent fully separate evidence. Second, it considers the evolutionary distance between each species and the species being analysed. Together, these signals are used to infer how compatible a nucleotide change may be with the constraints acting on that genomic position.
This approach is intended to provide a more transparent alternative to large sequence models. In the preprint, the authors contrast PHACTn with machine-learning classifiers, which they say can be affected by data circularity and demographic bias, and with genomic language models that require substantial computational resources and can be difficult to interpret. These comparisons are part of the authors' motivation for developing the method.
Evaluation on non-coding variants
The authors evaluated PHACTn on non-coding variants curated from ClinVar, a database of variants with clinical interpretations, and on non-coding variants potentially associated with selected Mendelian diseases curated from OMIM. In the reported comparisons, PHACTn outperformed all evaluated tools on both sets.
The preprint also reports state-of-the-art performance for variants within the informative range of alignment-based inference. In this setting, the available cross-species sequence alignment contains enough evolutionary information to support a useful comparison. The result suggests that a relatively small, explicitly defined phylogenetic model can capture constraint signals that the authors say were not recovered as effectively by the larger sequence-based approaches tested.
PHACTn could make evolutionary constraint analysis more accessible because it avoids the training datasets and specialised hardware often associated with modern genomic machine learning. Its interpretable parameters may also make the basis of a prediction easier to examine. For clinical interpretation, however, a computational tolerance estimate would be one evidence stream among several, alongside patient data, population observations and functional studies.
The evidence currently comes from a bioRxiv preprint and a computational benchmark. The supplied abstract reports performance rankings but not numerical scores, dataset sizes or results for individual comparison tools. The authors' strongest alignment-based claim is also limited to variants within the informative range of the available evolutionary comparisons.