PHACTn enables training-free, context-independent inference of nucleotide variant tolerance across the genome

PHACTn enables training-free, context-independent inference of nucleotide variant tolerability across the human genome, outperforming evaluated tools on non-coding variants with only 4 interpretable parameters.

• Why it matters: Predicting SNV tolerability genome-wide, especially in non-coding regions, is hindered by data circularity, demographic bias, and the compute demands and interpretability limits of genomic language models.
• What they did: PHACTn traverses the mammalian phylogenetic tree to model evolutionary independence of observed substitutions and their distance from the query species, using 4 parameters with no training and no GPU requirement.
• The result: It outperforms evaluated tools on non-coding variants curated from ClinVar and OMIM-linked Mendelian disease variants, and reaches state-of-the-art performance within the informative range of alignment-based inference.