What do within-site comparisons of AlphaGenome Atlas predictions measure? An audit at ClinVar sites in 11 genes

Background: AlphaGenome Atlas precomputes AlphaGenome predictions for every possible single-nucleotide change in the human genome. What comparisons of these predicted profiles measure, allele-specific information or locus, consequence and ascertainment context, is not established. Results: At ClinVar pathogenic or likely pathogenic (P/LP) sites we compared the ClinVar substitution with the two other substitutions, for 22 Atlas scorers in three distance geometries, with permutation nulls and Benjamini–Hochberg correction. In a four-gene cohort (DNM1, PCCA, PCCB, MMUT; 120 coding sites), the ClinVar substitution lay farther from a background profile than expected in 17 of 66 comparisons; post hoc, most of the enrichment sat at the 34 nonsense and start-loss sites, and a model-free consequence rule selected the ClinVar allele nearly as often as the AlphaGenome Variant Impact (AVI) score, which uses loss-of-function annotations (49% against 52%). Two replications were pre-specified, the second in a publicly deposited plan: 384 held-out positions of the four genes' full P/LP pool, and 1,713 positions in seven further genes. In the seven genes AVI ranked the ClinVar allele first at 88% of nonsense and start-loss positions against 38% of missense positions, four of the six tested scorers without consequence inputs showed the same contrast, and, post hoc, the consequence rule matched AVI at coding positions (82% against 80%). Under a consequence-matched null 10 of 66 comparisons were significant (37 uniform), above the pre-specified bound of a quarter of 37. At 246 missense positions AVI singled out the ClinVar allele under a mutation-rate-aware null (38% against 33%; q = 0.001), variably between genes, as did, post hoc, its input AlphaMissense. At 63 other intronic positions, where all three substitutions share a label, 13 of 21 scorers singled out the ClinVar allele (splicing scorers 73–75% first; q < 0.001). At 157 canonical splice positions none did. Conclusions: Much of the within-site preference for ClinVar alleles reflects predicted consequence class, but a small missense signal, matched post hoc by AlphaMissense, and a large signal at other intronic positions survive consequence matching, both where ClinVar classifications can rest on computational evidence. Evaluations should hold consequence fixed, use a mutation-rate-aware null and report a consequence-only baseline. Preprint; not peer reviewed. Files: the manuscript and Additional files 1–5. Code, cached predictions and all results: https://doi.org/10.5281/zenodo.23537469. Pre-specified plan for the seven-gene replication: https://doi.org/10.5281/zenodo.23524782. Terms. The manuscript is licensed CC BY 4.0. Its numbers, figures and tables are derived from AlphaGenome Atlas outputs, which are subject to the AlphaGenome Output Terms of Use. By using this information, you agree to AlphaGenome Output Terms of Use found at http://deepmind.google.com/science/alphagenome/output-terms. The modifications made to the outputs are listed in LEGALLY_BINDING_TERMS_OF_USE.txt in the deposit.