Preprint authors tested AlphaGenome Atlas predictions on 111,000 DNA variants across 75 Alzheimer's disease risk loci: correlation with variant-to-disease association strength was weak, and predicted cell types for candidates within individual loci diverged
Preprint authors tested AlphaGenome Atlas predictions on 111,000 DNA variants across 75 Alzheimer's disease risk loci: correlation with variant-to-disease association strength was weak, and predicted cell types for candidates within individual loci diverged
On October 11, authors from the Korea Advanced Institute of Science and Technology (KAIST) posted a preprint on medRxiv, an unreviewed manuscript. They evaluated AlphaGenome Atlas, a database of precomputed predictions about the consequences of DNA substitutions, designed to guide which variants and cell types to prioritize for experimental follow-up.
A genome-wide association study (GWAS) compares the DNA of affected and unaffected individuals and identifies genomic regions where a variant is more frequent among cases. Neighboring variants are inherited together, so the disease association is visible across the whole group. Statistical fine-mapping estimates the probability that each variant is causal, but typically leaves several candidates, and the conclusion about which cell type is relevant depends on which of them is truly causal.
Google DeepMind released AlphaGenome Atlas and computed an AVI score for every DNA substitution, a composite of five feature groups: gene expression (production of RNA copies of a gene), regulatory features (DNA accessibility to gene-regulatory proteins), splicing (assembly of an RNA transcript from segments), protein impact, and cross-species conservation of the region.
The first question was whether the predicted effect increases with the strength of disease association. Within each locus, the authors compared variant rankings by predicted effect and by association strength (partial Spearman correlation), adjusting for variant frequency, linkage disequilibrium with neighbors, and distance to the nearest gene; 0 means no correlation, 1 means perfect rank agreement. For AVI the result was +0.013 (under random shifts of scores along the locus, 95% of results did not exceed +0.005); for expression, +0.018; for splicing, +0.028. Conservation and regulatory features dominate the AVI composite (47% and 34%), yet expression and splicing were the components that correlated with disease association. Variants in or near genes could receive high scores simply because of their position, so the authors controlled for it. Expression dropped to +0.011 and remained above the random baseline; splicing fell to +0.010 and became indistinguishable from it. In the authors' interpretation, the splicing signal largely reflects position within a gene.
The predictions were then compared with statistical fine-mapping performed without AlphaGenome data, which identified 95 credible sets, groups of variants where the causal variant is included with 95% probability. In 87 of these sets, the lead variant (the most probable causal variant) had tightly linked neighbors that fell outside the credible set. The lead's expression rank (position among all 111,000 variants, scaled from 0 to 1) was on average 0.044 higher than that of these neighbors, and higher in 67% of the 87 comparisons. For splicing, no difference was observed.
The final question was which cell type the causal variant acts in. Atlas predicts effects across 285 cell and tissue types; the authors collapsed these into 12 classes and labeled each variant with the class where its effect was most pronounced. When candidate labels agree, the choice of causal variant does not alter the cell type conclusion. Label diversity was quantified by entropy: higher entropy indicates more heterogeneous labels. Across 50 loci with multiple candidates, mean entropy was 1.63, compared with 1.53 for the same candidate sets shifted to different positions within the locus (a control). In 39 of the 50 loci, candidate labels were more heterogeneous than the control, and in none were they notably more uniform. Candidate labels within a single locus spanned brain, immune system, liver, and vasculature: the predicted cell type depends on which candidate is chosen.
The authors recommend applying the predictions to all fine-mapping candidates and reporting them alongside the uncertainty about which variant is causal and in which cell type.