Abstract
Genomic foundation models promise to infer regulatory consequences of genetic variants directly from DNA sequence, but how well these predictions translate to clinical utility remains unclear. To probe this question, we evaluate AlphaGenome (AG) for understanding amyotrophic lateral sclerosis (ALS) genetics without any disease-specific fine-tuning across three settings of increasing difficulty: recovering measured variant effects, reconstructing a known splicing mechanism, and generating de novo regulatory hypotheses at unresolved loci. At population scale, AG failed to recover measured effect sizes across 1,254 spinal-cord splicing quantitative trait loci (QTL) (Spearman = 0.041, Pearson = 0.104) and was precise only at negligible recall (precision = 1.00 at 1.3% recall). At a mechanistically resolved locus of UNC13A cryptic-exon event, AG predicted the intronic risk variant's splice-site usage opposite its known risk effect (Delta approx. -2 x 10^-6), showed near-identical effects in disease-relevant and disease-irrelevant tissues (cerebellar effects reached approximately 60% of neural magnitude), and recovered only a partial cis-regulatory signature. Finally, AG showed limited gene-level resolution across seven ALS genome-wide association study (GWAS) loci, with predicted effects declining sharply as a gene lay farther from the variant (a 10-fold increase in distance cost 0.95 standard deviations of signal; p = 2.7 x 10^-23). In summary, AG can identify a small set of high-confidence regulatory variants, but cannot reliably rank variants by biological effect, or assign regulatory signals to the correct effector gene at unresolved loci. Thus, although AG captures meaningful local regulatory effects, these predictions do not resolve the effect sizes and effector genes needed to connect ALS-associated variants to their causal mechanisms.