Healthcare OperationsOct 7, 2026
Preprint finds AlphaGenome's zero-shot predictions fall short on ALS genetics, sometimes in the wrong direction
A bioRxiv preprint posted on 7 October 2026 tests Google DeepMind's AlphaGenome on ALS genetics with no disease-specific fine-tuning. Across 1,254 spinal-cord splicing QTLs it failed to recover measured effect sizes (Spearman 0.041, Pearson 0.104) and was precise only at 1.3% recall. At the well-studied UNC13A cryptic-exon locus it predicted the risk variant's splice-site effect in the opposite direction to the known one, and showed near-identical effects in disease-relevant and irrelevant tissues.
What it means Genomic foundation models are being pitched for variant interpretation; this is independent evidence that zero-shot outputs can be wrong in direction, not just in size, so validate before relying on them.
Where it came from bioRxiv