Google DeepMind published AlphaGenome Atlas on 8 September, a precomputed set of predictions for every possible one-letter change to human DNA, about nine billion variants in all. The dataset runs to roughly a petabyte, some thirty times the AlphaFold protein database. For each variant it estimates the effect on gene regulation across hundreds of human and mouse cell types, and folds in the earlier AlphaMissense scores for protein-altering changes.
The point is to remove a step. Until now, getting these predictions meant running the AlphaGenome model directly, which takes code and compute. The Atlas is a lookup table instead, free through a web portal for academic use, with a paid Google Cloud tier for commercial users still to come. The model reads a one-million-base-pair stretch around each variant and returns one impact score meant to flag whether the change matters.
That score is where the caution sits. Carl de Boer of the University of British Columbia called AlphaGenome the field’s leading model and said the Atlas could speed up basic biology and drug work. But a single number from a system with this many moving parts, he added, is likely to be easily misinterpreted. It also cannot see how variants combine, which is how much common disease works, and misses regulators acting from beyond its window.
DeepMind says the Atlas is not validated for clinical use and should not guide diagnosis or treatment. What it does not settle is whether an easier reference changes outcomes or only convenience. AlphaMissense covered 71 million protein-altering variants in 2023 and is widely cited. Three years on, which non-coding changes actually cause disease is still mostly open, and a petabyte of scores does not close that alone.
