Google released AlphaGenome Atlas, a resource that predicts the consequences of roughly nine billion possible single-base changes in the human reference genome. The human reference genome contains about three billion DNA bases. For each position, the system evaluated the three alternative bases that do not appear in the reference sequence. Noncoding DNA contains regulatory elements along with large amounts of sequence that may have little functional effect.
AlphaGenome is designed to predict functions and effects in noncoding DNA. Protein-coding sequence makes up less than three percent of the human genome. Existing experimental projects such as ENCODE supplied information used in model training. A prediction can help prioritize experiments but does not replace laboratory or clinical validation.
The model evaluates signals involving gene expression, transcription, chromatin and RNA processing. The atlas can provide an immediate prediction when a researcher finds a single-base variant. Precomputing all substitutions turns repeated model inference into a searchable resource.
Its current training and evaluation focus on human and mouse sequences and a limited set of well-studied cell types. Many possible substitutions will never appear in a person or have a meaningful biological effect. The harder test is performance on cell types, species or sequence contexts that were not represented in training. Independent researchers have not yet established how much the atlas adds beyond its training data across varied real-world uses.
The atlas returns model predictions rather than observed outcomes for nine billion substitutions. Researchers can use those scores to select a much smaller set of variants for laboratory testing or comparison with patient data. Validation will need to separate performance on familiar genomic regions from performance where training data are sparse, especially because a high-confidence prediction can still reflect correlations learned from existing datasets.
What to watch: External validation in underrepresented cell types and disease studies. Evidence that atlas predictions lead to confirmed biological discoveries.