Google's AI genome system evaluates every possible one-base change

Much of the human genome lies outside protein-coding regions, and a large fraction of that non-coding DNA consists of evolutionary leftovers such as remnants of ancient viral insertions, parasitic sequences, and genes disabled by mutation. Distinguishing the small subset of functional regulatory elements from this background is difficult because the proteins that bind DNA often tolerate mismatches and can act differently in various cell types. In some loci, the overall density of binding motifs matters more than any single site, further complicating interpretation. To tackle these probabilistic, context-dependent problems, Google developed AlphaGenome, an AI system trained to evaluate DNA sequences for likely functional roles. The model attempts to identify a broad set of regulatory and transcriptional features, including gene expression and transcription initiation signals, chromatin accessibility and histone marks, transcription factor binding patterns, chromatin contact maps, and indicators of splicing such as splice site usage and splice junction strength. AlphaGenome is currently restricted to human and mouse sequences and was trained on a limited collection of cell types that have been well characterized by researchers. That narrow training set limits the range of cell-type–specific predictions the system can make, but within those domains the tool can still be informative. For example, when investigators encounter variants in non-coding regions near a gene of interest, AlphaGenome can offer a hypothesis about whether the change is likely to affect regulation and suggest a possible mechanism. According to the developers, AlphaGenome’s output is generally as accurate as or better than that from specialized existing tools for the tasks it was assessed on. While it does not replace experimental validation, the system provides a scalable way to prioritize candidate variants and generate mechanistic ideas that researchers can test in the lab.