AlphaGenome: AI Mapping the Regulatory Genome

Most of the human genome does not encode proteins. Yet this so-called non-coding DNA contains regulatory instructions that influence when genes are activated, where they operate and how strongly they are expressed. Understanding those instructions is essential to explaining why apparently small genetic variants can contribute to disease.

AlphaGenome, developed by Google DeepMind, is an artificial intelligence model designed to map this regulatory genome. Rather than analysing isolated fragments, it can process DNA sequences of up to one million base pairs, giving it the wider genomic context in which regulatory signals interact.

The model generates predictions across multiple biological functions, including gene expression, RNA splicing, chromatin accessibility, transcription-factor binding and three-dimensional DNA contacts. Most outputs are produced at single-base-pair resolution. This unified approach allows researchers to compare how a mutation might alter several layers of gene regulation at once, rather than relying on separate specialist models.

In its Nature publication, the DeepMind team reported strong performance on benchmarks for regulatory variant effects, including non-coding mutations linked to disease. AlphaGenome can estimate how a single DNA change may shift molecular signals, helping researchers prioritise variants for laboratory investigation.

That promise needs careful framing. A prediction is not the same as a biological explanation, and performance on benchmarks does not establish clinical validity. Regulatory biology is highly dependent on cell type, developmental state and environmental context. Experimental validation remains necessary before such predictions can support diagnosis or treatment.

AlphaGenome’s significance lies less in replacing genomics experiments than in making the search space more navigable. Its research code and API give scientists tools for testing hypotheses across vast stretches of DNA. If those predictions continue to withstand independent validation, AI may turn the genome’s regulatory “dark matter” from an obstacle into a tractable research landscape.