Google DeepMind released AlphaGenome on January 28th. It is an AI model that predicts how DNA sequences translate into biological functions, processing up to 1 million base pairs at a time and outperforming existing models on 25 of 26 variant effect prediction benchmarks.
The model, published in Nature and detailed on the DeepMind blog, represents a significant advance in computational genomics. Whereas previous models required separate systems for different prediction tasks, AlphaGenome handles everything from gene expression to chromatin access in a single integrated architecture.
“AlphaGenome can interrogate long stretches of DNA to predict where important regulatory elements are located and their downstream effects on gene expression,” the DeepMind team said in the announcement. The model’s one-million-token context window allows us to capture long-range interactions between distant DNA regions that influence turning genes on and off.
structure
AlphaGenome combines two neural network architectures: a Borzoi-style 1D convolutional network for processing raw DNA sequences, and a U-Net architecture adapted from image segmentation. This hybrid approach allows the model to handle both the continuous nature of DNA and the complex spatial relationships between regulatory elements.
The training data spans approximately 7,000 genomic tracks from the ENCODE and FANTOM consortiums, large-scale collaborations that catalog functional elements across the human genome. The model learns to predict signals from experimental assays that measure gene expression, DNA accessibility, protein binding, and chromatin modification.
For researchers, the practical value lies in predicting mutational effects. When a patient’s genome contains a mutation, clinicians need to know whether the mutation is significant. AlphaGenome can predict how single nucleotide changes affect overall regulation, potentially discovering disease-causing mutations that are missed by current methods.
The model achieved excellent results in benchmarks testing its ability to predict how genetic variation affects gene expression and activity of regulatory elements. For expression quantitative trait loci (eQTLs) (variants known to affect gene expression levels), AlphaGenome matched or outperformed specialized models trained specifically for these tasks.
Open source availability
DeepMind released AlphaGenome’s source code to GitHub for non-commercial use, continuing the lab’s pattern of making basic biological tools available to the public. The repository contains model weights, inference code, and documentation for performing predictions on custom sequences.
This open release follows the model established by AlphaFold, DeepMind’s protein structure prediction tool, which has been used by over 3 million researchers since its release in 2021. AlphaGenome addresses complementary problems. AlphaFold predicts what proteins look like, while AlphaGenome predicts when and where genes will produce those proteins.
Google DeepMind CEO Demis Hassabis positions biology as a key application domain for the lab’s AI capabilities. The genomics effort extends DeepMind’s ambitions beyond the conversational AI and language models that power products like Gemini, applying similar architectural innovations to scientific problems.
why is this important
The human genome contains approximately 3 billion base pairs, but only about 1.5% directly encode proteins. The remaining 98.5%, long ignored as “junk DNA”, contains regulatory elements that control when, where, and how much genes are expressed. Mutations in these non-coding regions cause disease, but it has been extremely difficult to identify which mutations are important.
Traditional methods require expensive and time-consuming experiments to test individual variants. Machine learning models like AlphaGenome can computationally screen thousands of variants and prioritize those worthy of experimental pursuit. In the diagnosis of rare diseases, where patients often carry novel variants of unknown impact, this capability could accelerate the process from sequencing to diagnosis.
The ability of the model to handle 1 million base pair contexts is particularly important. Genetic regulatory elements are located hundreds of thousands of base pairs away from the genes they control and communicate through complex 3D folding of DNA. Previous models with short context windows were unable to capture these long-range dependencies.
AlphaGenome joins a growing ecosystem of AI tools that are transforming biological research. Protein structure prediction, drug discovery, and now gene regulation are increasingly tractable problems for machine learning. For the genetics research community, the open availability of these models democratizes access to computational power that was previously limited to well-funded laboratories.
The limitations of this model are also evident from DeepMind’s presentation. Although AlphaGenome is excellent at predicting experimental measurements, additional validation is required to translate those predictions into clinical outcomes. There remains a large gap between predicting chromatin accessibility and predicting disease risk.
For now, AlphaGenome serves as a research tool that could advance our understanding of how genomes work, although clinical applications are still years away. 3,000 scientists in 160 countries are already using the model, suggesting that the research community sees immediate value in what DeepMind has built.
