[An abstract visualization of AI mapping human biology. Photo credit to Pixabay]

Google DeepMind has unveiled AlphaGenome, a cutting-edge artificial intelligence system designed to analyze long stretches of DNA and predict how genetic variations influence gene regulation.

This marks the company’s most ambitious expansion beyond protein structure research into the broader field of genomics.

Described in a paper published in Nature, AlphaGenome addresses one of biology’s most challenging problems: interpreting the function of non-coding DNA, which makes up roughly 98 percent of the human genome but plays a central role in controlling when and how genes are activated.

Gene regulation refers to the biological processes that dictate when, where, and how intensely genes are turned on or off within a cell.

Developed by the same Google DeepMind team behind AlphaFold, AlphaGenome is designed to predict the regulatory impact of subtle DNA sequence changes, including variants linked to disease, by modeling genome function directly from raw DNA.

According to DeepMind, AlphaGenome can process DNA sequences up to one million base pairs long and predict thousands of molecular properties related to gene regulation, such as where genes start and stop, how RNA is spliced, how accessible DNA is within cells, and how actively genes are transcribed across different tissues.

This long-range context addresses a key limitation of earlier genomic AI models, which typically had to choose between high resolution and large sequence coverage, limiting their ability to capture distant regulatory interactions.

AlphaGenome evaluates predictions from mutated and unmutated DNA sequences to estimate the effect of a variant, a process DeepMind reports can be completed in roughly one second per variant, enabling large-scale analysis.

The model focuses on interpreting non-coding variants, which are notoriously difficult to study because they do not directly alter protein sequences but instead influence complex regulatory mechanisms that can vary by cell type and tissue.

Many rare genetic disorders, including some forms of spinal muscular atrophy and cystic fibrosis, arise from errors in RNA splicing rather than protein-coding mutations, and AlphaGenome explicitly models splice junction locations and expression levels solely from sequence data.

Researchers not involved in the work say the model could significantly narrow the search space for disease-causing variants.

Marc Mansour, a professor at University College London focusing on hematological malignancies, noted, “determining the relevance of non-coding variants at scale has been extremely challenging”, and described AlphaGenome as a tool that could help connect genetic variation to diseases such as cancer.

AlphaGenome was trained on large public datasets from international consortia including ENCODE, GTEx, 4D Nucleome, and FANTOM5, which collectively measure gene regulation across hundreds of human and mouse cell types.

Despite its scale, DeepMind stresses that AlphaGenome is not intended for direct personal genome interpretation and has not been validated for predicting individual disease risk.

The authors note that while the model predicts molecular outcomes, translating those predictions into complex traits or clinical disease involves broader biological processes, including development and environmental factors, that remain outside the model’s current scope.

Limitations also persist in capturing very long-distance regulatory effects, such as interactions between DNA elements separated by hundreds of thousands of base pairs, and in accurately modeling rare or highly specific cell types.

Independent experts echo those concerns, noting that AlphaGenome’s reliance on bulk tissue data may limit its generalization to uncommon cell populations or early developmental stages.

Nevertheless, researchers describe the system as a state-of-the-art tool for prioritizing which genetic variants are most likely to matter, reducing the burden of experimental validation.

DeepMind states that AlphaGenome is already being used to study cancer-driving mutations, rare genetic diseases, and the design of synthetic DNA sequences with tailored regulatory functions.

DeepMind has made AlphaGenome freely available for non-commercial academic research, continuing its pattern of promoting broad scientific access, as seen with Alphafold.

The launch also aligns with a wider strategy at DeepMind to build an integrated platform of biological AI models spanning protein structure, mutation prediction, protein design, and now genome regulation.

Company executives emphasize that combining these models could accelerate drug discovery and diagnostics by enabling more complete computational representations of biological systems.

While AlphaGenome is unlikely to replace experimental biology, researchers say it could reshape how scientists explore the regulatory genome, offering a powerful new way to generate testable hypotheses from DNA sequence alone.