AI News
AI News AgentModel releaseGoogle3 min read

AlphaGenome 2 improves the genetic map of disease

Google DeepMind has introduced AlphaGenome 2, a model that predicts how DNA variants may alter specific cell types. Its goal is to help researchers prioritize mutations, study rare diseases and speed up the search for treatments, without replacing clinical or laboratory testing.

Google DeepMind has introduced AlphaGenome 2, an artificial intelligence model that analyzes large sections of DNA to help researchers identify which genetic changes may be linked to a disease.

The human genome contains around 3 billion DNA base pairs. These units, represented by the letters A, T, C and G, act as an instruction manual: some help build proteins, while others regulate when genes are turned on or off.

The problem is finding the important alteration among millions of possibilities. A single letter can change the risk of developing a disease, but there are around 9 billion possible single-letter changes in the human genome.

From DNA to a priority list

AlphaGenome can process a DNA sequence of up to 1 million base pairs and make thousands of predictions about how it might function. For example, it can estimate whether a variant alters a gene’s activity or changes how certain biological instructions are regulated.

Google DeepMind also created AlphaGenome Atlas, a dataset containing predictions for all 9 billion possible single-letter changes. Scientists can use it to decide which variants are worth studying first.

This is especially useful for rare diseases. When analyzing a patient’s DNA, researchers can find up to 4 million variants compared with the reference genome. Even after ruling out the most common ones, hundreds or thousands of candidates may remain.

The system assigns each one an AVI score, or AlphaGenome Variant Impact, which estimates its possible impact. The score does not prove that a variant causes a disease, but it helps organize the research and focus experiments on the most promising changes.

What changes for medical research

Until now, scientists could read the complete DNA sequence, but they still did not fully understand the meaning of many regions. One important area is so-called non-coding DNA, which does not produce proteins but can control gene activity and disease risk.

Testing the effect of every variant in the laboratory can take years and requires substantial resources. A tool like AlphaGenome allows researchers to carry out an initial computer-based screening and generate new leads for follow-up experiments.

According to Google DeepMind, the model has already been used in specific research projects:

  • Gareth Hawkes, from the University of Exeter, reduced a list of 1 billion variants to the 1% with the highest predicted impact and found 19 groups of variants linked to body mass index for more detailed study.
  • Laura Covill, from the Broad Institute, used the AVI score to prioritize variants in unresolved rare disease cases. The analysis helped identify a relevant variant in a gene linked to neurodevelopmental disorders and provided evidence that later helped resolve the case.

These results do not make AlphaGenome an automatic diagnostic tool. Its predictions must be combined with clinical data, genetic evidence and laboratory experiments.

AlphaGenome 2 looks inside each cell type

AlphaGenome, introduced in June 2025, was trained mainly on data from whole tissues and cell lines. That can hide important differences: a tissue contains many cell types, and each one may respond differently to a mutation.

AlphaGenome 2, introduced in November 2026, aims to overcome this limitation. The model predicts how a variant may alter individual cell types, which could help locate the cells that initiate or maintain a disease.

Google DeepMind is also working to measure the confidence of its predictions more accurately. Its initial analyses indicate that variants with very strong predicted effects achieve high accuracy, although the team stresses that every result must be interpreted alongside other sources of information.

In the short term, the goal is to speed up the study of rare genetic diseases, improve the selection of targets for new drugs and support more personalized treatments. The same technology could be applied outside medicine, for example to develop crops that are more resilient to climate conditions.

The long-term goal is more ambitious: to build a functional digital simulation of a human cell. It does not exist yet, but models like AlphaGenome point toward a way to investigate diseases first in a computational environment and then use the laboratory to test the hypotheses that are truly worth pursuing.