Only approximately 2 per cent of the human genome encodes proteins, the rest being left to serve as an elaborate control system- one that harbors much of the variants associated with common disease. That has been an unbalance that has influenced genetics over decades. DNA can be read inexpensively, but so far deciphering most of it has been obstinately slow, even when a single-letter change is located anything distant relative to any gene, and has nonetheless altered biology in significant ways. AlphaGenome is an AI model developed at Google DeepMind that aims to bridge that gap, predicting the relationship between the DNA sequence and function, especially in the huge non-coding region otherwise known as the “dark genome.”
AlphaGenome is a sequence-to-function model: It takes stretches of DNA and tries to predict several biological implications instead of making text. It is capable of processing one million DNA letters at a time, which is a scale that is designed to capture long-range interactions, which can separate over a long distance. Practically, that will allow it to predict locations where genes are most likely to be regulated and approximate downstream consequences, including whether the activity of genes will increase or decrease and whether the splicing of RNA will be affected, processes capable of reorganizing the production of proteins by a cell.
More importantly, it is also capable of scoring the effect of a single change in the code. In cases where researchers are investigating conditions in which many of the implicated variations do not lie within genes, for example, obesity, diabetes, heart disease, autoimmune disorders and some mental health issues, this can be used to filter the number of candidates down to a set of lab experiments. Statistical geneticist Gareth Hawkes of the University of Exeter has said that the capacity to know what the non-coding majority is up to is “a big step forward,” especially as genome-wide biomarkers of risk identify variants of uncertain role.
The design of AlphaGenome is in line with an emerging trend in the engineering of biology: the development of general-purpose models, which consolidate tasks previously performed by distinct tools. The system is trained on extensive publicly available data on human and mouse experiments in its technical description, generating predictions on numerous molecular “tracks,” such as gene expression signals and chromatin features. The objective is not to substitute wet-lab work, but to come up with hypotheses as quickly as possible so that experiments can be selected based on their probability of providing the correct answer.
Cancer has been of some of the greatest immediate interest. The model has been applied in helping to distinguish between the mutations that actively promote tumour growth and those that just accumulate as collateral damage- this is important when researchers are attempting to identify the changes that may be a realistic treatment target. Marc Mansour, a clinical professor of paediatric haemato-oncology at UCL termed it a “step change” in attempts to identify genetic drivers of cancer.
AlphaGenome is also not launched as a clinical diagnostic system. Some forms of prediction of this type have been found to continue to be challenging to models in this class, such as very long-range regulation and changes in individual gene activity. The team at DeepMind has recognized itself that it needs to work on accuracy across tissues- since a neuron and a heart muscle cell cell have the same DNA, but run on different regulatory programs.
There has been an anomalously wide adoption of specialised genomics model. AlphaGenome has been accessed by almost 3000 scientists in 160 countries since mid 2025, with scaling by API. DeepMind has gone further to increase its accessibility by making the code of the model available in GitHub, allowing it to be reviewed and modified by the rest of the research community.
In the case of modern genomics, the headline is not about a particular model “solving” the DNA but rather a practical change in the workflow: transforming the genome’s most difficult-to-interpret regions into ranked, testable predictions that can be validated or invalidated and even built upon by the laboratory.
