technology
AlphaGenome
Also known as AlphaGenome API, AlphaGenome Atlas, AlphaGenome 2
AlphaGenome is a Google DeepMind AI model, announced on 25 June 2025, that predicts how single variants in human DNA affect the processes that regulate genes.[1] It reads up to 1 million DNA base pairs at a time and predicts thousands of molecular properties. The Nature paper of January 2026 reports that it matches or beats the strongest external models in 25 of 26 variant-effect evaluations.[2][3] It is for non-commercial research only and is not validated for clinical use.[4]
Key facts
AlphaGenome is an AI model from Google DeepMind for reading the parts of DNA that control genes. Google DeepMind announced it on 25 June 2025.[1] It builds on DeepMind’s earlier Enformer model and complements AlphaMissense, which classifies variants in protein-coding regions.[5]
What it predicts
AlphaGenome takes up to 1 million DNA base pairs as input and predicts thousands of molecular properties that describe their regulatory activity.[2] It then scores a variant by comparing predictions for the original and the mutated sequence.[2] Google DeepMind says it is the first model to predict the location and expression level of RNA splice junctions directly from DNA sequence.[6]
The focus is the non-coding genome. Google DeepMind says protein-coding regions make up only 2% of the genome. The remaining 98% controls gene activity and contains many disease-linked variants.[7]
How it was built
The model uses convolutional layers to find short patterns in the sequence and transformers to pass information across the whole input.[8] It was trained on experimental data from public consortia including ENCODE, GTEx, 4D Nucleome and FANTOM5, covering hundreds of human and mouse cell types and tissues.[9] Google DeepMind says training one model took four hours and half the compute used for its original Enformer model.[10]
Accuracy
In the June 2025 preprint results, AlphaGenome beat the best external models on 22 of 24 single-sequence evaluations. It matched or beat them on 24 of 26 variant-effect evaluations.[11] The peer-reviewed paper, published in Nature on 28 January 2026, reports that it matches or exceeds the strongest external models in 25 of 26 variant-effect evaluations.[3]
Access and limits
AlphaGenome is available for non-commercial use through an API. Google DeepMind says its predictions are for research only and were not designed or validated for clinical use.[4] Pre-trained weights are published on Kaggle and Hugging Face, and downloading them requires accepting non-commercial model terms.[12]
Google DeepMind says the model was not designed or validated for predicting individual people’s genomes. It also says the model does not explain on its own how variants lead to complex traits or diseases.[13] Capturing the effect of regulatory elements more than 100,000 DNA letters away remains a challenge.[14]
2026 additions
Google DeepMind says it has used AlphaGenome to predict the molecular impact of all nine billion possible single-letter changes in the human genome. It publishes the results as the AlphaGenome Atlas.[15] It also describes AlphaGenome 2, trained on more than 100 million single-cell profiles across 68 tissues. Google DeepMind says the new model identifies nearly twice as many cell-type-specific expression-linked variants as the first.[16]
For a different approach to genome modelling, see evo-2. For DeepMind’s protein structure model, see alphafold.
Read next
AlphaFold
AlphaFold is Google DeepMind's AI system for predicting protein and biomolecular structures. Versions, accuracy, database and limits.
Questions readers ask
What does AlphaGenome do?
It predicts how single variants or mutations in human DNA affect the biological processes that regulate genes, such as where genes start and end, how they are spliced and how much RNA is made.[1][2]
Why does it focus on non-coding DNA?
Google DeepMind says protein-coding regions are only 2% of the genome. The other 98% controls gene activity and contains many disease-linked variants, and AlphaMissense already covers coding variants.[7][5]
How accurate is AlphaGenome?
The January 2026 Nature paper reports that it matches or exceeds the strongest available external models in 25 of 26 variant-effect evaluations.[3]
Can doctors use AlphaGenome to read a patient's genome?
No. Google DeepMind says its predictions are for research only, were not validated for clinical use, and the model was not designed or validated for predicting individual people's genomes.[4][13]
Can anyone download AlphaGenome?
Researchers can use a non-commercial API, and the pre-trained weights are on Kaggle and Hugging Face for those who accept non-commercial model terms.[4][12]
What is the AlphaGenome Atlas?
Google DeepMind says it used AlphaGenome to predict the molecular impact of every possible single-letter change in the human genome, all nine billion of them, and published the results as the Atlas.[15]
Sources
· 16 claims from 3 publishers
Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.
- [1]
Google DeepMind announced AlphaGenome on 25 June 2025 as an AI tool that predicts how single variants in human DNA affect the biological processes that regulate genes. confirmedas of 2025-06-25
- AlphaGenome: AI for better understanding the genome · Google DeepMind · Primary source · 2025-06-25 · Blog post dated June 25, 2025 (retrieved 2026-11-14)
- [2]
AlphaGenome takes up to 1 million DNA base pairs as input, predicts thousands of molecular properties describing their regulatory activity, such as where genes start and end, splicing and RNA levels, and scores variants by comparing predictions for mutated and unmutated sequences. confirmedas of 2025-06-25
- AlphaGenome: AI for better understanding the genome · Google DeepMind · Primary source · 2025-06-25 (retrieved 2026-11-14)
- AlphaGenome: AI for better understanding the genome · Google DeepMind · Primary source · 2025-06-25 (retrieved 2026-11-14)
- AlphaGenome: AI for better understanding the genome · Google DeepMind · Primary source · 2025-06-25 (retrieved 2026-11-14)
- [3]
AlphaGenome was published in Nature on 28 January 2026; the paper reports that it matches or exceeds the strongest external models in 25 of 26 variant-effect evaluations. confirmedas of 2026-01-28
- Advancing regulatory variant effect prediction with AlphaGenome · Nature · Primary source · 2026-01-28 · Published online 28 January 2026 (retrieved 2026-11-14)
- AlphaGenome: AI for better understanding the genome · Google DeepMind · Primary source · 2025-06-25 (retrieved 2026-11-14)
- [4]
AlphaGenome is available for non-commercial use through an API, and Google DeepMind says its predictions are for research only and not validated for clinical use. confirmedas of 2025-06-25
- AlphaGenome: AI for better understanding the genome · Google DeepMind · Primary source · 2025-06-25 (retrieved 2026-11-14)
- AlphaGenome: AI for better understanding the genome · Google DeepMind · Primary source · 2025-06-25 (retrieved 2026-11-14)
- [5]
AlphaGenome builds on DeepMind's earlier Enformer model and complements AlphaMissense, which classifies the effects of variants in protein-coding regions. confirmedas of 2025-06-25
- AlphaGenome: AI for better understanding the genome · Google DeepMind · Primary source · 2025-06-25 (retrieved 2026-11-14)
- [6]
Google DeepMind says AlphaGenome is the first model to predict the location and expression level of RNA splice junctions directly from DNA sequence. confirmedas of 2025-06-25
- AlphaGenome: AI for better understanding the genome · Google DeepMind · Primary source · 2025-06-25 (retrieved 2026-11-14)
- [7]
Google DeepMind says protein-coding regions make up 2% of the genome and the remaining 98%, the non-coding regions, orchestrate gene activity and contain many disease-linked variants. confirmedas of 2025-06-25
- AlphaGenome: AI for better understanding the genome · Google DeepMind · Primary source · 2025-06-25 (retrieved 2026-11-14)
- [8]
AlphaGenome uses convolutional layers to detect short sequence patterns and transformers to pass information across the whole input sequence. confirmedas of 2025-06-25
- AlphaGenome: AI for better understanding the genome · Google DeepMind · Primary source · 2025-06-25 (retrieved 2026-11-14)
- [9]
AlphaGenome was trained on data from public consortia including ENCODE, GTEx, 4D Nucleome and FANTOM5, covering hundreds of human and mouse cell types and tissues. confirmedas of 2025-06-25
- AlphaGenome: AI for better understanding the genome · Google DeepMind · Primary source · 2025-06-25 (retrieved 2026-11-14)
- [10]
Google DeepMind says training one AlphaGenome model took four hours and half the compute budget used for its original Enformer model. confirmedas of 2025-06-25
- AlphaGenome: AI for better understanding the genome · Google DeepMind · Primary source · 2025-06-25 (retrieved 2026-11-14)
- [11]
In the June 2025 preprint results, AlphaGenome beat the best external models on 22 of 24 single-sequence evaluations and matched or beat them on 24 of 26 variant-effect evaluations. confirmedas of 2025-06-25
- AlphaGenome: AI for better understanding the genome · Google DeepMind · Primary source · 2025-06-25 (retrieved 2026-11-14)
- AlphaGenome: AI for better understanding the genome · Google DeepMind · Primary source · 2025-06-25 (retrieved 2026-11-14)
- [12]
Google DeepMind publishes AlphaGenome's pre-trained weights on Kaggle and Hugging Face, and downloading them requires accepting non-commercial model terms. confirmedas of 2026-11-14
- google-deepmind/alphagenome_research · Google DeepMind (GitHub) · Primary source (retrieved 2026-11-14)
- google-deepmind/alphagenome_research · Google DeepMind (GitHub) · Primary source (retrieved 2026-11-14)
- [13]
Google DeepMind says AlphaGenome was not designed or validated for predicting individual people's genomes and does not explain on its own how variants lead to complex traits or diseases. confirmedas of 2025-06-25
- AlphaGenome: AI for better understanding the genome · Google DeepMind · Primary source · 2025-06-25 (retrieved 2026-11-14)
- AlphaGenome: AI for better understanding the genome · Google DeepMind · Primary source · 2025-06-25 (retrieved 2026-11-14)
- [14]
Google DeepMind says capturing the effect of regulatory elements more than 100,000 DNA letters away remains a challenge for AlphaGenome. confirmedas of 2025-06-25
- AlphaGenome: AI for better understanding the genome · Google DeepMind · Primary source · 2025-06-25 (retrieved 2026-11-14)
- [15]
Google DeepMind says it has used AlphaGenome to predict the molecular impact of all nine billion possible single-letter changes in the human genome, published as the AlphaGenome Atlas. confirmedas of 2026-11-14
- AlphaGenome: using AI to understand the human genome · Google DeepMind · Primary source (retrieved 2026-11-14)
- [16]
Google DeepMind says AlphaGenome 2, trained on more than 100 million single-cell profiles across 68 tissues, identifies nearly twice as many cell-type-specific expression-linked variants as the first model. confirmedas of 2026-11-14
- AlphaGenome: using AI to understand the human genome · Google DeepMind · Primary source (retrieved 2026-11-14)
Revision history (1)
- Page created.
Created Nov 14, 2026. Last reviewed by an editor on Nov 15, 2026. Next scheduled review: Nov 14, 2027.
Cite this page
"AlphaGenome." ContentLora, updated Nov 14, 2026. https://contentlora.com/wiki/alphagenome
Spotted an error? Suggest a correction or emailcorrections@contentlora.com.