Google DeepMind has introduced AlphaGenome 2, a next-generation artificial intelligence model designed to predict how genetic variations influence gene activity across hundreds of human cell types. Announced on November 11, 2026, the model aims to help researchers understand why the same DNA mutation can have different biological effects depending on the type and condition of a cell.
The new system builds on the original AlphaGenome model but introduces a major change: instead of primarily predicting genetic activity across broad tissues, AlphaGenome 2 can focus on individual cell types and states. Trained on more than 107 million individual cell profiles from human and mouse samples, the model could help scientists identify disease-related genetic variants that conventional approaches might overlook.
What Is Google DeepMind’s AlphaGenome 2?
AlphaGenome 2 is a deep-learning model that studies how DNA sequences influence biological processes, particularly the regulation of genes.
Every cell in the human body contains essentially the same genetic blueprint. However, different cells activate different genes to perform specialized functions. A heart muscle cell, for example, uses a different set of gene activity patterns from an immune cell responding to an infection.
Small changes in DNA, known as genetic variants, can influence these processes. Some variants have little measurable effect, while others can alter gene regulation and contribute to disease.
Understanding which variants matter and where they act is a major challenge in genetics. AlphaGenome 2 is designed to help researchers investigate this problem by predicting how DNA changes affect specific biological conditions.
Google DeepMind says the model can identify nearly twice as many variants that influence gene expression in cells compared with its previous state-of-the-art model.
How AlphaGenome 2 Improves on the Original Model
The most important difference between AlphaGenome and AlphaGenome 2 is how they process information about cell types.
The original AlphaGenome model generated predictions across a collection of tissues and cell lines simultaneously. This approach was useful for studying broad biological patterns but could struggle to distinguish closely related cell types.
AlphaGenome 2 uses a conditional architecture. Researchers provide a DNA sequence alongside a specific biological condition, such as a tissue, cell type or cellular state.
The model then predicts how that particular biological environment influences the interpretation of the DNA sequence.
This matters because a genetic variant may affect gene activity in one cell type but have little effect in another.
For example, a mutation associated with an immune-related disease might influence gene regulation in a particular immune cell without producing the same effect in other cells from the same tissue.
By examining these differences, AlphaGenome 2 can provide a more detailed picture of how genetic variations operate inside the human body.
Trained on More Than 107 Million Cell Profiles
AlphaGenome 2 was trained using a large collection of biological data covering human and mouse samples.
According to Google DeepMind, its training dataset includes more than 107 million individual cell profiles across 68 tissues.
The model uses information from both bulk biological experiments and single-cell datasets. Single-cell data allows researchers to study gene activity in individual cells rather than averaging measurements across a mixture of different cells.
This distinction is important because tissue samples often contain multiple cell populations with different biological functions.
The larger training dataset allows AlphaGenome 2 to learn more detailed relationships between DNA sequences, cell identity and gene regulation.
| Model characteristic | AlphaGenome 2 |
|---|---|
| Developer | Google DeepMind |
| Announcement date | November 11, 2026 |
| Training data | More than 107 million cell profiles |
| Biological coverage | 68 tissues |
| Species represented | Humans and mice |
| DNA input sequence | Up to 1 million DNA letters |
| Main purpose | Predicting cell-specific genetic variant effects |
| Additional capability | Predicting unmeasured molecular properties |
The model also adds DNA methylation predictions to its existing range of molecular measurements, which includes gene expression, RNA splicing, histone modifications and chromatin accessibility.
AlphaGenome 2 Nearly Doubles Variant Detection Performance
One of the most important results reported by Google DeepMind comes from the TenK10K dataset, which contains information covering 28 primary immune cell types.
The researchers evaluated how effectively the model could identify genetic variants associated with changes in gene expression.
At 90% precision, AlphaGenome 2 recovered 30% of putatively causal expression-linked variants. The previous AlphaGenome model, after being fine-tuned on single-cell data, recovered 16%.
The new model also improved its ability to predict the direction of gene expression changes.
| Evaluation measure | AlphaGenome 1 | AlphaGenome 2 |
|---|---|---|
| Causal variant recall at 90% precision | 16% | 30% |
| Directional gene-effect recall at 90% accuracy | 23% | 43% |
These results suggest that AlphaGenome 2 can identify a larger proportion of relevant genetic signals while maintaining the specified precision threshold.
However, these measurements should not be interpreted as proof that the model can independently diagnose genetic diseases. They represent performance on research datasets and provide candidates for further scientific investigation.
The Model Can Study Cell States It Has Not Seen Before
Another important capability is AlphaGenome 2’s ability to work with new cell types and states without requiring conventional model fine-tuning.
Researchers can provide single-cell RNA sequencing data, which measures gene activity in individual cells. The model uses this information to create a representation of the biological condition and make predictions about genetic effects.
This could help scientists investigate cells affected by diseases, infections or vaccination.
Google DeepMind described an example involving data from a COVID-19 mRNA vaccine study. AlphaGenome 2 distinguished immune T cells activated by vaccination from resting T cells.
The model also predicted that a genetic variant associated with severe COVID-19 had a stronger effect on gene activity in activated immune cells.
Such differences could be important because genetic effects are not always constant. A variant may become more influential when a cell enters a particular biological state.
How AlphaGenome 2 Could Help Disease Research
Many genetic variants associated with human diseases are located in non-coding regions of DNA.
These regions do not directly encode proteins but help regulate when, where and how strongly genes are activated.
Identifying the precise biological consequences of these variants is difficult. Researchers often know that a particular region of DNA is associated with a disease but cannot easily determine which cell type, regulatory mechanism or target gene is responsible.
AlphaGenome 2 could help bridge this gap by connecting genetic variants with the cells and regulatory processes in which they are most likely to act.
Google DeepMind highlighted several potential research applications:
- Rare disease research: Prioritizing genetic variants that could contribute to unexplained conditions.
- Complex disease studies: Investigating the genetic mechanisms behind conditions such as type 2 diabetes and inflammatory bowel disease.
- Drug discovery: Identifying genes and regulatory pathways that could become therapeutic targets.
- Cell biology: Understanding how different cell types respond to genetic changes.
- Population genetics: Connecting genetic variations with biological traits across larger groups.
The model could also help researchers identify enhancer-gene relationships. Enhancers are regulatory DNA sequences that can influence the activity of genes located far away along the genome.
Understanding these connections could help scientists identify which genes are affected by disease-associated variants.
Availability: Research Access Through the AlphaGenome API
Google DeepMind has made AlphaGenome 2 available for non-commercial research through the AlphaGenome API.
Commercial researchers can access the model through a private preview, while Google plans to expand its integration with scientific development tools.
The company has also provided interactive tutorials to help researchers understand how to use expression conditioning and explore the model’s capabilities.
Importantly, AlphaGenome 2 is intended for research use. Google DeepMind has explicitly stated that the model has not been validated or approved for clinical or diagnostic purposes.
That means its predictions should not be used as a substitute for medical testing, clinical genetic interpretation or professional medical advice.
The Bigger Picture
AlphaGenome 2 represents an important development in the use of artificial intelligence for biological research. Instead of treating the genome as a fixed sequence with broadly predictable effects, the model examines how genetic information behaves in different cellular environments.
This is particularly important as scientists build increasingly detailed maps of individual human cells. Combining these datasets with predictive AI models could help researchers investigate genetic mechanisms that were previously difficult to identify using conventional experimental methods.
However, computational predictions still need experimental validation. Identifying a potentially important variant is only one step toward understanding a disease, developing a treatment or establishing a causal biological mechanism.
Looking Ahead
The next stage for AlphaGenome 2 will depend on how researchers use it to investigate real biological questions. Its ability to analyze new cell states without conventional fine-tuning could make it useful as single-cell datasets expand across more tissues, diseases and developmental stages. Wider scientific adoption will also help establish how consistently its predictions translate into experimentally confirmed discoveries.
For the biotechnology and pharmaceutical industries, the model could eventually support more precise genetic research and improve the process of identifying potential drug targets. Although it is not a clinical diagnostic system, AlphaGenome 2 demonstrates how AI could help scientists move from identifying disease-associated DNA variations toward understanding the specific cells and molecular processes through which those variations influence human health.
SEO Title: Google DeepMind AlphaGenome 2 Predicts DNA Effects Across Human Cells
Meta Description: Google DeepMind’s AlphaGenome 2 uses more than 107 million cell profiles to predict how DNA variants affect hundreds of human cell types and states.
URL Slug: google-deepmind-alphagenome-2-dna-variant-effects
Focus Keyphrase: AlphaGenome 2
Secondary Keywords: Google DeepMind AlphaGenome, DNA variant prediction, AI genomics, genetic variant effects, human cell types, single-cell biology, AI drug discovery
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
