AI News

DeepMind releases AlphaGenome 2, refining the analysis of genetic variant effects down to single-cell types

cnBetaUpdated Nov 12 00:39

AI summary

Google DeepMind has released AlphaGenome 2, refining the granularity of its analysis of genetic variants' effects on gene activity from the tissue level down to specific cell types and cell states. The model was trained on more than 100 million single-cell data points, covering 68 human tissues, and can identify roughly twice as many genetic variants that may alter cellular gene expression as previous state-of-the-art models, and can make predictions for cell types and states not directly present in training. AlphaGenome 2 will be opened to non-commercial researchers through the AlphaGenome API starting November 11, with commercial organizations having only limited private preview access; it will not replace the previous-generation model, which remains in use for large-scale genome-wide predictions.

Why read this

AI score 65

The original text explains the scale of the training data, the cell-level prediction granularity, and the open-access approach, allowing readers to judge its positioning in genetics research.

Topics

Tags

Body · AI translation

AlphaGenome 2 is a successor to the AlphaGenome model launched last year. The core task of both is to study the impact of genetic variants on gene activity, but the new model further refines the analysis granularity to specific cell types and cell states.

Almost all cells in the human body carry the same DNA, but different cells perform different tasks. For example, cardiomyocytes, skin cells, and immune cells responsible for defending against infection, although they possess essentially the same genetic information, activate or shut down different genes according to their own functions. Even for the same type of cell, gene activity may change during growth, differentiation, infection, or other physiological states.

This means that the same genetic variant may produce completely different effects in different cells. If genetic data are analyzed only at the level of whole tissues or the whole human body, abnormalities that appear only in specific cells may be missed. AlphaGenome 2 is designed precisely to solve this problem, with the goal of determining which cells a given DNA change affects and what consequences it may produce under specific biological states.

To achieve this more refined prediction, Google DeepMind used more than 100 million single-cell data records to train the model, covering 68 human tissues. The researchers hope that with these massive cell atlases, the model will no longer merely identify general associations between DNA sequences and gene activity, but will be able to predict the possible effects of genetic variants based on the biological characteristics of specific cells.

DeepMind says AlphaGenome 2 can identify nearly twice as many genetic variants that may alter cellular gene expression as previous advanced models. This means researchers have the opportunity to discover genetic signals that were easily overlooked in broader analyses in the past.

Another important capability is that AlphaGenome 2 can make predictions for cell types and biological states that did not directly appear in the training data, without having to retrain the model for each new cell type. This generalization ability helps researchers explore cell types for which data are relatively scarce, and can also help them identify gaps in existing experimental data, thereby deciding which directions to pursue in the next round of experiments.

Adam Gayoso, a Google DeepMind research scientist involved in the project, believes that the key advance of this work is to link human genetics more closely with cell biology. Researchers can first discover suspicious variants from large-scale population genetic studies, then use AlphaGenome 2 to further analyze the possible effects of these variants at the cellular level, and then look for their connections to specific diseases.

Žiga Avsec, head of genomics at Google DeepMind, pointed out that this fine-grained cellular-level analysis is especially important for drug development. Drugs usually act on specific proteins and cell types, so understanding which cells a given genetic variant affects and what changes occur in those cells may help researchers more accurately identify therapeutic targets.

Progress in this field has also benefited from the development of single-cell sequencing and other biotechnologies in recent years. Researchers can now capture individual cells and their biological states at higher resolution, and accumulate sufficiently large datasets to provide the foundation for training such artificial intelligence models.

DeepMind also plans to continue expanding the scale of training data, especially to supplement currently underrepresented cell states. Avsec said that future versions may even introduce data from multiple species, enabling the model to learn patterns from a broader range of biological systems. In the longer term, such models may become an important foundation for building virtual cells, that is, digitally simulating and reproducing the biological activities inside real cells.

However, AlphaGenome 2 will not immediately replace the previous-generation model. More refined cell-level predictions require more computing resources and are more expensive to run. AlphaGenome Atlas, which DeepMind launched in September 2026, has already used the original AlphaGenome to precompute about 9 billion possible single-base changes in the human genome and their molecular effects. If the more computationally expensive AlphaGenome 2 were used to complete such a massive task, efficiency and resource consumption would not be ideal.

Therefore, at this stage the two generations of models will still undertake different tasks: AlphaGenome is suitable for large-scale, genome-wide predictions, while AlphaGenome 2 is more suitable for in-depth analysis of specific genetic variants and cell types.

Google announced that AlphaGenome 2 will be opened to non-commercial researchers through the AlphaGenome API starting November 11, while commercial research institutions will be offered limited private preview access. This means researchers can begin using it for genetics and cell biology research, but commercial applications remain in a restricted-access phase.

It should be emphasized that AlphaGenome 2 is currently still a scientific prediction tool, not a clinically validated diagnostic or treatment system. The genetic variants identified by the model and their potential effects still need to be further confirmed through experimental research and clinical evidence, and cannot be directly used to determine whether a person has a certain disease.

From AlphaGenome to AlphaGenome 2, and then to AlphaGenome Atlas, which covers about 9 billion genetic variants, Google DeepMind is trying to build an artificial intelligence research system that spans DNA sequences, genetic variants, and cellular activity. If data from different tissues, cell states, and even multiple species can be further integrated in the future, such models are expected to help scientists more systematically understand how genetic information is translated into real biological functions, and provide new research tools for studying disease mechanisms and developing new drugs.

Source

cnBeta