Abstract:
Google DeepMind has launched a new generation of genomic artificial intelligence model AlphaGenome 2, which aims to more accurately predict how DNA genetic variations affect different types of cells in the human body and the activities of cells in different biological states. This technology hopes to help researchers identify key changes related to disease from massive genetic information, further reveal the mechanism of disease, and find new targets for future drug development.
AlphaGenome 2 is the follow-up to the AlphaGenome model launched last year. The core mission of both is to study the impact of genetic variation on gene activity, but the new model further refines the granularity of analysis 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, heart muscle cells, skin cells, and immune cells responsible for fighting off infections all have essentially the same genetic information but turn different genes on or off depending on their function. Even within the same cell, gene activity may change during growth, differentiation, infection, or other physiological conditions.
This means that the same genetic variant can have completely different effects in different cells. If genetic data is analyzed only at the level of the entire tissue or body, certain abnormalities that appear only in specific cells may be missed. AlphaGenome 2 is designed to solve this problem, hoping to determine which cells a certain DNA change will affect and what consequences it may have in a specific biological state.
To achieve this more refined prediction, Google DeepMind used more than 100 million single-cell data for model training, covering 68 types of human tissues. Researchers hope that with the help of these huge cell maps, the model will no longer just identify the general correlation between DNA sequences and gene activities, but will be able to predict the possible impact of genetic variations based on the biological characteristics of specific cells.
DeepMind stated that AlphaGenome 2 can identify almost twice the number of genetic variations that may alter cell gene expression than previous advanced models. This means researchers have the opportunity to discover genetic signals that have been overlooked in larger analyzes in the past.
Another important capability is that AlphaGenome 2 can predict cell types and biological states that have not directly appeared 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. It also helps them discover gaps in existing experimental data and decide where to conduct experiments next.
Adam Gayoso, a Google DeepMind research scientist who participated in the project, believes that the key progress of this work is to more closely connect human genetics and cell biology. Researchers can first discover suspicious variants from large-scale population genetic studies, and then use AlphaGenome 2 to further analyze the possible effects of these variants at the cellular level, and then find links between them and specific diseases.
Žiga Avsec, head of genomics at Google DeepMind, pointed out that this kind of fine analysis at the cellular level is particularly important for drug development. Drugs often target specific proteins and cell types, so knowing exactly which cells a certain genetic variant affects and what happens to 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 are now able to capture individual cells and their biological states at higher resolutions and amass data sets large enough to provide the basis for training such artificial intelligence models.
DeepMind also plans to continue to expand the scale of training data, especially to supplement cell states that are currently underrepresented. Avsec said that future versions may even introduce data from multiple species, allowing the model to learn patterns from a wider range of biological systems. In the longer term, such models may become an important basis for building virtual cells, that is, reproducing biological activities inside real cells through digital simulation.
However, AlphaGenome 2 will not immediately replace the previous generation model. More refined cell-level predictions require more computing resources and higher running costs. DeepMind's previously launched AlphaGenome Atlas in September 2026 has used the original AlphaGenome to pre-calculate approximately 9 billion possible single-base changes and their molecular impacts in the human genome. If AlphaGenome 2, which is more computationally expensive, is used to complete such a huge task, the efficiency and resource consumption will not be ideal.
Therefore, the two generations of models will still undertake different tasks at this stage: AlphaGenome is suitable for large-scale, genome-wide predictions, while AlphaGenome 2 is more suitable for in-depth analysis of specific genetic variations and cell types.
Google announced that AlphaGenome 2 will be available to non-commercial researchers through the AlphaGenome API starting on November 11, while providing limited private preview access to commercial research institutions. This means that researchers can start using it for genetics and cell biology research, but commercial applications are still limited.
It needs to be emphasized that AlphaGenome 2 is still a scientific research prediction tool, not a clinically proven 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 the AlphaGenome Atlas covering about 9 billion genetic variations, Google DeepMind is trying to establish an artificial intelligence research system from DNA sequences, genetic variations to cellular activities. If data from different tissues, cell states, and even multiple species can be further integrated in the future, this type of model is expected to help scientists more systematically understand how genetic information is converted into real biological functions, and provide new research tools for the study of disease mechanisms and the development of new drugs.
Comments