Editor丨&
Different cells in the human body carry essentially the same DNA, yet they perform completely different functions. Cardiac muscle cells need to maintain contraction, immune cells need to recognize and respond to foreign threats, and neurons are responsible for transmitting information. One of the key factors determining these differences is how different cells read the genome and selectively activate or deactivate specific genes.
This also presents a challenge for genetic disease research: even if a DNA variant associated with a disease is identified, researchers still need to determine which gene it affects, through what regulatory mechanism, and in which cell type it exerts its influence.
On November 11th, Google DeepMind released the next-generation genome model AlphaGenome 2, aiming to further integrate DNA sequence prediction with single-cell biology. Compared to the previous version, it advances the prediction focus from broader tissue-level averages to specific cell types and cell states, and attempts to identify genetic variations that function only in particular cells.
Original article link: https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/
Targeted design modifications
Initially, AlphaGenome was designed to input a DNA sequence of up to about one million bases and simultaneously predict its molecular activities across various tissues and experimental conditions. This approach allows efficient coverage of a large number of genomic functional reads, but there is a problem: regulatory differences between different cell types are placed within the same set of parallel prediction tasks, making it difficult for the model to fully capture subtle expression differences between similar cells.
AlphaGenome 2 made adjustments to this design. It employs a conditional model, incorporating specified biological conditions as model inputs alongside the DNA sequence, and utilizing this condition information at various stages of the network. This condition can be tissue type, cell type, or a specific cell state. As a result, the model can predict the functional behavior of the same DNA sequence under a specific cellular context.
Figure 1: Conditional processing model AlphaGenome 2.
The new model's training data includes bulk and single-cell data from humans and mice, with the official claim that it covers over 107 million cell profiles across 68 tissues. In terms of prediction content, in addition to the reads such as gene expression, histone modification, and chromatin accessibility that the previous model already covered, AlphaGenome 2 also incorporates DNA methylation.
Another change is cross-modality data imputation. Different experiments typically measure only part of a cell's molecular features, leading to uneven coverage of data across different tissues and cell types. AlphaGenome 2 can use an existing measurement to predict other unmeasured molecular reads. For example, after researchers provide single-cell gene expression information, the model can further predict features such as chromatin accessibility, offering clues for subsequent experiments.
Enhanced identification capabilities
An important evaluation of AlphaGenome 2 was predicting the effects of genetic variations in immune cells on the TenK10K dataset. This dataset covers 28 primary immune cell types, providing single-cell level data for studying gene expression differences across various immune cells.
Results published by DeepMind showed that in this test, AlphaGenome 2 identified a causal variant proportion of 30%, compared to 16% for the previous AlphaGenome fine-tuned with single-cell data; for predicting the direction of gene expression changes, the new model achieved 43%, while the previous model reached 23%. Both metrics are nearly double that of the previous version.
Figure 2: AlphaGenome 2's predictions are closer to observed gene expression.
DeepMind also presented a case related to inflammatory bowel disease (IBD). A study previously published in Nature found that a genomic regulatory region distant from the target gene can regulate the ETS2 gene via enhancers and influence the inflammatory activity of intestinal macrophages.
Related link: https://www.nature.com/articles/s41586-024-07501-1
After predicting across different intestinal immune cells, AlphaGenome 2 reproduced this mechanism: the model identified the relevant variants and predicted that they would selectively increase ETS2 expression in intestinal macrophages, while having relatively limited effects on other cell types and genes.
When confronted with researchers' own single-cell experimental data, AlphaGenome 2 can autonomously adapt to new cell states without requiring retraining.
Figure 3: Post-training procedure.
For example, in single-cell data from a COVID-19 mRNA vaccine study, the model distinguished between immune T cells in an activated state and those in a resting state, and predicted a genetic variant associated with severe COVID-19 that would more prominently perturb gene activity in activated T cells. Without distinguishing cell states, this difference might be masked by average signals from tissues or cell populations.
Connecting genetics with cellular regulation
AlphaGenome 2 also aims to address a long-standing challenge in genome-wide association studies (GWAS): the difficulty in determining which variants are truly functional, the target genes they affect, and the cell types in which they operate.
To this end, DeepMind precomputed the predicted outcomes of a large number of trait-related genetic variants across hundreds of human cell and tissue conditions, supporting two types of analysis: one for identifying regulatory relationships between enhancers and target genes, and the other for screening potential causal variants and their acting cells from GWAS signals.
For example, many disease-related variants are located in enhancers away from gene coding regions. AlphaGenome 2 extends the prediction of enhancer-gene relationships to more primary human cell types, including astrocytes in the brain, enabling researchers to more specifically analyze which genes the regulatory variants might affect and in which cell types these changes occur.
This step links genomic association signals to cellular regulatory mechanisms. For drug development, researchers not only need to know which gene is associated with a disease, but also wish to understand in which cells it functions and whether changes in its expression constitute an intervenable mechanism. The predictive results provided by AlphaGenome 2 can help narrow down candidate variants, offering direction for subsequent functional experiments and target research.
The development direction of genome AI models is shifting from predicting what molecular signals DNA sequences might generate to explaining in which cells and under what conditions these signals occur. It has the potential to help researchers identify more specific cellular and molecular mechanisms from complex genetic associations. The DeepMind team will subsequently release the related development platform to accelerate research efforts.
Related links: https://www.nature.com/articles/d41586-026-03673-0
https://x.com/GoogleDeepMind/status/2120532824465805749
