AlphaGenome 2 Released: Predicting Variant Effects by Cell Type

From "One Genome" to "One Genome per Cell Type"

Google DeepMind has released AlphaGenome 2, a conditional DNA sequence model capable of predicting regulatory activity and variant effects by cell type. According to MarkTechPost's report, the model is provided through the AlphaGenome API and is open for non-commercial research use.

The significance of this news can only be fully appreciated against a long-standing problem in genomics. All cells in the same person share nearly identical DNA sequences, yet liver cells, neurons, and immune cells behave vastly differently—the difference lies not in the sequence itself but in which segments are "turned on." Variants in non-coding regions are difficult to interpret precisely because their effects are highly dependent on the cellular context in which they occur. A variant may disrupt a regulatory switch in one cell type while having no effect whatsoever in another.

The core proposition of AlphaGenome 2 is to explicitly model this "context" as a condition: input a sequence along with a specified cell type, and output the corresponding regulatory activity and variant effect predictions.

Non-coding Regions and Regulatory Elements

Only about 1%–2% of the human genome directly encodes proteins; the vast majority belongs to non-coding regions. These regions are not "junk DNA" but are densely populated with regulatory elements: promoters sit near gene transcription start sites and recruit RNA polymerase; enhancers can act at distances of hundreds of thousands of bases from their target genes, contacting promoters through three-dimensional chromatin folding; silencers suppress transcription. Non-coding variants are difficult to interpret because they do not alter protein sequences but may disrupt or create transcription factor binding sites or change chromatin accessibility, with effects strongly dependent on whether that element is active in a particular cell. Understanding this is the prerequisite for understanding why "conditional modeling" is necessary.

What Conditional Modeling Changes

Traditional variant effect predictors often produce a single cross-tissue average score. This approach implicitly assumes that a variant's effects are largely consistent, but the tissue specificity of regulatory elements directly refutes this assumption. The conditional model approach makes cell type one of the model's input variables rather than splitting results by tissue after the fact.

The direct benefit of this design is improved interpretability: researchers can pose hypotheses for specific cell types—for example, "does this non-coding variant affect enhancer activity in this cell?"—and obtain predictions within the same framework. For disease genetics, this means signals identified in genome-wide association studies can be examined in more specific cellular contexts.

It should be noted that the source material does not disclose the model architecture, training data scale, or specific performance metrics, so this article does not make quantitative judgments on these dimensions. Only two things can be confirmed: the model is positioned as a conditional sequence model, and it is distributed via API rather than through local weights.

Genome-Wide Association Studies (GWAS) and Fine-Mapping

GWAS identifies genomic loci statistically associated with diseases or traits by comparing genotypes and phenotypes across large numbers of individuals. However, GWAS hits are typically located in non-coding regions, and due to linkage disequilibrium, a single signal peak often contains dozens to hundreds of highly correlated variants, making it difficult to directly identify causal variants. Combining GWAS signals with regulatory annotations and expression quantitative trait locus (eQTL) data is a common fine-mapping strategy. The value of a conditional variant effect model is that it can provide predicted regulatory impacts for each candidate variant within a signal peak in disease-relevant cell types, thereby providing a basis for prioritization and narrowing the scope of subsequent experimental validation.

Trade-offs of API Distribution

AlphaGenome 2 is open for non-commercial research via API, a choice worth discussing separately.

On one hand, the API model lowers the barrier to entry. Genomics laboratories may not have the computing resources needed to run large-scale sequence models, and on-demand access minimizes upfront investment while allowing DeepMind to uniformly control model versions and inference behavior. For methodology researchers, this means the model can be more quickly integrated into existing analysis pipelines.

On the other hand, the API model also brings constraints. Inference must be completed online, and the cost and throughput of large-scale batch variant scanning are limited by the interface rather than local hardware; the model's internal representations cannot be directly inspected, limiting the depth of interpretability research; and the licensing boundaries of "non-commercial research" also need to be clarified by users themselves. This differs from the ecosystem logic of open-weight models—which path to choose depends on research objectives.

The Real Test for Such Models

The value of predicting variant effects by cell type lies not in demonstrations but in whether it can withstand wet-lab validation. Predicted regulatory impacts must ultimately be verified through reporter gene experiments, CRISPR perturbations, or population genetic data. The additional dimension introduced by conditional models also means that validation work must be carried out separately for each cell type, correspondingly increasing the experimental burden.

Another issue to watch is the coverage and granularity of cell types. How cell type resolution is defined and whether rare or difficult-to-culture cell types are within the supported range will determine the model's applicability in practical research. The source material does not address these details, and we must wait for technical reports or documentation.

For readers interested in AI for Science, AlphaGenome 2 represents a trend of sequence models moving from "predicting a number" to "predicting a number in a specific context." Whether this direction holds will ultimately be decided by experimental data.

Wet-Lab Validation Methods

Reporter gene experiments (such as luciferase assays) attach candidate regulatory sequences upstream of a reporter gene and read their ability to drive transcription through fluorescence or luminescence intensity, suitable for validating individual enhancer or promoter fragments. CRISPR perturbations operate in the endogenous genomic context: using CRISPRi/a to modulate chromatin state at target regions, or base editing to precisely introduce candidate variants, followed by RNA sequencing or single-cell readouts to observe changes in target gene expression. Population genetic data provides another validation path—if the direction of effects predicted by the model is consistent with allele frequencies, eQTL effects, or trait associations for that variant in the population, this can be considered indirect support. The three types of evidence are complementary, but all require experiments designed separately for each cell type.

Frequently Asked Questions

What is AlphaGenome 2? A conditional DNA sequence model released by Google DeepMind that predicts regulatory activity and variant effects by cell type.

How to access it? Via the AlphaGenome API, for non-commercial research use.

How does it differ from ordinary variant effect predictors? It takes cell type as a conditional input rather than outputting cross-tissue average predictions.

One-line Summary

DeepMind releases AlphaGenome 2, a conditional DNA sequence model predicting regulatory and variant effects by cell type, available via API for non-commercial research.

Paragraph Summary

Google DeepMind has released AlphaGenome 2, a conditional DNA sequence model that takes a sequence and specified cell type as input and outputs corresponding regulatory activity and variant effect predictions, available via the AlphaGenome API for non-commercial research. The article notes that the same genome behaves very differently across cells, with differences arising from which sequences are "turned on," so the effects of non-coding variants are highly dependent on cellular context—while traditional predictors produce cross-tissue average scores that obscure this tissue specificity. Conditional modeling allows researchers to pose hypotheses for specific cell types and ground GWAS signals in specific cellular contexts. API distribution lowers the computing barrier but limits batch throughput and inspectability of internal representations. The model's ultimate value depends on wet-lab validation and cell type coverage; the source material does not disclose architecture or performance details.