AlphaGenome 2: Cell-Specific DNA Prediction Explained
See how AlphaGenome 2 uses cell-specific DNA predictions to help researchers connect genetic variants to disease—and understand the model’s limits.

AlphaGenome 2: What Cell-Specific DNA Prediction Changes
A DNA variant can be linked to a disease without making clear how it causes harm. Researchers still have to figure out which gene it affects, which cells show the effect, and whether it changes that gene's activity. The catch is that the same DNA sequence can be interpreted very differently depending on the cell - like a heart cell versus an immune cell, or the same cell in a different state.
AlphaGenome 2 is built to help with that interpretation. It pairs a long DNA sequence with a specific biological context - such as a cell type or state - to predict molecular activity and estimate how genetic variants may alter it. The model is trained on more than 107 million individual cell profiles, and the reported results suggest better cell-specific predictions than its predecessor.
That could make it useful for research, but it's not a shortcut from genetic association to diagnosis or treatment. This article looks at what the model predicts, how its conditional setup differs from broader tissue-level approaches, what the reported benchmarks show, and how research teams can judge whether it fits their work. It also covers the practical limits that matter if you're using AI-generated predictions to plan experiments.
What AlphaGenome 2 predicts - and why cell type matters
Every cell in the body has essentially the same DNA. What changes is how cells use it. A cell's type and state influence which genes are active and how strongly they're expressed. Regulatory DNA helps control that activity, sometimes from far away: an enhancer can affect a gene without sitting next to it in the genome.
That creates a real problem for genetic research. Genome-wide association studies (GWAS) can point to DNA regions tied to traits or disease, but an association doesn't automatically reveal the causal variant, the affected gene, or the cell where the effect happens. It's also easy for broad tissue averages to hide signals that show up only in a specific cell population - like a certain immune cell type in the gut.
AlphaGenome 2 takes a DNA sequence of up to one million letters and makes predictions conditioned on a biological context. That context can be a tissue, cell type, or cell state. The model predicts several molecular readouts, including gene expression, chromatin accessibility, histone modifications, and DNA methylation. Those measurements don't all tell the same story, but together they describe different ways DNA is organized and used by cells.
The key difference isn't just "more outputs." It's that the predictions are tied to a specified cell condition. A variant might look basically inactive in one cell type and show a stronger effect in another. If you focus on the relevant condition, you can prioritize hypotheses that a tissue-level average would never bring to the surface.
AlphaGenome 2 is also set up to compare predicted effects across different DNA sequences. Researchers can use those comparisons to ask whether a variant might change gene activity, which genes are plausible targets, and which cell contexts are most consistent with the effect. The output is support for follow-up work - not proof that a variant causes disease.
How conditional modeling changes the workflow
The earlier AlphaGenome approach generated predictions across many tissues and cell lines at once. That can be efficient when you want a broad overview, but it also means the model has to cover many conditions simultaneously. The developers say that setup made it harder to capture subtle differences between closely related cell types.
AlphaGenome 2 changes the setup by treating the biological condition as an input throughout the model. In practice, a researcher provides the DNA sequence and specifies the cell type or state they care about. The model then predicts how that sequence is likely to behave in that context. This is conditional modeling: the prediction depends on both the sequence and the condition.
Using newly profiled cell states
The condition doesn't have to come only from a fixed list. A cell encoder can turn a researcher's single-cell experiment into a representation of a cell state, which the model can then use as context. The announcement says this can work with a relatively small number of cells and doesn't require retraining AlphaGenome 2 for each new dataset.
That matters because datasets often describe cell states a model hasn't been trained on directly. Disease, infection, and vaccination can shift how cells behave. If researchers have measurements from those states, they can ask how variants might act there - rather than relying on predictions for a resting or generic cell.
Predicting measurements that were not collected
This conditional design can also support cross-modality imputation. For example, if a dataset contains one measurement (like single-cell gene activity), the model may predict other properties that weren't measured (like chromatin accessibility). That can help teams generate candidate questions and decide what experiments might be worth running.
Imputation still isn't a replacement for collecting the missing measurement. It's a way to make better use of limited experimental data and to triage what to do next. The value depends on how well the model generalizes to the cells, conditions, and assays in the research team's project. Teams should treat imputed outputs as hypotheses to validate - especially when the downstream experiments are expensive or high-stakes.
What the reported results show
AlphaGenome 2's reported improvements focus on cell-specific gene regulation and variant effects. The developers trained the model on bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles across 68 tissues. The goal is to move beyond tissue averages and model behavior in more specific cell types.
Outside its training data, the model reportedly explained more of the differences in gene expression between cell types within a tissue. One example is blood-cell development: expression shifts as stem cells mature into red blood cells and immune lineages. Modeling that progression can be more informative than treating the tissue as a single uniform category.
On the TenK10K dataset, a benchmark compared predictions across 28 primary immune cell types. AlphaGenome 2 reportedly identified 30% of causal variants in the evaluation, compared with 16% for the original AlphaGenome fine-tuned on single-cell data and other specialized tools, as described by the developers. For directional gene effects - whether a variant increases or decreases gene activity - the reported figures were 43% and 23%, respectively.
Those results are encouraging, but they're not a promise. Benchmarks reflect a particular dataset, a particular set of cell types, and a particular definition of what counts as success. Performance can shift depending on the cell populations studied and how labels are defined. Researchers should check whether the benchmark setup resembles the scientific question they're trying to answer.
The source also describes a case involving inflammatory bowel disease. A variant in a distant regulatory element was reported to increase activity of the ETS2 gene in gut macrophages. AlphaGenome 2 reproduced the cell-specific pattern computationally, identifying the variant and its predicted effect in macrophages while showing little effect in other cell types and genes. It's a good example of the intended workflow: connect a regulatory variant to a target gene and a plausible cellular context, then use that hypothesis to guide evidence gathering.
How researchers can connect genetics and disease
The core opportunity is to turn a GWAS hit into a more specific, testable biological hypothesis. A GWAS can point to a broad genomic region. A prediction model can help narrow down variants within that region, suggest which genes they may affect, and identify cell types where the effect is most plausible.
The developers say they precomputed predictions for trait-associated variants across hundreds of human cell and tissue conditions. The idea is to give researchers a starting point without requiring each team to run large-scale computations themselves. The predictions are aimed at two common challenges.
Linking enhancers to genes in relevant cells
Many disease-associated variants fall in enhancers - regulatory regions that can affect genes from a distance. Mapping enhancer-gene relationships experimentally is often hard. CRISPR screens frequently use stable, lab-grown cell lines, which don't always capture the regulatory wiring of primary cells in the body.
Predictions in primary cell types may help fill that gap. For instance, a researcher studying brain biology could look at candidate enhancer-gene links in astrocytes, cells that support and protect neurons. The model can help prioritize which relationships to test, but experiments are still needed to show that a predicted link is real and relevant to disease.
Prioritizing variants and cellular contexts
A long list of GWAS-associated variants doesn't automatically translate into a clear mechanism. By scoring variants across conditions, AlphaGenome 2 may help narrow the search to a smaller set of candidate variants, genes, and cell types. A predicted direction of effect - whether activity goes up or down - can also tighten the constraints when researchers build a biological explanation.
The same idea can apply to newly observed cell states. In the reported example based on data from a COVID-19 mRNA vaccine study, the model distinguished activated immune T cells from resting ones. It predicted that a variant associated with severe COVID-19 had a stronger effect on gene activity in the activated state. A model evaluated only in resting cells could miss that difference.
For a research team, the practical benefit is prioritization, not automated discovery. Predictions can help decide which variant to investigate, which cell type to model experimentally, and which assay might test the hypothesis most directly. That can stretch limited lab time and funding - so long as the team keeps predictions, evidence, and causation clearly separated.
What teams should evaluate before adopting it
A model's performance matters, but it's only one piece of whether it will be useful. Teams should start by defining the decision they want to improve. Is the bottleneck selecting variants for follow-up, choosing a cell type, linking enhancers to genes, or deciding what measurement to collect next? A tool is easier to evaluate when the job it's meant to do is narrow and measurable.
A practical assessment can follow four steps:
- Match the model to the biological question. Check whether the relevant tissue, cell type, state, and molecular readout are represented well enough for the project.
- Compare with a baseline. Test whether predictions add value beyond existing approaches, such as standard variant prioritization or project-specific models.
- Validate the highest-value predictions. Use independent data or experiments to test candidate effects before treating them as established mechanisms.
- Track failure modes. Keep a record of where predictions don't match observed data, especially across cell states or populations that may be underrepresented.
There are trade-offs. A conditional model can be more specific, but that depends on the quality of the condition you supply. A small single-cell sample may not capture all relevant cell states. Imputed measurements can reduce the amount of data you need to collect upfront, but they also introduce uncertainty that should factor into experimental planning.
There are also operational and ethical questions to think about. Genomic and single-cell data can be sensitive, so teams need to understand data handling, access controls, and the terms of the service they use. If the results influence a commercial research pipeline, teams should document how predictions contributed to decisions and keep human review in the loop. This isn't just "process paperwork" - when decisions are traceable, it's easier to see why a candidate moved forward and whether the model's contribution was justified.
At the time of the announcement, AlphaGenome 2 was available for non-commercial research through the AlphaGenome API, with commercial research in private preview. Its intended use is research. The developers state that AlphaGenome and AlphaGenome 2 have not been validated or approved for clinical or diagnostic purposes. That boundary should shape product planning: a research prediction is not a clinical result.
Key takeaways
- AlphaGenome 2 conditions DNA predictions on a cell type or state, instead of treating cell context as just a broad output category.
- Its reported benchmark gains look promising, but they apply to specific evaluations and shouldn't be assumed to transfer to every project.
- The most useful role is prioritizing hypotheses and experiments, with validation staying separate.
- It's a research tool, not a diagnostic system or a substitute for clinical evidence.
Conclusion
AlphaGenome 2 addresses a real limitation in genetic research: DNA variants don't act in a vacuum, and their effects can depend on the cell type or state where the DNA is read. By conditioning predictions on that context and using inputs from new single-cell experiments, the model offers researchers a more targeted way to connect genetic signals with plausible molecular mechanisms.
The reported improvements are promising, especially for variant prioritization and cell-specific gene regulation. But they don't remove the need for experimental evidence, careful interpretation of benchmarks, or responsible handling of genomic data. The real question for a research team is whether the model improves a specific decision in their workflow - and whether the hypotheses it suggests can be tested. Use it to narrow the search, not to declare an answer.