What does AlphaGenome 2 actually improve?
AlphaGenome 2 predicts how DNA changes affect gene regulation in specific human cell types. On Google's reported TenK10K evaluation, it identified 30% of relevant expression-altering variants versus 16% for a comparison model. It is a research tool, not a clinically validated test or approved diagnostic.
- Google DeepMind announced AlphaGenome 2 on November 11, 2026. The model conditions a roughly one-million-letter DNA sequence on the cell type or state and was trained using over 107 million single-cell profiles across 68 tissues.
- On the TenK10K immune-cell dataset, Google reports recall of 30% versus 16% for variants affecting gene expression and directional effect results of 43% versus 23% against an AlphaGenome 1 comparison and specialized alternatives. These are evaluation-specific metrics, not clinical sensitivity claims.
- AlphaGenome 2 is available through an API for non-commercial research and private preview for commercial work. Google states explicitly that AlphaGenome products are not validated or approved for clinical or diagnostic purposes.
The Signal
An individual DNA variant can behave differently in a liver cell, an immune cell or a particular disease state. Many genomic prediction tools average signals over broad tissues, potentially hiding the cell where a genetic change matters. Identifying a variant is only the first step; researchers also need to explain which gene and cell state it affects.
On November 11, Google DeepMind introduced AlphaGenome 2, which conditions its regulatory predictions on a specific cell type or biological state rather than producing a single fixed stack of tissue outputs. The researchers trained it on a large collection of human and mouse data that included more than 107 million individual cell profiles spanning 68 tissues.
Google's principal numerical claim is a nearly twofold recall gain on a particular held-out genetic evaluation. On TenK10K data involving 28 primary immune cell types, the model identified 30% of expression-associated variants against 16% for a comparison AlphaGenome configuration. It also improved a directional effect measure from 23% to 43%. The denominators, test split and comparator matter.
What the evidence changes
The architecture changes how the model answers a biological question. Researchers provide a long DNA sequence plus a condition representing a cell type or cell state. The prediction concerns gene expression, chromatin accessibility, histone modifications or methylation within that condition. Researchers can even provide a small new single-cell dataset to characterize a previously unseen state without retraining the full sequence model.
In a published illustration, the team revisited an inflammatory bowel disease association involving regulation of ETS2 in gut macrophages. AlphaGenome 2 reportedly singled out the known variant's cell-selective effect computationally. Reproducing an established result can support biological plausibility, but it is not equivalent to discovering and validating a new causal mechanism in a prospective wet-lab experiment.
The reported recall increases are meaningful yet incomplete. Identifying 30% still leaves many variants missed, and a ranking metric alone does not specify false-positive burdens or the cost of following weak leads. Training inputs, cell-state representation, population diversity and data leakage controls are necessary to interpret performance on genuinely novel samples.
The Cortist perspective
Cortist draws a firm line between predicting a molecular effect and diagnosing a patient. A highly ranked regulatory variant may help design an experiment; it cannot, by itself, prove disease causation, drug efficacy, risk for an individual or a medically actionable result. Google expressly limits the product to research use.
Independent reproduction requires a pinned model and API version, openly specified comparisons, and test variants drawn from populations or cell states sufficiently separate from training. Blind prospective CRISPR or other perturbation studies could measure whether model-chosen variants are confirmed more often, faster or at lower cost than alternatives.
This update complements Cortist's coverage of structRFM, BioBasicsBench and agentic cheminformatics. Those reports address different steps of computational biology: molecular representation, interpreting noisy lab evidence and constructing research workflows. The hard bridge for all of them is moving from a well-scored prediction to a independently measured physical outcome.
- External reruns of TenK10K and independent datasets with disclosed denominators, false positives, statistical intervals and proper held-out genetic regions.
- Prospective perturbation assays that test predictions in previously unseen cell states, with experimental failures counted.
- Whether non-commercial API users can pin and reproduce output behavior across model updates without hidden dataset overlap.
- Any eventual clinical validation, regulatory review or demonstrated improvements in patient-relevant outcomes; none is established by the November announcement.
A nearly twofold improvement in one cell-specific variant recall metric is a substantive research signal. It is not a validated diagnosis, a discovered therapy or a proven downstream improvement in laboratory decisions. The next evidence belongs to independent replication and prospective perturbation tests.
Frequently asked questions
Is AlphaGenome 2 a medical diagnostic tool?
No. Google DeepMind says AlphaGenome 2 is for research only and has not been validated or approved for clinical or diagnostic purposes.
What does 30% versus 16% measure?
It is Google's reported recall for variants affecting gene expression in the evaluated TenK10K immune-cell dataset. It does not mean 30% diagnostic accuracy across patients.
How is AlphaGenome 2 different from AlphaGenome 1?
AlphaGenome 2 conditions predictions on a specific cell type or state and uses large single-cell training data. The earlier model mainly predicted molecular activity across a fixed set of broader tissues and cell lines.
Original sources and verification limits
Original publisher results are reported with explicit boundaries. Cortist has not independently rerun the evaluations or validated clinical or industrial deployment claims. A source link is attribution, not an endorsement or a verification certificate.
Google DeepMind: AlphaGenome 2 and cell-type genetic effectsOriginal research announcement, November 11, 2026AlphaGenome research API and tutorialsOfficial research access and documentation, cited by GoogleSee research methodology, editorial standards and the Cortist Evidence Graph.