AlphaGenome 2 and the New Tooling Layer for Biotech Founders

Google DeepMind announced AlphaGenome 2 on November 11, 2026, a next-generation DNA sequence model that predicts how genetic variants affect gene regulation in specific human cell types. It was trained on more than 107 million individual cell profiles across 68 tissues, and it is available now for non-commercial research through the AlphaGenome API, with commercial research access in private preview. The model is for research use only and has not been validated or approved for clinical or diagnostic purposes.

The advance is a specific one, and it matters because it attacks the central problem in genomics: most disease-linked DNA variants do not sit inside genes. They sit in the regulatory regions around them, and their effects depend entirely on which cell is reading the DNA. A variant that matters in a gut macrophage may do nothing in a skin cell. Until now, models mostly predicted effects averaged across whole tissues — which meant the signal from the one cell type that mattered was diluted by all the others.

What AlphaGenome 2 Does

AlphaGenome 2 is what DeepMind calls a conditional model. It takes a DNA sequence of up to a million letters plus one condition at a time — a tissue, a cell type or a cell state — and predicts how that stretch of DNA is read in that condition. Scientists can choose a condition from a reference catalog, or generate one on the fly from their own single-cell experiment, using as few as a few dozen cells, without retraining the model.

Its readouts cover the core experimental measures of DNA function: gene expression, RNA splicing, chromatin accessibility and histone modifications, with DNA methylation — a widely requested readout — added for the first time. The model was trained on bulk and single-cell data from human and mouse samples, including over 107 million individual cell profiles.

The performance gains are substantial. DeepMind says the model identifies nearly twice as many variants that change gene expression in cells as its previous state-of-the-art model. On TenK10K, a single-cell map of 28 primary immune cell types, AlphaGenome 2 identified 30% of putatively causal expression-linked variants at 90% precision, compared with 16% for the original AlphaGenome fine-tuned on single-cell data, and got the direction of the gene effect right 43% of the time versus 23%.

The Tooling Layer Opportunity

For biotech founders and investors, AlphaGenome 2 is best understood as infrastructure, not as a product. DeepMind is giving away the engine and keeping the research interface; the value for startups sits in what gets built on top of it.

The most immediate opportunity is target identification. Drug discovery starts with a question: which gene, in which cell, is driving this disease? A model that can answer that question cell by cell shortens the most expensive part of the pipeline — the years spent figuring out what to drug. Startups that combine AlphaGenome 2's predictions with their own disease data, patient cohorts or experimental validation can offer pharmaceutical companies a faster route from a genetic association to a credible drug target.

A second opportunity is variant interpretation. Genetic testing companies and rare-disease diagnostics already sit on mountains of variants of uncertain significance — DNA changes found in patients that nobody can interpret. A tool that can say, with a probability, what a variant does in the relevant cell type turns an unanswered report into an actionable one. The regulatory boundary matters here: AlphaGenome 2 is not validated for clinical or diagnostic use, and its terms bar using API outputs for clinical decision-making. The startups that succeed will be the ones that treat the model as a research hypothesis generator and build their own validation around it, not the ones that present its outputs as diagnoses.

The Risks Founders Should Not Ignore

The paper itself is candid about the limits. Even at high precision, AlphaGenome 2 recovers only a minority of the expression-linked and disease-associated variants that statistical analyses nominate as likely causal. It can underestimate effects outside accessible chromatin, at very distant regulatory elements, or through mechanisms that act after transcription. Its outputs, DeepMind warns, contain predictions with varying levels of confidence, and caution should be exercised in their use. Some predicted tracks are marked exploratory, are not benchmarked, and should be treated as hypotheses.

That is the honest frame for the whole field. Genomics AI is moving from tissue averages to cell-type resolution, and that is a genuine step change. But a prediction is not a drug, and a model that is right 30% of the time on causal variants is a powerful research tool, not a substitute for biology.

Angeled.com connects founders with investors.