DeepMind Launches AlphaGenome 2, a DNA Model That Reads Genomes Cell by Cell

Google DeepMind announced AlphaGenome 2 on November 11, 2026, a next-generation artificial intelligence model that predicts how DNA sequences and their variations affect specific human cell types. It is the successor to AlphaGenome, DeepMind's first DNA model. A genetic variant can matter enormously in one specialised cell while showing no signal when averaged across a whole tissue; AlphaGenome 2 is designed to find those cell-specific effects.

What Happened

DeepMind announced the model in a blog post on November 11, 2026, written by researchers Adam Gayoso, Natasha Latysheva, Pushmeet Kohli and Žiga Avsec. The model takes a DNA sequence of up to one million letters, together with a chosen tissue, cell type or cell state, and predicts the molecular readouts for that condition: gene expression, RNA splicing, chromatin accessibility, histone marks and — new in this version — DNA methylation.

The model was trained on human and mouse data including more than 107 million individual cell profiles across 68 tissues. On the headline benchmark DeepMind highlights, a large single-cell immune dataset called TenK10K spanning 28 cell types, AlphaGenome 2 recovered 30 percent of putatively causal expression-linked variants at 90 percent precision, against 16 percent for AlphaGenome 1 fine-tuned on single-cell data. In other words, it identified nearly twice as many likely causal variants as its predecessor.

AlphaGenome 2 is available from today for non-commercial research through the AlphaGenome API,. Commercial research access is in private preview. DeepMind also published two precomputed resources: enhancer–gene links across 640 primary single-cell reference conditions and 39,323 genes, and scores for more than 6.8 million unique variants across 510 human conditions.

DeepMind is explicit about the limits. Both AlphaGenome versions are intended for research use only and have not been validated or approved for clinical or diagnostic purposes. The model weights have not been released, unlike version 1, whose weights sit on GitHub for non-commercial use.

What the Model Does

Most genetic variants linked to disease by genome-wide association studies sit outside protein-coding genes, in the regulatory DNA that controls when and where genes are switched on. Predicting what those variants do has been hard, because the same variant can have different effects in different cells — and most models predicted an average effect across a tissue, diluting the signal from the one cell type where it mattered.

AlphaGenome 2 redesigns the approach around the cell. The cell type or state steers every layer of the model's predictions, so a variant can be scored separately in, say, a gut macrophage and an activated T cell. That matters for drug discovery, where knowing which cell a variant acts in points researchers toward the right tissue, the right mechanism and the right experiment.

Why It Matters

Genomics is one of the fields where AI's impact is most concrete, and AlphaGenome 2 shows the direction of travel: from general-purpose sequence models toward models that understand biological context. The near-doubling of causal-variant recall on DeepMind's benchmark is a genuine advance, though it comes with an important caveat from DeepMind's own paper — even at high precision, the model still recovers only a minority of the variants that statistical analyses nominate as likely causal. At 90 percent precision, seven in ten causal variants are still missed.

The access model is as important as the model itself. Free non-commercial API access with Colab tutorials puts a state-of-the-art genomics model within reach of academic labs and small research teams that could never train one themselves. Commercial access through a private preview, with no public price, keeps the most valuable use cases — pharmaceutical target discovery — under DeepMind's control. The decision not to release the weights reflects how strategically valuable these models have become.

What This Means for Founders and Investors

For founders in computational biology and drug discovery, AlphaGenome 2 lowers the cost of a credible starting hypothesis. A startup that once needed its own genomics team and compute cluster can now query DeepMind's model through an API, test whether a variant acts in a relevant cell type, and focus its limited resources on the most promising targets. The constraint is dependence: the model is an API, not an open download, and commercial terms are set by DeepMind.

For investors, the release is a reminder that the genomics AI stack is stratifying. At the top sit foundation models owned by DeepMind and a handful of others; below them, a growing layer of companies building applications — target identification, patient stratification, drug repurposing — on top of those models. The value in that application layer depends on access terms, data moats and regulatory positioning, not on owning the base model.

Angeled.com connects founders with investors.