Google DeepMind announced AlphaGenome 2 on Wednesday, November 11, a next-generation DNA sequence model that predicts how genetic variants affect individual human cell types and cell states, rather than averaging their effects across whole tissues. The model is available now for non-commercial research through the AlphaGenome API, with commercial research access offered in private preview.
Every cell in the body carries the same DNA blueprint, yet a heart muscle cell and an immune cell switch very different genes on and off to do their jobs. That is why a variant that matters in one specialized cell can disappear in a tissue-wide average. DeepMind's first AlphaGenome, launched last year, predicted how variants affect molecular activity across broad tissues, and its Atlas covers all 9 billion possible single-letter changes in the human genome. AlphaGenome 2 zooms in from those averages to the cells where disease actually unfolds.
The redesign is fundamental. Instead of predicting thousands of experiments in parallel from a DNA sequence of up to one million letters, AlphaGenome 2 is a conditional model. It takes one biological condition at a time, defined as a tissue, a cell type, or a cell state, and uses that condition as an input at every stage of its architecture alongside the sequence. It was trained on bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles across 68 tissues. It predicts gene expression, histone modifications, and chromatin accessibility, and it adds DNA methylation, a widely requested readout.
On the TenK10K dataset, a large single-cell map of 28 primary immune cell types, AlphaGenome 2 identified 30 percent of causal variants, compared with 16 percent for the original AlphaGenome fine-tuned on single-cell data and other specialized tools. It also predicted the direction of gene effects correctly 43 percent of the time, compared with 23 percent. DeepMind says the model reproduced a landmark inflammatory bowel disease finding computationally, singling out the variant that overactivates the ETS2 gene in gut macrophages while leaving other cell types and genes largely unaffected.
The model can also work with cell states it has never seen, without retraining. A researcher can supply a single-cell RNA snapshot from as few as a few dozen cells, and the model's cell encoder compresses that snapshot into a representation it can condition on. Given unseen data from a COVID-19 mRNA vaccine study, it distinguished vaccine-activated immune T cells from resting ones, and showed that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state. It can also predict unmeasured properties, such as chromatin accessibility, from a single measurement.
To connect genetics with cell biology at scale, DeepMind has precomputed AlphaGenome 2 predictions for trait-associated variants across hundreds of human cell and tissue conditions, covering traits from height to type 2 diabetes, and it has extended enhancer-to-gene mapping to primary human cell types such as brain astrocytes. Interactive Colab tutorials are available now, and AlphaGenome 2 Skills are planned soon for Google Antigravity, the company's agentic development platform.
The limits matter as much as the results. Even at high precision, the model recovers only a minority of the causal variants that statistical studies nominate, and every performance figure in the announcement is company-reported. DeepMind states plainly that AlphaGenome 2 is intended for research use only and has not been validated or approved for clinical or diagnostic purposes. Within those bounds, it gives genetics researchers a sharper question to ask: not just what a variant does, but in which cell, and in which state, it does it.