● What: a DNA sequence model that predicts how variants behave in specific human cell types, not averaged across tissue.
● Trained on: over 107 million individual cell profiles across 68 tissues, human and mouse.
● The jump: causal variants identified 30% of the time against 16% for the previous model. Direction of gene effect 43% against 23%.
● Access: free for non-commercial research through the AlphaGenome API. Commercial access is in private preview.
Google DeepMind released AlphaGenome 2 on 11 November. The first AlphaGenome took a one-million-letter DNA sequence and predicted thousands of experiments across many tissues at once, which spread its capacity thin. The second is a conditional model: it predicts for one condition at a time — a tissue, a cell type, a cell state — and carries that condition through every stage of the architecture.
On TenK10K, a single-cell map covering 28 primary immune cell types, that change roughly doubles accuracy. AlphaGenome 2 identifies causal variants 30% of the time against 16% for the original fine-tuned on the same single-cell data, and gets the direction of a gene effect right 43% of the time against 23%. DeepMind also reports it flags nearly twice as many variants that change gene expression as the previous state of the art.
Two additions matter for working researchers. The model now outputs DNA methylation alongside its other predictions, and it can be conditioned on new single-cell data without retraining — so a lab with its own cell atlas can point the model at it directly. DeepMind has also published precomputed predictions for trait-associated variants across hundreds of cell and tissue conditions, and says AlphaGenome 2 Skills for Google Antigravity are coming.
The limitation DeepMind states plainly: research use only. The model has not been validated for clinical or diagnostic use, and nothing in the release changes that.
Sources