AlphaGenome 2 brings cell-specific context to DNA research

Google DeepMind announced AlphaGenome 2 on November 11, 2026. According to the company, the research model for studying DNA variation in a cellular context was trained on more than 107 million cell profiles across 68 tissues. In its TenK10K comparison, causal-variant identification rose from 16% to 30%. These are developer-reported research results, not clinical success rates. The announcement introduced non-commercial research access and a private preview for commercial research. Google DeepMind announcement

The same DNA in different cellular contexts

The official guide explains that each prediction targets a selected biological condition: a tissue, cell type, cell line or the researcher's own cell population. Users can query the reference catalogue, search conditions by name and inspect which outputs have measured training data. Selecting a DNA sequence therefore does not fully specify the research question. The cellular context being investigated also needs to be defined explicitly. Condition-selection guide

Working with data from a new experiment

Researchers with single-cell RNA-sequencing data can condition the model on their own human cell samples without fine-tuning it. The documentation requires raw integer UMI counts with Ensembl gene identifiers and permits up to 256 cells per call. Normalized or log-transformed measurements are not equivalent inputs. Although a single cell can be supplied, at least 16 are recommended when available to reduce sampling noise. The tutorial demonstrates comparing custom cell data with a reference condition while investigating variant effects. Single-cell tutorial

Not simply a drop-in replacement

The migration guide describes AlphaGenome 2 as taking DNA sequence, species and condition information together. New capabilities include single-cell gene-expression, chromatin-accessibility and DNA-methylation outputs. The first model remains useful for reproducing earlier research or specific tasks such as PRO-cap predictions. Version two returns raw variant scores rather than the previous model's quantile scores. For comparisons with an existing analysis, changing the model name alone may therefore be insufficient. Migration documentation

Access does not remove usage restrictions

The documentation homepage says the API is free for non-commercial use, with query rates varying according to demand. Live model inference is positioned for analyses involving thousands of predictions rather than millions. Precomputed Atlas lookups support higher rates. AlphaGenome documentation

The FAQ restricts direct API outputs and Atlas scores to non-commercial research and prohibits their use for machine-learning training or clinical decisions. It directs commercial workflows to separate Google Cloud access. Research access is therefore not an automatic commercial licence. The documentation also warns that the model may miss variation when input data differ substantially from its training distribution. Outputs that have not been benchmarked should be treated as hypotheses, not as a patient's diagnosis or treatment decision. Usage restrictions and limitations

Dji Banner Inside Article - 728x90