AlphaGenome 2 opens research API for cell specific DNA analysis
Google DeepMind released AlphaGenome 2 on November 11, 2026, giving researchers a more specific way to predict how DNA changes affect different kinds of cells. Noncommercial research access is available through the AlphaGenome API. Commercial research remains in private preview.
The practical change is the biological context supplied with a sequence. Rather than producing fixed outputs across many tissues and cell types at once, AlphaGenome 2 takes a specific condition for each prediction. The result is still a prediction that needs validation, with important access and migration limits.
What changes when the cell type becomes an input
DeepMind describes AlphaGenome 2 as a conditional model: it takes a DNA sequence plus the biological condition being studied. The documented output changes include DNA methylation predictions.
Research access, commercial preview and practical scale
The current API documentation says noncommercial use is free, with query rates varying according to demand. Live inference is intended for analyses involving thousands of predictions rather than millions. AlphaGenome Atlas offers faster lookups of precomputed variant effects, which is a different workflow from requesting new model predictions.
| Route | What to expect |
|---|---|
| AlphaGenome 2 research API | Free noncommercial access, an API key and demand dependent query limits. |
| Commercial AlphaGenome 2 | Private preview in the November 11 announcement. Confirm eligibility and terms separately. |
| AlphaGenome Atlas | Precomputed variant lookups through a website or API, rather than a fresh prediction for a custom condition. |
Do not equate a downloadable Python client with downloadable model weights.
November 12 research access update: DeepMind invites community use cases through the form linked in its announcement by November 30, 2026. Selected projects may qualify for extra quota and credits; neither selection nor support is guaranteed. This is a use-case submission date, not an announced cutoff for API access. Noncommercial research API access and the separate commercial research private preview remain distinct routes.
Existing AlphaGenome code needs more than an import change
Google’s migration guide keeps both generations in the same Python package. Version 2 calls require one reference condition or an expression conditioning input. Batch requests need a corresponding condition for each item. Scripts written to score all tissues at once therefore need an explicit choice about which conditions to compare.
The output schema changes too. AlphaGenome 2 returns raw scores without the earlier quantile score column. Contact maps are not directly comparable with the first generation’s outputs. PRO-cap prediction remains a reason to use AlphaGenome 1, alongside reproducing the earlier paper’s results.
For a migration, our recommendation is to preserve a small known analysis before changing the client. Record the selected model, condition, requested output and scoring method together. A script completing successfully does not establish that an old filtering threshold or comparison still means the same thing.
A documented example provides a bounded starting point
The official getting started example predicts and scores a variant’s effect on GLB1L expression. It offers a reference condition using CD14 positive blood monocytes, or expression conditioning from the researcher’s own human single cell RNA sequencing data. The latter accepts up to 256 cells per call. An API key and the alphagenome package are required.
This is a developer supplied walkthrough, not a ByteForward experiment. Start by reproducing its documented inputs and checking the resulting metadata before adapting it to a new research question. That separates basic setup problems from differences caused by changing the biological condition.
Read the output labels and research restrictions
The AlphaGenome FAQ distinguishes tracks measured in training from supported imputations and exploratory imputations. The exploratory category has not been benchmarked and should be treated as a hypothesis. Expression conditioning can also miss variation when the supplied data differ substantially from the training distribution.
The same FAQ says direct API outputs and Atlas scores are limited to noncommercial research, cannot be used to train machine learning models, and cannot be used for clinical decisions. Human and mouse reference genomes underpin the models. Predictions for other species, synthetic constructs or large structural variants are not benchmarked.
For related context, ByteForward’s SpliceAI2 research guide explains why variant prediction benchmarks, licensing and diagnostic use need separate checks. Neither a better score nor easier API access removes those distinctions.
For prospective users, the immediate value is a more targeted research workflow: specify the relevant cells, inspect which predictions have supporting evidence, and plan independent validation. Keep access terms, biological validity and software compatibility as separate decisions.
Featured image: AI generated editorial illustration, not a scientific diagram or experimental result. Google is a trademark of Google LLC. ByteForward is not affiliated with or endorsed by Google.
