AlphaGenome 2 predicts how the same DNA behaves in different cells
AlphaGenome 2 adds cell-specific context to DNA prediction and improves variant-effect recall in its authors’ tests; a non-commercial research API is available now.
A genetic variant can matter in one cell type and barely register in another.
AlphaGenome 2, released November 11, combines DNA sequence with cellular context to predict how a variant may change gene regulation. Researchers can condition the model on their own single-cell expression data, rather than relying only on a fixed catalogue of familiar tissues. [1][3]
The useful improvement is more true effects found at a chosen confidence threshold. In the paper’s blood-cell evaluation, recall for causal variants rises from 16% to 30% at 90% precision. Recall for the direction of an effect rises from 23% to 43% at 90% sign accuracy. Those are different prediction tasks; neither figure means that the model diagnoses disease with 90% accuracy. [2]
This makes a concrete research tool worth trying: use a known variant and a relevant cell population, then compare its predictions with existing evidence before expanding the analysis. The API is available for non-commercial research, while commercial research is in private preview. It has not been validated or approved for clinical or diagnostic use. [1][3]
Take it further
Explore the API and cell-context tutorial Non-commercial research access, with account/API key and variable quotas. Commercial research is private preview. Not for clinical decisions; no deadline.
Sources & details 3 sources
Discuss in your voice app
Take the reporting, visual explanations and original sources into a conversation. Copy the brief, paste it into your preferred assistant, then start its voice mode.
Paste the brief into your voice app to discuss this reporting.Reporting sources
- Google DeepMind — AlphaGenome 2
Conditional DNA model incorporates cellular context; available non-commercial API and commercial private preview. Research only, no clinical or diagnostic validation.
- AlphaGenome 2 — original paper, Figure 2
Original 78-page paper downloaded after barrier; Figure 2 on page 5 rendered and inspected. TenK10K: 6,346 causal sc-eQTLs in 28 primary blood immune cell types; causal-task 1:1 negatives/positives. At 90% precision, recall is 30% versus 16%; at 90% sign accuracy, recall is 43% versus 23%. These are recall values at fixed operating points, not overall accuracy. 107 million human single-cell profiles in post-training; human and mouse data in broader training. Author evaluation, not external replication.
- AlphaGenome — API and tutorials
Docs list AlphaGenome 2 and condition-on-own-scRNA-seq tutorial, free non-commercial API subject to varying quotas. Commercial eligibility is not established by holding a consumer Google subscription. Terms link rendered only shell, not full terms; avoid claiming broader rights.
Images & figures
At fixed confidence thresholds, AlphaGenome 2 retrieves more true variant effects in the reported blood-cell evaluation.
Euangelion · Source · Usage terms