What happened
The Google DeepMind AlphaGenome repository lists v2.0.0 as its latest release, published on 11 November. The principal addition is an AlphaGenome 2 model API under alphagenome.models.v2. According to the release notes, the API supports predictions under reference conditions and predictions conditioned on single-cell expression. It also supports variant scoring, interval scoring, cell embeddings, and input gradients.
The release is not only a model namespace change. It adds GeneExpressionData, a container for gene-level expression predictions. For users studying regulatory relationships, it introduces enhancer-to-gene scoring and calling utilities in alphagenome.interpretation.enhancer_scoring. The release notes identify these as new library capabilities, but do not attach accuracy figures, throughput measurements, or comparisons against prior AlphaGenome interfaces.
Version 2.0.0 also broadens the package’s result-inspection surface. It adds GeneExpression, OverlaidGeneExpression, and EnhancerArcs plotting components. These are joined by a filter_to_observed helper on both TrackData and JunctionData. In practical terms, the release groups model access, gene-expression data handling, regulatory scoring, and visual components in one versioned package rather than limiting the update to a single prediction endpoint.
Google DeepMind’s release notes also list tutorials, vignettes, and user guides for AlphaGenome 2. That documentation is part of the release alongside the code additions, giving users an indicated path for learning the new interface. The notes do not describe the contents of each tutorial, nor do they state which workflows or user categories each one targets.
There is a compatibility change for existing users. Deprecated alphagenome.models.* aliases for v1 have been removed, and the release notes direct developers to import v1 functionality from alphagenome.models.v1 instead. The release page does not state a price, an access policy, supported hardware, service limits, or a detailed migration procedure beyond that import direction. It also does not claim that v2 replaces every v1 workflow.
What the archive adds
New Runtime’s earlier coverage on 9 September described AlphaGenome Atlas as a predictive map of possible single-letter DNA changes in the human genome. Version 2.0.0 is a new tooling and API event around the same research program, rather than a restatement of the Atlas announcement.
Trend signal
This does not yet fit a pattern New Runtime tracks. The update expands a specialized genomic modeling library, but the release notes alone do not establish a broader shift in agent architecture, interoperability, runtime security, or AI delivery practice.
Why it matters
For teams building research workflows around genomic prediction, the useful change is the newly bundled path from model invocation to analysis. A team can evaluate reference-condition or single-cell expression-conditioned predictions, score variants or intervals, work with gene-level expression data, inspect enhancer-to-gene relationships, and render selected outputs through components named in the same release.
That does not make the package a general-purpose AI development tool. Its immediate audience is researchers and developers already working on genomic modeling. Still, versioned interfaces and companion analysis utilities matter in scientific software because they define the workflow surface that downstream notebooks, pipelines, and internal applications must integrate.
The removal of the old v1 aliases is the operational detail to treat seriously. Teams with AlphaGenome code should check imports before updating, then test the revised alphagenome.models.v1 paths in a controlled branch. Teams evaluating v2 should separate the availability of new capabilities from evidence that those capabilities are suitable for a particular biological question.
What happens next
- Downstream AlphaGenome projects will begin updating imports away from deprecated v1 aliases. (3 months, 70%). Watch for: issues, pull requests, or examples adopting
alphagenome.models.v1. - Users will test the v2 API and new analysis utilities in domain-specific genomic workflows. (6 months, 65%). Watch for: published notebooks or repository examples using expression-conditioned predictions and enhancer scoring.
Limits and uncertainty
The release notes describe what was added but do not report benchmark results, predictive accuracy, latency, compute requirements, or comparisons with earlier AlphaGenome workflows. They do not say whether every new function works under the same input constraints or access conditions. Pricing, quotas, and infrastructure requirements are also not stated on the release page.
The practical effect of the removed aliases on downstream projects is not independently measured. Changing an import path may be straightforward for some users, while integrations relying on undocumented assumptions could require more work. The release notes also do not establish how the new prediction modes perform on representative research tasks.
Skill to build now
If your team already uses AlphaGenome, create a small compatibility check this week. Search for imports beginning with alphagenome.models., replace deprecated v1 aliases with alphagenome.models.v1 where appropriate, and run a known workflow before adopting v2 functionality. Keep prediction validation separate from the upgrade: use a predefined set of genomic cases and compare outputs under your own research criteria.
Reader takeaway
AlphaGenome 2.0.0 is a substantive but specialized research-software release. It adds a v2 API and associated scoring, data, plotting, and documentation tools for genomic prediction workflows. Existing users should audit removed v1 aliases; prospective users should test the new interfaces against their own biological and operational requirements.
Open question
How do AlphaGenome 2’s new prediction and scoring interfaces compare with prior workflows on representative genomic tasks?