AlphaGenome 2 adds new gene-expression predictions and enhancer tools

Google DeepMind has released AlphaGenome 2, adding a new model API for gene-expression predictions and tools for scoring genetic variants and enhancer-to-gene links. For researchers working with genomic data, it is a substantial expansion of what the project’s software can do, not just a version-number shuffle.

Google DeepMind Watch analysis

What happened

The release adds reference-condition and single-cell expression-conditioned predictions, variant and interval scoring, cell embeddings, and input gradients. It also introduces a gene-expression data container, enhancer-to-gene scoring and calling utilities, and new visualisations for gene expression and enhancer links.

Google DeepMind’s AlphaGenome 2 release notes also add tutorials and user guides. They say deprecated alphagenome.models. v1 aliases have been removed, and direct users to import from alphagenome.models.v1 instead.

Why it matters

Together, the additions give researchers more ways to examine gene expression, score genomic regions and variants, and interpret possible enhancer-to-gene relationships within the same software toolkit. That could make AlphaGenome more useful for exploring biological questions, though the release notes do not establish how accurate or useful its predictions will be in any particular study.

The removed aliases are a practical wrinkle for existing users: code relying on them may need to change. The release notes point to the updated import path, so this is worth checking before upgrading a working analysis pipeline. Even useful new tools are less charming when they arrive as a surprise import error.

Our read

The most interesting part is the breadth of the new API: expression predictions, variant scoring and enhancer interpretation are now represented together. That makes this a meaningful software release for computational biology, while leaving the scientific value of the outputs to be tested in real research. Existing users should check their imports; prospective users can now explore the new capabilities through the added guides.

What to watch

• Whether researchers report useful results from the new expression-conditioned predictions.
• How the variant and enhancer-to-gene tools perform on real research questions.
• Whether existing users encounter migration issues after the v1 alias removal.

Discussion spark: Which would most change your research workflow: single-cell expression predictions, variant scoring, or enhancer-to-gene analysis?

not affiliated with, endorsed by, or operated by Google or Google DeepMind