AlphaGenome 2 API: What to Change Before Migrating Your Analysis

Google DeepMind released AlphaGenome 2 on November 11, with free API access for non-commercial research and a private preview for commercial research. Its practical change is cell-specific DNA prediction: migrating an existing analysis now means choosing biological conditions and revisiting how the pipeline reads scores.

That is more consequential than the headline benchmark improvement. A notebook can keep familiar prediction methods and still answer a different biological question after an upgrade, especially if it previously scored a variant across every available tissue at once.

AlphaGenome 2 versus AlphaGenome 1: choose the analysis first

The original AlphaGenome predicts many experimental readouts across tissues in parallel. AlphaGenome 2 takes a DNA sequence and a biological condition, such as a tissue, cell type or cell state, and predicts for that condition. Both support DNA sequences up to roughly one million base pairs. The new model can also condition on a researcher's own human single-cell RNA sequencing data.

The official migration guide recommends AlphaGenome 2 for most analyses, including cell-specific variant scoring, methylation and cross-modality imputation. It explicitly retains reasons to use AlphaGenome 1: reproducing the original paper, scoring across all tissues and cell types without selecting a condition, and predicting PRO-cap tracks.

Keep the old model where reproducibility is the job. A better new model does not make an existing published result reproducible under different outputs and assumptions. Preserve the version, condition selection and score interpretation alongside the analysis rather than replacing an import and treating the result as equivalent.

What changes in the API

Both versions are available through the alphagenome Python package. The new client lives under alphagenome.models.v2; the original remains under alphagenome.models.v1. Core prediction and scoring method names remain familiar, but three migration points deserve explicit checks.

First, an individual AlphaGenome 2 prediction or scoring call requires one condition. Use a reference condition, or supply expression-conditioning input and its label. The documented custom-data route accepts up to 256 human cells per call, using raw unique molecular identifier counts with Ensembl gene IDs. That is a data requirement, not an invitation to send an arbitrary normalized expression table.

Second, batch methods expect one condition or expression-conditioning input per item. Comparing the same variant across several cell types therefore requires separate condition-specific items or a loop. When porting score_variants, Google's guide specifically says to pass conditions and variant_scorers as keyword arguments.

Third, metadata discovery changes. AlphaGenome 2 uses supported_conditions to identify available conditions and observed_metadata to identify measured tracks. Downstream code that expects the old ontology_curie column should account for biosample_curie and the new per-track imputation labels.

A practical first migration: one variant, two cell conditions

Start with a small analysis whose biological question is already clear. For example, suppose a research workflow compares a variant's predicted effect on gene expression in two immune-cell conditions. The migration should preserve the same variant and genomic interval while making the two condition choices explicit.

The getting-started notebook provides a reference-condition example using CD14-positive monocytes in blood and a variant associated with a GLB1L expression analysis. It shows the sequence of choosing a condition, defining the interval and variant, predicting reference and alternate tracks, scoring the variant, and plotting the outputs. Use that documented example to establish the new response structure before substituting research data.

For the two-condition comparison, run the same variant and interval under each selected condition, then retain the condition identifier with every result. Inspect the requested output type and track metadata before comparing the results. This is a proposed validation workflow, not a report of a test I have run or a prediction about which cell type will show the larger effect.

Only after that small comparison works should the workflow expand to a larger variant set. The API documentation says query rates vary with demand and describes live predictions as suitable for thousands of predictions rather than millions. For precomputed human-variant lookups, AlphaGenome Atlas offers higher query rates and may avoid unnecessary live model calls.

Why old score thresholds and plots need review

AlphaGenome 2 returns raw_score without the original model's quantile_score column. A pipeline that filters on quantile_score will need more than a renamed import, and silently substituting raw_score would change the meaning of the filter. Review the scoring method before carrying an old threshold into the new analysis.

Output shapes and meanings change too. Splice-site outputs become condition-specific with five channels per condition, rather than the original four condition-agnostic tracks. Contact-map targets are no longer coarse-grained, so their values are not directly comparable to the original model's contact maps. PRO-cap is removed from AlphaGenome 2; the guide points to CAGE for transcription-start-site activity, while retaining AlphaGenome 1 for analyses that specifically require PRO-cap.

The new model also labels tracks as observed, imputed_supported or imputed_exploratory. These labels should travel with exported results and plots. An imputed track is not an additional experiment. Presenting every returned channel as equally supported would conceal an important distinction from the person interpreting the analysis.

What the launch benchmark does and does not establish

In its launch announcement, DeepMind reports identifying 30% of causal variants on the TenK10K immune-cell dataset, compared with 16% for its stated comparison, and 43% of directional gene effects compared with 23%. The comparison includes the original AlphaGenome fine-tuned on single-cell data and other specialized tools.

Those are task-specific research results reported by the developer. They are not a general diagnostic accuracy score, a guarantee for a new dataset or proof that a predicted variant causes a particular patient's disease. DeepMind explicitly says AlphaGenome and AlphaGenome 2 are research tools that have not been validated or approved for clinical or diagnostic use.

Commercial research is also a separate access path, currently in private preview. The free non-commercial API should not be treated as an unrestricted commercial license. AlphaGenome 2 Skills for Google Antigravity are announced as coming soon, so they should not be a dependency in a workflow that must run today.

The useful upgrade is the ability to ask a more precise cellular question. Make that question explicit in the code and the saved results, and the model change becomes an interpretable research decision rather than an invisible change underneath an old notebook.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams turn complex models, developer workflows and research updates into clear stories people can understand and use.

Book a 30-minute call.