No. 117
2026-11-13
the AI pulse — brewed, not raw
 AI CRITIQUE

Google DeepMind unveils AlphaGenome 2, which models how cells read DNA

Read the article    English (automatic translation)

Summary

Google DeepMind has released AlphaGenome 2. This DNA sequence model predicts how genetic variants affect gene expression, chromatin accessibility, and epigenetic marks in user-specified cell types or states. The model was trained on over 107 million single-cell profiles across 68 tissues. The tool detects approximately twice as many expression-altering variants compared to the previous version and is available for free for non-commercial research. However, it is explicitly noted that the model has not been validated for clinical use.

Why it matters

Most genetic variants associated with disease are effective only in one or a few cell types, making them invisible in mixed-tissue analyses. AlphaGenome 2's cell-type-specific predictions can help researchers identify the truly causal variants and where they exert their effects. The tool requires only a small measured expression dataset to adapt to a new cell state, as demonstrated with mRNA vaccine T-cell data. This feature lowers the barrier for laboratories without large-scale epigenomic infrastructure. If the pre-computed variant catalog maintains consistency, it could accelerate the mapping of non-coding variants to target genes in hundreds of diseases.

Highlights

  • Trained on 107 million single-cell profiles from human and mouse across 68 tissues; predicts expression, histone marks, chromatin accessibility, and methylation from a 1 Mb sequence.
  • In the TenK10K immune cell benchmark, it recovers 30% of causal variants, compared to 16% for the original AlphaGenome and other tools.
  • Adapts to a new cell state using only dozens of measured cells without retraining; this was validated on T cells activated by mRNA vaccines and resting T cells.

Critical take

The 30% causal variant recovery rate on TenK10K, while double the baseline, implies that the model misses approximately two-thirds of true causal variants in a relatively narrow immune cell panel. The paper does not provide independent reproducibility or out-of-sample generalization data beyond DeepMind's own benchmarks. The 'not for clinical use' warning, coupled with the lack of mention of prospective validation in clinical cohorts, underscores that the tool remains a hypothesis generator and that its real-world utility has not yet been proven outside of DeepMind's own test sets.

This critique was written by the local AI model Qwen3.8-27B. I chose the story; I do not review each critique before it is published.
Back to the issue