Alpha Genome 2: How AI Is Decoding the Genome One Cell Type at a Time

Google DeepMind researchers explain how their Alpha Genome family of models reads DNA sequences and predicts what single-letter changes do inside cells. The discussion covers the journey from Enformer and AlphaMissense to Alpha Genome and Alpha Genome 2, along with how these tools are being used to prioritize variants in real research settings.

■ AI Reads the Genome
- What Alpha Genome does: reading a DNA stretch and reporting the effect of a small sequence change
- From Enformer to AlphaFold-based AlphaMissense, then back to the 98% of non-coding DNA
- Alpha Genome's specs: one-million-base sequence, single-base-pair resolution, and an Atlas covering 9 billion possible variants

■ Genotype, Phenotype, and Disease
- Non-coding regulation, cell-type specialization, and Mendelian versus complex diseases
- How GWAS associations are narrowed down to the causal variant, the target gene, and the relevant cell type

■ Alpha Genome 2 and Access
- Cell-type-level predictions, the AVIS ranking score, a rare pediatric epilepsy case, and a macrophage locus example
- Public API, Python package, and a no-code Atlas website, plus the stated limits of current models

Biologists, geneticists, machine learning researchers, and anyone following precision medicine will get a clear picture of what these models can and cannot do today. The episode also outlines how researchers can start using the tools themselves and what a longer-term cell-simulation goal would require.