Google DeepMind's AlphaGenome 2 interface displays intricate genomic data, symbolizing AI-driven biological discovery.

Editorial illustration for Google DeepMind Unveils AlphaGenome 2 Trained on 100 Million Cell Profiles

Google DeepMind Unveils AlphaGenome 2 Trained on 100...

By AI Daily Post Edited by Brian Petersen, Editor-in-Chief

November 11, 2026 • 3 min read

Google DeepMind put out AlphaGenome 2 this week, a DNA sequence model built to answer a question its predecessor couldn't: what does a genetic variant actually do in one specific cell type, not just on average across a tissue sample. The original AlphaGenome, released earlier, predicted thousands of genomic tracks at once but blurred everything into bulk-tissue averages. That's fine for broad patterns, useless if you're trying to figure out why a variant matters in, say, a particular immune cell subtype but not its closest neighbor.

AlphaGenome 2 changes the setup entirely. It takes a 1 megabase stretch of DNA, human or mouse, and pairs it with a target cell state pulled from raw single-cell RNA sequencing data. No fine-tuning required, even on datasets the model never saw during training.

DeepMind trained it on 2,048 TPU v4 chips and ran it against benchmarks built from single-cell eQTLs and enhancer-gene links, the kinds of tests where bulk-tissue models tend to fall apart. The model is live now through DeepMind's API, limited to non-commercial research, with weights promised only after "final publication."

The main result is on TenK10K single-cell eQTLs across 28 immune cell types. At 90% precision, AlphaGenome 2 recovered 30% of causal variants. AlphaGenome 1 fine-tuned on single-cell data recovered 16%.

Google DeepMind Releases AlphaGenome 2 for Cell-Type-Specific Variant Effects - MarkTechPost

Why this matters

For researchers mapping disease variants to specific cell types, averaging across bulk tissue has always been a blunt instrument. AlphaGenome 2's bet is that a frozen model plus a swappable single-cell encoder, trained on over 100 million scRNA-seq profiles from CELLxGENE, can resolve that without retraining the whole network for every new dataset. That's a real architectural choice, not just a bigger training run, and it matters for anyone building variant-interpretation tools on top of it: you're not fine-tuning, you're passing raw counts and getting a condition embedding back.

The 256-cells-per-call limit in the docs is worth flagging early. It caps how granular a single query can get, which will shape how founders design pipelines around this rather than treating it as infinite-resolution genomics. And the non-commercial research license means anyone with product ambitions needs a plan for what happens after the research phase, not just during it.

Worth watching: whether the "unseen single-cell datasets without fine-tuning" claim holds up against data that looks nothing like CELLxGENE's composition.