Dominic Chiappe · People, capability & transformation

Thinking about how organisations perform in an AI-enabled world

AyEye Today ·

The model knows which cell to ask. Biology still has to answer

AlphaGenome 2 conditions DNA predictions on specific cell types and states, doubling recall from 16% to 30% at 90% precision on one single-cell benchmark. That can improve experimental triage, but the model still misses most variants and its observational predictions are neither causal proof nor clinical guidance.

A sweeping ultramarine current gathers one chartreuse and oxblood knot from dispersed dark marks across scraped warm ground.
Original artwork: The Specific Context. Chiappe × OpenAI.

Google DeepMind released AlphaGenome 2 on 11 November as a model for predicting how changes in DNA affect particular human cell types and states. The important advance is not merely a larger catalogue of predictions. It is a change in the question the model can answer.

The first AlphaGenome could predict molecular activity across broad tissues. AlphaGenome 2 adds a biological condition, such as an immune-cell type or an activated cell state, alongside the DNA sequence. DeepMind reports that this design doubled recall of high-confidence single-cell expression variants at 90% precision on one held-out benchmark, from 16% to 30%. The same paper also says that, even at high precision, the model still recovers only a minority of the variants it is trying to find.

That combination matters. A model that asks where a variant acts can make a research queue more relevant. It does not make the queue conclusive. In genomic medicine, better context should change which experiment happens next, not remove the need for an experiment.

The same DNA can mean something different in another cell

Most cells in a person contain essentially the same genome, but they use different parts of it. A stretch of non-coding DNA may influence gene activity in a gut macrophage while doing little in a neighbouring cell type. A tissue average can blur that difference.

AlphaGenome 2 is a conditional sequence-to-function model. “Sequence-to-function” means that it takes a DNA sequence and predicts molecular readouts such as gene expression, chromatin accessibility and DNA methylation. “Conditional” means that it makes those predictions for a specified tissue, cell type or cell state rather than producing one fixed set of outputs for every context.

The architecture uses that condition throughout the network, starting in its early layers. It was trained on bulk genomics data and thousands of single-cell-derived tracks, then post-trained with profiles sampled from 107 million human cells across 68 tissues. Researchers can also provide gene-expression measurements from a new single-cell experiment to create a representation of a state the model did not see during training.

This is a real capability shift because many disease-associated variants do not sit inside protein-coding genes. They lie in regulatory regions whose effect depends on which cell is using them. The model is intended to connect three questions that genome-wide association studies often leave unresolved: which variant is likely to matter, which gene it affects and in which cell state the effect appears.

The benchmark gain is large and bounded

On the TenK10K dataset, which maps genetic effects across 28 primary immune-cell types, the paper reports that AlphaGenome 2 recovered 30% of high-confidence expression variants at 90% precision. Fine-tuned AlphaGenome 1 recovered 16%. At 90% sign accuracy, recall was 43% for AlphaGenome 2 versus 23% for the fine-tuned original AlphaGenome.

Recall asks how many relevant cases a system finds; precision asks how many of its selected cases are relevant. Holding precision at 90% makes the comparison useful because it reduces the temptation to find more true cases simply by admitting many more false positives. The result suggests a substantially better research filter.

It is not an independent clinical validation. The model, benchmark analysis and headline comparison come from DeepMind’s own preprint. The “causal” variants used in the benchmark are high-confidence statistical fine-mapping results, not all direct laboratory demonstrations of mechanism. The paper reports held-out genomic intervals and datasets for key evaluations, but its authors also chose the architecture, training corpus and evaluation design.

The denominator is equally important. At 30% recall, seven in ten benchmark variants remain unfound at that precision. The authors say the model can underestimate effects outside accessible chromatin, at very distant regulatory elements or through post-transcriptional mechanisms. Transient, stimulated and disease states are underrepresented, and fine-grained states such as neuronal subpopulations remain difficult.

A worked research decision

Consider a hypothetical inflammatory-bowel-disease team studying a non-coding variant near no obvious protein-coding target. The team has samples from several gut and immune-cell populations, but laboratory perturbation across every variant, gene and cell type would be slow and expensive.

The team could use AlphaGenome 2 to rank a set of hypotheses. The model might predict that the variant raises activity of one gene specifically in gut macrophages, while showing little effect in epithelial cells or other immune populations. That would change the next experiment: test the variant and target gene in the predicted macrophage context first, while retaining alternatives that the model scored less highly.

The decision record should separate four layers:

  • Association: human genetic data links a region to disease.
  • Model prediction: a particular variant, gene and cell state are prioritised.
  • Experimental test: perturbation checks whether changing the variant or regulatory element changes the gene in that context.
  • Clinical evidence: separate studies ask whether acting on that mechanism is safe and beneficial for patients.

DeepMind reports that AlphaGenome 2 computationally reproduced part of a published inflammatory-bowel-disease finding involving a distant regulatory switch, the ETS2 gene and gut macrophages. That retrospective example shows that the model can recover a biologically meaningful pattern. It does not establish how often prospective predictions will survive experiments, or that the system is ready to guide diagnosis or treatment.

Conditioning is not causation

The strongest competing interpretation is that much of the gain comes from better coverage and statistical matching rather than a deeper causal model of biology. AlphaGenome 2 uses far more cell-resolved training data than its predecessor and learns condition embeddings from observational, steady-state profiles. A strong prediction may therefore reflect recurring correlations in the data without identifying what would happen after a gene knockout, drug treatment or dynamic transition.

The paper is explicit about this limit. Its condition representations are correlational, and it remains unclear whether the model can predict responses to interventions such as gene knockout or overexpression. That boundary matters because scientific and clinical decisions are usually about interventions, not descriptions.

A useful falsifier follows. If the model’s highest-ranked variant-gene-cell hypotheses do not validate prospectively more often, or with fewer experiments, than strong existing methods, the new conditioning layer is an impressive benchmark improvement without a corresponding discovery advantage. If it consistently reduces wasted experiments while preserving unexpected findings, it has earned a place in the research workflow.

Revisited: reality still has to decide

AyEye argued on 28 October that AI can fill a queue of plausible designs faster than laboratories can test them. AlphaGenome 2 sharpens that view. Better context can improve the queue itself by identifying the cell state in which a hypothesis is most informative.

The earlier judgement is confirmed, not overturned. Experimental reality still decides. What changes is the quality of triage: the model can help choose a more discriminating first test, especially when tissue averages conceal a narrow biological effect.

That makes the practical governance question more precise. Research organisations should record the model version, condition data, confidence threshold, alternatives retained and the experiment that could overturn the prediction. A prediction should be traceable to a research decision, and the result should flow back into the model’s assessment rather than disappear as an inconvenient negative.

AlphaGenome 2 is available for non-commercial research through an application programming interface, with commercial research in private preview. The paper says the system and its outputs are for research, have varying confidence and have not been validated or approved for clinical or diagnostic use.

The advance is therefore consequential and properly bounded. AI can now ask a more specific biological question: what does this DNA change do in this cell state? The answer can make the next experiment smarter. Biology still gets the final word.

Sources and method

This edition treats DeepMind’s launch post and preprint as producer evidence. It distinguishes held-out benchmark performance from independent replication, retrospective recovery from prospective discovery and research prioritisation from clinical validation. Nature’s contemporaneous news report was reviewed for external context, but the quantitative claims here follow the originating paper.

Review window: 11 November 2026 06:45 to 12 November 2026 06:45 Europe/London.

Chiappe × OpenAI. Primary-source review and research with OpenAI; editorial argument, evidence distinctions and final judgement by LecXie.