AlphaGenome 2 and the Risk of One Genetic Shortlist

Thirty per cent, read carefully

Google DeepMind’s AlphaGenome 2 announcement contains a number worth pausing over. On the TenK10K dataset, which maps genetic activity across 28 immune-cell types, the company says the new model identified 30% of causal variants. The earlier AlphaGenome model, fine-tuned on single-cell data, identified 16%. AlphaGenome 2 also predicted directional gene effects at 43%, versus 23%. It improved substantially in this reported evaluation. It also did not identify 70% of the causal variants in that test. ^[1]

That last sentence is not a dismissal. Thirty is more than sixteen, and sometimes more is enormously useful. But percentage gains have a linguistic habit of making the remainder disappear. “Nearly twice as many” directs the eye toward progress; “seven in ten not identified” directs it toward uncertainty. Both describe the same evaluation. A responsible understanding needs both views in the room. ^[1]

The demonstrated change is specific: AlphaGenome 2 made better cell-specific predictions than its predecessor and other specialised tools on DeepMind’s reported tests. Trained on more than 107 million individual cell profiles across 68 tissues, it examines how a DNA sequence or variant may affect a chosen cell type or state. A variant is simply a difference in DNA. The model estimates how that difference may alter processes such as gene activity or access to genetic instructions. ^[1]

DeepMind also showed the model retrospectively reproducing part of a known inflammatory-bowel-disease result. AlphaGenome 2 singled out an implicated variant, connected it to the ETS2 gene and placed its effect in gut macrophages, a type of immune cell. That is meaningful computational evidence. It is not a prospective discovery, a newly validated experiment or a treatment result. The model found a known biological address; the harder test is whether it can reliably help researchers choose unknown destinations. ^[1]

Independent experiments, dependent attention

What interests me is not simply whether one laboratory gets a better shortlist. It is what happens if many laboratories get the same shortlist. Research groups can run separate experiments, use separate equipment and publish separate papers, yet still begin from one shared ranking of what appears promising. The experiments remain independent in execution. Their attention becomes dependent in selection. That small reversal matters because repeated interest can look like emerging consensus even when the original curiosity came through one computational doorway.

This does not mean AlphaGenome 2 is secretly choosing the scientific agenda. Researchers choose experiments; funders choose grants; journals choose papers. Yet a ranking tool can influence all three without making any formal decision. A highly scored variant may be easier to justify because it arrives with a quantitative aura. A low-ranked candidate may require someone to spend scarce money explaining why the model could be wrong. Convenience is not coercion, but it can become a remarkably effective committee member.

The benefit could be considerable. DeepMind says the model can distinguish effects that broad tissue averages miss and can use data from a newly profiled cell state without retraining. It is available through an API for non-commercial research, with commercial research in private preview. If prospective studies confirm that its rankings improve experiment selection, researchers could avoid some unproductive work and direct laboratory resources toward better-supported candidates. That promise remains conditional: API access alone does not supply expertise, samples, funding or a functioning laboratory. ^[1]

The serious counterargument is that shared tools already organise science. Researchers consult common databases, statistical methods and previous papers. A better predictor might reduce guesswork, help understaffed teams investigate difficult biology and expose mechanisms hidden by averages. Different laboratories could also challenge the model rather than obey it. I think this is persuasive. Rejecting a useful filter because it influences attention would be odd; every microscope influences what can be seen. The question is whether we mistake one lens for the whole landscape.

Make disagreement measurable

The practical risk is correlated omission. If ten teams independently test the same ten highly ranked variants, we gain replication of those experiments. We do not gain ten independent searches of the wider possibility space. Candidates scored poorly by the shared model may receive little attention, including candidates associated with biological contexts poorly captured by its training or evaluation data. The hidden cost is not necessarily a wrong answer. It may be an unasked question, which is harder to audit because it leaves no failed experiment behind.

NIST’s AI risk guidance offers a useful discipline here. It says validity should be assessed for a specific intended use, with realistic test sets and evidence of generalisation beyond training conditions. It also notes that accuracy, transparency, interpretability and fairness can pull against one another, and that consequential systems require contextual human judgment. For AlphaGenome 2, a strong benchmark is evidence about prediction. It is not yet evidence that laboratories using its rankings will choose better portfolios of experiments across diseases, populations and institutions. ^[2]

DeepMind explicitly says AlphaGenome 2 is for research and has not been validated or approved for clinical or diagnostic use. So when a future headline says an AI “found” a disease target, ordinary readers should separate three stages: the model ranked a possibility; scientists tested it; independent work confirmed what it means. Those stages may eventually move faster. They do not become interchangeable merely because the first one is impressive. ^[1]

A useful safeguard would make the shortlist part of the scientific record. Researchers could report which candidates the model ranked highly, which low-ranked candidates they sampled anyway, and whether conventional statistics or another model produced a different list. That would not eliminate model influence. It would make the influence visible—and preserve genuine disagreement before agreement hardens around a shared filter. AlphaGenome 2 may help decide where the pipette goes next. We should also record where it did not go, and why.