AlphaGenome 2 Raises Parkinson’s Genetic Effect Prediction From 17% to 41%
AlphaGenome 2, DeepMind’s model for predicting how DNA changes affect gene activity in different cell types, identified more mutations linked to altered gene activity and performed better on an unseen Parkinson’s disease study than a specialist model trained on that study, according to Two Minute Papers presenter Zsolt Hunyadi Zsolnai-Fehér. He says the results could help researchers investigate disease-related mutations that are difficult to test experimentally, while stressing that the model still misses many mutations and struggles with rare cell types.

A DNA variant only matters in the cells where it changes gene activity
A DNA sequence contains roughly three billion letters, but the sequence alone does not explain when and where genes are switched on. Regulatory switches embedded in DNA help control that activity, and mutations in those switches can contribute to disease. The challenge is that a mutation’s effect depends not just on which letter changes, but on where the change occurs and what kind of cell is reading it.
Károly Zsolnai-Fehér describes AlphaGenome 1 as able to predict the effects of changing a DNA letter, including whether a mutation might activate a disease-related gene. But it struggled with the cell-specific differences that make the problem harder: brain and skin cells contain the same DNA, yet use it differently. With billions of possible single-letter mutations and hundreds of cell types, testing every combination experimentally is impractical.
AlphaGenome 2 is designed to predict those interactions. Its input can pair a DNA sequence with a cell type, such as a piece of pancreatic DNA and a pancreatic beta cell; its output estimates how active genes are in that context. Scientists have already measured gene activity in many cells, giving the model input-output examples to learn from. The task is not simply to identify what a stretch of DNA can do in general, but to predict what it does in a particular cellular setting.
One example in the presentation concerns mutations associated with excessive insulin production. AlphaGenome 2 predicted that the mutations switch on the HK1 gene in insulin-producing pancreatic beta cells, while having almost no effect in other pancreatic cells. The same mutation can therefore have different predicted effects depending on which cells are involved.
The model’s reported results separate detection from generalization
The results presented address two questions: whether AlphaGenome 2 can identify mutations that affect gene activity, and whether it can generalize beyond its training data. The reported figures answer different questions, so the detection result and the cross-study comparison should be read separately.
In one evaluation, AlphaGenome 2 identified twice as many mutations known to affect gene activity, at 90% precision. In practical terms, Zsolnai-Fehér says, about nine out of every ten mutations the model marked were likely to affect gene activity. The chart alongside this result compares cell-type-pooled precision and recall: Scooby scored 0.69, AlphaGenome 1 scored 0.72, and AlphaGenome 2 scored 0.76. Those chart scores describe the models’ performance on that precision–recall test; they are distinct from the separate claim that AlphaGenome 2 identified twice as many relevant mutations.
The generalization example comes from a Parkinson’s disease study. A previous specialist model could correctly predict whether gene activity went up or down for 17% of the known genetic effects. AlphaGenome 2 had never trained on that study. On the same measure, it reached 41%, with both models operating at the same stated level of accuracy.
As Zsolnai-Fehér presents the comparison, AlphaGenome 2 did better on this study’s information than a specialist model that had been trained on it. The result is a test of whether the model can carry its predictions to a study it had not seen, rather than a repeat of the mutation-identification evaluation.
The promise comes with limits—and an open research route
Károly Zsolnai-Fehér is explicit that AlphaGenome 2 remains far from perfect. Even its best predictions miss many disease-related mutations, and it is still not very good at rare cell types. The work does not mean genetic diseases have been solved.
The potential significance, as he frames it, is that sometimes the difference between healthy and sick is a single DNA letter. Predicting what that letter does in a particular cell could help researchers understand disease-linked changes. Zsolnai-Fehér hopes that this work will help save people from serious diseases in the future; that is the hoped-for impact, not an outcome established by the predictions described.
AlphaGenome 2 is available to researchers for free through an API, and DeepMind plans to release the code and model weights as well. Zsolnai-Fehér congratulates the team, including Demis Hassabis and Pushmeet Kohli, and says the work recalls the promise he saw in AlphaFold. The immediate contribution he describes is a model for predicting how genetic changes affect gene activity across cell types—a tool researchers can use, despite its remaining gaps, to investigate mutations that are difficult to test exhaustively in the lab.
