The cell becomes an input
For a scientist investigating a genetic variant, choosing a cell is part of designing the experiment. The same DNA sequence is associated with different gene activity in different cells. Google DeepMind’s AlphaGenome 2 makes that choice an input to computation too: a tissue, cell type or cell state accompanies roughly 1 million bases of DNA. The model opened to researchers on 11 November. Its interesting architectural move is to bring biological context inside the prediction, making the selected cell part of what the model computes.[1]
The technical preprint describes how this condition modifies sequence representations from the network’s early layers onward. A researcher can select a reference cell state or define one from their own single-cell RNA measurements. For a cell with measured gene activity, the model can predict an unmeasured property such as chromatin accessibility. This is a research tool for proposing results in the gaps between costly experiments. Making the question of which cell to test explicit is part of its scientific value.[1]
How the test is assembled
To understand the size of that contribution, examine how the test is assembled. The report’s TenK10K evaluation compares 6346 high-probability variant–gene pairs with 6346 low-probability pairs across 28 immune-cell types. Positive and negative examples are matched one to one. This design evaluates discrimination within a controlled sample. That deliberately constructed mixture does not describe the natural distribution of every variant a scientist might encounter in a real sample.[1]
It would therefore be a mistake to read precision in this test as the probability that an unexamined variant is involved in disease. Where positive examples are rarer, false positives make up a different share of all alerts. Moreover, the comparison labels come from observational genetic associations. Distinguishing cells could reflect sensitivity to biological context; alternatively, the model could be reproducing correlations attached to cell identities well represented in training. Separating those explanations requires an experiment that directly perturbs the variant in the target cell.[1]
The next experimental question
The authors’ discussion matters here. Training atlases consist largely of observational profiles, and some transient or disease-related states remain underrepresented. Defining a state as an input does not guarantee that every biological process in that state has been learned. Prediction from DNA to gene activity also leaves processes occurring after transcription incompletely covered. This gives a concrete boundary to a research tool that has not been validated for clinical or diagnostic use: moving from a predicted molecular association to a useful patient decision requires further measurements.[1]
The next strong test, in my view, would follow a predetermined set of variants rather than a showcase of successful predictions. The same variants would be experimentally altered in different cell states, with measured gene activity compared against directions and effects predicted in advance. Such a design would specify both the chosen cells and what counts as failure before the experiment. AlphaGenome 2 keeps the scientist’s responsibility for selecting a cell visible. Bringing that responsibility into the computation helps formulate a sharper experimental question.[1]