Adding cellular context to DNA
The same DNA change can have different consequences in different cells. AlphaGenome 2 gives researchers a way to put that context inside the question. The model Google DeepMind introduced on November 11 takes a million-letter DNA sequence alongside a target tissue, cell type or state. A prediction can therefore specify the biological setting in which a genetic variant is being examined. A researcher’s own single-cell gene-activity data can also supply that condition.[1]
That design helps turn a broad gene–disease association into a more specific experimental question. In the announcement’s ETS2 example, the team computationally revisits a variant whose effect in gut immune cells was already established experimentally. The useful move is to examine gene activity in the relevant cell, rather than seek one answer for every tissue. I read this as a tool for narrowing experimental candidates: which variant, cell and target gene deserve to be tested together?[1]
Describing a cell and intervening in it
A more detailed cell context still carries the boundaries of the training data. The methods in the preprint, which has not yet undergone peer review, describe selecting healthy, unperturbed samples for the single-cell training corpus. Disease samples and cells subjected to genetic interventions were removed from that corpus. Accepting a researcher’s own cell data provides useful flexibility. Adapting a prediction to a description of a state and explaining how that state arose remain different scientific objectives.[1]
The authors’ discussion makes the distinction explicit: relationships learned from observational, steady-state cell profiles are correlational. Predicting the consequences of switching off a gene or increasing its activity remains uncertain. Better description of a cell is one explanation for improved prediction; a competing explanation is that the model has learned a regulatory mechanism that generalizes to interventions. Distinguishing them requires comparing gene activity measured after an intervention with a prediction made beforehand.[1]
A narrower experimental target
That boundary leaves substantial research value. Choosing candidates for genetic experiments is useful work in its own right. AlphaGenome 2 can predict a missing measurement in cells where gene activity is available but another measurement is absent. Such an output could help a laboratory decide where to spend its experimental resources. It could also mislead if the chosen cell context is inappropriate or poorly represented. Selecting the biological context therefore remains a research responsibility rather than something delegated entirely to the model.[1]
An API for non-commercial research makes that use concrete. Commercial research is in private preview, and the model has no validation or approval for clinical or diagnostic use. The meaningful next threshold, in my view, is independent experimental testing of intervention predictions in cells chosen in advance. A well-specified cellular question sharpens the experimental target. Evidence for a decision about a patient remains a separate outcome to measure.[1]