Google DeepMind presents AlphaGenome Atlas as a catalogue of predictions about single-letter DNA changes. Its AVI score combines estimates of effects on gene regulation and proteins to rank variants by predicted impact. The announcement on 8 September places the tool at a decisive stage in research: choosing which candidates deserve the next experiments. Source: Google DeepMind
A map becomes useful before anyone reaches a destination. If a laboratory has more candidates than it can examine, better ranking can change where it spends time and materials. Ordering a list does not, however, establish what causes a disease. The atlas’s value depends on its predictions leading to checks that produce additional knowledge.
One letter, several questions
DNA contains information whose use depends on cellular context. The National Human Genome Research Institute describes gene expression as the process turning that information into function: where, when and how much RNA and protein are produced matters. Studying a variant therefore involves more than asking whether it changes a protein’s sequence. Source: National Human Genome Research Institute
A difficulty for readers arises when one number condenses different processes. A single ranking helps decide where to start. It can also conceal that two candidates receive similar priority for different biological reasons. To design an experiment, a researcher needs to recover those reasons rather than retain only each variant’s position in the list.
Consider a hypothetical comparison. A laboratory can examine twenty candidates out of many thousands. One method selects those most similar to known examples; another identifies less familiar but plausible candidates. The first might perform better in an evaluation dominated by known cases and prove less useful for discovering a new mechanism. The relevant comparison depends on the question the laboratory wants to answer.
Prediction and experiment need each other
In 2023 the IGVF Consortium set out a programme combining cellular maps, genomic perturbations and predictive models to investigate the relationship between variation and function. Its abstract proposes comparisons of experimental and computational approaches and an open resource integrating the results. That agenda is independent of the Atlas announcement and shows why predictions belong within a wider research effort. Source: IGVF Consortium / arXiv
A useful test of a prioritisation tool should examine both its successes and what it leaves out. If experiments are conducted only on candidates at the top of the list, researchers learn about those successes but have more difficulty identifying opportunities discarded along the way. Reserving some experiments to examine alternative selections would help assess that cost.
The comparison should also be fixed before the result is known. A ranking can look extraordinary against an arbitrary selection and offer less improvement over the procedure an expert team already uses. Announcing a capability therefore requires explaining which previous work it replaces or complements, and under what conditions.
Who can use the map?
Accessibility is a different consideration from accuracy. An intelligible tool can broaden the range of teams able to ask questions; a complex one may require specialists many laboratories lack. Assessing usefulness also involves documentation, computing requirements and the ability to reproduce a query without depending on a supplier’s private interpretation.
The move towards clinical applications adds another question: whether the available evidence is sufficient for a decision about a particular person. A result that guides research does not automatically answer it. Confusing those purposes could turn an experimental priority into a certainty the model was never in a position to provide.
Atlas’s most interesting promise is more precise than “decoding DNA”: helping researchers make better choices about what to investigate. Judging that promise will require looking at which discoveries survive experimental testing and which results other teams can reproduce. A map earns authority by leading to places that can be checked.