Discussion
The eight texts answer the paper's question in a way that is easy to miss because no single sentence states it. They do not say when a predicted structure may stand in for an experimental one. They say who must decide that, and with what. The developers built a method, tested it blind and attached a confidence score to every residue. The stewards of the database published the predictions for everyone, set the scores in bands, wrote guidance for each band and placed the duty to interpret on the user. The independent assessments then showed why the duty cannot be discharged once: the same prediction may serve one purpose and fail another, a high score does not rule out a local error, and a protein that changes shape or binds a partner may not be the protein the prediction shows.
The title of the comparison with density maps compresses the position of the independent assessments into a phrase: predictions are valuable hypotheses and accelerate, but do not replace, experimental structure determination. Its closing paragraph says what the authors expect next, and the expectation is theirs:
The accuracies of these predictions and of the uncertainties associated with them are very likely to improve continuously as additional factors are included and as databases of sequence and structural information expand.
Read against the allocation this paper has described, the expectation does not move the decision anywhere. Better predictions and better estimates of their uncertainty would give the user better instruments. They would not, on this paper's reading of the texts, take the choice from the user, because the choice depends on the purpose, and the purpose is the user's. The community assessment says the same of experiment in its own terms:
AF2 will not do away with experimental studies, but the combination of experimental data collection plus artificial intelligence is likely to be a growing trend.
The allocation has a cost the texts acknowledge only in part. A decision handed to every user is made as many times as there are users, with as much care as each brings. The second database paper concedes that many researchers may not sufficiently understand the strengths and limitations of predicted structures and answers with training. Whether training is enough is a question the texts leave open, and this paper leaves it open too. From the conditions the texts name, this paper reads off what a record of such a decision would need to hold if it were ever reopened, without recommending that anyone keep one: the version of the prediction read, the confidence of the residues relied on, the purpose the coordinates served, the partners and states left out, and any experiment that later bears on it. Each item on that list corresponds to one of the conditions under which the texts say a use should be looked at again.
Contributions
This paper has made three contributions, stated at the start and restated here with what the reading added.
First, it has located the decision. Across the eight texts, the decision whether a predicted structure may stand in for an experimental one belongs to the user, for a residue and a purpose, with confidence measures as instruments. The developers' blind test settles whether the method works; it does not settle any use. No text gives the decision to a journal or a funder, and the boundary between the archive of experimental structures and the archive of predictions is drawn by the stewards without the texts saying how a reader is kept from mistaking one for the other.
Second, it has shown what the texts' dates do. Training cut-offs, test windows and database versions date the evidence on which a decision rests. None of them is a deadline. The only pressure of time the texts state is the price of the alternative: months to years for an experimental structure, against seconds to fetch a stored prediction (or potentially hours to compute one) and about a day for an experiment that starts from one.
Third, it has gathered the conditions for looking again that the texts write themselves: new experimental data, a change of purpose, a partner or a state the prediction left out, and a new version of the predictor or the database. These are not this paper's recommendations. They are the occasions the texts themselves name, in their own words, on which a substitution made earlier stops being safe to assume.
Limits
The paper read eight texts, chosen by a rule that took every openly licensed text the study had saved in full. That makes the set small and leaves out much that bears on the question: the many studies that are not openly licensed, which the paper knows only through short notes on their abstracts, and the policies of journals, archives and funders, which none of the eight texts states. A finding that no text gives the decision to a journal or a funder is a finding about these texts, not about the world.
The texts are mostly about one family of methods, from one developer and its partners. Other predictors appear only through a note on an abstract. The four independent assessments each test a few uses, a few protein classes or a few hundred structures, and their authors say so; the comparison with density maps, for example, notes that its sample is small and dominated by high-confidence residues. The paper reports their findings as findings of those studies.
The paper is AI writing from public texts. It does not judge whether any method is right, and it gives no advice on any protein, drug or project. The readings through the works named are this study's interpretation.
Further research
Three questions follow from what the texts leave open. The first is how the archives that now hold both experimental and predicted structures mark the difference, and whether readers who meet a prediction inside a primary resource read it as one; the texts report the integration but not its safeguards. The second is how journals and funders, absent from these texts, in fact decide when a prediction is enough; their written policies would be the next texts to read in the same way. The third is whether the conditions for looking again can be recorded alongside a use of a prediction, so that a later reader knows which version, which residues and which purpose a decision rested on. Each of these could be studied from public texts with the method used here.