Conclusion

Discussion

The question this paper put to eleven texts was whether, when, by whom and on what condition a computed result may stand in for evidence from animals or people. The texts answer the middle two parts of it and leave the first and last to be settled case by case. That division is not an accident of drafting. It follows from the way the Cures Act built its qualification process: a tool is qualified for a context of use, on the scientific merit of a package, by the Secretary, against a standard of evidence the act declares unchanged. Every later text inherits that shape. The FDA Modernization Act 2.0 widens the list of tests the process can weigh. The bill for a 3.0 would rewrite the agency's regulations in the new words. The draft guidance gives sponsors a way to show credibility in proportion to risk. The 2025 announcement says the agency will encourage, pilot and update. None of them moves the decision from the place where the Cures Act put it, and none replaces the case-by-case judgment with a general rule.

Seen from the debate in the introduction, this shape gives each of the four lines a part of what it asks for and none the whole. The credibility line, for which Viceconti and others speak, gets a process in which the checking required rises with what rests on the model; that is close to the draft guidance's own language of model risk and context of use. The data line, for which Bender and Cortes-Ciriano speak, gets a standard of evidence that does not bend for the newcomer; but nothing in the statutes asks whether the data a model was trained on could ever answer the question it is offered for. Their sharpest statement of that worry is a sentence about order rather than about technique:

Read beside the statutes, the sentence names what the qualification process checks last and the data critique checks first. The process starts from a requestor's proposed use and asks whether the tool is fit for it; the critique starts from the question and asks whether any tool could be. This is this study's reading of the two texts together; neither says it of the other.

The translational line, for which Zushin and others speak, gets a statute that no longer names animals as the default; but the same statute keeps animal tests on the list, and the decision to prefer one nonclinical test over another remains the agency's. The line that sits between enthusiasm and scepticism, in Schneider and others' phrase, is the one whose question the texts leave most fully open, because they defer the choice to a process that can say yes for one use and no for another.

The deadlines tell a similar story. Every date in the statutes binds the agency to write something down: guidance, a framework, a taxonomy, a report, a rule. The timing of a single review is handed to guidance the agency writes for itself. A requestor waiting for a model to be accepted therefore depends on two clocks the texts set in different places: the statutory clock for the agency's rules, which is public and fixed, and the clock for the review itself, which is administrative and was left to the agency. The texts read here say nothing about how long such reviews have taken.

The reopening conditions are the most varied part of the answer. The statute's condition, new information that calls a qualification's basis into question, is observable but passive: it waits for someone to bring the information. The engineer's condition is active but local: it sets the rigour of checking by the model's influence and warns against trusting predictions outside the inputs a model was tested on. The data scientist's condition is the most demanding: that past performance on proxy data may not predict what happens in a body, and that, in the authors' words, data of the right types are not currently available. Writing all three into one record would mean saying who watches for new information, what range of inputs the acceptance covers, and what prospective evidence would count against the model. None of the texts asks for all three, and none forbids it.

Contributions

The paper makes three narrow contributions, the same three it set out at the start. First, it shows that the texts answer the question of who decides with one voice: the regulator, for one context of use at a time, on the merit of what a requestor brings, under a standard of evidence that the Cures Act declares unaltered. The widening of the category nonclinical test, which has drawn most of the attention, decides nothing within the category. Second, it shows that the dates in the texts bind the regulator's procedures, not any act of acceptance, and that the clock for an individual review was left to the regulator's own guidance. Third, it sets side by side three kinds of reopening condition, legal, engineering and data-scientific, and shows that each is observable by someone but that no text assigns the watching.

The paper also clarifies the object of the question. In-silico drug discovery, in the narrow sense of finding and designing molecules by computer, is outside the draft guidance's scope and, on this study's reading, asks nothing of a regulator. The decision this paper studies begins only when a computed result is offered as evidence about safety, effectiveness or quality.

Limits

The study has the limits of its rule. It reads only texts whose licence lets them be quoted, which means United States government works and four open-access articles. It does not read the draft guidance itself, only its notice and landing page; it does not read the roadmap the 2025 announcement mentions, the guiding principles the topic page lists, or any guidance from other regulators, including the European Medicines Agency, whose reflection paper it holds only as a bibliographic record. It reads the bill for an FDA Modernization Act 3.0 as introduced and knows nothing of its later course. Its scholarly voices are few, and four of the works it cites are known to it only through their abstracts. It reports how the texts frame the decision, not how any application was decided, and it cannot say whether any computed result has in fact been accepted in place of an animal or clinical study. Nothing in it is medical, clinical, regulatory or investment advice, and it does not recommend any course to a regulator or a developer.

Further research

Three lines of work would test what this paper can only describe. The first would read the agency's published list of qualification submissions and determinations, which the Cures Act requires it to keep and update, and ask how many concern computational models and in what contexts of use. The second would read the guidance and roadmap that this study holds only by description, and ask whether they set the clock for individual reviews or name who watches for the conditions under which an acceptance should be revisited. The third would set the United States texts beside those of other regulators, to see whether any assigns the watching that these texts leave unassigned. Each would answer, with documents rather than argument, the part of Meridian's question that the texts read here leave open: not whether a model may stand in for an animal, but who will notice, and when, if it should not have.