Falsification works in this field because a claim and a test can be made to touch. A model that predicts an assay result asserts something checkable about cases it has not seen; when the cases come back the other way, the model is wrong, and the wrongness is the useful part. On the premise the piece sets out, that link loosens [4]: what is under test is a representation rather than a single prediction, and any benchmark reaches the representation only through some downstream task. Ten task scores tell you about ten tasks.
The reference list does hold the case where community assessment settled something. The entry for CASP round XIV is annotated as the independent evidence that AlphaFold2 reached near-experimental accuracy for many single-protein structure prediction targets [7]. Note the qualifiers: many, and single-protein. It worked because each target had an experimental answer that existed independently of the model. The DREAM assessment and the crowdsourcing tradition sit in the same list [10]. So do the systems now in question: AlphaFold 3, Boltz-2, AlphaGenome, Evo 2, a general-purpose pathology model, generative transformers for the natural history of human disease, and a roadmap for building a virtual cell [11]. For those, the abstract never says what would play the role of the experimental answer.
The accessible text is a statement of the problem rather than a resolution of it. It commits to discussing epistemological value, refutation versus verification versus utility, guiding principles, and the community's role [5], and it asks outright how the limitations of these models can be tested [6]. Cited alongside is Kapoor and Narayanan on leakage, the contamination that makes held-out scores look better than the model is [8].
On the community's role, the arithmetic is awkward. Reading the piece costs $39.95 as a single article, against the $21.58 per issue the same page quotes for a $259 annual subscription [12], roughly 1.85 times as much [13]. Four of the twenty references visible are preprints, two of which describe models in the debate [14].
My reading, and I will hold it conditionally: utility is a legitimate criterion, and a sufficient one for a deployment decision, but it is not evidence about mechanism, and scoring a model only on utility shows what it does, not that it has learned biology. The bar that earns the stronger claim is the CASP bar, which is targets whose answers are generated after the model is frozen and scored by people who did not train it. Whether its authors believe that bar is reachable for a virtual cell is left unanswered by the abstract [5].
