What the F1-beta case actually shows is that the closed basin was already sitting inside AF3's learned landscape. The default sampler just always lands in the deepest one, because the diffusion process moves atoms toward higher-probability, which in energy terms means downhill [5]. Raise the energy around the coordinates it already found [6], and the same weights walk somewhere else, including states between the two endpoints [9].
That accounting has a price attached. The bias term in prediction number k has to know the outputs of predictions 1 through k-1, so ten conformations means ten AF3 passes run in order, not one pass and a wider readout [14]. This is inference spend, and it scales with how many states you want, which is a different budget line from the one-off cost of training a model.
The thing this does not tell you is how much of the time a protein sits in each state it visits. Diversity here is bought by deliberately deforming the landscape the model learned, so the frequency with which AF3-ReD returns a conformation reflects the bias, not AF3's own probability for it [15]. Anything that turns on populations or free-energy differences between states still needs simulation, which is where Ohnuki and Okazaki point next: they suggest running molecular dynamics from the predicted structures to study how a protein crosses between them, and describe AF3-ReD itself as predicting diverse conformations rapidly and accurately [12]. That last phrase is the authors' characterization, and the material behind it is thin on the numbers you would want. One worked protein is shown, the rest are described as a variety, and there is no error metric, no count, and no comparison against experimental ensembles or long trajectories [16][7].
The retooling claim, though, holds up on mechanism. As described, the repulsive force acts during structure prediction rather than on the network's parameters [17], which means a pipeline that already calls AF3 can produce candidate alternative states by changing what happens between calls. The authors also note that diffusion generative models now show up in the design of new proteins and drug candidates [13], and the same penalty logic would apply there; the source material stops mid-sentence on that extension, so it is a direction, not a result.
For a group deciding whether to spend the extra passes, the useful test is not whether the sampled set resembles a trajectory. It is whether one of the newly reachable conformations is a state you can go and check at the bench, given that AlphaFold's single-answer habit is precisely what has limited its use in work like drug design [11].
