Resolving power as a measure of protein stability prediction

Rank correlation and discriminative resolution are distinct properties of a stability predictor, yet current benchmarks measure only the first. Here we introduce resolving power: the {Delta}{Delta}G separation at which a predictor orders two variants correctly three times in four. Applied to BioEmu across twelve wild-type domains and three supervised regressors on an independent twenty-seven, resolving powers cluster near 1 kcal mol -1 --- comparable to the median variant-pair separation the benchmarks ask these models to rank, so most pairs cannot be ordered. A closed form ties this resolution to the calibration slope ; Spearman correlations account for only 31% of its variance. BioEmu extracts {Delta}G from the ratio of folded and unfolded populations, so what it samples of each becomes a question. Its folded populations reproduce experimental radii of gyration within half an [a]ngstrom, whereas those it labels unfolded are collapsed chains roughly 20 to 40% below a sequence-specific random-coil reference. The ensemble still carries variant differences on panels where the sampled unfolded fraction approaches zero, but the log-odds counting readout runs out of range. Scored by a learned energy, those same conformations recover resolution on ten of twelve panels, with the gain separated from noise on four, without retraining.