The fold-switching benchmark for multi-state structure prediction is 53% near-isomorphic: on the audited strict tier, AlphaFold2's two-state recovery is 0/23

Abstract

Evaluating multi-state structure prediction requires a reference set that can distinguish genuine two-state recovery from a single-fold output that happens to match two similar reference states — and the de facto standard, the 92 experimentally characterized fold-switching proteins on which the field has evaluated since 2024, has never been certified for this. We computed the whole-chain TM-score directly between each entry's two annotated reference states (a composition check the original curation does not provide; its published per-pair scores are region-restricted and used to bin targets by difficulty): 48 of 92 entries (52.2%, reported as "53%") are near-isomorphic (refsTM ≥ 0.70, unchanged under a region-restricted recompute) — on these entries a model that outputs a single conformation "recovers both states" by construction. We deliver a stratified, homology-confirmed reference set (23 strict pairs, 37 under a broader definition, an operational positive-control band, and two reporting calibers) with per-entry provenance, together with its first certified measurement: on the audited set the two-state-hit rate — both reference states independently recovered — falls from 46/48 in the near-isomorphic band to 0/23 in the strict tier (Wilson 95% upper bound 14.3%), and the more permissive dominant-state matching falls in parallel from 47/48 to 4/23 (65/92 over the full collection under the same rule). The collapse is a graded dose-response in structural dissimilarity (Cochran–Armitage Z = 6.95, p = 3.6 × 10⁻¹²), unchanged at tau = 0.65 and 0.60 and across three MSA-depth regimes — no sampling depth of AlphaFold2 recovers a strict pair's second state. The zero has a location rather than being generic: given the fold-2 sequence, the fold-2 conformation never appears (0 of 11, robust across every tau tested), and for a literature-verified 9 of 11 pairs the fold-2 state is populated in vivo only under partner, membrane or oligomeric conditions that the sequence-plus-MSA input does not encode — a pattern consistent with missing conditioning information rather than with insufficient sampling depth alone — while the model's own confidence carries no specific warning about the unsampled state. All 92 entries fall within the models' training window, within which there is no deposition-recency gradient. The audited reference set, per-sample classification tables and a 490-entry post-cutoff inventory are released with this paper; a conditional-responsiveness protocol is pre-registered.