All figures below are stated as measured. The three-target permutation test is
reproducible from this repository (python examples/run_targets.py); the
remaining results are reported from the full study on a family-declustered
protein panel.
Blind prediction from the apo structure. Permutation test with NPERM = 20000,
seed 1234, background = distal (≥ 8 Å from the active site), non-anchor,
non-ground-truth residues.
| Target | Permutation p | Significant | Top-5 residues (author numbering) |
|---|---|---|---|
| KRAS G12C | 0.006 | yes | 156, 159, 93, 23, 80 |
| BCR-ABL1 | 0.001 | yes | 321, 320, 322, 306, 323 |
| Cardiac myosin | 0.035 | yes | 123, 670, 125, 698, 122 |
Significant targets: 3 / 3.
Five-fold, protein-grouped, out-of-fold cross-validation over ~210–290 family-declustered held-out proteins; a protein is scored by a model that was not trained on it. Fraction of proteins reaching permutation significance (p < 0.05):
| Method | Out-of-fold significant | Keeps all 3 targets |
|---|---|---|
| Chance | ~5% | — |
| Classical heat-diffusion | ~57% | no (misses myosin) |
| Plain CTQW (quantum walk) | ~57–60% | no (misses myosin) |
| Unified noisy-or (this model) | ~58% | yes |
| Learned combiner (variant) | ~66% | no |
The unified noisy-or model keeps all three targets significant while generalizing comparably to the classical baseline. A learned combiner reaches a higher out-of-fold rate (~66%) but does not preserve all three targets.
Replacing each quantum operator with its same-graph classical-diffusion analog
(unitary exp(-iHt) → diffusive exp(-Lt)), the classical analog is not
significant on the two hard pockets, while the quantum model is:
| Target | Quantum p | Classical-analog p |
|---|---|---|
| Cardiac myosin (collapsed) | 0.002 | ~0.35 |
| BCR-ABL1 (buried) | 0.0002 | ~0.30 |
| KRAS G12C (accessible) | 0.02 | ~0.01 |
The quantum walk is load-bearing specifically for the collapsed and buried pockets. On the accessible KRAS pocket, classical diffusion also succeeds.
Replacing the domain-supplied active site with one predicted from the apo structure alone (geometry-based pocket detection plus burial, network centrality, catalytic-residue identity and sequence conservation), and re-seeding the model from the predicted site, all three targets remain significant:
| Target | p (true anchor) | p (predicted anchor) |
|---|---|---|
| KRAS G12C | 0.006 | 0.024 |
| BCR-ABL1 | 0.001 | 0.0002 |
| Cardiac myosin | 0.035 | 0.026 |
On a 24-protein held-out set the significance rate is unchanged under predicted versus true anchors (0.79 = 0.79). The allosteric prediction is robust to the exact anchor: predicted-anchor recall of the annotated active site is only ~23%, yet the downstream result is unaffected.
Every learned component is fit only on a family-declustered training set; the three targets are held out of all folds; generalization is measured strictly out-of-fold; the classifiers use apo-only features (no bound structure, no ground-truth leakage). Results were checked by independent reproduction.
Keeping the apo-collapsed myosin pocket carries a small generalization cost: its regime is not separable from the apo structure (separability probe AUC ≈ 0.57, unchanged after enlarging the training set), so keeping it appears to require a non-selective quantum lift. A model tuned purely for out-of-fold generalization ties or marginally exceeds plain CTQW; no universal quantum advantage over classical diffusion is claimed on the broad panel.