On the SARS-CoV-2 Mpro retrospective benchmark (753 ChEMBL compounds: 257 measured actives and 496 measured inactives at pChEMBL<5, receptor PDB 7VU6), AutoDock Vina's below-chance ranking (AUROC 0.427) is a SCORING failure and not a pose-SAMPLING failure. Therefore: substituting DiffDock-L pose generation while retaining Vina scoring will change AUROC by less than 0.05, whereas rescoring poses with the gnina CNN scoring function will raise AUROC by at least 0.10 over the 0.427 baseline. Neither arm will reach the 0.763 AUROC obtained from seven physicochemical descriptors alone on the same 753 compounds.

On the fixed retrospective benchmark of 753 ChEMBL-annotated compounds (257 actives, pChEMBL≥5; 496 inactives, pChEMBL<5) docked against SARS-CoV-2 Mpro (PDB 7VU6, single rigid receptor conformation), three arms will be compared holding the compound set and receptor fixed: (A) AutoDock Vina pose generation + Vina scoring (baseline, expected AUROC 0.427); (B) DiffDock-L pose generation + Vina rescoring of the top DiffDock-L pose per ligand; (C) AutoDock Vina pose generation + gnina CNN rescoring of the top Vina pose per ligand. Claim 1 (sampling is not the bottleneck): AUROC(B) − AUROC(A) < 0.05 in absolute value. Claim 2 (scoring is the bottleneck): AUROC(C) − AUROC(A) ≥ 0.10. Claim 3 (ceiling): AUROC(B) < 0.763 and AUROC(C) < 0.763, where 0.763 is the AUROC of a logistic regression on seven physicochemical descriptors (MW, logP, HBD, HBA, TPSA, rotatable bonds, formal charge) alone on the same 753 compounds.

Minimum viable test: single-receptor, single-conformation, 753-compound benchmark, 3 arms (A/B/C) plus descriptor-only baseline (D, already established at 0.763 — replicate to confirm), n=1 run per arm with fixed random seeds, bootstrap resampling (2,000 iterations) for confidence intervals on AUROC differences. Full validation adds: 3 independent seeds per stochastic arm (DiffDock-L, gnina CNN ensemble), a second receptor conformation (alternate Mpro PDB, e.g. 6LU7) as a robustness check, and pose-quality RMSD-to-crystal validation for the top-N compounds with available co-crystal structures.

# Pseudocode
compounds = load_chembl_benchmark(n=753)  # SMILES, pChEMBL, label
receptor_7vu6 = prepare_receptor("7VU6.pdb")
receptor_6lu7 = prepare_receptor("6LU7.pdb")  # robustness arm

def arm_A(compounds, receptor):
    poses = vina_dock(compounds, receptor, exhaustiveness=32, seed=42)
    scores = [p.top_vina_score for p in poses]
    return poses, auroc(scores, labels)

def arm_B(compounds, receptor, vina_poses):
    diff_poses = diffdock_L_infer(compounds, receptor, n_samples=10)
    top_poses = select_top1_by_confidence(diff_poses)
    rescored = vina_score_only(top_poses, receptor)  # no re-docking
    return auroc(rescored, labels)

def arm_C(vina_poses, receptor):
    cnn_scores = gnina_rescore(vina_poses, receptor, ensemble="default")
    return auroc(cnn_scores, labels)

def arm_D(compounds):
    descriptors = rdkit_descriptors(compounds, features=7)
    model = LogisticRegression(cv=5).fit(descriptors, labels)
    return auroc(model.cv_predict(), labels)

for receptor in [receptor_7vu6, receptor_6lu7]:
    poses_A, auroc_A = arm_A(compounds, receptor)
    auroc_B = arm_B(compounds, receptor, poses_A)
    auroc_C = arm_C(poses_A, receptor)
    auroc_D = arm_D(compounds)  # receptor-independent, compute once
    bootstrap_ci([auroc_A, auroc_B, auroc_C, auroc_D], n=2000)
    evaluate_claims(auroc_A, auroc_B, auroc_C, auroc_D)