Cross-Validation of Three Neural Network Potentials for MMP-1 Zn Active-Site Inhibitor Ranking: A Computationally-Driven Repositioning Study of Vorinostat and Indapamide

Abstract

Correction (new version, 2026-07-21) — integrity correction. Correction (new version, 2026-07-21) - fabricated ligand-panel annotations and withdrawn repositioning narrative. A primary-source audit (2026-07-16) established that the 15-ligand panel underlying this work carries fabricated compound names and potency values (all seven named entries are a different molecule than named; 14 of 15 structures are unknown to PubChem), and a quantitative-claims audit found the paper's descriptor correlations and its repositioning identifications to be unreproducible. This version is a ground-up reconstruction that makes no claim about the identity, potency, or therapeutic candidacy of any compound. The original title's "repositioning study of Vorinostat and Indapamide" framing is withdrawn in full: the two compounds were misidentified and their supporting analyses do not reproduce. What survives, and all this version claims, is the reproducibility characterisation of the co-fold -> xtb -> three-NNP stack (per-cell sigma_E, conformal coverage, two-arm numeric floor, NNP redundancy). Full scope: preprints/_metadata/FABRICATED_PANEL_SCOPE_2026_07_16.md; removed-claims ledger: section 6.1 of the corrected manuscript. Co-folding stacks are now fast enough to generate 10⁴–10⁵ poses per campaign, which makes reproducibility, not throughput, the binding constraint on using them for triage. We characterise the reproducibility of a three-stage stack — Boltz-2 protein–ligand co-fold, GFN2-xTB single-point and tight-optimisation refinement, and a three-engine neural-network-potential (NNP) consensus — on the catalytic Zn²⁺ site of matrix metalloproteinase-1 (MMP-1), using a 15-ligand panel of structurally diverse active-site binders and a separate 140-ligand cohort. We make no claim about the identity, potency, or therapeutic candidacy of any compound. That is not modesty: a primary-source audit found the panel file's compound names and potencies to be fabricated — all seven named entries are a different molecule than the drug they name, and 14 of 15 structures are unknown to PubChem — so the panel is used strictly as chemistry. The reproducibility question we ask is unaffected, because whether a computation repeats on a fixed input does not depend on what that input is called. The three-engine NNP consensus reaches Pearson r = 0.914 (1000-bootstrap 95 % CI [0.826, 0.971]; leave-one-out r = 0.914 ± 0.010, n = 15) — but we report this as the stack's floor, not its headline: it is the lowest of the three pairwise correlations, and the other two engines (MACE-OMol-0 and Orb-v3) agree at r = 0.9992 because both are trained on OMol25. The "three-engine" design therefore carries approximately two independent signals, and we quantify rather than assert its diversity. Wrapping the per-cell σ_E sept-matrix in a normalised split-conformal layer yields distribution-free intervals whose empirical coverage matches nominal over 200 random splits (80 % → 80.01 ± 0.20 %, 90 % → 90.00 ± 0.14 %, 95 % → 95.00 ± 0.10 %, n = 1,755 cells). A two-arm numeric-floor control separates conformational signal from floating-point/SCF noise by 6–9 orders of magnitude, foreclosing the "σ is just BF16 noise" objection. σ_E's outlier ordering survives a GBSA↔ALPB solvent-model change and a leave-one-out/James-Stein/held-out-cross-fitting treatment of the optimizer's curse. On the 140-ligand cohort, ligands with high polar surface area, high heteroatom content and high long-lag intrinsic-state autocorrelation show systematically larger relaxation energy ΔE_relax (Spearman ρ = +0.55 to +0.64, p < 1e-11, each robust to deletion of the three extreme ligands). We report rank statistics because three ligands at ΔE_relax ≈ 495–600 kcal/mol against a median of 11.65 drive spurious |r| > 0.99 product-moment correlations — a leverage artifact that a large n does not protect against. The resulting triage rule is a prior for concentrating refinement effort, not a classifier, and its prospective hit rate is untested. A PoseBusters v2 audit of 45 co-folded structures gives a 93.9 % mean pass rate over 11 checks, with 33 % of structures perfect (11/11) and a floor of 10/11. The contribution is a reproducibility protocol and an explicit statement of its limits: this stack's ensembles are stable, and stability is not accuracy. No claim of predictive or experimental validity is made anywhere in this work. Keywords: reproducibility; co-folding; Boltz-2; GFN2-xTB; neural network potentials; conformal prediction; matrix metalloproteinase-1; zinc coordination