The OMol25 Paradox: Training-Data Domain Specificity Outweighs Architecture in Neural Network Potential Performance on Zn²⁺ Metalloenzyme Conformer Energies

Correction (new version, 2026-07-18) — integrity correction. Corrected version (2026-07-18), fabricated ligand-panel annotations removed. A primary-source audit (2026-07-16) established that the 15-ligand calibration panel used here carried fabricated compound names, potencies and literature attributions (all seven named entries are a different molecule than named; 14 of 15 structures are unknown to PubChem). This version replaces the Section 2.1 ligand table with structure-only descriptors, deletes the drug names and the reported-IC50 column, retains ChEMBL accessions only as unverified structure keys, and withdraws the pIC50 biology-validation (Sections 3.5/4.5) in full. The headline OMol25-paradox result and all cross-NNP agreement statistics are unaffected — they are properties of computations on fixed structures, independent of compound identity or potency. Versioned correction, not a retraction. In silico only. Machine-learning interatomic potentials (MLIPs) trained at high-level DFT are widely assumed to dominate semi-empirical and materials-trained alternatives on molecular tasks. We test this assumption on the conformational energy landscape of fifteen Zn²⁺-binding matrix metalloproteinase-1 (MMP-1) inhibitors drawn from ChEMBL, generated by Boltz-2x cofold sampling with physicality-steering (--use_potentials). Across thirteen independent Boltz-2x diffusion replicates (Boltz seeds 42–54, conditions v11–v23), we evaluate ten energy functions on 100 conformers per ligand per replicate (19,500 conformers per replicate; 253,500 conformer single points total): three semi-empirical (GFN1-xTB, GFN2-xTB, GFN-FF), one Acellera-trained NN force field (AceFF-2), two molecular NN potentials (AIMNet2-NSE, ANI-2x), three universal MLIPs trained on materials physics (Orb-v3 OMat, Orb-v3 OMol25, MatterSim-5M), and two classical force fields (MMFF94, UFF). Per-ligand intra-conformer Pearson correlations are computed within each ligand (size-invariant by construction), then averaged across the 15 ligands and 13 replicates. Three clusters emerge with high cross-replicate stability (SD ≤ 0.038). Cluster I (QM-like) tightly groups GFN1, GFN2, AIMNet2, AceFF-2, Orb-v3 OMat, and MatterSim-5M (intra-cluster r = 0.66–0.99). Cluster II (force-field) groups MMFF94 and UFF (r = 0.66), weakly anti-correlated with Cluster I (r = −0.13 to −0.25). The headline finding is the OMol25 paradox: the same Orb-v3 architecture trained on Meta FAIR's OMol25 molecular dataset (ωB97M-V/def2-TZVPD) drops to r = 0.374 ± 0.025 against GFN1-xTB, versus r = 0.886 ± 0.009 for Orb-v3 OMat — a 0.512 Pearson gap with ≈ 68σ statistical significance across 13 replicates. The paradox is corroborated by a second materials-trained MLIP, MatterSim-5M, which agrees with Orb-v3 OMat at r = 0.936 ± 0.006 (the tightest NN–NN pair in our matrix) but drops to r = 0.410 ± 0.026 against Orb-v3 OMol25. Cluster placement is set by training-data domain, not by architecture. We discuss the mechanistic explanation — OMol25's self-reported weakness on "long-range interactions" (Levine et al., arXiv:2505.08762) coincides with the dominant physics of MMP-1 hydroxamate–Zn²⁺ chemistry — and we explicitly retract a preliminary biology-validation claim that was confounded by ligand size. We recommend that practitioners choose materials-trained or broadly-trained MLIPs (Orb-v3 OMat, MatterSim, AIMNet2) over molecular-only-trained MLIPs (Orb-v3 OMol25, SevenNet OMol25) for Zn²⁺ metalloenzyme conformer-ensemble work, pending higher-level DFT validation. Word count (abstract): 388 words. v5h revision (2026-05-15): Section 2.6 added: extended-cascade pipeline timing reproducibility (chains v_v77–v_v93, 17 independent ensembles) 15-cycle SUSTAINED baseline reported for the four-stage chain backbone (SP CV 1.9%, OPT CV 0.4%, HESS CV 0.94%, GFN-FF CV 1.0%) GFN-FF cache 5-cycle plateau finding (C88–C92 = 10.7 min, C93 break to 11.0 min) Chain × concurrent-ADMET fair-share linear regression model derived from C80 (8 ADMET chains, +52% saturation cap) and C93/C94 (1 chain, +10–13% per stage): regression% ≈ 0.34 × overlap_fraction × N_ADMET_chains Section 4.6 Limitation #6 added (timing-baseline ADMET caveat) Original cross-NNP cluster results (Sections 3.1–3.9) and OMol25 Paradox conclusion UNCHANGED PDF rebuilt with typst engine to fix WeasyPrint table-row drop bug (v0.1 PDF had 14/16 ChEMBL panel rows + 3/4 timing rows missing)