End-to-end de novo design of Zn²⁺ metallohydrolase binders: an open-source canonical pipeline anchored by LigandMPNN's metal-coordination recovery

Correction (new version, 2026-07-18) — integrity correction. Corrected version (2026-07-18), a fabricated-panel identity misstatement removed. A primary-source audit (2026-07-16) established that the referenced calibration panel's annotations are fabricated. This version removes the false statement that 'CHEMBL406 [is] a known MMP-1 inhibitor in the ChEMBL database' and reframes all references to CHEMBL406 and to the '15 ChEMBL MMP-1 ligands' as calibration-panel structures of unverified identity. The headline result (LigandMPNN roughly doubles Zn-coordinating-residue recovery on 1HFC, 95.3% vs 46.4%, with ESM-C corroboration), the de novo library and MAP-Elites statistics, and the HETATM silent-fallback finding are all unaffected — none use the panel. Versioned correction, not a retraction. In silico only. De novo design of binders against Zn²⁺ metallohydrolases (matrix metalloproteinases, carbonic anhydrases, thermolysins, and related catalytic-metal enzymes) remains one of the most demanding stress tests for modern generative protein modeling. The catalytic geometry of these enzymes depends on a small set of coordinating residues (typically His/Asp/Glu/Cys) whose identity and side-chain rotamer states must be preserved through every stage of an end-to-end design pipeline. We present a fully open-source canonical pipeline that integrates four publicly released components — RFdiffusion3 for backbone generation, LigandMPNN for metal-aware inverse folding, FlowPacker for side-chain refinement, and AlphaFold3 / Boltz-2x / Chai-1 for cofold validation — into a reproducible workflow we apply to the matrix metalloproteinase-1 (MMP-1) catalytic domain. The pivotal stage is sequence design: on the 1HFC reference scaffold (157 residues, 2× Zn²⁺ + 1× Ca²⁺), LigandMPNN recovers 95.3% of the six Zn-coordinating positions versus 46.4% for plain ProteinMPNN. The disparity is most pronounced at the structural-Zn triad (His183/Asp185/His196), where ProteinMPNN scores 0% versus LigandMPNN's 90.6%. An orthogonal ESM-C 600M zero-shot likelihood oracle independently confirms that LigandMPNN sequences are more native-like (mean perplexity 2.85 vs 3.03). We document a silent failure mode — when HETATM lines are stripped during preprocessing, LigandMPNN reports use_ligand_context=True but quietly degenerates to ProteinMPNN behavior — and provide a preflight check. The pipeline composes naturally with neural network potential (NNP) ranking (paper_A) and physicality-steered cofold validation (paper_B, --use_potentials). We argue that this open canonical stack now matches or exceeds the design quality of closed alternatives (AlphaProteo) at zero licensing cost for academic users.