Systematic over-binding in ZAFF-AMBER absolute binding free energy calculations for zinc-metalloenzyme inhibitors: a 14-compound accuracy benchmark on MMP-1

Absolute binding free energy (ABFE) calculation on zinc metalloenzymes stress-tests fixed-charge zinc force fields such as ZAFF-AMBER. We benchmark the ZAFF-AMBER + GAFF-2 + AM1-BCC ABFE protocol against experimental affinity for fourteen MMP-1 inhibitors drawn from PubChem bioassay records (IC50 0.78 nM to 98 µM; eleven hydroxamates, two carboxylates, one phosphonate), each carrying its PubChem CID, assay AID and retrieval timestamp. The protocol over-binds every compound, by 0.6 to 58.1 kcal/mol, and the over-binding is governed by the chelation strength of the warhead rather than by the measured potency: a weak (50 µM) hydroxamate still over-binds by 43 kcal/mol while a mid-potency (77 nM) phosphonate over-binds by only 4 kcal/mol. The rank correlation is weak and imprecise (Spearman 0.587, 95% CI [0.041, 0.847]; Pearson 0.631) with no sign-flips. Within-run statistical precision (0.5 to 0.9 kcal/mol) is not accuracy. We also document two pipeline failure modes and their fixes (a ReplicaExchangeSampler swap-all deadlock, and a production NaN crash requiring an explicit warmup) and contrast ABFE single-compound instability with population-level xtb rank stability. We recommend that fixed-charge ZAFF-AMBER ABFE not be used to rank zinc-chelating metalloenzyme inhibitors without warhead-stratified error budgeting. All data, code and provenance are attached and released at https://github.com/crazat/genesis_medicine (Apache-2.0). Provenance note: this is a fresh accuracy benchmark on a real, PubChem-sourced panel. An earlier draft of this work used a fabricated calibration panel and was withdrawn before any deposit; none of its compound identities, potencies or accuracy claims are carried into this record.