Co-Folding Confidence Dispersion Is a Coarse Binder Filter, Not an Affinity Ranker: Boltz-2 sigma-iptm Evaluated Against Potency Rank, Same-Chemotype Weak Binders, and Property-Matched Decoys on Matrix Metalloproteinase-1

Abstract

The dispersion of co-folding confidence across an inference ensemble is widely used to triage protein-ligand predictions, but what it measures has not been fixed against experiment. This work evaluates Boltz-2 sigma-iptm, the standard deviation of iptm over 100 diffusion samples, on the Zn(II) metalloprotease MMP-1 across three cohorts of increasing negative-distance, all co-folded under one protocol. Among 93 ChEMBL actives spanning 3.62 log units of IC50 it does not rank potency: the Spearman correlation with the model's own potency-rank error is -0.005 (95% CI [-0.21, +0.20]), which bounds a small effect rather than excluding one. Among 121 PubChem compounds spanning 5.10 log units, 36 of them weaker than 10 micromolar, it separates the arms at every threshold tested (AUC 0.68 to 0.74, all CIs excluding 0.5; Spearman with pIC50 -0.443, essentially unchanged at -0.440 after controlling for heavy-atom count). Against 90 ZINC decoys property-matched to 52 sub-micromolar actives on six descriptors yet topologically distinct (ECFP4 Tanimoto below 0.35), separation is near-perfect (AUC 0.992 [0.979, 1.000]; heavy-atom-residual 0.990). The resulting gradient, null then 0.68 to 0.74 then 0.99, is monotonic in how different the negatives are, which is the signature of a coarse confidence filter rather than of an affinity model. The energetic analog computed on identical poses, pose-ensemble sigma_E under GFN2-xTB and Orb-v3 OMol25, is weaker at every level despite the two potentials agreeing at r = +0.982. The measurement basis is a systematic 161-reseed audit of a 15-ligand panel that resolves per-ligand sigma and surfaces two orthogonal outlier classes in double dissociation: one ligand carries the panel's largest ligand-energy dispersion while ranking fifth of fifteen on confidence dispersion, and another is the largest confidence outlier while its energy dispersion is mid-panel. The framework is offered as a detection protocol; the upstream interventions it proposes for either axis are not validated in this work. sigma-iptm is therefore licensed to prioritise a screening queue, between a floor near 0.68 against hard same-chemotype negatives and a ceiling near 0.99 against easy off-chemotype ones, and is not a potency ranker among binders. All cohort data, analysis code and persisted outputs are attached; the code repository is https://github.com/crazat/genesis_medicine (Apache-2.0). Provenance note: the 15-ligand reseed panel is used for structural reproducibility only. No compound identity or potency claim is made for any of its members, following the 2026-07-16 primary-source audit that withdrew a fabricated calibration panel from this project. Every potency figure in this record comes from the PubChem-sourced and ChEMBL-sourced cohorts, each row of which carries its own source identifier and retrieval timestamp in the attached files.