End-to-end protein-ligand structure prediction with affinity heads (Boltz-2, AlphaFold3-class cofolding) has become a routine first-pass tool for virtual screening, but the rate at which these models emit physically implausible poses — and how that rate depends on inference-time configuration — remains incompletely characterised. We evaluate six published cofold configurations on 15 ChEMBL MMP-1 zinc-hydroxamate active-site ligands (1500 cofold poses per condition; 9000 total), using GFN2-xTB single-point energies on the predicted ligand geometries as a physical-plausibility readout. Standard Boltz-2 without the --useₚotentials (Boltz-2x) steering flag emits catastrophic outliers (per-ligand population σ up to 14. 27 kcal mol⁻¹ on CHEMBL94487, with one pose reaching an implausible +8911 kcal mol⁻¹ relative energy). Enabling --useₚotentials reduces σ on the same ligand to 3. 18-4. 29 kcal mol⁻¹ across three independent seeds, with zero positive-energy outliers in 4500 samples (0. 022%, vs. 0. 188% without the flag, 9× reduction). A recently released community fork of Boltz-2 (Volgin et al. , March 2026) that fixes a silent wrong-answer bug, a metal-ion C-alpha filter, and a bfloat16 dtype path does not eliminate the outliers when run without the steering flag (σfiltered = 6. 66 kcal mol⁻¹; 2/100 catastrophic poses on the canary ligand). Adding --useₚotentials to the fork removes the outliers but the within-population precision remains ~2× worse than standard Boltz-2x (σfilt 6. 98 vs. 3. 18 kcal mol⁻¹). The operative factor is therefore the steering-potential flag, not the bug fixes. We recommend standard Boltz-2 + --useₚotentials as the canonical cofold protocol for protein-ligand affinity prediction, especially for metalloprotein active sites; a residual ≤0. 025% catastrophic-failure rate remains, so xtb-based filtering of cofold poses is still mandatory downstream.
Cheongwoo Han (Tue,) studied this question.