openadmet // structure track // pose inspector

Right place, wrong way round

Caffeine bound to CYP3A4. Below: the crystal structure and two predictions of it — same molecule, same model, same settings, same batch. One is essentially correct, one is wrong, and the model cannot tell you which.

detail

plain language, no jargon

overlay · drag to rotate

Move

crystal ▸ prediction · 0%

Measured

crystal best worst heme Fe

Both predictions put caffeine in the right pocket. Their centres sit 0.38 Å and 0.53 Å from the crystal — a rounding error apart. One scores 0.908 and the other 0.320, because the molecule is turned the wrong way inside the site. Generation is not the bottleneck; choosing between what was generated is. Boltz's own confidence ranks these at chance (within-ligand ρ = −0.033), so picking the most confident pose is worse than random.

Why this pocket defeats co-folding

6ththe only free coordination site Iron already holds four pyrrole nitrogens and the Cys442 thiolate. Ligand density on the proximal face is physically impossible, whatever the model reports.

2.20 Åmedian Fe–donor in crystals 1.94 / 2.20 / 2.38 Å at p5/p50/p95 over 116 deposited CYP3A4 entries. 83 of 101 unique ligands coordinate the iron directly.

F/Gthe loop that moves OpenADMET's cryoEM work names F/G-loop remodelling as why co-folding does poorly here — the region left unmodelled in many X-ray structures.

Type II inhibition, in one paragraph

An sp2 nitrogen — imidazole, triazole, pyridine — donates its lone pair into the iron's empty d_z² orbital at about 2.0–2.3 Å. A dative bond with a hard directional requirement, not a soft contact, and the single most CYP-specific signal available to a scorer. Caffeine is not a Type II binder, which is why its closest approach here is 3.3–3.7 Å rather than 2.2 Å.

What we tried at the iron, and why it failed

Forcing the Fe→donor contact during generation is a null: oracle −0.0023, selected −0.0278. About 84% of poses already land inside any reasonable coordination window, so an acceptance test on it is nearly inert — it moved selection by −0.0002. Coordination is real signal as a scorer (+0.14 LDDT-PLI for coordinating poses) and useless as a filter. On the challenge compounds, though, only 41% / 28.5% coordinate — there it has real variance, which is the opposite of what our own crystal-set finding would have predicted.

Campaign record

Sixteen findings, most of them negative. The noise floor here is +0.0138; anything under +0.020 is indistinguishable from a random feature. Verified against a formal multiple-comparisons correction — the two routes agree to 0.0008.

#	finding	number	status

Ideas borrowed from other fields