xX-its-amit-Xx/pxr-structure-pose-pool

PXR Structure Challenge — Full Multi-Model Pose Pool (184 ligands)

Every protein–ligand pose generated during the OpenADMET PXR (pregnane X receptor / NR1I2) structure-prediction challenge, released openly with per-pose labels so the community can reuse the compute already spent — and, we hope, crack the problem this data makes visible.

What's here

• poses/<model>/<SID>.pdb — one best pose per (model, ligand). Protein chain A + ligand (resname LIG). 15 models, up to 184 ligands each (2,246 poses total). Filenames use an uppercase leading X (e.g. X00011-1.pdb); the canonical ligand id in manifest.csv is lowercase (x00011-1). Map with sid = filename.lower(). (The Hub rejects paths beginning x00…, which it reads as a null-byte escape — hence the uppercase.)
• manifest.csv — the labels, one row per pose:

column	meaning
sid	ligand id (xNNNNN-1)
model	predictor (boltz1, boltz1x, of3ft, protenix_v2, decaf, esm2_full/shallow/none, esmfold2fast, rf3, apo, af3_iface, af3_massive, chai, restr)
smiles	ligand SMILES
population	fragment (73) or drug_like analog (111)
series	congeneric series id
mw, clogp, tpsa, rotb, heavy	ligand descriptors
pEC50	measured PXR activation where available
lig_plddt	model confidence (ligand pLDDT surrogate) where reported
rmsd_to_medoid	this pose's distance from the cross-model consensus (A)
disagreement_mean_pw_rmsd	ligand-level cross-model disagreement (A)
consensus_frac, medoid_model, n_models	consensus metadata

The open problem this data exists for

Post-hoc analysis showed the pool contains sub-2 A poses for most ligands, but no confidence signal — and not even a GT-trained ML model on physics features — can reliably select them (the "selection wall"). The full autopsy, including an interactive 3D viewer, is at the project's GitHub Pages report. If your method can learn to pick the good pose already sitting in this pool, that is exactly the gap this release targets.

Ground truth

No crystal structures were released for the 184 blind test ligands. RMSD-style labels here are cross-model consensus geometry, not crystal RMSD. For true-RMSD work, use the homologous PXR holo crystal panel (also in the project repo).

Acknowledgment — LatchBio 💚

The final leg of this campaign, including additional AlphaFold3 pose sampling, was sponsored by LatchBio ($500 compute credits). Those poses deepened the pool exactly where models struggled most and are included in this open release. Thank you, LatchBio, for backing open, reproducible structural science.

Citation

OpenADMET PXR Structure Challenge — multi-model pose pool. Generated with Boltz-1/2, OpenFold3, AlphaFold3, Chai, Protenix v2, ESMFold2, RoseTTAFold3, apo templates. Released CC-BY-4.0.