Structure184 five-method PXR cofolding dataset

This repository contains 92,000 PXR–ligand cofolded models:

• 184 ligands
• 20 seed positions
• 5 samples per seed
• Boltz2, Chai-1, ESMFold2, OpenFold3, and Protenix
• 18,400 models per method

Models use the governed prepared top solution-state ligand representation. Coordinates are harmonized to the 1NRL PXR frame using a 162-Cα core. The original generated coordinates are preserved up to that rigid alignment.

Repository contents

models.parquet has 119 columns. structure_archive and structure_member locate each CIF.

Load the tables

from datasets import load_dataset

models = load_dataset("dargason/structure184-five-method-cofolding", "models", split="models")
posebusters = load_dataset("dargason/structure184-five-method-cofolding", "posebusters_full", split="models")

Or with pandas:

import pandas as pd

models = pd.read_parquet("hf://datasets/dargason/structure184-five-method-cofolding/data/models.parquet")

Extract structures

tar --use-compress-program=unzstd -xf structures/boltz2.tar.zst

Archive members follow method/ligand_id/model_id.cif. Join them through structure_member.

PoseBusters

PoseBusters was recomputed from scratch using generated ligand coordinates, authoritative prepared-state topology, and generated receptor coordinates.

• Computed: 91,734
• Explicit topology quarantine: 266
• Passed all configured checks: 75,668
• Failed one or more configured checks: 16,066

Quarantined rows remain in models.parquet with posebusters_result_status=not_run_topology_quarantine.

Important interpretation notes

• PoseBusters is physical-validity QC, not pose-accuracy evidence.
• protein_ligand_iptm is the primary model-native confidence field; iptm_source_key records the engine-specific source.
• Columns beginning official_ost_, plus truth-relative RMSD and centroid-distance fields, are evaluation labels. Do not use them as blind model-selection inputs.
• ESMFold2 models are the campaign's no-MSA, 50-step protocol, not upstream-default ESMFold2.
• The public tables remove machine-local paths. File hashes and source-group identifiers are retained.
• This release does not claim that every engine parameter can be reconstructed from the public tables alone.

Source and attribution

The ligand set and challenge context come from the OpenADMET PXR Induction Blind Challenge and its public challenge dataset. Please cite the OpenADMET challenge and this dataset repository when using these models or derived tables.

Repository size

The packaged repository is approximately 4.6 GB compressed. The Parquet tables and .tar.zst structure archives are configured for Git LFS through .gitattributes.

License

The tables, generated/aligned structures, truth-relative metrics, aligned truth structures, packaging, and documentation in this repository are provided under CC BY 4.0. Model-generating software remains under its respective license. See LICENSE.