Pyrite: A Benchmark for Protein-Ligand Pose Ranking Agents

Pyrite is a pairwise pose-ranking benchmark. Each item is a pair of docked poses — Pose A and Pose B — of the same protein–ligand complex, and the task is to pick the pose closer to the experimental (crystallographic) binding mode. The ground-truth label is the closer pose.

• 750 pose pairs over 177 crystal structures / 120 proteins, all from public RCSB PDB entries (177 unique PDB IDs).
• Each pose carries its 3D ligand coordinates, a protein–ligand interaction fingerprint, and physics/geometry descriptors (MM-GBSA energy, Mogul torsion validity, torsion strain, PoseBusters validity, pose-shape descriptors).
• Each pair is annotated for temporal split (released_after_af3_cutoff), difficulty (pair_kind), UniProt family, ligand metadata, and water/ion context.
• A reasoning-LLM baseline (Qwen3.8-27B-GRPO) ships with its per-pair predictions.

Scoring convention used in the accompanying work: mean per-structure accuracy — per-pair correctness averaged within each structure_id, then averaged across structures (equal weight per structure), reported overall and split by released_after_af3_cutoff.

Label leakage: columns computed against the crystal structure (metric_rmsd_*, pose_rmsds_*, metric_rmsd_lt*, metric_lddt_pli_*, metric_gt_interaction_*, gt_frac/a_frac/b_frac, and correct/MM_correct) encode the answer. Use them as labels/analysis only — never as model inputs.

Files

• data/pyrite_v6.1_curated.csv / .parquet — the benchmark: 750 pairs × 113 columns.
• data/pyrite_grpo_predictions.csv / .parquet — Qwen3.8-27B-GRPO predictions per pair.
• viewer.html — self-contained interactive viewer (see below).

Columns suffixed _A / _B give the value for Pose A / Pose B respectively.

Interactive viewer

viewer.html is a fully self-contained, offline browser viewer to aid visualization of the benchmark — just open it in any modern browser (3Dmol.js and all data are embedded; no network needed). For each of the 750 pairs it renders Pose A and Pose B side by side in 3D: the ligand pose as sticks inside a lite binding pocket (only pocket residues within ~6 Å of the ligand, drawn as a transparent cartoon so the poses stay in focus). It shows the Qwen3.8-27B-GRPO prediction and its full reasoning trace for each pair.

The ground truth is hidden by default — use the Show ground truth toggle to reveal which pose is crystal-closer (RMSDs to the crystal pose, which pane is correct, and whether GRPO's pick was right). Toggling it also overlays the experimental crystal ligand in green in both panes, so you can see directly how each docked pose compares to the reference. This lets you inspect the model's reasoning blind, then check the answer.

pyrite_v6.1_curated columns

Identifiers & label

• pair_id — unique pose-pair identifier.
• structure_id — crystal-structure identifier (<PDB>_<ligand code>).
• correct — ground-truth label: which pose (A/B) is closer to the crystal pose.
• pair_kind — difficulty: good_bad (one clearly better pose) or good_maybe (both plausible).
• A_pose_num, B_pose_num — index of each pose within the generated pose set.

Ligand & complex metadata

• pdb_id — source RCSB PDB entry.
• ligand_soi — ligand of interest (chain:resname:instance).
• num_ligands_in_structure — number of ligands in the PDB entry.
• canon_smiles — canonical SMILES of the ligand.
• is_drug_like — boolean drug-likeness flag.
• uniprot_id, uniprot_name, uniprot_description — target UniProt accession, name, and description.
• uniprot_family — UniProt "belongs to …" family (or Pfam fallback); grouping key above protein.

Temporal split & provenance

• deposition_date, release_date — PDB deposition / release dates.
• released_after_af3_cutoff — boolean; True = released after the AlphaFold3 training cutoff (held-out / novel-complex split).

3D structure & interactions

• pose_A, pose_B — 3D ligand coordinates (atoms, elements, and x/y/z), compact text format.
• pose_interactions_A, pose_interactions_B — protein–ligand interaction fingerprint, ;-separated RESIDUE-{BB,SC}-<type> tokens (HBD/HBA, SALT, PI_STACK, T_STACK, …).

MM-GBSA energy

• mm_step1_binding_energy_A, mm_step1_binding_energy_B — MM-GBSA binding free energy (lower = more favorable).
• mm_gap — absolute MM-GBSA energy difference between the two poses, |A − B|.
• MM_correct — 1 if the lower-MM-GBSA-energy pose is the crystal-closer one (label-derived convenience flag).

Pose accuracy vs crystal (LABEL-LEAKING — analysis only)

• metric_rmsd_A, metric_rmsd_B — ligand RMSD to the crystal pose (Å).
• pose_rmsds_A, pose_rmsds_B — ligand RMSD to the crystal pose (Å; alternate computation).
• metric_rmsd_lt2_A/B, metric_rmsd_lt1_A/B — indicator: pose RMSD < 2 Å / < 1 Å.
• metric_pb_valid_A, metric_pb_valid_B — PoseBusters validity (1 = passes all physical checks).
• metric_rmsd_lt2_pb_valid_A/B, metric_rmsd_lt1_pb_valid_A/B — pose is < 2 Å / < 1 Å and PoseBusters-valid.
• metric_lddt_pli_A, metric_lddt_pli_B — lDDT-PLI (protein–ligand interaction lDDT) vs crystal.
• metric_gt_interaction_recall_A/B, metric_gt_interaction_iou_A/B — recall / IoU of the pose's interactions vs the crystal interaction set.
• metric_gt_interaction_conserved_count_A/B, metric_gt_interaction_missing_count_A/B, metric_gt_interaction_extra_count_A/B, metric_gt_interaction_count_A/B — conserved / missing / extra / total interaction counts vs crystal.

Electron-density support — 2Fo-Fc map (LABEL-LEAKING — analysis only)

Fractions of a ligand's heavy atoms that sit in strong experimental electron density, scored against the entry's 2Fo-Fc map (map value z-scored over the whole map, threshold ≥ 1.0 σ). The better-density-supported pose is the crystal-closer one, so these leak the label.

• gt_frac — density-support fraction of the deposited crystal ligand (reference-quality check).
• a_frac, b_frac — density-support fraction of Pose A / Pose B.
• map_available — whether an experimental 2Fo-Fc density map (<pdb>.ccp4) was retrievable (yes/no/error). Constant yes here — the curated set keeps only entries with a usable map.
• C1, C2, C3 — density-based exclusion flags used to build the set (all False here; every pair passes). C1 = gt_frac < 0.75 (deposited ligand poorly supported by its own density); C2 = correct-pose density fraction < 0.50; C3 = |correct − wrong density fraction| < 0.10 (density cannot discriminate the two poses).

Pose-intrinsic validity & strain (usable as inputs)

• metric_pb_valid_A/B — see above (PoseBusters validity is pose-intrinsic, not crystal-derived).
• metric_mogul_num_unusual_torsions_A/B, metric_mogul_percent_unusual_torsions_A/B, metric_mogul_total_torsions_A/B — Mogul torsion-outlier count, percent, and total torsions checked.
• metric_mogul_torsion_penalty_A/B — aggregate Mogul torsion-geometry penalty.
• metric_mogul_num_unusual_rings_A/B — Mogul ring-geometry outlier count.
• metriclog_mogul_unusual_torsions_details_A/B — per-torsion Mogul detail records (JSON).
• metric_total_backbone_torsion_strain_A/B, metric_average_backbone_torsion_strain_A/B — ligand backbone torsion strain (total / per-torsion).
• metric_total_sidechain_torsion_strain_A/B, metric_average_sidechain_torsion_strain_A/B — ligand sidechain torsion strain (total / per-torsion).
• metric_total_torsion_strain_A/B, metric_average_total_torsion_strain_A/B — total ligand torsion strain (total / per-torsion).

Pose shape descriptors (usable as inputs)

• metric_geom_Rg_A/B — radius of gyration.
• metric_geom_asph_A/B — asphericity.
• metric_geom_extent_A/B — spatial extent.
• metric_geom_n_close_A/B — intramolecular close-contact count.
• metric_geom_n_atoms_A/B — ligand heavy-atom count.
• metric_geom_n_res_A/B, metric_geom_n_sc_res_A/B — protein residues / sidechain residues in contact.

Pose-pair comparison (A vs B)

• pose_overlap_shape, pose_overlap_color, pose_overlap_combo — shape / pharmacophore(color) / combined overlap between Pose A and Pose B.
• protein_rmsd_AB — protein RMSD between the two poses' complexes (induced-fit / flexibility proxy).
• ligand_rmsd_AB — ligand RMSD between Pose A and Pose B.
• ligand_max_atom_disp_AB — maximum single-atom displacement between A and B (Å).
• ligand_n_atoms_disp_ge2_AB — number of ligand atoms displaced ≥ 2 Å between A and B.

Water / ion context

• water_class — protein_ligand_water (crystallographic water bridges protein–ligand) or no_water.
• missing_ion_coordinating — boolean: a deposited metal/halide ion within 3.0 Å (inner-sphere coordination) of the crystal ligand is missing/under-represented in the reference structure. missing_ion_nearby — same, within 6.0 Å.
• missing_ion_elems — element(s) of any such missing ion.
• missing_ion_min_dist — distance to the nearest missing ion (Å; −1 if none).
• n_ions_deposited_near_lig, n_ions_reference — ion counts near the ligand / in the reference structure.

pyrite_grpo_predictions columns

• pair_id — matches pyrite_v6.1_curated.pair_id.
• structure_id — crystal-structure identifier.
• correct — ground-truth label (A/B).
• released_after_af3_cutoff — temporal-split flag (see above).
• grpo_pred — the model's predicted pose (A/B).
• grpo_correct — 1 if grpo_pred == correct.

Reference model: Qwen3.8-27B-GRPO

A reasoning LLM (Qwen3.8-27B) RL-tuned with GRPO to pick the crystal-closer pose from the tabular features and interaction fingerprints (no raw coordinates).

metric	value
per-pair (pooled) accuracy	0.669
mean per-structure accuracy — overall	0.611
mean per-structure accuracy — pre-AF3	0.575
mean per-structure accuracy — post-AF3 (held-out / novel complexes)	0.687

Provenance & license

All structures are public RCSB PDB entries (see pdb_id). Poses were generated by cofolding/docking models; the metrics are computed descriptors. No proprietary or partner data is included. Released under CC BY 4.0 (see LICENSE); please also cite the relevant PDB entries where appropriate.