PLI-Parallax

Predicted protein-ligand complexes are used as training data at considerable scale today. Recent distillation sets contain several hundred thousand cofolded BindingDB systems, and the filter applied to them is usually the predicting model's own confidence score. That filter is a self-assessment, and cofolding models have been shown to be confidently wrong in ways their own confidence does not reveal.

This dataset provides the coordinates and derived distance labels of four structure-prediction teacher configurations on the same systems, together with the experimental structures on the crystal tier and annotations describing how the teachers relate to them. Agreement between separately built methods is a different signal from any one model's confidence in itself, and where experimental structures exist it can be calibrated against them.

The four teacher modes cover a current range of approaches. Two are cofolding models, one of them in two configurations, once from the single sequence and once with a multiple-sequence alignment. The fourth is rigid docking into a predicted receptor. Their predictions coincide closely on some systems and diverge widely on others, and both outcomes are recorded rather than resolved.

Dataset structure

Two tiers, five teacher slots, and 3D coordinates as the primary artifact.

tier	systems	ground truth	receptor
corpus	31,746	none by construction	AlphaFold model, one accession
crystal	19,350	experimental	deposited asymmetric unit

Teachers are Chai-1, Boltz-2 single-sequence, Boltz-2 with a ColabFold MSA, and smina docking, plus the experimental structure on the crystal tier.

• stores/corpus.h5, stores/crystal.h5 hold per-system coordinates: an alpha carbon, a side-chain centroid and a backbone carbonyl carbon for every residue, plus all receptor heavy atoms within 15 angstrom of the ligand, plus every teacher's ligand pose. This is the artifact everything else derives from.
• labels/, 18 Parquet files, 307 million rows. Residue-to-ligand-atom distances at a fixed 15 angstrom cutoff. A materialised view of the store, not an independent measurement.
• metadata/ carries the calibrated per-system reliability, per-residue teacher support, a ligand-independent pocket ensemble over 20,404 receptors, and the teacher-agreement geometry.
• splits/, 646 leakage-controlled configurations across the Park and Marcotte pair-input classes, each with measured shift diagnostics rather than an asserted guarantee.

Quick start

from huggingface_hub import snapshot_download
import os, h5py, numpy as np, pyarrow.parquet as pq

D = os.path.join(snapshot_download("ThorKl/PLI-parallax", repo_type="dataset"),
                 "deposit")

f = h5py.File(f"{D}/stores/crystal.h5", "r")
g = f["10mh_SAH_A"]
cen, sc = g.attrs["centroid_xyz"], g.attrs["scale"]

lig = g["ligand_coords"][1, 0] / sc + cen        # chai1, pose 0
ca  = g["protein/chai1/ca"][:] / sc + cen        # alpha carbons
d   = np.linalg.norm(lig[:, None, :] - ca[None, :, :], axis=2)

A worked notebook covering loading, the derivation check, the three-bead representation and its visualisation is at notebooks/pli_parallax_usage_examples.ipynb.

Verify the download with cd deposit && sha256sum -c MANIFEST.sha256.

How the teachers relate to each other

metadata/teacher_agreement.parquet reports, for each system carrying two or more valid poses, how far apart those poses sit once the receptors are superposed. Superposition is not optional: each teacher's ligand lies in its own receptor frame, so comparing stored coordinates directly measures the frame offset rather than the pose.

Median pairwise ligand-centroid separation is 18.2 angstrom on the corpus tier and 11.5 angstrom on the crystal tier. Every teacher pair falls within 2 angstrom on 2.9 per cent of systems and within 5 angstrom on 9.4 per cent. On the corpus tier that is often a different binding site rather than a different pose, which is the expected outcome for non-cognate pairs where no teacher has an experimental structure to anchor on.

metadata/system_reliability.parquet turns the same information into a per-system estimate. An isotonic regression fitted on the crystal tier maps inter-teacher contact-set agreement onto measured accuracy against ground truth, and that fit is transferred to the corpus tier. Agreement correlates with accuracy at Pearson 0.99 on the crystal tier, and per-residue precision rises from 0.22 at one asserting teacher to 0.73 at two and 0.94 at three, over 147,966, 45,268 and 16,200 residues respectively.

Three things to know before using it

Contact rows are censored, not exhaustive. A row exists only where the minimum heavy-atom distance is at most 15 angstrom, which covers 23.5 per cent of a system's residues on average. An outer join filling zero produces a distance matrix that is wrong on most of its entries.

Coordinates need both decode terms. They are int16 at 0.01 angstrom about a per-system centroid. Omitting the centroid is silent and wrong by a median of 47 angstrom.

affinity is a docking score, not a measurement. No experimental binding affinity is deposited in this record, despite BindingDB being an upstream source of the corpus tier.

What it is not

Not an all-atom resource. The store holds three points per residue plus the heavy atoms of standard amino acids within 15 angstrom of the ligand, so side chains outside the shell, hydrogens and every non-standard residue are absent, and the shell itself contains only carbon, nitrogen, oxygen and sulfur. PoseBusters and OpenStructure cannot run on it, so the crystal tier is an index and annotation layer rather than a drop-in evaluation set for a cofolding method. Recent work has argued that all-atom resolution is unnecessary for pose and affinity prediction specifically, which is the setting this format is aimed at, but the limitation is real for anything that needs full geometry.

Four of the five teacher poses are predictions. Training against a cofolded or docked pose as though it were experimental is distillation from that method and inherits its failure modes.

The reliability annotation is a calibrated prediction rather than a measurement. The isotonic fit was made on the crystal tier and transferred to the corpus tier under an exchangeability assumption the two tiers do not satisfy exactly. 4,151 systems sit at the fit's floor of 0.1604, which is the regression's lower bound and not a measured accuracy, and the conformal interval is marginal and constant rather than per-system.

The smina teacher is a different construction from the three cofolders. It docks rigidly into a predicted receptor, so part of its disagreement with them reflects the receptor conformation it was given rather than the docking itself.

Relation to prior work

Multi-method predictions have been deposited before. PoseBench ships predicted poses from seven methods over four benchmark sets of a few hundred experimentally solved systems, and Runs N' Poses provides inputs, ground truth and per-method accuracy tables for four cofolding methods over 2,600 systems. Both are evaluation benchmarks built entirely on systems where the answer is known.

What this resource adds is the corpus tier, where no experimental structure exists and the agreement between teachers has to carry the signal on its own, together with the calibration that makes that agreement interpretable. SAIR and similar distillation sets cover comparable ground at larger scale but with a single predictor. CrossDocked2020 varies the receptor rather than the method. PLINDER remains the reference for split methodology, and the split families here follow its lead.

Documentation

deposit/README.md is the technical reference and is long. FIELDS.json and FIELDS.md define every column once, with unit, null convention and caveats. SPLITS.md covers what each split family licenses and where its guarantee stops. CHAIN_AUDIT.md documents an input-construction artefact that excluded 9,565 of 19,350 crystal systems from cofolding, of which 8,454 would have fitted under a per-chain rather than a fused length cap.

The field dictionary is machine-readable and is regenerated from a single source, so FIELDS.json, FIELDS.md, the README schema section and the Croissant descriptions cannot diverge.

Citation and licence

CC-BY-4.0. Cite the accompanying Data Descriptor and this record's version DOI: 10.5281/zenodo.21560088.

Upstream sources carry their own terms, listed in full in the deposit README. Downstream users must satisfy the terms of every source they rely on.