plinder-org/plinder

plinder – Protein & Ligand Interaction Dataset & Evaluation Resource

What it is – plinder is a curated, large‑scale dataset of protein‑ligand (and protein‑protein) complexes together with rich annotations (ligand chemistry, binding‑site similarity, clusters, binding‑affinity data, etc.). It is packaged as a Python library that lets you download a specific monthly release, query the tables, and load individual systems as biophysical objects (structures, RDKit molecules, etc.). The repo also ships an evaluation harness for benchmarking structure‑prediction or docking models.

Why it matters – Training or testing modern ML models for drug‑discovery‑related tasks (docking, binding‑affinity prediction, co‑folding, generative ligand design) needs high‑quality, non‑redundant interaction data. Existing datasets are smaller or have hidden redundancy. PLINDER provides >400 k interactions, systematic de‑duplication, and carefully stratified train/validation/test splits that minimise information leakage, making results more trustworthy.

Key components

• plinder.data – code that builds a release from the RCSB PDB, applies the curation pipeline, and writes index tables, cluster assignments, and optional large files (ligand coordinates, alignments, scores).
• plinder.core – a user‑friendly API (PlinderSystem, Entry, Chain, etc.) that loads a system and gives you:
  • Biotite AtomArray for the receptor,
  • RDKit Mol objects for ligands,
  • pre‑computed similarity clusters for pockets, ligands, and proteins.
• Evaluation harness – optional extra dependencies (plinder[eval]) add PoseBusters validation and OpenStructure‑backed metrics (lDDT, RMSD) for comparing predicted structures against the reference.
• Versioning – dataset releases are month‑stamped (e.g., 2026‑09) and can be fetched with the CLI plinder_download. The library caches files locally and validates them, retrying interrupted downloads.

How to get started

1. Install the package: pip install plinder (add [eval] for evaluation tools).
2. Pull the index for a release: plinder_download --release 2026-09.
3. Load a system:

   from plinder.core import PlinderSystem
   sys = PlinderSystem.from_id('PDBID_LIGAND')
   receptor = sys.receptor_structure   # Biotite AtomArray
   ligands  = sys.ligand_mols          # list of RDKit Mol objects
   

4. Use the provided split tables to train models or run the evaluation scripts on your predictions.

Who’s behind it – A collaboration between the University of Basel, SIB Swiss Institute of Bioinformatics, Proxima (formerly VantAI), NVIDIA, and MIT CSAIL, with regular updates and a public Discord community.

License & citation – Data are Apache‑2.0; ligand/affinity sources retain their original CC‑BY‑4.0 or CC‑BY‑SA‑4.0 licenses. Cite the bioRxiv preprint (2024) when using the dataset.