plinder-org/plinder

Protein Ligand INteraction Dataset and Evaluation Resource

What it solves

PLINDER addresses the need for high-quality, annotated datasets to train and evaluate protein-ligand docking algorithms. It provides a standardized resource to move beyond simple memorization in structural biology by offering curated splits that minimize information leakage and support realistic inference scenarios like hit discovery and optimization.

How it works

PLINDER combines a massive dataset of over 400,000 protein-ligand interaction (PLI) systems with an automated curation pipeline that syncs with the Protein Data Bank (PDB). It provides over 750 annotations per system, including protein and ligand properties and similarity scores. The resource includes a Python package (plinder.core) for easy data interaction and a data generation package (plinder.data). It also links unbound (apo) and predicted AlphaFold2 structures to their corresponding holo systems to help models handle realistic structural variations.

Who it’s for

This resource is designed for researchers and developers working on machine learning for structural biology, specifically those building protein-ligand docking models, co-folding algorithms, or tools for generating novel ligand structures.

Highlights

  • Massive Scale: Over 400k PLI systems across 11k SCOP domains and 50k unique small molecules.
  • Robust Benchmarking: Includes curated train-val-test splits stratified by novel ligands, pockets, or proteins to prevent data leakage.
  • Comprehensive Annotations: 750+ annotations per system and 14 different PLI metrics.
  • Integrated Structures: Links experimental apo structures and AlphaFold2 predictions to holo systems.
  • Evaluation Harness: Provides a dedicated framework to simplify and standardize performance comparisons between different models.

Related

  • Project
    thinking-machines-lab/tinker-cookbook

    A library of abstractions and recipes for fine-tuning language models via the Tinker API, supporting SFT, RL, and multimodal training.

  • Project
    google-deepmind/alphafold3

    An implementation of the AlphaFold 3 inference pipeline for the accurate 3D structure prediction of biomolecular interactions.

  • Project
    sokrypton/ColabFold

    A tool that makes protein folding predictions accessible via Google Colab and local installations, integrating models like AlphaFold2 and AlphaFold3 to predict 3D protein structures.

  • Dispatch
    Deep Learning with Proteins – Hugging Face guide to protein language models and folding

    Hugging Face announced a tutorial series showing how to fine‑tune protein language models and use ESMFold for protein folding, demonstrating that transfer learning techniques from NLP can be applied directly to protein sequence tasks.

  • Project
    RosettaCommons/foundry

    A tooling and infrastructure framework for protein design, providing a unified interface for generative models like RFdiffusion3 and structure prediction models like RosettaFold3.