Dataset Card for BALM-Benchmark

BALM-Benchmark is a curated collection of datasets designed to advance machine learning and deep learning model research for protein-ligand binding affinity prediction. This benchmark consolidates several key datasets including BindingDB, LP-PDBBind, and specific protein-ligand systems like USP7, MPro, SYK, HIF2A, and MCL1, each chosen for its distinct data characteristics and evaluation.

This dataset collection has been refined and standardized, making it readily accessible for deep learning model training and testing on Hugging Face, providing a structured foundation for advancements in target-based drug discovery.

• Dataset Repository: https://huggingface.co/datasets/BALM/BALM-benchmark
• Code Repository: https://github.com/meyresearch/BALM
• Paper: https://www.biorxiv.org/content/10.1101/2024.11.01.621495v1
• License: CC-BY-4.0

Dataset Details

To benchmark our models, we utilized several publicaly available datasets, encompassing diverse protein-ligand interactions and binding affinity values. Key datasets include BindingDB (1D data with protein sequnces and SMILES), LP-PDBBind (containing 3D complexes), and other target-specific datasets such as USP7, MPro, and three targets from the protein-ligand free energy benchmark (SYK, HIF2A, and MCL1). These datasets capture a wide range of binding affinity measurements, allowing us to evaluate and compare model performance against traditional docking and free energy methods. All datasets have been meticulously cleaned and are available on Hugging Face as BALM-Benchmark.

BindingDB

BindingDB provides experimental binding affinity data (Kd values) for protein-ligand interactions. We focused on K_d values due to inconsistencies in other affinity types. After filtering for computational efficiency and data consistency, the dataset comprises around 25,000 interactions with ~1,070 unique targets and 9,200 ligands. We implemented four data splits (Random, Cold Target, Cold Drug, and Scaffold) to evaluate generalizability on test set with splits based on unseen proteins, ligands and ligand scaffolds, guided by the Murcko scaffold approach.

LP-PDBBind

Derived from PDBBind v2020, LP-PDBBind is a curated collection of ~20,000 protein-ligand structures with experimental binding data. This dataset was reorganized to reduce similarity across splits and cleaned to remove covalently bound ligands and rare atomic elements. To ensure model reliability, we used Clean Level 1 (CL1) for training and the higher-quality CL2 data for validation and testing as recomended here. Here we provide 1D data, for 3D complexes please download from here.

USP7

The USP7 dataset, developed by Shen et al., contains binding data for USP7 inhibitors from ChEMBL. After processing to remove assay limits, it includes 1,799 ligands with experimentally measured affinities, provided as IC50 values and converted to pIC50 for consistency.

MPro

Collected as part of the COVID Moonshot project, the MPro dataset focuses on inhibitors targeting the SARS-CoV-2 main protease. The final cleaned dataset includes 2,062 ligands with IC50 values, converted to pIC50 for stability in training.

Protein-Ligand Free Energy Benchmark

Selected from the protein-ligand free energy benchmark by Hahn et al. with 21 target systems, we selected three targets to evaluate the deep learning model: MCL1, HIF2A, and SYK. These targets offer diverse interactions, allowing for robust comparison with alchemical free energy methods. The datasets contain 37, 25, and 43 ligands, respectively, for benchmarking model predictions against established free energy methods.

Dataset Columns

• BindingDB_filtered:
  • Index (string): Index of the ligand-target pair.
  • Drug_ID (string): Index of the ligand from the TDC.
  • Drug (string): Ligand sequence (i.e., SMILES string).
  • Target_ID (string): Index of the target protein from the TDC.
  • Target (string): Protein sequence (i.e., sequence of amino acids).
  • Y (float32): binding affinity value in pKd.
• Mpro:
  • Index (string): Index of the ligand-target pair.
  • Y (float32): binding affinity value in pIC50.
  • Drug (string): Ligand sequence (i.e., SMILES string).
  • Target (string): Protein sequence (i.e., sequence of amino acids).
• USP7:
  • Index (string): Index of the ligand-target pair.
  • Y (float32): binding affinity value in pIC50.
  • Drug (string): Ligand sequence (i.e., SMILES string).
  • Target (string): Protein sequence (i.e., sequence of amino acids).
• LeakyPDB:
  • Index (string): Identifier for each ligand-target pair in the dataset.
  • pdb_id (string): Unique identifier for the protein structure in the Protein Data Bank (PDB).
  • Drug (string): SMILES string representing the ligand's chemical structure.
  • category (string): Classification category for the ligand-protein complex.
  • Target (string): Protein sequence, represented as a sequence of amino acids.
  • resolution (float32): Structural resolution of the protein-ligand complex, typically measured in angstroms.
  • date (string): Date of structural determination or deposition in the PDB.
  • type (string): Type or family classification of the protein.
  • new_split (string): Specifies the split category for the LP-PDBBind dataset.
  • CL1 (bool): Boolean indicating whether the complex belongs to Clean Level 1 (CL1) in the LP-PDBBind dataset.
  • CL2 (bool): Boolean indicating whether the complex belongs to Clean Level 2 (CL2) in the LP-PDBBind dataset.
  • CL3 (bool): Boolean indicating whether the complex belongs to Clean Level 3 (CL3) in the LP-PDBBind dataset.
  • remove_for_balancing_val (bool): Boolean indicating if the entry is excluded for balancing in validation sets.
  • kd/ki (string): Original binding affinity measurement (Kd or Ki) with units (uM or nM).
  • Y (float32): Binding affinity value provided in log scale (pKd).
  • covalent (bool): Boolean indicating if the ligand is covalently bound to the protein.
• HIF2A, MCL1, and SYK:
  • Index (string): Index of the ligand-target pair.
  • Y (float32): binding affinity value in pKi (for MCL1) and pIC50 (for HIF2A, and SYK).
  • Drug (string): Ligand sequence (i.e., SMILES string).
  • Target (string): Protein sequence (i.e., sequence of amino acids).

Dataset Sources

• BindingDB_filtered: Derived from Therapeutics Data Commons (TDC), with additional filtering and cleaning to enhance consistency and computational efficiency.
• LeakyPDB: Collected from the LP-PDBBind repository and described in this publication.
• HIF2A, MCL1, and SYK: Sourced from the protein-ligand benchmark dataset available on GitHub and detailed in the LiveCoMS journal.
• Mpro: Data for SARS-CoV-2 main protease (Mpro) inhibitors sourced from Science.
• USP7: Collected from ChEMBL and curated as described in this Journal of Cheminformatics article.

Uses

BALM-Benchmark was initially created as a part of the BALM project (https://github.com/meyresearch/BALM) which fine-tunes Protein and Ligand Language Models by optimizing the distance between protein and ligand embeddings in a shared space using the cosine similarity metric that directly represents experimental binding affinity. Nevertheless, BALM-Benchmark can be used by itself, just like any other HuggingFace dataset:

from datasets import load_dataset

# For instance, you want to load SYK data. Change the second argument into SYK
syk_data = load_dataset("BALM/BALM-benchmark", "SYK", split="train")

Notice that all datasets only have one split (train). This is intentional such that the users can define their own splits, and can experiment with more random seeds for robustness. We highly recommend checking out different strategies for splitting the data (e.g., BindingDB) in our BALM code repository.

Citation

BibTeX:

@article{Gorantla2024,
  author = {Gorantla, Rohan and Gema, Aryo Pradipta and Yang, Ian Xi and Serrano-Morr{\'a}s, {\'A}lvaro and Suutari, Benjamin and Jim{\'e}nez, Jordi Ju{\'a}rez and Mey, Antonia S. J. S.},
  title = {Learning Binding Affinities via Fine-tuning of Protein and Ligand Language Models},
  year = {2024},
  doi = {10.1101/2024.11.01.621495},
  publisher = {Cold Spring Harbor Laboratory},
  journal = {bioRxiv}
}

Dataset Card Contact

• Rohan Gorantla (rohan.gorantla@ed.ac.uk)
• Aryo Pradipta Gema (aryo.gema@ed.ac.uk)
• Antonia Mey (antonia.mey@ed.ac.uk)