license: mit
language:
- en
tags:
- chemistry
- biology
- medical
pretty_name: HuggingLigand Embedding Dataset
datasets:
- RSE-Group11/Hugging-Ligand-Embeddings
code_repository: https://codebase.helmholtz.cloud/tud-rse-pojects-2025/group-11
HuggingLigand Dataset
Overview
HuggingLigand is a deep learning pipeline developed to predict the binding affinity between proteins and ligands. This prediction task is essential in fields such as drug discovery, biophysics, and computational biology, where determining how strongly a small molecule ligand binds to a protein target is a key step in understanding molecular interactions and prioritizing drug candidates.
The dataset provides precomputed embeddings for proteins and ligands, enabling efficient training and testing of machine learning models for binding affinity prediction.
Embedding Models Used
- BindingDB: A Dataset contains proteins, ligands, and their affinities.
- ProtT5: A transformer-based protein language model pretrained on millions of protein sequences.
- ChemBERTa: A transformer-based molecular language model trained on SMILES representations of chemical compounds.
Both models generate high-dimensional embeddings from raw sequence and molecular data. These embeddings are concatenated and used as input to a customizable regression model, which predicts continuous binding affinity values (e.g., Kd, Ki, or IC50).
Applications
- Structure-free binding affinity prediction
- Virtual screening and hit prioritization in drug discovery
- Data-driven biophysical modeling
- Benchmarking molecular embedding models
Dataset Structure
The dataset contains:
- Protein embeddings generated by ProtT5
- Ligand embeddings generated by ChemBERTa
- Binding affinity labels (e.g., Kd, Ki, IC50)
Citation
If you use this dataset, please cite: