CheapNet: Cross-attention on Hierarchical representations for Efficient protein-ligand binding Affinity Prediction

This is the official repository for "CheapNet: Cross-attention on Hierarchical representations for Efficient protein-ligand binding Affinity Prediction".

• Paper
• Project Page
• OpenReview

We propose CheapNet, a novel interaction-based model that integrates atom-level representations with hierarchical cluster-level interactions through a cross-attention mechanism. By employing differentiable pooling of atom-level embeddings, CheapNet efficiently captures essential higher-order molecular representations crucial for accurate binding predictions. Extensive evaluations demonstrate that CheapNet not only achieves state-of-the-art performance across multiple binding affinity prediction tasks but also maintains prediction accuracy with reasonable computational efficiency.

CheapNet: Cross-attention on Hierarchical representations for Efficient protein-ligand binding Affinity Prediction
Hyukjun Lim, Sun Kim, and Sangseon Lee† († indicates corresponding author)
Published in The Thirteenth International Conference on Learning Representations, 2025. (ICLR 2025)

Overview of the Architecture

Key Contributions

• We propose a hierarchical model that integrates atom-level and cluster-level interactions, improving the representation of protein-ligand complexes.
• Our model incorporates a cross-attention mechanism between protein and ligand clusters, focusing on biologically relevant binding interactions.
• CheapNet achieves state-of-the-art performance across multiple binding affinity prediction tasks while maintaining computational efficiency.

Key Features

• Hierarchical Representations: Higher-level representations to enhance the understanding of protein-ligand interactions that act as groups.
• Cross-Attention Mechanism: Leverages cross-attention to capture significant interactions between protein and ligand clusters.
• Performance: Achieves high accuracy of protein-ligand binding affinity prediction across various datasets.
• Efficiency: Designed to be memory-efficient, requiring minimal memory and computation compared to other attention-based models.

Installation

To install and use CheapNet, follow these steps:

1. Clone the repository:

   git clone https://github.com/hyukjunlim/CheapNet.git
   cd CheapNet

2. Set up the environment:

   # For Cross-dataset Evaluation
   conda env create -f cheapcross.yaml
   conda activate cheapcross
   
   # For Diverse Protein Evaluation, and LEP
   conda env create -f cheapdivlep.yaml
   conda activate cheapdivlep

Dataset

The original dataset can be found at GIGN, and ATOM3D.
If you want to download the preprocessed datasets, run the following commands:

# For Cross-dataset Evaluation
cd cross_dataset
gdown 1wuLBRgPUSshmhE33UigLTJ_GsDyayXhX
unzip cross_dataset.zip
rm -f cross_dataset.zip
cd ../

# For Diverse Protein Evaluation
cd diverse_protein
gdown 1HxKLtCX3VCuHfXXWj1KlMB-S0PXQ_Hjl
unzip diverse_protein.zip
rm -f diverse_protein.zip
cd ../

# For LEP
cd lep
gdown 1_3FJo4eWm7IHLQpxdUuRSFYG_sM5tQm1
unzip LEP.zip
rm -f LEP.zip
cd ../

Usage

Training the Model

To train CheapNet for each dataset, use the train.py script:

# For Cross-dataset Evaluation
python train.py

# For Diverse Protein Evaluation
## LBA 30%
python train.py --seqid 30 --num_epochs 15 --learning_rate 1.5e-3 --use_scheduler 0 --data_dir $LMDBDIR

## LBA 60%
python train.py --seqid 60 --learning_rate 1e-3 --data_dir $LMDBDIR 

# For LEP
python train.py --learning_rate 15e-4 --data_dir $LMDBDIR

Prediction

To predict the binding affinity of protein-ligand complexes, use the predict.py script:

# For Cross-dataset Evaluation
python predict.py
python predict_casf.py

# For Diverse Protein Evaluation
## LBA 30%
python predict.py --seqid 30

## LBA 60%
python predict.py --seqid 60

# For LEP
python predict.py

Evaluation

To evaluate the model's performance, use the evaulate.py script:

python evaluate.py

Citation

If the code have been helpful in your research, please cite CheapNet:

@inproceedings{
    lim2025cheapnet,
    title={CheapNet: Cross-attention on Hierarchical representations for Efficient protein-ligand binding Affinity Prediction},
    author={Hyukjun Lim and Sun Kim and Sangseon Lee},
    booktitle={The Thirteenth International Conference on Learning Representations},
    year={2025},
    url={https://openreview.net/forum?id=A1HhtITVEi}
}