CheapNet: Cross-attention on Hierarchical representations for Efficient protein-ligand binding Affinity Prediction
Hyukjun Lim, Sun Kim, Sangseon Lee†
Inha University Inha University Logo
Motivation

Problem Domain:

  • Accurate protein-ligand binding affinity prediction for drug discovery

Existing Challenges:

  • Traditional lab methods are slow and costly
  • Existing models often rely solely on atom-level interactions, which suffer from:
    • Noise from irrelevant atoms & Computational inefficiency
    • Difficulty in determining meaningful atom clusters that drive interactions
    • Previous cluster-level approaches relied on predefined clusters
Protein-ligand interaction visualization
Figure 1: Illustration of protein-ligand complex and interaction patterns
Method Overview

CheapNet Architecture:

CheapNet architecture
Figure 2: CheapNet architecture with hierarchical representations and cross-attention

Key Components:

  1. Atom-Level Embedding:
    • Graph Neural Network learns atom-level embeddings
  2. Hierarchical Clustering:
    • Differentiable pooling learns soft assignment matrices
    • Extracts higher-level representations that act as groups based on the cluster-level embeddings
  3. Cross-Attention:
    • Applies cross-attention between protein and ligand cluster embeddings
    • Captures important interactions between protein and ligand clusters

Advantages:

  • Reduces noise by focusing on meaningful atom groups dynamically
  • Captures higher-order interactions through end-to-end training
  • Improves computational & memory efficiency
  • Enhanced interpretability through cluster-attention mechanism
Results

Ligand Binding Affinity (LBA) - Cross-dataset Evaluation:

Model Params # v2013 core set v2016 core set v2019 holdout set
RMSE ↓ Pearson ↑ RMSE ↓ Pearson ↑ RMSE ↓ Pearson ↑
IGN 1.66M 1.428 0.807 1.269 0.821 1.410 0.630
EGNN 1.59M 1.498 0.782 1.289 0.816 1.399 0.628
GIGN 0.62M 1.380 0.821 1.190 0.840 1.393 0.641
GAABind 17.95M 1.488 0.772 1.297 0.803 - -
DEAttentionDTA 2.32M 1.470 0.800 1.266 0.827 - -
CheapNet 1.33M 1.262 0.857 1.107 0.870 1.343 0.665

Ligand Binding Affinity (LBA) - Diverse Protein Evaluation:

Model Params # LBA 30% LBA 60%
RMSE ↓ Pearson ↑ Spearman ↑ RMSE ↓ Pearson ↑ Spearman ↑
Atom3D-GNN 0.38M 1.601 0.545 0.533 1.408 0.743 0.743
ProNet 1.39M 1.463 0.551 0.551 1.343 0.765 0.761
LEFTNet 0.85M 1.366 0.592 0.580 - - -
GET 0.69M 1.327 0.620 0.611 - - -
BindNet >47.61M 1.340 0.632 0.620 1.230 0.793 0.788
CheapNet 1.39M 1.311 0.642 0.639 1.238 0.794 0.789

Ligand Efficacy Prediction (LEP):

Model Params # AUROC ↑ AUPRC ↑
Atom3D-GNN 1.21M 0.681 0.598
TorchMD-NET 0.29M 0.717 0.724
GET 1.60M 0.761 0.751
Uni-Mol 47.61M 0.782 0.695
BindNet >47.61M 0.882 0.870
CheapNet 1.45M 0.935 0.924

Memory Efficiency & Interpretability:

Memory footprint comparison
Figure 3: Memory footprint comparison
Interpretability case study
Figure 4: Visualization of cluster-attention weights
Conclusion & Future Work

Key Contributions:

  • Novel integration of hierarchical cluster-level representations with cross-attention
  • State-of-the-art performance with improved computational efficiency
  • Enhanced interpretability through cluster-attention mechanism visualization

Future Directions:

  • Integration with more advanced SE(3)-equivariant encoders
  • Improving robustness to noisy predicted 3D structures (e.g., from AlphaFold3)
  • Application to related drug discovery tasks
Contact us: hyukjunlim@snu.ac.kr
Project Page QR Code
Scan for Project Page