H-DrugNet

H-DrugNet is a deep learning framework designed to predict the frequency of drug side effects. By integrating a Hierarchical Representation strategy with Differentiable Pooling (DiffPool), the model effectively captures multi-level structural dependencies within drug molecules and their pharmacological associations.

---

🛠 Requirements

• torch==1.12.1+cu113
• torch-geometric==1.7.2
• torch-cluster==1.6.0
• torch-scatter==2.0.9
• torch-sparse==0.6.15
• rdkit==2022.9.4
• numpy==1.22.4
• pandas==1.5.1
• scipy==1.10.1
• scikit-learn==1.2.1
• networkx==2.8.8

---

Files

1. Data

This folder contains the standardized drug-side effect datasets and preprocessing matrices.

• frequency_data.txt The standardized drug side effect frequency classes used in our study.

• drug_SMILES_750.csv The SMILES representations of 750 drugs.

• raw_frequency_750.mat The original matrix of drug-adverse effect frequencies (highly sparse).

• side_effect_label_750.mat The encoded features of side effects.

• mask_mat_750.mat The mask matrix for ten-fold cross-validation in a warm-start scenario.

• blind_mask_mat_750.mat The mask matrix for ten-fold cross-validation in a cold-start scenario.

---

2. Experimental Scenarios

• warm-scene_data/ Contains side effects and drug data specific to the warm-start scenario.

• cold-scene_data/ Contains side effects and drug data specific to the cold-start (blind) scenario.

---

3. Code Implementation

• Net.py Defines the main architecture of the H-DrugNet model.

• diffpool_layer.py Implements the Differentiable Pooling layer:

  • Learns node clustering for hierarchical representation
  • Computes auxiliary Link Prediction Loss and Entropy Loss

• vector.py Converts drug SMILES into graph representations (nodes and edges) for GNN processing.

• utils.py Provides utility functions for:

  • Data loading
  • Evaluation metrics
  • General preprocessing

• warm-scene.py Execution script for the warm-start scenario (750 drugs, 994 side effects)

• cold-scene.py Execution script for the cold-start scenario (750 drugs, 994 side effects)

---

Methodology: Hierarchical Pooling

The core innovation of H-DrugNet lies in modeling drugs at multiple levels of granularity.

Key Mechanisms

• Aggregate Atoms Learns an assignment matrix to group atoms into higher-level functional motifs.

• Structural Integrity Uses Link Prediction Loss to preserve the topological structure of the original graph after pooling.

• Deterministic Assignment Applies Entropy Loss to reduce uncertainty in node-to-cluster assignments.

---

Contact

If you have any questions or suggestions regarding the code or the research:

Changming Yao China University of Geosciences (Wuhan)

📧 Email: [yaochangming@cug.edu.cn] 🆔 ORCID: 0009-0003-5928-2647