HyperDFI

HyperDFI is a PyTorch-based graph learning project for drug-food interaction prediction. The code builds a bipartite drug-food graph, learns drug and food representations with graph and hypergraph propagation, and predicts interaction labels with a classifier.

Task

The current codebase is adapted for:

• Drug-food interaction prediction
• Bipartite graph representation learning

The dataset loader expects interaction triples in CSV format:

drug_id,food_id,relation

Project Structure

HyperDFI/
├── Code/
│   ├── Main.py
│   ├── Model.py
│   ├── DataLoader.py
│   ├── Params.py
│   └── Utils/
│       ├── RunLogger.py
│       └── Utils.py
└── Data/
    └── DrugBank-df/
        └── transductive/
            ├── train.csv
            └── test.csv

Core Files

• Code/Main.py: training and evaluation entry point
• Code/Model.py: model definition
• Code/DataLoader.py: dataset loading, ID remapping, adjacency construction
• Code/Params.py: command-line hyperparameters
• Code/Utils/RunLogger.py: simple runtime logger
• Code/Utils/Utils.py: loss and normalization helpers

Model Overview

The model contains three main parts:

1. Embedding layer Initializes learnable embeddings for drugs and foods.
2. Graph propagation Uses a GCN-style sparse propagation layer on the bipartite adjacency matrix.
3. Hypergraph propagation Uses hypergraph-style transformations to capture higher-order structure.
4. Classifier Concatenates drug and food embeddings and predicts the interaction label.

The final embedding is the sum of propagated representations across layers.

Data Processing

Code/DataLoader.py performs the following steps:

• Reads train.csv and test.csv
• Renames columns internally as drug_nodes, food_nodes, and relations
• Maps raw drug and food IDs to contiguous indices
• Builds the label matrix
• Creates sparse training and testing interaction matrices
• Converts the bipartite adjacency matrix to a CUDA sparse tensor
• Wraps samples into PyTorch DataLoader objects

Environment

Recommended environment:

• Python 3.9+
• PyTorch
• NumPy
• pandas
• SciPy
• scikit-learn
• wandb

You can install the common dependencies with:

pip install torch numpy pandas scipy scikit-learn wandb

How To Run

Run from the Code directory:

cd Code
python Main.py

Example with explicit parameters:

python Main.py --data DrugBank-df --epoch 100 --batch 1024 --lr 1e-3 --gpu 0

Training

The training script:

• loads the dataset
• initializes the model and optimizer
• trains for multiple epochs

Reported metrics include:

• Accuracy
• Precision
• F1
• AUC
• AUPR

Notes

• The code currently uses CUDA tensors in several places, so GPU execution is expected by default.
• The dataset path is resolved relative to the Code directory.

Citation

If you use this repository in your work, cite the corresponding paper or project source that this implementation is based on, and describe your drug-food adaptation clearly.