# DRCMDA

**Repository Path**: Tomhappy/DRCMDA

## Basic Information

- **Project Name**: DRCMDA
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-08-08
- **Last Updated**: 2026-08-08

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# DRCMDA

DRCMDA is a novel drug repositioning framework that integrates cluster-aware structured mask reconstruction with diffusion-based metapath augmentation.

## Repository Structure

```text
DRCMDA_release/
├── train.py                         # Main cross-validation training entry
├── drcmda/
│   ├── data.py                      # Dataset loading, preprocessing, graph construction
│   ├── model.py                     # DRCMDA prediction model
│   ├── layers.py                    # HGT layers, fusion layers, contrastive losses
│   ├── metrics.py                   # Evaluation metrics
│   ├── sampling.py                  # Metapath and negative sampling utilities
│   ├── modules/
│   │   ├── csmr/                    # Cluster-guided similarity matrix reconstruction
│   │   └── metapath/                # Diffusion-based metapath augmentation
│   └── pretraining/sgmae/           # SGMAE encoder used to precompute entity embeddings
├── scripts/
│   └── precompute_embeddings.py     # Regenerate SGMAE embeddings
├── data/                            # Benchmark datasets
├── embeddings/                      # Precomputed drug, disease, and protein embeddings
└── requirements.txt
```

## Installation

Create a clean Python environment and install the required dependencies:

```bash
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
```

DGL wheels are tied to the local PyTorch and CUDA setup. The included `requirements.txt` uses the CPU-compatible DGL package tested for this release. For GPU training, install the DGL build that matches your CUDA and PyTorch versions before running the model.

## Quick Start

Run the default 10-fold experiment on the C-dataset:

```bash
python train.py --dataset C-dataset --device 0
```

Use CPU explicitly when a GPU is not available:

```bash
python train.py --dataset C-dataset --device cpu
```

For a fast smoke test:

```bash
python train.py \
  --dataset C-dataset \
  --device cpu \
  --epochs 1 \
  --k-fold 2 \
  --diffusion-epochs 1 \
  --diffusion-timesteps 10 \
  --diffusion-batch-size 1
```

Training prints fold-level AUROC, AUPR, accuracy, precision, recall, F1, MCC, and final mean metrics across folds.

## Regenerating Embeddings

Precomputed SGMAE embeddings are provided under `embeddings/`. To regenerate them for a dataset:

```bash
python scripts/precompute_embeddings.py --dataset C-dataset --device 0
```

The generated `drug.csv`, `disease.csv`, and `protein.csv` files are written to `embeddings/<dataset>/`.

## Key Arguments

| Argument | Description | Default |
| --- | --- | --- |
| `--dataset` | Benchmark dataset: `F-dataset`, `C-dataset`, or `B-dataset` | `C-dataset` |
| `--device` | `auto`, `cpu`, or a CUDA index such as `0` | `auto` |
| `--epochs` | Number of DRCMDA training epochs per fold | `300` |
| `--k-fold` | Number of cross-validation folds | `10` |
| `--neighbor` | KNN size for similarity graphs | `5` |
| `--n-neg` | Number of two-hop negative candidates | `30` |
| `--diffusion-epochs` | Training epochs for metapath diffusion augmentation | `500` |
| `--diffusion-timesteps` | Diffusion timesteps | `1000` |
| `--l-con-rate` | Contrastive regularization weight | `1e-5` |

Run `python train.py --help` for the full configuration.

## Reproducibility Notes

The training script fixes Python, NumPy, and PyTorch random seeds through `--seed`. Reported performance may still vary slightly across hardware, PyTorch/DGL versions, and CUDA kernels. For manuscript-scale reproduction, use the default 10-fold setting and a GPU-enabled DGL installation.

## Citation

If this code supports your research, please cite the corresponding DRCMDA manuscript after publication.