# pathway-RGCNMDA

**Repository Path**: Tomhappy/pathway-RGCNMDA

## Basic Information

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

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-10-14
- **Last Updated**: 2026-10-14

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# pathway-RGCNMDA

Pathway-bridge relational graph convolutional network for human miRNA--disease association prediction.

This repository contains only the files required to run the final pathway-RGCNMDA model:

- `code/`: model and training/evaluation code.
- `dataset/`: processed HMDD strict-human internal benchmark and RNADisease v4.0 external benchmark.
- `pathway/`: processed pathway-bridge edges used by the six-relation RGCN.
- `params/`: locked hyperparameters for random, cold-disease, and cold-miRNA settings.
- `scripts/`: minimal runnable examples.
- `preprocessing/`: optional scripts for reconstructing the processed benchmarks from raw database files.

Scripts and retained result artifacts for the principal revision analyses are provided in the archived release where applicable.

## Directory layout

```text
pathway-RGCNMDA/
  code/
    run_pathway_bridge_rgcn.py
    rgcnmda_source.py
    rank32_model.py
  dataset/
    HMDD31296/common/
    RNADisease24140/common/
  pathway/
    HMDD31296_PATHWAY_BRIDGE_external_edges_for_rgcn.csv
    RNADisease24140_PATHWAY_BRIDGE_external_edges_for_rgcn.csv
  params/
    random.json
    cold_disease.json
    cold_mirna.json
  scripts/
    run_hmdd_random.sh
    run_rnadisease_random.sh
  preprocessing/
    preprocess_hmdd.py
    preprocess_rnadisease.py
    README.md
```

## Data

The processed benchmarks are stored under `dataset/`.

- `HMDD31296`: strict-human HMDD benchmark.
- `RNADisease24140`: processed RNADisease v4.0 benchmark.

Each dataset contains a `common/` directory with processed mappings, matrices, protocol audit files, and fixed five-fold splits. All fold-specific MDA-derived features are reconstructed from training associations only by the running code.

## Pathway bridge

The files in `pathway/` contain the external pathway-bridge graph used by the six-relation RGCN. The directed relation types are:

1. miRNA to disease, reconstructed from training MDA edges only;
2. disease to miRNA, inverse training MDA edges only;
3. miRNA to pathway;
4. pathway to miRNA;
5. disease to pathway;
6. pathway to disease.

The external pathway edges do not use held-out validation or test MDA labels.

## Preprocessing

The processed benchmark files under `dataset/` can be used directly. For transparency, optional scripts for reconstructing these processed benchmarks from raw database files are provided in `preprocessing/`. Raw database files are not redistributed in this repository.

## Example commands

From the repository root:

```bash
mkdir -p results
bash scripts/run_hmdd_random.sh
bash scripts/run_rnadisease_random.sh
```

To run a different split, replace `--split random` and `--params params/random.json` with one of:

- `--split cold_disease --params params/cold_disease.json`
- `--split cold_mirna --params params/cold_mirna.json`

## Main Python entry point

```bash
python code/run_pathway_bridge_rgcn.py \
  --source code/rgcnmda_source.py \
  --runner code/rank32_model.py \
  --common_dir dataset/HMDD31296/common \
  --bridge_file pathway/HMDD31296_PATHWAY_BRIDGE_external_edges_for_rgcn.csv \
  --params params/random.json \
  --output results/HMDD31296_pathway_rgcn_random.json \
  --dataset HMDD31296 \
  --split random \
  --model pathway_rgcn \
  --decoder_mode full \
  --negative_ratio 1 \
  --train_batch_size 65536 \
  --folds 5 \
  --epochs 200 \
  --patience 20 \
  --mf_init_steps 300 \
  --seed 42 \
  --device cuda
```

## Notes

- Fixed splits and negative samples are included in the processed dataset directories.
- The model uses validation AUC for epoch selection.
- No validation or test associations are used to construct fold-specific MDA-derived features.