📢 Accepted at ECML PKDD 2026

Welcome to the official repository for:

PathogenKG: Cross-Species Drug Repurposing via Heterogeneous Knowledge Graph Link Prediction

European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases

📍 Naples, Italy | 🗓️ September 7-11, 2026

PathogenKG is a heterogeneous knowledge graph embedding framework for cross-species antimicrobial drug repurposing. It integrates protein–protein interaction (PPI) networks from 31 bacterial pathogens (STRING), COG orthology, Gene Ontology annotations, and validated drug–target associations (DrugBank) into a unified graph of over 3 million triples. Heterogeneous GNN encoders (R-GCN, CompGCN) with a DistMult decoder are trained to predict new drug–target interactions (DTIs) via link prediction on the TARGET relation.

---

Repository Structure

PathogenKG/
├── train_and_eval.py           # Main training & evaluation script (Section 3.2)
├── drug_eval.py                # Compound-centric repurposing evaluation (Section 3.3)
├── drug_eval_results.py        # Process and summarise drug evaluation outputs
├── tuning_hyperparameter.py    # Bayesian HPO via Weights & Biases sweeps (Section 3.2)
├── tuning_dataset.py           # HPO on PathogenKG dataset
├── tuning_dataset_drkg.py      # HPO on DRKG benchmark (Section 3.2, pipeline validation)
├── model_eval.py               # Standard model evaluation utilities
├── evaluate_ranking.py         # Ranking metric computation
├── kg_stats_visualization.py   # KG statistics and figures (Table 2, Figure 2)
├── build_pathogenkg.py         # Build organism-specific KGs from raw data (Section 3.1)
├── merge_pathogen_kgs.py       # Merge organism KGs into PathogenKG (Section 3.1)
├── generate_all_pathogenkg.py  # Batch-build all 31 pathogen KGs
├── get_pathogenkg.py           # Download pre-built datasets from HuggingFace
│
├── src/
│   ├── utils.py                # Data loading, splitting, negative sampling, metrics
│   ├── evaluation_metrics_filtered.py  # Type-constrained filtered evaluation (Section 3.2)
│   ├── hetero_compgcn.py       # CompGCN encoder (Section 3.2)
│   ├── hetero_rgcn.py          # R-GCN encoder (Section 3.2)
│   ├── hetero_rgat.py          # R-GAT encoder (experimental)
│   ├── models_params.json      # Best hyperparameter configurations
│   ├── bio_utils.py            # Biological ID mapping utilities
│   └── downloaders/            # Scripts to fetch raw data from STRING, DrugBank, UniProt, GO, COG
│
├── dataset/
│   ├── PathogenKG_n31_core.tsv.zip   # Pre-built PathogenKG (31 species, 3M triples)
│   └── DRUGBANK/                     # Auxiliary DrugBank mappings
│
├── figures/                    # Figures used in the paper
├── requirements.txt
└── environment.yml

---

1. Environment Setup

Requirements: Python 3.10, CUDA-capable GPU (≥16 GB VRAM recommended for CompGCN with hidden dim 200).

# Create conda environment and install dependencies
conda create -n pathogenkg python=3.10 -y
conda activate pathogenkg
pip install -r requirements.txt

If torch-sparse installation fails (common on some platforms), install PyTorch and PyG extensions separately:

pip install --index-url https://download.pytorch.org/whl/cu128 torch torchvision torchaudio
pip install --no-cache-dir --only-binary=:all: \
  pyg_lib torch-geometric torch-scatter torch-sparse torch-cluster torch-spline-conv \
  termcolor torcheval \
  -f https://data.pyg.org/whl/torch-2.7.1+cu128.html

Adjust the CUDA version (cu128) to match your local installation.

---

2. Dataset

Option A: Use the pre-built dataset (recommended)

The default dataset PathogenKG_n31_core.tsv.zip (31 core STRING species, ~3M triples) is included in dataset/. No further action is needed.

Additional datasets can be downloaded from HuggingFace:

python get_pathogenkg.py

This downloads both PathogenKG variants and the raw source data to dataset/.

Option B: Build from scratch (Section 3.1)

To reproduce the KG construction pipeline from raw biological databases:

# 1. Download source data (STRING PPI, DrugBank, COG, GO)
#    Scripts in src/downloaders/ fetch from public APIs.

# 2. Build organism-specific KGs for all 31 pathogens
python generate_all_pathogenkg.py

# 3. Merge into a single multi-organism graph
python merge_pathogen_kgs.py

Dataset format

PathogenKG is stored as a TSV of triples: head, interaction, tail.

• Entity types: ExtGene (pathogen proteins, UniProt accession) and Compound (drugs, PubChem CID)
• Relation types (13): 8 PPI types (ASSOCIATION, GENE_BIND, ACTIVATION, REACTION, CATALYSIS, PTMOD, INHIBITION, EXPRESSION), ORTHOLOGY, 3 GO types (BiologicalProcess, MolecularFunction, CellularComponent), and TARGET

---

3. Reproducing the Results

3.1 Training and Evaluation (Table 4)

Train a CompGCN model on PathogenKG with the best hyperparameters (stored in src/models_params.json):

# Single run with default settings (CompGCN, TARGET task, PathogenKG_n31_core)
python train_and_eval.py

# 12-run evaluation for statistical comparison (as in Table 4)
python train_and_eval.py --model compgcn --runs 12 --epochs 400 --task TARGET

# R-GCN comparison
python train_and_eval.py --model rgcn --runs 12 --epochs 400 --task TARGET

Key arguments:

Argument	Default	Description
--model	compgcn	GNN encoder: compgcn, rgcn, or rgat
--tsv	dataset/PathogenKG_n31_core.tsv.zip	Path to dataset
--task	TARGET	Relation type(s) to predict
--runs	1	Number of independent runs
--epochs	400	Training epochs
--patience	20	Early stopping patience (with --early_stopping)
--alpha	0.25	Focal loss alpha
--gamma	3.0	Focal loss gamma
--alpha_adv	2.0	Adversarial negative weighting temperature
--oversample_rate	5	TARGET triple repetition factor
--undersample_rate	0.5	Fraction of non-TARGET edges to keep
--negative_rate	1	Negatives per positive
--pretrain_epochs	0	Multi-relational pretraining epochs

The script produces:

• Console output: AUROC, AUPRC, MRR, Hits@1/3/10 per run and aggregated
• Saved model: models/target_<dataset>_<timestamp>/ (weights, config, entity/relation mappings)

3.2 DRKG Pipeline Validation (Section 3.2)

To validate the training pipeline on DRKG before applying it to PathogenKG:

# Download DRKG
python get_pathogenkg.py

# Run HPO or train directly
python tuning_dataset_drkg.py

3.3 Hyperparameter Optimisation (Section 3.2)

Bayesian HPO via Weights & Biases:

# Edit ENTITY in tuning_hyperparameter.py with your W&B username
python tuning_hyperparameter.py

The sweep optimises the composite metric M = 0.2·AUROC + 0.4·AUPRC + 0.4·MRR. Search space and best configurations are documented in src/models_params.json.

3.4 Compound-Centric Drug Repurposing (Section 3.3, Table 5)

After training, run the compound-centric evaluation to rank all pathogen proteins for each drug:

# Evaluate all compounds (produces top-k ranked lists per compound)
python drug_eval.py --model_folder models/<your_model_folder>

# Evaluate a single compound
python drug_eval.py --model_folder models/<your_model_folder> \
    --compound "Compound::Pubchem:2764"  --topk 20

# Process results into summary tables
python drug_eval_results.py

Key arguments for drug_eval.py:

Argument	Default	Description
--model_folder	(required)	Path to trained model directory
--compound	all	Compound ID or all
--topk	20	Number of top predictions to return
--batch_size	4096	Scoring batch size

This produces per-compound ranked lists of predicted protein targets across all 31 pathogen species, corresponding to the biological plausibility analysis in Section 5 and Table 5 of the paper.

---

4. KG Statistics and Visualisation

To reproduce the dataset statistics (Table 2) and schema figures:

python kg_stats_visualization.py

Output figures are saved to figures/.

---

Citation

📌 Citation information will be available upon publication. The paper has been accepted at ECML PKDD 2026 (September 7-11, Naples, Italy). BibTeX and additional citation formats will be provided once the proceedings are published.

License

This project is released under the MIT License.