Welcome to MVP-MP: a Multi-view Graph Learning Framework Augmented by Pathway-aware Representation Losses for Multi-label Metabolic Pathway Prediction

Accurate inference of metabolic pathway categories for unannotated small molecules is a key step for pathway-centric metabolomics interpretation and is increasingly relevant to drug discovery and lead optimization. In this study, we propose MVP-MP, a multi-view graph learning framework augmented by pathway-aware representation losses for multi-label metabolic pathway prediction. MVP-MP integrates a molecular graph branch and a fingerprint branch through a Bi-Gate fusion module for adaptive cross-view integration. In the graph branch, MVPool scores node importance from feature, structure, and diffusion perspectives, selects a top-k induced subgraph, and performs subgraph-to-global refinement, enabling explicit substructure focusing and propagating substructure evidence into global representations. At the objective level, we train the model end to end with a multi-label classification loss together with two pathway-aware regularizers. A pathway-aware contrastive loss exploits label overlap to structure compound relations in the embedding space, while a prototype-aware loss forms pathway prototypes to promote pathway-discriminative, pathway-centric embeddings. In performance evaluation, MVP-MP consistently outperforms existing state-of-the-art methods on the independent test set. Moreover, prototype-centered latent analysis and MVPool-driven substructure interpretation provide pathway-centric interpretability by linking predictions to enriched substructures and reusable pathway-associated chemotypes.

🔧 Installation instructions

1. Clone the repository

git clone https://github.com/GGCL7/MVP-MP.git
cd MVP-MP

2. Set up the Python environment

conda create -n mvpmp python=3.10
conda activate mvpmp
pip install -r requirements.txt

Model Training

Train the model from scratch:

  python main.py \
    --csv_path Data/kegg_dataset.csv \
    --index_path Data/data_index.txt \
    --num_labels 11 \
    --batch_size 64 \
    --epochs 200 \
    --lr 1e-3 \
    --seed 42 \
    --lambda_contrast 0.7 \
    --lambda_proto 0.4 \
    --select_metric f1 \
    --ckpt_path best_model.pth

The training script will automatically save the model with the best validation F1 to best_model.pth.

Model Evaluation

Evaluate the trained model:

python evaluation.py

The script reports the following metrics:

• Accuracy
• Precision
• Recall
• F1 score

Prediction (Inference)

Run prediction:

python predict.py \
  --smiles "NC(=O)c1ccc[n+](C2OC(COP(=O)(O)OP(=O)(O)OCC3OC(n4cnc5c(N)ncnc54)C(O)C3O)C(O)C2O)c1" \
  --checkpoint best_model.pth \
  --threshold 0.5

Example output::

[00] Carbohydrate metabolism                       prob=0.0008  pred=0
[01] Energy metabolism                             prob=0.9593  pred=1
[02] Lipid metabolism                              prob=0.0010  pred=0
[03] Nucleotide metabolism                         prob=0.0000  pred=0
[04] Amino acid metabolism                         prob=0.0293  pred=0
[05] Metabolism of other amino acids               prob=0.0055  pred=0
[06] Glycan metabolism                             prob=0.0000  pred=0
[07] Cofactors and vitamins                        prob=0.9271  pred=1
[08] Terpenoids and polyketides metabolism         prob=0.0011  pred=0
[09] Biosynthesis of other secondary metabolites   prob=0.0013  pred=0
[10] Xenobiotics biodegradation and metabolism     prob=0.6966  pred=1