MoPE-MOI

Pathway-Guided Mixture-of-Experts for Multi-omics Integration

MoPE-MOI is a deep learning framework designed for multi-omics integration in cancer research. It utilizes a biologically informed, sparse Mixture-of-Experts (MoE) architecture to provide state-of-the-art predictive performance while maintaining strict biological interpretability and patient-level mechanistic traceability.

Key Features

• Pathway-Guided Experts: Employs 50 MSigDB Hallmark pathways as structural priors to prevent "black-box" predictions.
• Asymmetric Routing: Uses a parsimonious RNA-only global gating network by default to filter dimensionality-induced noise, while local experts process full multi-omics tensors.
• Stochastic Routing (Noisy Gating): Injects tunable Gaussian noise during routing to prevent expert collapse and preserve inter-patient tumor heterogeneity.
• Non-Competitive Activation: Replaces traditional zero-sum Softmax with independent Sigmoid activation, faithfully capturing concurrent hyperactivation of oncogenic hallmarks.
• Robust Evaluation: Built-in nested 5-Fold cross-validation ensemble with automated hyperparameter tuning via Optuna.

Environment Setup

Create a virtual environment and install the required dependencies:

conda create -n pathmoe python=3.9 -y
conda activate pathmoe
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install pandas numpy scipy scikit-learn optuna

Data Preparation

All multi-omics datasets (RNA, CNV, Methylation) should be processed and placed in the designated data directory. The default path configured in the scripts is /data/zliu/Path_MoE/data/.

Ensure you have the MSigDB Hallmark gene set downloaded:

• h.all.v2023.1.Hs.symbols.gmt

Quick Start: Subtype Classification

The run_pathmoe_subtype.py script handles the complete pipeline, including Optuna hyperparameter searching and 20-seed ensemble evaluation.

Standard Run (Asymmetric RNA-only Router + Noisy Gating + Sigmoid):

python run_pathmoe_subtype.py \
    --cancer BRCA \
    --seeds 20 \
    --noise_std 0.2 \
    --save_gating \
    --save_results_json ./results/BRCA_standard.json

Ablation Run 1: Symmetric Tri-omics Router:

python run_pathmoe_subtype.py \
    --cancer BRCA \
    --seeds 20 \
    --use_tri_gating \
    --noise_std 0.2 \
    --save_results_json ./results/BRCA_triomics.json

Ablation Run 2: Traditional Softmax Activation (Competitive):

python run_pathmoe_subtype.py \
    --cancer BRCA \
    --seeds 20 \
    --use_softmax \
    --noise_std 0.2 \
    --save_results_json ./results/BRCA_softmax.json

Repository Structure

PathMoE/
├── src/
│   ├── model_moe_pm50.py             # Core TopKPathMoE architecture (Subtyping task)
│   ├── model_moe_.py                 # Core TopKPathMoE architecture (Survival task)
│   ├── run_pathmoe_subtype.py        # Main training/evaluation script (Subtyping task)
│   ├── run_pathmoe_survival.py       # Main training/evaluation script (Survival task)
│   ├── dataset_subtype.py            # PyTorch Dataset for multi-omics loading (Subtyping task)
│   ├── dataset_survival.py            # PyTorch Dataset for multi-omics loading (Survival task)
│   ├── utils.py                      # Utility functions (Mask generation, etc.)
└── README.md

Outputs

Running the pipeline will automatically generate three directories:

• predictions_subtype/: Contains detailed per-sample predictions.
• gating_subtype/: Contains the sparse routing weights for biological interpretation.
• checkpoints_subtype/: Stores model weights for each fold.

Contact & Support

For technical issues, bug reports, and code-related questions: Please reach out to the lead developer:

• Author: Zhe Liu, Seoul National University
• Email: lizzie_liu@snu.ac.kr

For academic correspondence and general inquiries: Please contact the corresponding author,

• Corresponding Author: Prof. Taesung Park, Seoul National University
• Email: tspark@stats.snu.ac.kr