PIMO: Pathway-based Interpretable Multi-Omics interactions for multi-omics integration

Abstract

We propose a Pathway-based Interpretable deep learning Multi-Omics interaction model, PIMO, that explicitly captures regulatory effects across omics layers. Experiments on multiple TCGA cancer datasets showed that PIMO consistently outperformed state-of-the-art baselines in survival analysis, up to 13% increase in the C-index. PIMO provides biologically interpretable analyses that identify important pathways, genes, and inter-omics interactions with DNA methylation and copy number alterations

Overview

Data Preparation

The tcga_data_preparation/ directory contains the following subfolders:

• raw_brca/ — raw input data
• processed_brca/ — processed multi-omics data
• splits_brca/ — train/validation/test splits
• kegg_pathway_data/ — KEGG pathway and gene annotations

In addition, this directory includes:

• tcga_data_preparation.ipynb — main notebook for data preprocessing
• data_preparation_pipeline.py — pipeline for dataset construction and splitting
• helper scripts for running preprocessing

Required Inputs

Users must download the following data from cBioPortal for Cancer Genomics:

• Gene expression
• DNA methylation
• Copy number alteration (CNA)
• Clinical data

These files should be placed in:

• raw_brca/ (for BRCA), or
• raw_<cancer>/ (for other cancer types)

The kegg_pathway_data/ folder (already provided) contains pathway definitions and gene mappings required for model input construction.

Processing Steps

Running tcga_data_preparation.ipynb will:

1. Integrate multi-omics data into a unified representation
2. Generate merged datasets stored in:
   • processed_brca/ or processed_<cancer>/
3. Apply preprocessing and pathway-based transformations

The data_preparation_pipeline.py script further:

• Splits the data into training, validation, and test sets
• Applies pathway ordering (PathCNN-style preprocessing)
• Saves outputs to:
  • splits_brca/ or splits_<cancer>/

---

Model Training

The root directory contains:

• run.py — main script for training the PIMO model
• run.sh — SLURM batch script for cluster execution
• run_command.txt — example command for submitting jobs

Running the Model (SLURM)

sbatch --export=gpu_num=0,cancer_type="brca",exp_num=1,learning_rate=0.00005,decay_rate=0.3,decay_epochs=100,weight_decay=0.9,num_epochs=500,num_kernels1=32,dropout1=0.58,num_kernels2=4,dropout2=0.025,activation_fn="tanh",fc_nodes=256,batch_size=256,min_epochs=200,patience=25 run.sh