Graph Attention Autoencoder for MultiOmics Integration, Risk Stratification, and Biomarker Identification in Cancer

This software is licensed under a Custom License Agreement. © 2024 Vatsal Patel. All rights reserved.

Overview

This repository contains the codebase for our research project titled "Graph Attention Networks for Biomedical Insights: MultiOmics Integration for Risk Stratification and Biomarker Identification". The project leverages advanced Graph Attention Networks (GATs) to integrate and analyze multiomic data types, such as gene expression, mutations, methylation, and copy number alterations, with a primary focus on head and neck squamous cell carcinoma (HNSCC).

Key Insights and Methodology

Cohort and Cancer Type:

• Patients: Derived from the TCGA-HNSCC dataset.
• Samples: Approximately 430 patient samples.

Multi-omics Features per Gene (Node Features):

Each gene is represented as a single node with a 17-dimensional feature vector comprising:

• Gene Expression: Standardized and normalized to [0,1].
• Copy Number Alteration (CNA): Encoded from -2 to 2, linearly mapped to [0,1].
• Mutations: Binary encodings for mutations, including frame shift, in-frame changes, missense, nonsense, silent, and other mutation types.
• Methylation Patterns: Values from six genomic regions (1st exon, 3’UTR, 5’UTR, gene body, TSS1500, TSS200).

Gene Node Processing

Construction of Patient-Specific Gene Networks

Methodology Overview

Encoder:

• Employs GATv2Conv layers to reduce the 17D features into a 1D latent embedding per gene.
• Uses an attention mechanism to prioritize biologically relevant gene-gene interactions.

Decoder:

• Reconstructs both node features and edge weights from the learned latent embeddings.
• Training is guided by a composite loss, combining:
  • Node reconstruction loss.
  • Edge presence reconstruction loss.
  • Edge weight reconstruction loss.

Outcome:

• Reduces multi-omics data into a single latent dimension per gene per patient (2,784 latent features per patient).
• Achieves a high cosine similarity (≥0.8) between input and reconstructed features.

Feature Extraction and Patient Clustering

• Latent Representation: Each patient is represented by a 2,784-dimensional latent vector.
• Clustering: K-means clustering identifies initial groups, later refined to three distinct survival-linked clusters based on statistical testing.

Risk Stratification and Biomarker Identification

• Attention Analysis: High-attention edges reveal significant gene-gene interactions across patients.
• Gene Selection: Attention-based filtering identifies 70 genes strongly associated with distinct survival outcomes.
• Predictive Modeling: Models trained on these genes achieve:
  • Multi-class classification AUC > 0.9.
  • Binary classification (high-risk vs. lower-risk) AUC ~0.98.

Results

Key Outcomes

• Integrated Multi-omics: Combines gene expression, mutations, CNA, and methylation into a unified latent representation.
• Patient Stratification: Identifies patient subgroups with significant survival differences.
• Biomarker Discovery: Highlights potential biomarkers for precision oncology.

Repository Structure

• data/: Contains necessary datasets (Note: Large files like .pth are not tracked).
• models/: Includes the models developed during the study.
• results/: Stores outputs from the models, including figures and summaries.
• graph_autoencoder.py: Implements the Graph Autoencoder model.
• HyperParameterSearch.py: Script for hyperparameter tuning of models.
• train.py: Main training script for the Graph Attention Network.
• OmicsConfig.py: Configuration file for multiomics data preprocessing and modeling.
• GraphAnalysis.py: Tools for analyzing graph structures and outputs.

Installation and Usage

Installation

1. Clone the repository:

   git clone https://github.com/your-username/Graph-Attention-Autoencoder-for-MultiOmics-Analysis.git
   cd Graph-Attention-Autoencoder-for-MultiOmics-Analysis

2. Install required dependencies:

   pip install -r requirements.txt

Training

Run the main training script:

python train.py

Example Usage

Below is an example of how to load and use the pre-trained model:

from torch_geometric.data import DataLoader
from model import GraphAutoencoder

# Load model
model_path = 'path_to_model.pth'
gae = GraphAutoencoder(in_channels=17, edge_attr_channels=1, out_channels=1, original_feature_size=17)
gae.gae.load_state_dict(torch.load(model_path))
gae.gae.eval()

# Example inference
output = gae.gae(data.x, data.edge_index, data.edge_attr)

Contact

For further information or queries, please contact:

• Vatsal Pravinbhai Patel ([email protected])

Citation

If you use this repository in your research, please cite as follows:

Patel, V. P., & Biswas, N. K. (2024). Graph Attention Networks for Biomedical Insights: MultiOmics Integration for Risk Stratification and Biomarker Identification. Zenodo. https://doi.org/10.5281/zenodo.14357409

---

Thank you for exploring our work. We hope this repository aids your research!