Multi-Omics GNN

Overview

Applied modified IntegrAO graph neural network methodology to the MMRF COMPASS cohort (N=655 samples) for multi-omics patient stratification in multiple myeloma, achieving 258% high-risk detection enhancement.

The Problem

Multiple myeloma patients show heterogeneous responses to treatment:

• Standard risk stratification misses important subgroups
• Multi-omics data underutilized for personalized treatment
• Need for actionable therapeutic vulnerability profiles

Technical Approach

Data Integration

Integrated four omics layers from MMRF COMPASS cohort (N=655 samples):

• SNV: Single nucleotide variants
• CNV: Copy number variations
• TME: Tumor microenvironment signatures
• WES: Whole exome sequencing features

Graph Neural Network

Modified IntegrAO architecture:

• Multi-layer graph construction from patient similarity
• Attention-based message passing across omics layers
• Interpretable node embeddings for clinical actionability

Results

Clinical Impact

Identified 18 distinct vulnerability profiles enabling:

• Targeted therapeutic recommendations
• Identification of patients likely to respond to specific agents
• Early intervention for high-risk subgroups
• 94% of patients now have actionable therapeutic targets

Technical Stack

Framework:      PyTorch Geometric
Architecture:   Modified IntegrAO
Dataset:        MMRF COMPASS (N=655)
Omics Layers:   SNV, CNV, TME, WES
Analysis:       Scanpy, Geneformer
Clustering:     Unsupervised ML

Publication

• Clinical Lymphoma Myeloma Leukemia 2025 - “Integrating Microenvironment with Tumor Multi-Omic Using Unsupervised Machine Learning” (Contributing Author)