Introduction
The analysis of "omics" data, such as transcriptomics and metabolomics, presents a significant challenge in bioinformatics—the "curse of dimensionality." These datasets typically feature thousands of variables (genes, metabolites) but only a few hundred samples, making it difficult for traditional machine learning models to distinguish signal from noise. Graph Neural Networks (GNNs) have emerged as a promising solution by incorporating relational context into the learning process. This study investigates the optimal way to structure these graphs for classifying Parkinson’s Disease (PD), comparing two fundamental strategies: connecting biological molecules or connecting patient samples.
Methodology
The researchers evaluated two primary topological strategies for graph construction:
Sample Similarity Networks (SSNs)
In the SSN approach, each node represents an individual patient sample. Edges are established between patients based on the cosine similarity of their omics profiles. This transforms the diagnostic task into a node classification problem, where the model predicts the disease state of a patient based on their profile and their similarity to others in the network.
Molecular Interaction Networks (MINs)
Conversely, the MIN approach treats individual molecular entities (genes or metabolites) as nodes. Edges represent known biological interactions sourced from databases like STRING or STITCH. Here, the entire molecular signal for a patient is mapped onto a fixed graph topology, and the task becomes graph classification.
To manage high dimensionality, the team employed LASSO for feature selection and tested various architectures, including traditional GNNs (GCN, GAT, ChebyNet) and advanced Graph Transformers (GPST, GTC).
Experimental Setup
The models were validated using two independent cohorts:
- PPMI (Parkinson’s Progression Markers Initiative): Transcriptomics data (14,548 features, 378 samples).
- LUXPARK (Luxembourg Parkinson’s Study): Metabolomics data (1,079 features, 1,136 samples).
Results and Discussion
The experimental results revealed several critical insights into graph-based omics analysis.
Superiority of Similarity-Based Graphs
A key finding was that SSNs consistently outperformed MINs across all tests. This performance gap likely stems from the incompleteness of current interaction databases—particularly for metabolites—and the fact that static interaction networks may not capture the dynamic pathological changes specific to PD.
Performance of Graph Transformers
Graph Transformers, specifically the GPST architecture, achieved the highest performance. On the LUXPARK metabolomics dataset, GPST reached an AUC of 0.84, significantly outperforming the transcriptomics-based models. The attention mechanisms in Transformers allow for dynamic weighting of neighbors, providing robustness against the noise inherent in similarity-based edges.
Architectural Considerations
The study also found that "deeper is not always better." Increasing GNN depth to 50 layers led to performance degradation, suggesting that shallow 2-layer models are more appropriate for the relatively small sample sizes typical of clinical cohorts. Furthermore, the inclusion of LASSO was vital; models trained without feature selection suffered from significant overfitting.
Conclusion
This research demonstrates that structuring omics data as Sample Similarity Networks, combined with Graph Transformer architectures, provides a robust framework for Parkinson’s Disease classification. While Molecular Interaction Networks offer biological interpretability, their current performance is limited by database gaps. Future work will focus on refining these interaction networks and using tools like GNN-Explainer to uncover specific molecular subnetworks that drive disease progression, potentially identifying new therapeutic targets.
Summary updated