Edge-Based Execution of Graph Neural Networks for Protein Interaction Network Analysis in Clinical Oncology

Graph Neural Networks (GNNs) are increasingly used to model protein–protein interaction networks in cancer biology but are commonly deployed using central ized, cloud-based GPU infrastructure. This study examines the computational feasibility and numerical stability of executing GNN training and inference on decentralized, GPU-accelerated edge computing platforms. Using a protein protein interaction network comprising 1,603 genes and transcriptomic data from TCGA breast and lung cancer cohorts, we evaluate convergence behavior, mem ory constraints, and inference latency on a Jetson-class Single Board Computer. To avoid conflating hardware validation with biological performance, experiments are explicitly separated into (i) synthetic graph signal stress tests for systems level evaluation and (ii) biologically unmodified TCGA benchmarks for baseline predictive assessment. Results demonstrate stable convergence, tightly bounded inference latency (˜15 ms), and baseline predictive performance consistent with prior network-based oncology studies. These findings indicate that edge-based platforms can support graph neural network execution for molecular network analysis without reliance on centralized computing infrastructure.