Can graph neural networks for biology realistically run on edge devices? (doi.org)

Recent research explores the feasibility of running Graph Neural Networks (GNNs) for modeling protein–protein interaction networks on edge devices, which are typically more decentralized than traditional cloud-based GPU systems. The study specifically tests these GNNs using a dataset of 1,603 genes linked to breast and lung cancer from the TCGA. Researchers focused on key factors such as convergence behavior, memory constraints, and inference latency on a Jetson-class Single Board Computer, conducting separate experiments to assess both system performance and biological accuracy. The results are promising, showcasing stable convergence and a low inference latency of approximately 15 milliseconds, while maintaining predictive performance that aligns with previous oncology studies. This indicates that edge computing platforms can effectively support the execution of GNNs for analyzing molecular networks, thereby reducing reliance on centralized infrastructure. This advancement could significantly enhance the accessibility and efficiency of computational biology, particularly in real-time applications where low latency is critical.