Edge-GNN: A Constraint-Aware Graph Neural Network Framework for Resource-Efficient Biological Interaction Modeling

Graph neural networks (GNNs) have become an effective framework for modeling biological interaction networks such as protein–protein interaction graphs and gene regulatory systems. However, many existing GNN approaches are designed primarily for high-performance computing environments and do not explicitly account for computational constraints such as memory usage or inference latency. In this study we introduce Edge-GNN, a constraint-aware training framework for graph neural networks designed to balance predictive performance with computational efficiency. The proposed approach incorporates a multi-objective optimization formulation that jointly minimizes predictive loss together with proxy measures of model complexity and computational latency. The framework is evaluated using several graph neural network architectures including Graph Convolutional Networks, GraphSAGE, and Graph Attention Networks. Experiments conducted on the PROTEINS benchmark dataset and transcriptomic data derived from The Cancer Genome Atlas integrated with protein interaction networks demonstrate that the proposed approach reduces computational cost while maintaining comparable predictive performance. These findings suggest that incorporating computational constraints directly into the training objective can improve the practicality of graph-based learning methods for biological network analysis in resource-constrained environments.