SEAGAN Improves Plant Physiology Analysis with Edge-Aware Graph Attention Networks

Graph Neural Networks (GNNs) offer a flexible framework for analyzing scientific data characterized by physical, biological, or functional relationships. In plant physiology, a significant challenge is accurately identifying the active biochemical limitation state along A-Ci curves, which relate net CO2 assimilation rate to leaf intercellular CO2 concentration. This identification is crucial for estimating photosynthetic parameters but is often a major source of uncertainty. This research formulates the limitation-state identification problem for A-Ci curves as a graph-based node classification task, where each point on the curve is treated as a node. Domain-specific graph representations are constructed using both distance-based k-nearest-neighbor (kNN) connectivity and auxiliary-signal-guided (ASG) connectivity, with edge attributes encoding pairwise relationships between points. The proposed model, SEAGAN (domain-Specific and Edge-Aware Graph Attention Network for Dynamic Plant Processes), integrates process-aware node features, edge attributes, kNN connectivity, and graph attention with a weighted cross-entropy loss. Evaluated on a large synthetic dataset with known ground-truth limitation states, SEAGAN achieved an F1-score of 0.857 and an accuracy of 0.882, outperforming conventional learning baselines and other graph-based architectures, especially near critical biochemical transition regions. This demonstrates that representing A-Ci curves as graphs, particularly with edge-aware attention over local kNN neighborhoods, provides a highly effective strategy for biochemical limitation-state analysis.