PlantMetBench: A Multimodal Knowledge Graph Benchmark for Plant Biosynthesis Prediction

Plant metabolism produces a vast diversity of molecules, yet identifying the enzymes that catalyse each reaction remains a major open problem. Solving it requires integrating heterogeneous biological and chemical information, but the field lacks standardized benchmarks that preserve these relationships for multimodal machine learning. We present PlantMetBench, a multimodal knowledge-graph benchmark built on PlantMetWiki, a curated graph of plant metabolism across 424 species. PlantMetBench frames enzyme--reaction prediction as a cross-species generalisation task and provides graph-native dataset splits, pre-computed multimodal node features, and evaluation protocols designed to expose structural shortcuts. We evaluate heterogeneous graph neural networks (e.g., HeteroSAGE, GAT, HGT, R-GCN) and a dual-encoder retrieval baseline, finding that the latter greatly outperforms all graph architectures and that benchmark design choices, such as split type, shortcut removal, and evaluation metrics, can strongly affect the interpretation of the model's performance. PlantMetBench provides a foundation for systematic and reproducible evaluation of multimodal biosynthetic data and graph architectures.