Explainable graph neural networks for link prediction in plant molecular interaction networksŠuštar, Klara (Avtor) Curk, Tomaž (Mentor) Bleker, Carissa Robyn (Komentor) Zrimec, Jan (Komentor) graph neural networkslink predictionmolecular interaction networksArabidopsis thalianabatch effect correctionvariational autoencoderscArchesexplainable artificial intelligenceCKNPlant molecular biology leans on curated interaction networks such as the Comprehensive Knowledge Network (CKN) of Arabidopsis thaliana, yet it is rarely tested whether building such a network into a learning system helps at all, and if so, where. This thesis integrates CKN with public expression data and asks that question directly, after revising the batch correction pipeline so that the choice of correction method cannot influence classifier evaluation. For sample classification the answer is negative: replacing CKN's wiring with a degree-matched random graph changes macro F1 by nothing measurable, so the models use little beyond the degree sequence. For link prediction it is positive but conditional: a relational graph neural network beats a parameter-matched lookup table by +0.023 mean reciprocal rank once both are tuned, and the advantage reaches an order of magnitude for sparsely connected genes while disappearing at hubs. Attributions over the trained classifiers mostly read out node connectivity, though their functional enrichment exceeds a degree-matched null. Aggregate metrics hide where the graph matters; stratified evaluation with explicit controls reveals it.20262026-09-16 08:15:52Magistrsko delo/naloga187899VisID: 164162COBISS_ID: 291184899sl