An important prerequisite to understanding how a plant functions and responds to the environment is to determine which gene expression patterns are associated with a specific tissue type and external perturbation response. Neural network-based methods can provide subsets of highly informative genes for such predictions using backpropagation-like methods. Here, we propose that integrating prior molecular knowledge related to gene expression within neural network architectures can lead to more reliable and insightful models, improving identification of tissue and perturbation-related gene sets. The Comprehensive Knowledge Network used to this end contains information on protein-DNA and protein-protein interactions. We first construct an Arabidopsis tissue- and perturbationspecific gene expression resource from published datasets and metadata. We then develop a pipeline comprising batch effect correction, prediction training, and model explanation. To address batch effects, we implement and evaluate several approaches, finding that Conditional Variational Autoencoders achieve the highest performance among tissue types, outperforming the other methods. We develop and train deep neural networks models to classify the underlying tissue types and perturbation groups using gene expression patterns as input. The knowledge graph-based models incorporating the prior molecular knowledge as additional network layers achieve similar classification performance as baseline models. However, the analysis of model explainability, by computing class specific relevance scores per gene, demonstrates that the knowledge graph-based models outperform baseline models by prioritising biologically relevant genes, known to be related to specific tissue types and molecular processes (e.g. phytohormone and stress-related responses). Our results thus demonstrate the applicability and reliability of knowledge graph-primed deep learning for identifying condition-specific genes and gene sets.