Högskolan i Skövde

Identification of effective therapeutic targets is an important part of the drug development process. Traditional pharmacological assays for target identification are expensive and timeconsuming. Therefore, the demand for efficient computational methods is persistent. Here we introduce DeepTarg, a network-based omics-oriented deep learning framework for therapeutic target discovery. The framework uses Graph Neural Networks, a specialized deep learning architecture for learning in graphs, to prioritize candidate therapeutic targets for diseases based on their gene expression signatures. To test the proposed framework, a Graph Neural Network model has been developed, trained, and tested to prioritize therapeutic targets for Alzheimer's disease - a neurodegenerative disorder with a distinct gene expression signature. The framework included constructing a heterogeneous knowledge graph consisting of six modules – perturbations, genes, biological processes, pathways, molecular functions, and tissues. A hypothetical "therapeutic" signature for the disease was generated based on the signaturereversion paradigm. Large-scale perturbation signature profiles induced in central nervous system cells were attributed as node features in the graph after dimensionality reduction. The model was trained to predict links between the “perturbations” and “genes” modules. Finally, the trained model was used to prioritize gene targets for Alzheimer's disease based on the estimated link scores between the therapeutic node and gene nodes using node embeddings. The trained model showed a relatively high performance in predicting links (AUC = 0.94). The predicted candidate targets showed high relevance to Alzheimer's disease based on the scientific literature and the functional enrichment analysis performed.