Multilayered Prioritization of Graph Attention Autoencoder–Derived Drug Repurposing Candidates for COVID-19

Background/Objectives: The COVID-19 pandemic highlighted the critical need for drug repurposing. While computational graph-based approaches show promise, single analytical methods often fail to capture the complex pathological mechanisms needed for clinical translation. We aimed to prioritize COVID-19 repurposing candidates by combining computational prediction with orthogonal lines of evidence. Methods: We developed a Graph Attention Autoencoder (GATE) framework with multilayered evaluation integrating computational prediction, real-world data, and phenome analyses. Network-based embeddings identified candidates through latent-space similarity within biomedical knowledge graphs. Clinical relevance was assessed by disproportionality analysis (DPA) of an adverse event reporting system, and biological plausibility through gene expression and pathway profiling. Results: Our GATE framework identified 16 candidate drug entities, including 12 already used or recognized in the context of COVID-19 or its symptoms, consistent with the model recovering known drug–disease associations. The remaining four had minimal prior COVID-19 associations (cilastatin, megestrol, drotrecogin alfa, and ethacrynic acid). Of these four, the multilayered evaluation prioritized two: DPA showed significant inverse reporting associations for COVID-19-related adverse event terms for cilastatin (reporting odds ratio [ROR]: 0.19, 95% confidence interval [CI]: 0.09–0.43) and megestrol (ROR: 0.63, 95% CI: 0.42–0.94). Pathway profiling indicated that cilastatin shares molecular signatures with drugs investigated in COVID-19 clinical trials and megestrol with compounds reported as COVID-19 therapeutic candidates; gene expression analyses suggest cilastatin may exert antiviral and anti-inflammatory effects through DPEP1 inhibition. Conclusions: By integrating GATE-based prediction, real-world evidence, and phenome-level pathway profiling, this study enabled multidimensional candidate prioritization that single analytical methods do not provide. These findings, including the proposed mechanisms of action, are computational predictions that await experimental confirmation. They nonetheless provide mechanistic working hypotheses that may streamline subsequent preclinical studies. Prioritizing clinically approved drugs with established safety profiles may facilitate subsequent clinical evaluation, although the risks specific to each drug would need to be weighed. The framework offers a generalizable strategy for rapid candidate identification during pandemics and diseases with unmet therapeutic needs.