Abstract: Objective This study aimed to predict the similarity in the mechanisms of action for different drugs, facilitate the efficient screening of candidate therapeutics for ischemic stroke, and enhance drug development efficiency through a modified variational autoencoder-based neural network. Methods A dual encoder-decoder variational autoencoder (VAE), termed hyperbolic contrastive VAE (HCVAE), was developed to extract features from training datasets comprising integrated drug-perturbed single-cell RNA sequencing (RNA-seq) gene expression profiles. Phenotypic effect features of drugs were mapped into a hyperbolic latent space. Drug repositioning was performed based on the distances between drug feature representations in the latent space to identify candidate compounds. Biological effects were preliminarily validated through bioinformatics analyses. Results The constructed HCVAE model demonstrated superior performance compared to standard VAEs in effectively extracting drug phenotypic features. Cerdulatinib was identified as a candidate therapeutic exhibiting effects comparable to edaravone in ischemic stroke models. Further analysis indicated that cerdulatinib,s phenotypic effects are associated with the regulation of key proteins RIPK1 and caspases within the cellular apoptosis pathway. Conclusion The HCVAE model effectively extracts“drug-disease”phenotypic features and can serve as a robust platform for candidate drug screening, providing novel insights for drug repositioning research.