A Siamese Neural Network framework for computational drug repositioning using literature-derived biomedical knowledge

Traditional drug development remains costly, time-consuming, and characterized by high attrition rates, motivating increased interest in drug repositioning as a strategy to identify new therapeutic indications for existing compounds. A central challenge in computational repositioning is the extraction and utilization of drug–disease associations embedded within large, unstructured biomedical literature. In this work, we present a data-driven computational framework for drug repositioning based on heterogeneous Siamese Neural Networks (SNNs). The proposed framework integrates literature-derived relational triples from SemMedDB with FastText-based word embeddings to learn pairwise similarity functions between biomedical entities. Through hyperparameter optimization and controlled validation, the model achieved a validation accuracy of 87.66% and a test accuracy of 83%, with balanced precision, recall, and F1 scores. These results indicate that the learned similarity space effectively captures semantic relationships relevant to therapeutic repurposing. The framework provides a scalable approach for prioritizing candidate drug–disease pairs, supporting early-stage pharmaceutical research and decision-making. By coupling representation learning with curated biomedical sources, this study contributes a reproducible pipeline at the intersection of machine learning and pharmaceutical systems engineering, enabling informed exploration of potential therapeutic alternatives.