Reviewing the Computational Landscape of Drug Repurposing: Evolution from Structure-Based Methods to LLM-Based Methods

Traditional drug discovery is a high-risk, time-consuming, and costly endeavor. Drug repurposing has emerged as a pivotal strategy to overcome these challenges by identifying new therapeutic indications for approved drugs, thereby significantly reducing development timelines, costs, and safety risks. This review aims to provide a comprehensive methodological survey of computational strategies for drug repurposing. It seeks to clarify the core principles, applicability, and limitations of various approaches, offering a clear technological landscape and valuable insights for future research directions. We categorize and elaborate on the prevailing methodologies, following a logical progression. The review begins with biological mechanism-driven methods, including structure-based, omics-based, fuzzy logic-based, and adverse event-based methods. It then details network-based methods that integrate multi-source data, encompassing graph mining and matrix factorization/completion techniques. Finally, we explore data-driven paradigms, tracing the evolution from traditional text mining-based methods to cutting-edge large language model (LLM)-based methods. Each methodological category presents unique advantages and challenges. While structure-based, omics-based, fuzzy logic-based, and adverse event-based methods provide deep mechanistic insights, network-based methods enable systematic prediction. Text mining unlocks information from vast literature, a potential greatly amplified by LLMs. This review highlights that the future of drug repurposing lies in the intelligent integration of diverse methodologies. In the future, we believe that network-based methods and data-driven methods will mark the beginning of large-scale drug repurposing, but ultimately, biological mechanism-driven methods will still be necessary for rigorous validation and explanation.