Abstract
Background
Artificial intelligence (AI)-assisted drug repurposing seeks to identify new therapeutic applications for existing drugs. Nevertheless, most of the existing studies are centered on prediction scores and candidate ranking, and there is still relatively little attention paid to the pharmacological mechanisms, safety and translational validation.
Objective
This critical review summarizes existing AI approaches to drug repurposing, identifies their mechanistic, safety, and translational barriers, and incorporates the emerging evidence gaps into a mechanism-driven and safety-proven translational pharmacology model.
Findings
The literature reviewed shows that machine learning, deep learning, network medicine, knowledge graphs, graph neural networks, and foundation models have the potential to significantly increase the number of candidates identified, but they have marked variations in terms of interpretability, biological support, safety integration, external validation, and clinical applicability. In all of these strategies, common shortcomings are inadequate mechanism validation, poor causal evidence, incomplete dose and safety evaluation, weak pharmacovigilance integration, and reliance on internal validation, and varying experimental or real-world confirmatory evidence. These persistent evidence gaps were incorporated into a sequence of steps that incorporates candidate predictions, mechanism mapping, determination of causal plausibility, safety and toxicity evaluation, pharmacovigilance, real-world validation, expert pharmacology review and evidence-based candidate prioritization.
Conclusion
AI-driven drug repurposing presents huge potential in the field of therapeutic innovation, but current applications require more clinical evidence and pharmacological mechanisms demonstrating safety, as well as clinical relevance. The proposed framework can support the more reliable prioritization of candidates for experimental testing, real world validation and future clinical testing.