AI-enabled drug repurposing in Alzheimer's disease: from computational nomination to pharmacological proof.

Artificial intelligence (AI) has expanded the scale and speed of drug repurposing in Alzheimer's disease (AD). Transcriptomic signature matching, network-based systems pharmacology, real-world data pharmacoepidemiology and multi-omics machine learning have generated hundreds of candidate drugs, while repurposed agents account for 35% (56 of 158) of drugs and 42% (15 of 36) of Phase 3 agents in the 2026 AD pipeline. Clinical translation, however, remains unproven. Randomised trials of GLP-1 receptor agonists illustrate the gap between prioritisation and therapeutic proof: semaglutide did not slow clinical progression in evoke and evoke+, while liraglutide did not improve the primary FDG-PET endpoint in ELAD. We examine four complementary evidence domains, defining what each can and cannot establish, and position large language models as a cross-cutting tool for evidence synthesis and knowledge-graph reasoning rather than an independent source of biological evidence. Rather than cataloguing methods and candidates, we appraise leading drugs at the individual-molecule level and propose explicit criteria for progression from computational nomination to clinical testing. The framework defines seven requirements for clinical progression: disease-stage matching, adequate site-of-action exposure, including CNS exposure for central mechanisms, mechanistic validation in human-relevant models, biomarker-stratified trial design, convergence across independent evidence domains, causal inference with explicit boundary conditions, and mechanism-aligned endpoints with predefined stopping rules. AI has improved candidate prioritisation in AD, but clinical value will depend on converting ranked hypotheses into pharmacologically plausible and prospectively testable programmes.