AI drug repurposing: the hurdles, the promise and what it means for patients
Photo: N43 and HermesDrug repurposing searches for new uses for medicines that already have human safety information. Artificial intelligence can connect disease biology, molecular targets and clinical evidence faster, but validation, regulation, incentives and patient safety remain the decisive hurdles.
01What drug repurposing is and why it matters
Drug repurposing, or repositioning, investigates whether an existing medicine can treat a different disease or patient population. Because pharmacology, manufacturing and some safety information may already be known, the route can be faster than discovering a molecule from scratch.
It is not a shortcut around clinical evidence. A dose that is safe for one condition may not be effective or safe in another, and a new population can expose interactions or adverse effects that were not visible in the original indication.
02How AI is accelerating drug repurposing
AI can search across biomedical literature, molecular graphs, gene-expression signatures, clinical records and trial registries. It can propose links between a disease mechanism and a compound, rank candidates and reveal evidence that is distributed across many databases.
The gain is primarily in prioritization. A model reduces the number of hypotheses researchers must investigate; it does not turn an association into a therapeutic effect. The best workflows combine machine ranking with pharmacology, disease expertise and experiments.
03The biggest hurdles in the process
Data quality is the first bottleneck. Biomedical datasets contain missing outcomes, inconsistent terminology, publication bias and populations that do not represent future patients. Models can be precise about an artifact if the artifact is embedded in the training data.
The second hurdle is translation. A drug may influence a pathway in cells but fail in a living organism because of exposure, tissue penetration, compensatory biology or toxicity. Repurposing still needs a plausible mechanism, a workable dose and a trial that can answer the question.
04Success stories of repurposed drugs
Well-known examples include thalidomide, which was later repurposed under strict controls for multiple myeloma and other diseases, and sildenafil, first studied for cardiovascular uses before its efficacy in erectile dysfunction was recognized. These stories show that new indications can emerge from unexpected evidence.
They also show why history should not be mistaken for an AI success rate. Each case depended on clinical observation, biological reasoning, safety controls and eventual trials. A retrospective example validates the concept, not every computational prediction.
05The regulatory pathway for repurposed drugs
A new indication generally requires evidence that the benefit-risk balance is favorable for the proposed use. Sponsors may need preclinical work, an investigational application, clinical trials and a regulatory submission, even when the active ingredient is already approved for something else.
Existing manufacturing and safety data can reduce duplication, but regulators still examine dose, formulation, population, endpoints and interactions. Off-label prescribing is not the same as formal approval or proof of efficacy.
06How AI identifies new uses for existing drugs
Common methods include knowledge graphs linking drugs, targets and phenotypes; transcriptomic signatures that compare disease states with drug responses; molecular simulations; and natural-language models that extract hypotheses from publications. Ensemble approaches can reduce dependence on one data source.
Researchers then test the ranking. They may use cell models, organoids, animal studies or retrospective clinical cohorts before starting a prospective trial. A transparent model that exposes evidence and uncertainty is more useful than an opaque score with false precision.
07What the future of drug repurposing looks like
AI-assisted repurposing is likely to become a standard hypothesis-generation layer in pharmaceutical and academic research. Its value will be measured by validated candidates, faster trials and better patient selection, not by the number of associations a model can produce.
The biggest opportunity is a learning loop in which trial results and real-world evidence improve the next search while protecting patient privacy. If incentives reward negative results and data quality, repurposing can turn existing medicines into new options without lowering the bar for proof.
