Drug Repurposing: How AI Is Finding New Uses for Old Compounds
Drug repurposing has produced some of medicine's most useful treatments. AI is accelerating the identification of candidates - but the gap between computational signal and clinical validation remains wide.
Sildenafil was originally developed as a treatment for angina and hypertension. It failed those indications in phase I trials but showed an unexpected side effect that participants were reluctant to report. Thalidomide was withdrawn from the market in 1961 after causing thousands of birth defects, then re-approved decades later for multiple myeloma and leprosy. Aspirin has been continuously repurposed since its introduction - from analgesic, to antiplatelet, to candidate cancer preventive. Drug repurposing is not new. What is new is the scale and speed at which computational approaches can generate candidate hypotheses.
The Economic and Regulatory Case
Developing a new molecular entity from scratch takes an average of more than a decade and costs, by widely cited estimates, well over a billion dollars - with a failure rate in clinical development that exceeds 90%. A repurposed drug enters clinical evaluation with an existing safety profile, known pharmacokinetics, and often an existing manufacturing process. Phase I safety trials can sometimes be abbreviated or eliminated for low doses in new indications. Regulatory pathways for repurposed drugs, including the FDA's orphan drug designation and the European Medicines Agency's procedures for established medicinal products, offer incentives that can make the economics of repurposing viable even for rare diseases with small markets.
The regulatory and economic case for repurposing is strong - but it is not sufficient on its own to validate a candidate. The history of repurposing includes many computational or observational signals that did not survive clinical testing.
How AI Generates Repurposing Candidates
AI-driven repurposing approaches fall into several broad categories. Network-based methods construct drug-disease networks using protein-protein interaction data, drug-target binding data, and disease gene associations, then identify drugs whose targets are connected to disease pathways through network proximity metrics. Transcriptomic approaches compare the gene expression signatures of diseases with the signatures of drug perturbations - a drug whose signature is the inverse of a disease's signature may reverse the disease state. This approach, popularised by the Connectivity Map (CMap) at the Broad Institute, has generated numerous testable hypotheses.
Electronic health record mining identifies drug-disease associations from real-world prescribing data and outcome patterns - patients on drug X for condition A who also have condition B show unexpectedly better outcomes in B. A 2020 paper in Nature Medicine demonstrated how EHR data could be used systematically to generate repurposing hypotheses that outperformed random selection in subsequent validation.
Real Examples and What They Show
Metformin is one of the most-discussed repurposing candidates in oncology. Observational data consistently shows that diabetic patients on metformin have lower cancer incidence and better cancer outcomes than those on other diabetes medications. The biological mechanism is plausible - metformin activates AMPK, which inhibits mTOR, a key driver of cancer cell proliferation. Computational models have flagged it repeatedly. Yet multiple randomised controlled trials in cancer indications have failed to show the expected benefit. The observational signal was real; the causal interpretation was probably wrong, or the effect was smaller than the confounded observational data suggested.
This pattern - a strong computational and observational signal that does not translate to RCT validation - is common enough in repurposing to be considered the baseline expectation rather than an exception.
The Evidence Gap
The fundamental challenge in AI-driven repurposing is that the data on which models are trained - gene expression databases, protein interaction networks, EHR records - reflects biological associations and real-world correlations, not causal mechanisms. A drug whose target is proximal to a disease pathway in a network model may not modulate that pathway at clinically relevant concentrations, or may modulate it in tissues other than the relevant one, or may have off-target effects that outweigh the on-target benefit in the new indication.
AI can narrow the search space from thousands of candidates to dozens, and that is genuinely valuable. What it cannot do is replace clinical validation - and the evidence gap between computational signal and clinical efficacy remains the primary bottleneck in repurposing pipelines.
What Evidence Synthesis Adds
Before a repurposing candidate enters clinical evaluation, a systematic review of existing evidence - preclinical data, observational studies, off-label use data, related indication trials - can provide a more honest assessment of the signal strength than the computational prediction alone. It can identify whether the observational associations that support the hypothesis are consistent across populations and data sources, or confined to specific settings. It can reveal whether mechanistic studies support the biological rationale, or whether the signal is purely statistical. And it can estimate the prior probability of success that should inform the design of the clinical trial - including whether a full phase II programme is warranted or whether a smaller signal-finding study is the more efficient next step.
At NousLab, we support repurposing teams at the evidence synthesis stage - the moment between computational hypothesis and clinical protocol where the quality of the evidence review most directly affects the quality of the investment decision. Reach out to discuss your repurposing programme.
