Drug Repurposing with AI: Finding New Uses for Old Medicines

Developing a new drug from scratch takes an average of 12 to 15 years and costs upwards of two billion dollars — and the odds of success hover below 10 percent. Yet sitting inside pharmacy warehouses and hospital formularies right now are thousands of molecules that have already cleared the most expensive hurdle in medicine: they are proven safe in humans. The question is whether any of those molecules can do something they were never asked to do.

Drug repurposing — identifying new therapeutic applications for existing approved or investigational compounds — has always had occasional breakthroughs, usually discovered by accident or by a perceptive clinician noticing an unexpected side effect. What has changed is the computational power now available to search for those matches systematically. Artificial intelligence can interrogate molecular databases, genomic atlases, electronic health records, and the entire published biomedical literature simultaneously, compressing a decade of serendipity into months of structured discovery.

Why Repurposing Is Faster Than Starting From Scratch

The Safety Advantage

The most time-consuming phases of drug development are early safety studies: establishing that a compound does not kill cells indiscriminately, does not cause organ toxicity at therapeutic doses, and does not interact catastrophically with common co-medications. A repurposed drug has already answered those questions. Phase I dose-finding trials are often shortened or skipped entirely because the pharmacokinetic profile is already characterised in humans. Regulators at the FDA and EMA have separate pathways — including repurposing supplements and 505(b)(2) applications — that acknowledge this prior evidence.

This safety shortcut is why repurposed drugs have historically reached patients roughly three to five years faster than novel compounds. When the COVID-19 pandemic created an urgent need for antivirals and anti-inflammatory agents, the fastest validated treatments — dexamethasone, baricitinib, remdesivir — were all repurposed or repositioned molecules. The speed was not coincidental; it was structural. Safety was already known, so trials could focus immediately on efficacy in the new indication.

Historical Repurposing Successes

Some of the most important medicines in use today reached their current indication through repurposing. Sildenafil was a cardiovascular compound before its now-primary use. Thalidomide, catastrophically misused in the 1960s for morning sickness, was later found to treat multiple myeloma through its anti-angiogenic and immunomodulatory properties and is now a standard-of-care agent. Metformin, prescribed for type 2 diabetes since the 1950s, is being actively investigated as a longevity and cancer-prevention compound. Aspirin was analgesic before it became the cornerstone of cardiovascular prevention. These discoveries happened through observation, not design — which is exactly the gap AI is positioned to close.

How AI Approaches Drug Repurposing

Network Pharmacology and Knowledge Graphs

Biological systems are not lists of molecules — they are networks of interactions. A drug that binds a single protein ripples outward through protein-protein interaction networks, gene regulatory cascades, and metabolic pathways. Network pharmacology models those ripple effects computationally. AI systems construct knowledge graphs in which nodes represent genes, proteins, diseases, drugs, and phenotypes, and edges represent experimentally validated relationships between them. When a known drug activates or inhibits a node that sits inside a disease network, the algorithm flags a potential therapeutic connection worth investigating.

This approach has particular power for complex diseases where no single target dominates. In Alzheimer's disease, for example, researchers have used knowledge graph methods to identify drugs approved for unrelated conditions — including some anti-inflammatory and metabolic agents — whose network footprints overlap substantially with the molecular signatures of neurodegeneration. The same logic underpins work in quantum drug discovery pipelines, where the combinatorial complexity of molecular interactions exceeds what classical optimisation can handle efficiently.

Genomic and Transcriptomic Signature Matching

One of the most powerful repurposing strategies compares disease gene expression signatures with drug-induced gene expression signatures. Databases like the Connectivity Map (CMap) and the Library of Integrated Network-Based Cellular Signatures (LINCS) contain transcriptomic profiles of thousands of cell lines treated with thousands of compounds. If a disease upregulates a set of genes and a drug downregulates exactly those genes — or vice versa — the drug may reverse the disease signature. Machine learning models trained on these datasets can now screen millions of drug-disease pairs in hours.

This signature-matching strategy aligns naturally with AI genomics and machine learning approaches to DNA analysis, where large language models trained on biological sequences can predict how a given compound will shift the transcriptional landscape of a target tissue before any experiment is run.

Protein Structure-Based Repurposing

The revolution in protein structure prediction — exemplified by AlphaFold and its successors — has opened a structural dimension to repurposing. If AI can predict the three-dimensional shape of a disease-relevant protein that was previously undrugged, it can then ask which existing drug molecules have binding pockets that geometrically complement that structure. Protein folding advances powered by quantum computing are extending this further, enabling conformational sampling of protein flexibility that static crystal structures cannot capture — revealing cryptic binding sites invisible to classical docking simulations.

Electronic Health Records as Repurposing Evidence

Real-World Data as a Discovery Engine

Clinical trial populations are small and carefully selected. Real-world patient populations are enormous and messy — but they contain something trials do not: unplanned natural experiments. When millions of patients are prescribed a drug for one condition, some of them inevitably have a second condition as a comorbidity. If the outcomes for that second condition are systematically better in drug-exposed patients than in unexposed patients, that signal is a repurposing hypothesis hiding inside routine care data.

AI systems trained on de-identified electronic health record (EHR) datasets can detect these patterns at scale. Propensity score matching and causal inference methods allow researchers to control for confounders — the fact that patients prescribed a drug are different from those who are not — and isolate genuine pharmacological effects. This approach identified metformin's potential anti-cancer properties from EHR data years before the prospective trials that are now testing it formally.

Pharmacovigilance and Adverse Event Inversion

Adverse event reports submitted to regulatory pharmacovigilance databases — the FDA's FAERS system, for example — are typically mined for safety signals. AI repurposing researchers have inverted this logic: if a drug causes an unexpectedly low rate of a particular complication in patients who take it for another reason, that protective effect might be a therapeutic signal. This adverse-event inversion approach has generated hypotheses for cardiovascular, neurological, and metabolic indications from drugs approved in completely unrelated areas.

Precision Repurposing: Matching the Right Drug to the Right Patient

Beyond One-Size-Fits-All Repositioning

Early repurposing efforts asked a blunt question: can Drug A treat Disease B? Precision medicine reframes the question: can Drug A treat Disease B in the subset of patients who carry Variant C in Gene D? This distinction matters enormously for efficacy. A repurposed drug that shows no average benefit across a heterogeneous patient population may show dramatic benefit in a genomically defined subgroup — and conversely, a drug that appears moderately effective on average may be causing harm in a subset while helping the rest.

AI makes stratified repurposing computationally tractable. By integrating genomic data with drug response data from cell lines, organoids, and patient cohorts, machine learning models can identify biomarkers that predict response to a repurposed candidate. This is the same logic that drives pharmacogenomics — the study of how genetic variation determines drug response — applied now to a much wider universe of potential drug-disease pairs than pharmacogenomics traditionally considered.

Oncology as a Model System

Cancer has become the proving ground for precision repurposing. Tumours are defined molecularly — by their mutational landscape, copy number alterations, epigenetic state, and gene expression signature — rather than solely by their tissue of origin. A drug approved for breast cancer that targets a specific kinase mutation may be equally effective against a lung cancer, a glioma, or a sarcoma that carries the same mutation. This basket trial logic, enabled by comprehensive molecular tumour profiling of the type described in work on precision oncology and tumour profiling, is now routinely identifying repurposing opportunities across cancer types in ways that tissue-centric oncology would have missed entirely.

Quantum Computing's Role in Molecular Repurposing

Simulating Binding Interactions at Quantum Accuracy

Classical AI can predict drug-protein binding affinities with impressive accuracy for well-characterised systems, but it relies on approximations when modelling the quantum mechanical behaviour of electrons during bond formation. Quantum computers running variational quantum eigensolver (VQE) algorithms can, in principle, simulate those electron correlations exactly — producing binding energy predictions that classical molecular dynamics cannot match. For repurposing, this matters when the drug-target interaction involves subtle electronic effects: unusual metal coordination, proton transfer reactions, or binding sites where small changes in electron density alter affinity by orders of magnitude.

Current quantum hardware is still in the noisy intermediate-scale quantum (NISQ) era, limiting the size of molecules that can be simulated exactly. But hybrid classical-quantum workflows — where quantum processors handle the most computationally demanding quantum chemistry steps and classical processors handle the rest — are already demonstrating advantages for small molecule fragments that appear in many approved drugs. As fault-tolerant quantum hardware matures, the scope of exact simulation will expand to cover full drug molecules and their target binding sites, as explored in depth in the comparison between quantum simulation and classical pharmaceutical methods.

Combinatorial Optimisation for Polypharmacology

Many diseases — particularly psychiatric conditions, metabolic syndrome, and neurodegenerative disorders — are not driven by a single aberrant protein but by dysregulation across multiple interacting pathways. Effective treatment may require modulating several targets simultaneously, a concept called polypharmacology. Finding the optimal combination of repurposed drugs — which compounds to combine, at what ratio, targeting which set of nodes in the disease network — is a combinatorial optimisation problem that grows exponentially with the number of candidates considered. Quantum optimisation algorithms, including the Quantum Approximate Optimisation Algorithm (QAOA), are well-suited to this class of problem and represent one of the near-term practical applications being developed by pharmaceutical companies actively investing in quantum hardware, a landscape surveyed in coverage of quantum computing investments by pharma companies.

Clinical Validation: From AI Prediction to Patient Benefit

The Gap Between Prediction and Proof

AI repurposing generates hypotheses — it does not generate treatments. The distance between a high-confidence computational prediction and an approved new indication remains substantial. Biological systems are deeply context-dependent: a drug that reverses a disease gene expression signature in a cancer cell line may have no effect in primary patient cells, may be metabolised differently in the organ system affected by the disease, or may achieve insufficient tissue concentrations at tolerable doses. These failures are not unique to AI-generated hypotheses, but they are a reminder that computational discovery accelerates the front of the pipeline without eliminating the need for rigorous experimental validation.

The field is addressing this validation gap through adaptive clinical trial designs — including platform trials that can test multiple repurposing candidates simultaneously against the same indication, with pre-specified interim analyses that allow early stopping for efficacy or futility. AI is contributing here too: predictive models of trial success probability, patient stratification algorithms that enrich for likely responders, and synthetic control arms generated from historical trial data can all reduce the cost and duration of repurposing validation studies. The intersection of AI and trial design is explored in detail in the context of quantum computing applications in clinical trials.

Regulatory Considerations

Regulatory agencies are adapting their frameworks to accommodate AI-generated repurposing evidence. The FDA's Real-World Evidence program, its Complex Innovative Trial Design meetings, and its emerging guidance on AI in drug development all create pathways for incorporating computational evidence into approval submissions — not as a substitute for randomised trial data, but as supportive evidence that can strengthen the overall package. In the European Union, the European Medicines Agency has similarly published guidance on the use of real-world data in regulatory decision-making. The institutional infrastructure for AI repurposing is being built in parallel with the science itself.

The Future of Drug Repurposing at Scale

Foundation Models for Drug Discovery

The same architectural advances that produced large language models for text — transformers trained on massive datasets with self-supervised objectives — are now being applied to biology. Foundation models pre-trained on protein sequences, molecular graphs, genomic data, and clinical records can be fine-tuned for repurposing tasks with relatively little labelled data. This transfer learning paradigm means that a model trained on the structural biology of thousands of proteins can be rapidly adapted to predict interactions with a newly characterised disease target, even when experimental data for that specific target is sparse.

These models are becoming multimodal — integrating sequence, structure, gene expression, clinical, and imaging data into unified representations. A single model can simultaneously process a patient's genomic profile, their tumour's transcriptomic signature, and a library of drug molecular graphs to rank repurposing candidates personalised to that individual. This is not a distant projection; early versions of such systems are running in academic medical centres and biotechnology companies today, though their clinical deployment remains limited by validation requirements and regulatory frameworks that are catching up to the technology.

Open Science and Data Sharing

The quality of AI repurposing predictions is directly proportional to the quality and quantity of training data. The field is being shaped by the degree to which biomedical data is shared openly — through initiatives like the Open Targets platform, the NIH National COVID Cohort Collaborative, the UK Biobank, and commercial data-sharing consortia between pharmaceutical companies. Each wave of new open data — a new proteomics atlas, a new patient cohort, a new cell line drug response database — expands the search space in which AI can find repurposing connections that were previously invisible.

The tension between data sharing and commercial incentive remains real. Companies that invest heavily in generating proprietary biological datasets have limited motivation to share them freely, even when doing so would accelerate discovery that could eventually benefit them through licensing or partnership. Federated learning architectures — where models are trained across distributed datasets without the underlying patient data ever leaving its source institution — offer a partial resolution, enabling AI to learn from data that cannot practically be centralised. The principles behind this approach connect directly to broader questions about federated learning in healthcare and how sensitive biological data can be used for discovery without being exposed.