AI Drug Repurposing: How Artificial Intelligence Finds New Uses for Existing Medicines
Artificial intelligence can analyze enormous networks of information linking drugs, proteins, genes, pathways, diseases, clinical records, biomedical publications and molecular signatures. These systems can identify relationships that may be difficult to see using conventional literature review or one-disease-at-a-time research.
But an AI prediction is not a treatment. The most important distinction is between a computationally interesting candidate and a clinically validated repurposed medicine. Recent reviews and research on AI drug repurposing emphasize this translational gap.
What Is Drug Repurposing?
Drug repurposing, also called drug repositioning, means finding a new therapeutic use for a medicine that already exists or has already been investigated.
The candidate may be:
- An FDA-approved drug being investigated for a completely different disease.
- An approved drug being studied in a different patient population.
- An investigational drug being redirected toward another indication.
- A medicine being evaluated at a different dose, schedule or combination.
The FDA defines drug repurposing as identifying potential new uses, such as a new indication or population, for FDA-approved drugs where safety and effectiveness data can support the proposed use. In 2026, the FDA also initiated a broader effort to gather input on prioritizing drug-repurposing opportunities for unmet medical needs.
Why Repurpose Existing Drugs?
Traditional drug development can require years of laboratory research, toxicology testing, clinical trials and regulatory review. Repurposing can start with a drug whose pharmacology, manufacturing and some safety characteristics are already known.
That does not mean a repurposed drug automatically bypasses clinical testing. A drug that is safe for rheumatoid arthritis, for example, may have a different benefit-risk profile when used in cancer, infectious disease or another population.
NCATS has long supported efforts to identify new therapeutic uses for existing drugs and compounds as a way to accelerate translation of biomedical discoveries into potential treatments.
Why Does AI Help?
The modern biomedical literature contains an extraordinary number of relationships:
| Data Type | Examples | How AI Can Use It |
|---|---|---|
| Drugs | Chemical structures, targets, indications | Compare molecular and therapeutic relationships |
| Genomics | Mutations, disease-associated genes | Identify molecular vulnerabilities |
| Transcriptomics | Gene-expression signatures | Find drugs that may reverse disease-associated patterns |
| Proteomics | Protein abundance and interactions | Map drug-target and pathway relationships |
| Clinical data | Electronic health records and outcomes | Search for real-world treatment-response signals |
| Literature | Papers, abstracts and biomedical knowledge | Extract relationships and discover hidden connections |
| Knowledge graphs | Drug-gene-disease networks | Predict previously unrecognized drug-disease links |
Modern AI systems can integrate these heterogeneous data sources rather than examining each one independently. This is one reason knowledge graphs and graph neural networks have become important in computational drug repurposing.
How AI Drug Repurposing Works
A simplified AI repurposing pipeline looks like this:
Biomedical data → disease signature → drug representation → AI model → candidate ranking → biological validation → clinical testing → regulatory decision
Step 1: Define the Disease
The first challenge is determining what the model is actually trying to treat.
A broad label such as "cancer" may be too vague. A modern repurposing model can potentially work with more specific concepts such as:
- A molecular subtype.
- A particular mutation.
- A dysregulated pathway.
- A treatment-resistant phenotype.
- A rare disease with limited treatment options.
- A specific clinical phenotype or patient population.
This matters because diseases that share the same conventional diagnosis can have very different molecular mechanisms.
Step 2: Build a Biological Representation
Researchers can represent a disease using multiple layers of information, including:
- Genes and mutations.
- RNA expression.
- Proteins and protein-protein interactions.
- Metabolic pathways.
- Cellular phenotypes.
- Clinical phenotypes.
- Published disease mechanisms.
The objective is to transform a complex biological condition into a structured representation that a computational model can analyze.
Step 3: Represent the Drug
The AI system can represent a drug using characteristics such as:
- Chemical structure.
- Molecular targets.
- Known mechanisms of action.
- Pathway effects.
- Known indications.
- Adverse effects.
- Pharmacological properties.
A drug therefore becomes more than a name. It becomes a multidimensional biological object.
Step 4: Search for Hidden Connections
This is where machine learning becomes especially useful.
Instead of asking only, "What drugs are already used for this disease?", the model can ask:
Which drugs interact with biological systems that resemble the disease's disrupted systems?
This is fundamentally a network problem.
Knowledge Graphs: The AI Map of Medicine
One of the most important concepts in modern AI repurposing is the knowledge graph.
Imagine a network containing:
Drug → Protein → Pathway → Gene → Disease → Symptom → Clinical phenotype
Each relationship becomes an edge in a graph.
A knowledge graph might contain a known relationship such as:
Drug A → inhibits Protein B
and:
Protein B → contributes to Pathway C
while:
Pathway C → is dysregulated in Disease D
An AI system can therefore investigate whether Drug A → Disease D is a plausible new relationship.
Knowledge-graph approaches have become a major area of computational repurposing research, with newer systems moving from conventional machine learning toward graph neural networks and other deep-learning approaches.
Graph Neural Networks
Graph neural networks (GNNs) are designed to learn from interconnected data.
Rather than treating a drug, gene or disease as an isolated variable, the model learns from the structure of the network around it.
This is particularly useful for diseases because biological mechanisms are rarely isolated. A single disease can involve numerous genes, proteins, signaling pathways and cellular processes.
A 2025 review of knowledge-graph methods described the progression from classical machine-learning approaches to graph neural networks for drug-repurposing prediction and highlighted both their predictive potential and their limitations.
AI Can Also Compare Molecular Signatures
Another important strategy is to compare gene-expression or molecular signatures.
Suppose a disease causes a characteristic expression pattern:
Gene A ↑
Gene B ↑
Gene C ↓
Gene D ↓
Researchers can search for medicines that produce a contrasting signature:
Gene A ↓
Gene B ↓
Gene C ↑
Gene D ↑
The hypothesis is that the drug may counteract some of the biological changes associated with the disease.
This is not proof of therapeutic benefit. It is a mechanism for generating testable hypotheses.
Machine Learning for Drug-Target Interaction Prediction
AI can also predict whether a drug may interact with a particular protein or biological target.
Models can incorporate:
- Molecular fingerprints.
- Protein sequences.
- Protein structures.
- Drug-target interaction databases.
- Known pharmacological relationships.
- Experimental assay data.
The output may be a predicted probability or ranking rather than a confirmed interaction.
Reviews of AI repurposing published in 2025 and 2026 describe the increasing use of machine learning, deep learning, natural-language processing, knowledge graphs and multi-omics data to prioritize candidate medicines.
Natural Language Processing: Mining the Medical Literature
Biomedical knowledge is not stored only in databases. A huge amount exists inside research papers.
Natural-language processing (NLP) can help extract statements such as:
- Drug X binds Protein Y.
- Protein Y is elevated in Disease Z.
- Pathway A contributes to treatment resistance.
- Patients receiving Drug X experienced outcome Y.
This allows researchers to construct or update knowledge graphs at a scale that would be difficult through manual curation alone.
The baricitinib COVID-19 example is illustrative: BenevolentAI described an AI-enhanced biomedical knowledge graph combined with machine-learning literature extraction and human-guided analysis to identify baricitinib as a candidate.
Foundation Models and Zero-Shot Drug Repurposing
One of the more interesting developments is zero-shot drug repurposing.
Traditional computational systems often perform best when the disease already has substantial biological and treatment information.
But what happens when a disease is poorly characterized or has few treatment options?
Researchers at Harvard and collaborators developed TxGNN, a graph-based foundation model designed to predict therapeutic candidates for diseases with limited treatment information.
The model was trained on a medical knowledge graph covering 17,080 diseases and used graph neural networks and metric learning to predict possible drug-disease relationships. The researchers also developed an explanatory component designed to identify knowledge paths supporting predictions.
This is an important direction for rare diseases and other areas where conventional clinical datasets may be sparse.
A Real-World Example: Baricitinib and COVID-19
One of the clearest examples of AI-assisted repurposing involved baricitinib, a drug originally approved for rheumatoid arthritis.
During the COVID-19 pandemic, BenevolentAI used an AI-enhanced biomedical knowledge graph and human expert analysis to identify baricitinib as a candidate for COVID-19 because of its potential relevance to both viral processes and inflammatory pathways.
The subsequent evidence did not consist of the AI prediction alone. Clinical trial data were generated, regulatory review occurred, and the FDA issued an emergency authorization in November 2020. In May 2022, FDA approved baricitinib for treatment of COVID-19 in hospitalized adults requiring specified levels of respiratory support.
This illustrates the correct interpretation of AI-assisted repurposing:
AI hypothesis → laboratory/clinical evidence → regulatory evaluation → clinical use
AI was part of the discovery process, not a substitute for clinical evidence.
AI Drug Repurposing Is Not the Same as AI Drug Discovery
These two concepts are increasingly confused.
| AI Drug Repurposing | AI New-Drug Discovery |
|---|---|
| Starts with an existing medicine or compound | Can start with a target and design a new molecule |
| Looks for new indications | Looks for new chemical entities or biologics |
| Can leverage existing pharmacology and safety information | Usually requires more extensive characterization |
| Often uses knowledge graphs and clinical/biological data | Often uses molecular design, virtual screening and structure-based modeling |
| Primary question: "Where else could this drug work?" | Primary question: "What new molecule might work?" |
Some modern platforms combine both strategies.
Where AI Can Be Especially Valuable
Rare Diseases
Rare diseases can have limited commercial incentives, small patient populations and fragmented evidence.
AI can search across biological relationships and identify candidates that would be difficult to prioritize manually.
Cancer
 |
| Read more: Top 30 Repurposed Drugs and Metabolic Interventions to Control Cancer |
Cancer is a particularly interesting application because different tumor types can share:
- Mutations.
- Signaling pathways.
- Metabolic vulnerabilities.
- DNA-repair defects.
- Resistance mechanisms.
Instead of asking only, "What drugs treat lung cancer?", an AI system can potentially ask:
Which existing drugs interact with this tumor's specific molecular vulnerabilities?
This may eventually support more granular repurposing strategies involving biomarkers, tumor phenotypes and mechanisms of resistance.
Neurological Diseases
Neurological disorders are another important area because many have complex, poorly understood biology and limited treatment options.
Multi-omics and network-based approaches may help identify connections between seemingly unrelated diseases and therapeutic targets.
Infectious Diseases
Rapid repurposing became particularly visible during COVID-19, when researchers urgently needed treatment candidates before entirely new therapeutics could be developed.
AI can rapidly screen large numbers of existing compounds against pathogen biology and host response pathways.
AI Can Analyze Contraindications Too
A major advance is the recognition that a useful repurposing model should not only ask "Could this drug work?"
It should also ask:
"Could this drug be harmful in this disease or patient population?"
TxGNN explicitly modeled both potential indications and contraindications, illustrating the importance of incorporating negative as well as positive predictions.
This is particularly important because an apparently attractive biological mechanism may coexist with a dangerous safety signal.
Why an AI Prediction Can Be Wrong
AI systems are powerful pattern-recognition machines, but biological systems contain enormous complexity.
Common failure modes include:
1. Poor-Quality Training Data
AI models learn from existing data. If the underlying data are incomplete, biased, duplicated or incorrectly annotated, model predictions may inherit those problems.
2. Data Leakage
A model may appear highly accurate because information from the test set was indirectly available during training.
3. Shortcut Learning
A model may learn an easy statistical relationship rather than the underlying biological mechanism.
The TxGNN researchers specifically discussed this problem in conventional repurposing benchmarks and designed more demanding zero-shot evaluations to reduce such shortcuts.
4. Correlation Is Not Causation
Two biological variables can be strongly associated without one causing the other.
An AI model can find the correlation. Experimental science must determine whether the relationship is causal and therapeutically useful.
5. Molecular Activity May Not Translate Into Clinical Benefit
A drug may interact with a target in a laboratory model but fail in humans because of:
- Pharmacokinetics.
- Tissue penetration.
- Protein binding.
- Dose limitations.
- Drug interactions.
- Off-target effects.
- Disease heterogeneity.
- Insufficient exposure at the target tissue.
6. The Disease Label May Be Too Broad
A treatment might work for a biomarker-defined subgroup but have no meaningful effect across the entire disease population.
The Biggest Bottleneck: Clinical Validation
This is arguably the most important point in the entire field.
AI can produce thousands of candidate hypotheses much faster than researchers can test them.
The bottleneck therefore shifts from finding ideas to proving which ideas are clinically meaningful.
A 2026 perspective in Nature Reviews Drug Discovery argued that although AI methods in drug discovery have advanced substantially, evidence of clinically relevant impact remains limited and that evaluation should focus more heavily on whether AI improves real-world decision-making rather than only computational benchmark performance.
An Evidence Ladder for AI Drug Repurposing
One useful way to evaluate an AI-generated repurposing claim is to separate the evidence into levels.
| Level | Evidence | What It Means |
|---|---|---|
| E0 | AI prediction only | Hypothesis generation |
| E1 | Computational + mechanistic support | Biologically plausible candidate |
| E2 | Cellular or laboratory validation | Experimental activity demonstrated |
| E3 | Animal/preclinical evidence | In-vivo evidence supports further development |
| E4 | Human clinical evidence | Safety and/or efficacy evaluated in humans |
| E5 | Regulatory approval for the new indication | The repurposed use has passed regulatory review |
This framework helps prevent a common internet error: presenting an E0 AI prediction as if it were an E5 approved treatment.
AI Repurposing vs Off-Label Use
These are not identical.
Off-label use occurs when a clinician uses an approved medicine for a purpose that is not included in its approved labeling, where permitted by applicable law and clinical practice standards.
Drug repurposing research is the broader development process of investigating whether a medicine has a scientifically supported new indication.
FDA's CURE ID platform, developed with NIH/NCATS, is one example of an initiative that captures clinical experiences involving novel uses of existing drugs, particularly in difficult-to-treat infectious diseases.
How Researchers Validate an AI Candidate
A sensible validation sequence may include:
1. Computational replication
Can an independent dataset or model reproduce the prediction?
2. Mechanistic evaluation
Does the proposed biological mechanism make sense?
3. Laboratory testing
Does the drug produce the predicted effect in relevant cells or experimental systems?
4. Pharmacological feasibility
Can the required exposure actually be achieved safely in humans?
5. Preclinical studies
Does the effect survive testing in appropriate animal or other in-vivo models when needed?
6. Clinical trials
Does the medicine improve outcomes in humans?
7. Regulatory assessment
Are the total benefits and risks sufficient to support a new indication?
What Makes a High-Quality AI Repurposing Prediction?
A strong candidate should ideally have several independent lines of support.
- A reproducible computational signal.
- A biologically credible mechanism.
- Evidence from more than one data source.
- Relevant experimental confirmation.
- A feasible human exposure range.
- A plausible safety profile for the proposed population.
- A testable clinical hypothesis.
In other words, researchers should not ask only whether the AI's prediction is statistically impressive. They should ask whether it generates a useful, falsifiable and clinically testable hypothesis.
The Role of Generative AI and Large Language Models
Large language models are beginning to add another layer to repurposing workflows.
They can potentially help researchers:
- Summarize large bodies of biomedical literature.
- Extract drug-target relationships.
- Connect findings across disciplines.
- Generate mechanistic hypotheses.
- Identify apparently conflicting research.
- Help formulate experimental questions.
However, language models have a major limitation: they can generate plausible-sounding statements that are not supported by evidence.
For medical repurposing, an LLM should therefore function as a research assistant layered on verified databases and primary literature, not as an autonomous medical decision-maker.
AI + Multi-Omics + Real-World Data
The next generation of drug repurposing is likely to combine multiple types of evidence.
For example:
Genomics identifies disease-associated variants.
Transcriptomics identifies abnormal gene-expression patterns.
Proteomics identifies altered proteins.
Metabolomics identifies altered metabolic states.
Clinical data identifies patient outcomes.
AI integrates these layers to prioritize candidate drugs.
This moves drug repurposing from a simple "drug versus disease" problem toward a more detailed:
Drug × Disease × Biomarker × Mechanism × Patient Population × Safety
model.
Personalized Drug Repurposing
One of the most interesting long-term possibilities is patient-specific repurposing.
Instead of asking:
"Does Drug X work for cancer?"
the question may become:
"Does Drug X have a plausible therapeutic mechanism for this patient's molecular profile and disease state?"
Potential inputs could include:
- Tumor mutations.
- Gene-expression signatures.
- Biomarkers.
- Proteomic profiles.
- Prior treatment history.
- Resistance mechanisms.
- Drug interactions.
- Patient-specific safety factors.
This approach is conceptually attractive, but clinical implementation requires rigorous validation because individualized predictions can be particularly vulnerable to data quality problems and selection bias.
Why Cancer May Become a Major AI Repurposing Battlefield
Cancer illustrates why a network approach is attractive.
A tumor can evolve around treatment pressure, activating alternative pathways or changing its dependence on specific biological systems.
An AI system could potentially analyze:
Tumor type → Biomarker → Altered pathway → Drug target → Existing medicine → Resistance mechanism
This creates the possibility of a repurposing matrix rather than a simple drug list.
However, oncology is also a field in which false positives can be especially dangerous. A compelling laboratory mechanism does not establish that a drug will control a patient's cancer.
What AI Drug Repurposing Cannot Do
AI cannot independently establish:
- That a drug is effective in humans.
- That a drug is safe for a particular individual.
- That one repurposed drug is better than standard therapy.
- That a laboratory effect will translate into survival benefit.
- That an association is causal.
- That an AI-ranked candidate should be self-administered.
Even when the original medicine is FDA-approved, its new proposed use still needs appropriate evidence and regulatory consideration. The FDA's current repurposing initiative specifically emphasizes new uses supported by safety and effectiveness evidence.
How to Read an AI Drug Repurposing Study
Before accepting a claim that "AI discovered a treatment," ask seven questions:
- What did the AI actually predict?
- What data trained the model?
- Was the test set truly independent?
- Was there experimental validation?
- Was the candidate tested in humans?
- Was the clinical endpoint meaningful?
- Has the new indication received regulatory approval?
A headline saying that AI "identified" a medicine can refer to anything from an in-silico ranking to a clinically validated therapy. Those are fundamentally different evidence levels.
The Future of AI Drug Repurposing
The field is moving toward increasingly integrated systems combining:
- Knowledge graphs.
- Graph neural networks.
- Foundation models.
- Multi-omics.
- Electronic health records.
- Real-world evidence.
- Biomedical literature mining.
- Protein-structure information.
- Virtual screening.
- Causal inference.
The central challenge is no longer simply building a model that can predict a drug-disease relationship.
The harder question is:
Can the AI prediction improve real-world therapeutic decisions?
That distinction is increasingly emphasized in the latest critical assessments of AI-driven drug discovery.
Bottom Line
AI drug repurposing is best understood as an evidence-generation engine.
Artificial intelligence can search enormous biomedical networks, compare molecular signatures, identify drug-target relationships, mine scientific literature and prioritize existing medicines for further investigation.
Knowledge graphs and graph neural networks are particularly important because they allow researchers to model the interconnected nature of biology. Newer foundation-model approaches such as TxGNN demonstrate how AI can extend repurposing predictions into diseases with limited existing treatment information.
The baricitinib experience during COVID-19 shows the appropriate pathway: an AI-assisted hypothesis can accelerate identification of a candidate, but clinical trials and regulatory evaluation remain essential.
The future of drug repurposing is therefore unlikely to be "AI replaces drug development."
It is more accurately described as:
Biology tests the mechanism.
Clinical trials test the treatment.
Regulators evaluate the evidence.
That separation between prediction, mechanism, clinical evidence and regulatory approval is essential when evaluating AI-generated claims about existing medicines.
Selected References
- Wei S, et al. The use of knowledge graphs for drug repurposing: From classical machine learning algorithms to graph neural networks. Computers in Biology and Medicine. 2025. PMID: 40784078.
- Zitnik M, et al. A foundation model for clinician-centered drug repurposing. Nature Medicine. 2024;30:3601–3613.
- Applications of Artificial Intelligence in Drug Repurposing. Advanced Science. 2025. PMID: 40047357.
- Artificial intelligence in drug research and development: a review of methods and applications in drug repurposing. 2026. PMID: 42166427.
- FDA. Drug Repurposing. Updated 2026.
- FDA. Drug Repurposing: Considerations for Selection Criteria and Prioritization. August 5, 2026.
- BenevolentAI researchers. Expert-Augmented Computational Drug Repurposing Identified Baricitinib as a Treatment for COVID-19. Frontiers in Pharmacology. 2021. PMID: 34393789.
- FDA. Emergency Use Authorizations for Drugs and Non-Vaccine Biological Products. 2026 update.
- The New York Times. Doctors Told Him He Was Going to Die. Then A.I. Saved His Life. 2025.
Medical information disclaimer: This article is for educational and research purposes only. An AI-generated drug-repurposing prediction is not evidence that a medicine is effective or appropriate for an individual patient. Repurposed or off-label medicines should be evaluated by qualified healthcare professionals using the available clinical evidence, approved indications, contraindications, drug interactions and patient-specific factors.
.png)
.png)
.png)
Comments
Post a Comment