- What AI drug repurposing is and why it's attractive
- AI approaches to drug repurposing
- Case studies in AI drug repurposing
- From computational signal to clinical validation
- Regulatory pathway for AI-identified repurposed drugs
- Limitations of AI drug repurposing
- What AI drug repurposing means for the field going forward
AI drug repurposing has moved from an opportunistic side project into a systematic discipline, using machine learning to scan approved molecules against thousands of disease indications at once. Drug repurposing has always been attractive because an approved compound already carries an established safety profile, which can shorten the path to clinical testing considerably. Computational drug repositioning now expands that search dramatically, but the field's central problem has not gone away: separating a genuine biological signal from a computational artifact that looks like one.
Key takeaways
- AI drug repurposing screens approved molecules against new disease targets far faster than manual literature review.
- Network medicine and knowledge graphs are the two dominant computational approaches behind most published repurposing candidates.
- Baricitinib's use in coronavirus disease 2019 (COVID-19) remains the most cited example of an AI-assisted repurposing signal reaching clinical validation.
- The FDA's 505(b)(2) pathway lets sponsors reference a drug's existing safety data while generating new efficacy evidence for the added indication.
- False positives, sparse validated interaction data, and weak mechanistic follow-through remain the field's biggest limitations.
What AI drug repurposing is and why it's attractive
Drug repurposing means finding a new therapeutic use for a drug that already has regulatory approval, or one that reached late-stage clinical testing for a different disease. Because existing records already document the molecule's pharmacokinetics, toxicology, and manufacturing profile, sponsors can often move directly into efficacy-focused trials rather than restarting the full nonclinical safety package that a novel chemical entity requires.
That head start translates into real time and cost advantages, which is why repurposing has drawn renewed attention from a field long accustomed to double-digit attrition rates during early-stage drug discovery. Long before AI entered the picture, repurposing already had a track record: sildenafil moved from a failed angina treatment to an approved therapy for erectile dysfunction and pulmonary arterial hypertension, and regulators later approved thalidomide, after its withdrawal for teratogenicity, for multiple myeloma and complications of leprosy under tightly controlled distribution programs. Those cases show the underlying economics have always favored repurposing; AI simply makes the search for the next such case systematic rather than accidental.
Deflazacort's approval for Duchenne muscular dystrophy, built on decades of use as a corticosteroid in Europe, further illustrates how an established safety record can support a new indication under the FDA's 505(b)(2) regulatory pathway. The FDA has formalized its interest in this approach, soliciting stakeholder input on priority disease areas, including neurodegenerative conditions and rare diseases, where existing molecules might address unmet need but lack commercial incentive for a full repurposing program, according to the FDA's drug repurposing page.
That combination of scientific plausibility and reduced financial risk is what makes repurposing an attractive complement to early-stage target identification and validation work rather than a substitute for it. A broader look at AI in drug discovery situates repurposing as one piece of a much larger computational shift spanning the full arc of pharmaceutical research, from initial target identification through clinical translation.
AI approaches to drug repurposing
AI approaches to drug repurposing generally fall into two overlapping categories: knowledge-graph and text-mining methods, and network-based or structural methods that model disease biology directly. Knowledge-graph platforms ingest structured databases and machine-read literature to connect drugs, targets, genes, and diseases, then use graph algorithms to surface indirect relationships a human reviewer would be unlikely to spot manually, effectively compressing years of literature review into a queryable map of biomedical relationships.
Network medicine, a separate but complementary approach, treats disease as a disruption localized to a specific neighborhood, or module, within the broader map of protein-protein interactions rather than the product of a single gene. Drugs whose molecular targets sit close to a disease module in that network qualify as better candidates for repurposing, a concept validated across conditions ranging from asthma to heart disease and formalized in network medicine platforms such as NeDRex.
Both approaches increasingly incorporate machine learning models trained on chemical structure, gene expression signatures, and electronic health record patterns to rank candidate drug-disease pairs, an area of active development across computational biology. These methods work alongside the phenotypic and target-based screening approaches long used in target identification and validation, adding a computational prioritization layer on top of experimental workflows rather than replacing them. AI drug repurposing pipelines typically blend both categories, using network proximity scores as one feature among many inputs to a broader predictive model rather than relying on a single method in isolation.
Case studies in AI drug repurposing
Baricitinib stands as the most thoroughly documented example of an AI-assisted repurposing signal that progressed all the way to regulatory authorization. An AI-enhanced biomedical knowledge graph, combined with expert-guided queries, flagged the rheumatoid arthritis drug as a candidate for COVID-19 based on its dual antiviral and anti-inflammatory mechanism, a process described in Frontiers in Pharmacology.
The ACTT-2 randomized phase three trial subsequently tested that computational hypothesis and showed baricitinib plus remdesivir improved time to recovery, supporting the FDA's initial emergency use authorization (EUA) in November 2020. The CoV-BARRIER trial then found a statistically significant mortality reduction with baricitinib, evidence that came after that initial authorization and went on to support both an expanded 2021 EUA and full FDA approval of baricitinib for hospitalized COVID-19 patients in 2022, as documented in a peer-reviewed analysis in Pharmaceuticals. The case illustrates the full pipeline: a computational hypothesis generated the signal, but randomized clinical trials, not the algorithm, established clinical efficacy in stages. It also shows why speed matters most during a public health emergency, when the months saved by skipping a fresh safety program can translate directly into lives affected by treatment availability.
Beyond individual molecules, curated screening resources have institutionalized systematic repurposing at scale. The Broad Institute's Drug Repurposing Hub, launched in 2015 and comprising more than 4,700 annotated approved drugs and clinical-stage compounds at its original publication, lets researchers run phenotypic screens against poorly understood disease biology and has expanded from its original oncology focus to tuberculosis, kidney disease, and other indications, according to a peer-reviewed report in Nature Medicine. Resources like this one give computational teams a well-annotated starting library, which reduces the chemical and mechanistic uncertainty that would otherwise accompany any newly generated hit list.
From computational signal to clinical validation
A computational hit is a hypothesis, not a finding, and every credible AI new indication discovery program builds in a structured path from algorithmic ranking to experimental confirmation. That path typically follows a consistent sequence regardless of which computational method generated the original signal.
- Computational prioritization narrows thousands of candidate drug-disease pairs to a shortlist ranked by predicted mechanism, network proximity, or literature support.
- In vitro or phenotypic assays test the top candidates against relevant cell models or organoid systems to confirm biological activity.
- Retrospective analysis of electronic health record or insurance claims data checks whether patients already taking the drug show altered outcomes for the target disease.
- Prospective clinical trials, often accelerated by the compound's existing safety record, generate the efficacy data regulators require.
This sequence mirrors, in compressed form, the computational triage used in adaptive trial design, where algorithmic ranking narrows a large candidate pool before committing resources to human testing. That parallel is not a coincidence: both fields depend on the same underlying premise, that algorithmic ranking is only useful to the extent it survives contact with real patient data.
Skipping any one of those steps, particularly the retrospective real-world check, is a common reason computational candidates fail to replicate when tested prospectively. Teams that treat the four-step sequence as a checklist rather than a formality tend to catch weak candidates earlier, before they consume budget in a formal clinical program.
Regulatory pathway for AI-identified repurposed drugs
The FDA's 505(b)(2) application is the primary regulatory route for repurposed drugs because it allows a sponsor to reference the existing nonclinical and safety data for an approved molecule while generating new evidence focused specifically on the additional indication. That structure means a repurposing program can often compress or, in some circumstances, bypass early-phase safety trials and move more directly into efficacy-focused studies, as outlined in the FDA's guidance on 505(b)(2) applications.
The agency has also built dedicated infrastructure around off-label and repurposed use outside the formal application process. The CURE ID program, described on the FDA's website, lets clinicians report real-world treatment experiences with repurposed drugs, generating observational signals that can support later, adequately controlled studies.
| Pathway or program | What it does | Governing body |
|---|---|---|
| 505(b)(2) application | Lets sponsors reference existing safety data while proving efficacy for a new indication | The FDA |
| CURE ID | Collects clinician-reported real-world repurposing experiences | The FDA |
| Orphan drug designation | Offers incentives for repurposing toward rare diseases with unmet need | The FDA |
For translational scientists, the practical takeaway is that regulatory strategy should be part of the earliest computational triage decisions in AI drug repurposing, not an afterthought applied once a candidate has already been chosen. A molecule with an expired patent but a strong AI-generated mechanistic rationale may still need a sponsor willing to navigate orphan drug incentives or public-private partnership funding to reach a 505(b)(2) filing.
Limitations of AI drug repurposing
The biggest limitation in AI drug repurposing is not computational power but validated data: confirmed drug-target interaction pairs remain comparatively scarce relative to the space of possible drug-disease combinations, which makes it difficult for models to learn reliable negative examples. Reviews of the field note that selecting true negative training data is genuinely difficult because a lack of reported activity does not confirm a lack of biological effect, which skews model confidence and complicates benchmarking across different repurposing algorithms.
Network-based tools also face a scaling problem as multi-omics datasets grow, since some analysis pipelines cannot process the resulting data volume without producing false-positive associations that look statistically significant but lack mechanistic grounding. Every computational hit therefore still requires experimental validation in cellular models and, eventually, clinical trials, a step that carries its own substantial cost and cannot be shortcut regardless of how researchers generated the candidate.
There is also a translational gap in AI drug discovery downstream of discovery itself. Even a well-validated repurposing candidate needs a sponsor willing to run the confirmatory trial, and molecules with expired patents or generic status often lack the commercial incentive that would fund that work, which is part of why the FDA has begun soliciting public input on prioritizing repurposing candidates in underserved disease areas. For computational biologists building these models, the practical implication is that model performance metrics reported on historical benchmark datasets do not guarantee prospective success, since decades of biased pharmaceutical priorities, rather than random sampling across all possible biology, selected the drug-disease pairs that made it into training data.
What AI drug repurposing means for the field going forward
AI drug repurposing has matured from an occasional serendipitous discovery into a structured discipline that pairs network medicine and knowledge-graph methods with the FDA's existing regulatory infrastructure for accelerated approval. The baricitinib case demonstrates what success looks like: a computational hypothesis, confirmed through phenotypic and mechanistic follow-up, validated in randomized controlled trials, and translated into an authorized therapy within a compressed timeline.
The technology's value depends entirely on the rigor applied after the algorithm produces its ranked list, since a computational signal is only as credible as the experimental and clinical work that follows it. For drug discovery scientists and translational researchers, that means AI new indication discovery is best treated as a triage tool that narrows the search space, not a replacement for the validation chain that separates a promising pattern from a proven treatment.
This article was produced under Drug Discovery News' AI Editorial Guidelines.
