AI’s growing role in cancer, rare disease and COVID-19 drug repurposing

For rare diseases, AI-driven repurposing fills a critical gap. With more than 7000 rare diseases and only a small percentage having approved treatments, repurposing offers a viable path where pharmaceutical development is often economically unattractive. Tools such as knowledge graphs, single-cell RNA models and heterogeneous label-propagation systems identify candidate drugs by linking molecular signatures to known drug profiles. Organ-on-chip systems enhanced by AI add physiologically relevant data to the pipeline, offering mechanistic validation without relying exclusively on animal models.

Artificial intelligence (AI) is accelerating discovery timelines and creating new pathways for treating high-burden diseases. A new study warns, however, that this scientific shift comes with complex technical, regulatory and ethical obstacles that must be resolved before AI-driven drug repurposing reaches full maturity.

Their paper, “AI-Driven Drug Repurposing: Applications and Challenges,” published in the journal Medicines in 2025, examines how artificial intelligence enables the rapid identification of new uses for existing drugs, reducing research costs and shortening timelines that traditionally span more than a decade. The review evaluates machine learning, deep learning, network science, signature-based methods, natural language processing and multimodal models, showing how these technologies extract meaningful insights from vast biomedical datasets. It also outlines how these tools have already supported major advances in oncology, rare diseases, infectious diseases, neurology and COVID-19 therapeutic discovery.

The authors argue that drug repurposing has become increasingly important as traditional drug development remains expensive, slow and vulnerable to high failure rates. By leveraging AI to identify new therapeutic opportunities in approved drugs with established safety profiles, researchers can reduce uncertainty and move rapidly toward clinical testing.

However, current AI systems face persistent challenges, including data scarcity, inconsistent model transparency and regulatory uncertainty. Without addressing these issues, the potential of AI-enabled repurposing may be held back despite promising scientific progress.

How AI is transforming drug repurposing across disease areas

Traditional repurposing depends on clinical observation, trial data or serendipitous findings. AI expands this scope by analyzing complex biological networks, gene expression signatures and chemical relationships that would be impossible for humans to interpret at scale.

The authors detail how machine learning and deep learning models examine patterns across chemical structures, protein sequences, transcriptomic data and clinical outcomes. By identifying similarities between disease mechanisms and drug actions, these algorithms surface candidate therapies more quickly and with greater precision than manual evaluation. Tree-based models, support vector machines, neural networks, convolutional networks and recurrent architectures each contribute different strengths, from feature learning to pattern recognition in large multidimensional datasets.

Network-based approaches form another powerful pillar of modern repurposing. These methods map interactions between drugs, proteins, genes and diseases, identifying areas where drug nodes overlap with disease modules. When a drug lies close to a disease cluster within a biological network, it may modulate relevant pathways and offer therapeutic potential. The review cites multiview learning frameworks, heterogeneous knowledge graphs and multimodal graph algorithms as emerging tools that unify disparate biomedical data types into single integrated models.

Signature-based approaches contribute additional insight by aligning disease gene-expression profiles with drug-induced transcriptional signatures. When a drug counteracts the molecular signature of a disease, it may reverse disease trajectories. The paper describes how large resources such as CMap and LINCS provide millions of transcriptomic signatures that AI models match with disease patterns. Deep learning frameworks enhance this approach by improving the accuracy of signature matching and reducing the noise inherent in large-scale genomic datasets.

Natural language processing and large language models add another dimension. They extract mechanistic insights from unstructured biomedical literature, clinical notes and trial registries, revealing previously overlooked drug–disease associations. Tools such as BioBERT and transformer-based architectures identify biological relationships by scanning vast scientific corpora that would otherwise remain underutilized. This gives researchers an automated way to track mechanistic hypotheses and identify drug candidates based on textual patterns rather than purely molecular features.

According to the review, these combined methodologies have helped identify promising repurposed candidates for cancer, genetic disorders, neurological diseases, metabolic diseases and infectious diseases. They have also played an essential role in pandemic-response drug discovery by enabling rapid screening of antiviral compounds and supporting combination therapy design.

Where AI-driven repurposing is already producing impact

The authors describe oncology as one of the most advanced fields in this domain. Drug–target interaction models, gene-expression algorithms and multimodal graph networks have highlighted new uses for drugs such as statins, disulfiram, celecoxib and pantoprazole. Neural networks and integrative models have expanded understanding of the multi-pathway effects of these agents, some of which have already progressed into clinical trials.

For rare diseases, AI-driven repurposing fills a critical gap. With more than 7000 rare diseases and only a small percentage having approved treatments, repurposing offers a viable path where pharmaceutical development is often economically unattractive. Tools such as knowledge graphs, single-cell RNA models and heterogeneous label-propagation systems identify candidate drugs by linking molecular signatures to known drug profiles. Organ-on-chip systems enhanced by AI add physiologically relevant data to the pipeline, offering mechanistic validation without relying exclusively on animal models.

During the COVID-19 pandemic, AI-enabled repurposing demonstrated urgent value. Deep learning models screened millions of molecules against SARS-CoV-2 targets, identifying compounds that were later tested clinically. Graph neural networks and knowledge graphs helped predict combination therapies, suggesting repurposing strategies that influenced further research. Signature-based approaches produced additional insights into antiviral and immunomodulatory drug behavior, contributing to therapeutic evaluation during the emergency phase of the pandemic response. AI also contributed to vaccine and de novo antiviral development by identifying epitopes and generating candidate molecules.

Neurological disorders are another area where repurposing benefits from AI. The complexity of diseases like Alzheimer’s and Parkinson’s requires tools capable of navigating multi-layered biological systems. Graph-based algorithms, transcriptomic signature models and integrative learning approaches have surfaced new therapeutic possibilities such as ibuprofen, ceftriaxone, cholecalciferol and gemfibrozil for neurodegenerative pathways. The authors emphasize that these insights strengthen the case for repurposing drugs that already have well-documented safety records.

In metabolic diseases such as diabetes, AI-assisted repurposing works in both directions: identifying non-diabetes drugs that benefit metabolic health, and revealing diabetic agents that may possess anticancer or cardioprotective properties. These findings underscore how repurposing bridges traditionally isolated therapeutic domains.

The review also presents examples in infectious disease and pediatrics. AI models support antiviral drug discovery, antibiotic resistance prediction and pediatric therapeutic identification. While pediatrics remains more heavily dependent on clinical observation, the authors foresee AI playing a growing role as datasets and digital infrastructure improve.

Why AI-driven drug repurposing still faces major barriers

Despite the scientific progress described in the review, the authors devote significant attention to the obstacles that continue to limit widespread adoption of AI-driven repurposing. At the technical level, data quality remains one of the largest challenges. Biomedical datasets are often incomplete, noisy or biased toward positive findings. Many repurposing algorithms rely on datasets that lack negative drug–target examples, creating distorted training patterns that reduce predictive reliability.

Mechanistic uncertainty is another barrier. AI models may identify associations without explaining underlying biological pathways, making clinical translation difficult. Disease behavior varies across cell types, tissues and patient populations, and models that do not capture biological context risk producing misleading results. When AI proposes candidates without mechanistic clarity, researchers must invest heavily in validation before moving toward clinical evaluation.

The paper also highlights the difficulty of zero-shot scenarios, where drugs or diseases lack prior examples in training data. In such cases, models struggle to generalize. Effective repurposing requires algorithms capable of handling sparse or incomplete information, yet many existing models depend on dense, well-labeled datasets to perform reliably.

Ethical and legal challenges further complicate deployment. Data privacy concerns arise when models rely on patient-level records or genomic data. Lack of transparency in deep learning systems reduces trust among clinicians and regulators. Weak intellectual-property incentives discourage companies from investing in clinical trials for repurposed drugs, especially when generic versions exist and exclusivity protections are limited.

Regulatory processes also lag behind technological progress. Current frameworks, including provisions in U.S. and European drug legislation, are not designed for workflows that depend on AI-derived hypotheses. Without clear pathways for approval, repurposed drugs identified through AI may face delays or barriers in reaching clinical testing despite promising scientific support.

The authors call for improved regulatory clarity, enhanced incentives and investment in explainability tools. Techniques such as SHAP and LIME may help interpret model outputs, but broader integration of explainability into the repurposing pipeline is needed. According to the review, successful AI-driven repurposing requires alignment between technical innovation, ethical safeguards and supportive policy structures.