A Narrative Review of Artificial Intelligence for Drug Repurposing: Lessons From COVID‐19 and Oncology (2020–2025)

ABSTRACT

Drug repurposing presents a cost‐effective and time‐efficient strategy to identify new therapeutic applications for existing drugs. Recent advances in artificial intelligence, including machine learning, deep learning, knowledge graphs, and natural language processing, have revolutionized this field by enabling automated discovery of drug‐disease associations. This review examines the role of artificial intelligence in drug repurposing, drawing insights from two critical case areas: Coronavirus disease 2019 and oncology. We explore current trends, methodological frameworks, and technological innovations in artificial intelligence‐driven drug repurposing, as well as challenges and emerging future directions. The findings of this paper underscore the transformative potential of artificial intelligence in biomedical research and justify its continued integration in pharmaceutical pipelines.

Study Highlights

• What is the current knowledge on the topic?

  • ○
    Artificial intelligence has transformed drug repurposing from a largely serendipitous process into a systematic, data‐driven strategy integrating machine learning, deep learning, knowledge graphs, graph neural networks, and natural language processing.
• What question did this study address?

  • ○
    This review provides a focused analysis of artificial intelligence–driven drug repurposing between 2020 and 2025, with emphasis on two high‐impact domains: coronavirus disease 2019 and oncology.
• What does this study add to our knowledge?

  • ○
    Knowledge graph–based and deep learning approaches enabled rapid identification and prioritization of repurposed candidates during the COVID‐19 pandemic, including baricitinib and other clinically evaluated therapeutics.
• How might this change drug discovery, development, and/or therapeutics?

  • ○
    In oncology, artificial intelligence–supported multi‐omics integration and graph‐based models improved precision repurposing by linking existing drugs to cancer subtypes, signaling pathways, and patient‐specific therapeutic vulnerabilities.

1. Introduction

The conventional drug development pipeline is notoriously protracted, expensive, and fraught with risk. On average, bringing a new drug to market takes over a decade and costs upwards of $1 billion. Drug repurposing, also known as drug repositioning, offers a pragmatic alternative by identifying new therapeutic uses for existing drugs with established safety profiles. This strategy significantly reduces development time and cost, bypassing early‐stage toxicity and pharmacokinetic testing [1, 2, 3, 4, 5].

Artificial Intelligence (AI) has emerged as a transformative force in drug repurposing, shifting the paradigm from serendipitous discovery to systematic, data‐driven exploration. AI techniques such as machine learning (ML), deep learning (DL), natural language processing (NLP), knowledge graphs (KGs), and graph neural networks (GNNs) enable the integration and analysis of vast biomedical datasets, including genomics, proteomics, electronic health records (EHRs), and scientific literature. These tools can uncover hidden drug‐disease relationships, predict drug–target interactions, and prioritize candidates for experimental validation [1, 2, 6, 7, 8, 9, 10, 11, 12, 13]. The urgency of the COVID‐19 pandemic catalyzed the adoption of AI in drug repurposing. For instance, BenevolentAI used a KG‐based approach to identify baricitinib as a potential treatment for COVID‐19, leading to rapid clinical evaluation and regulatory authorization, with subsequent approval in defined settings. Similarly, representation‐learning models demonstrated the ability to map disease‐drug relationships by analyzing embeddings derived from real‐world patient data [7, 14, 15, 16, 17, 18, 19, 20, 21, 22].

AI‐driven drug repurposing is not limited to infectious diseases. In oncology, AI models have been used to match existing drugs to specific cancer subtypes based on multi‐omics data, enabling precision medicine approaches. GNNs,e.g., have shown promise in identifying multi‐target drug candidates by modeling complex biological networks and anticipating polypharmacy effects [2, 6, 13, 23, 24, 25, 26, 27, 28, 29, 30].

Despite its promise, AI in drug repurposing faces several challenges. These include data quality and standardization issues, model interpretability and mechanistic plausibility, and regulatory hurdles across the medicine's lifecycle. Ethical concerns around data privacy and algorithmic bias also persist. Nevertheless, the convergence of AI with high‐throughput experimental platforms, predictive safety analytics, and the growing availability of curated biomedical databases continues to accelerate innovation in this field [1, 2, 6, 9, 11, 26, 31, 32, 33, 34].

This study reviews the evolving landscape of AI‐driven drug repurposing from 2020 to 2025, focusing on two critical domains: COVID‐19 and oncology. By examining methodological frameworks (KGs, GNNs, structure‐based DL, and real‐world evidence mining), performance.

Through metrics and real‐world case studies, we aim to elucidate the impact and future potential of AI in reshaping pharmaceutical research, highlighting opportunities at the intersection of large language models and multimodal learning, alongside validation, safety, and regulatory guardrails for responsible adoption [12, 34, 35, 36, 37, 38].

1.1. Review Approach / Literature Identification and Selection

This narrative review covers studies published between 2020 and 2025 on artificial intelligence approaches for drug repurposing, with emphasis on applications to COVID‐19 and oncology. We identified relevant literature by searching major biomedical and multidisciplinary databases (including PubMed, Scopus, Web of Science, and Google Scholar) and by screening reference lists of highly cited reviews and primary studies. Where appropriate, we also considered preprints (e.g., arXiv and medRxiv) only to provide context on emerging methods, clearly distinguishing these from per‐reviewed evidence.

Search strings combined terms related to repurposing (e.g., “drug repurposing”, “drug repositioning”) with AI methods (e.g., “machine learning”, “deep learning”, “graph neural network”, “knowledge graph”, “natural language processing”, “large language model”). We included articles that (i) applied AI/ML/DL/KG/NLP methods to identify, prioritize, or mechanistically explain repurposing candidates; and (ii) reported sufficient methodological detail to understand inputs, outputs, and validation. We excluded papers that were outside the 2020–2025 window, were not focused on repurposing (e.g., de novo discovery only), or provided purely speculative claims without methodological description or supporting evidence.

1.2. AI Methodologies and Performance Evaluation

The period between 2020 and 2025 has seen the maturation and application of a diverse suite of AI methodologies for drug repurposing. These techniques are not mutually exclusive and are often combined into powerful hybrid pipelines [4, 5, 8, 10, 11, 12, 13].

1.2.1. Key AI Methodologies

AI techniques provide valuable assistance in data analysis, complex decision‐making, and pattern recognition. These techniques encompass machine learning, deep learning, network pharmacology, and knowledge graphs. Each approach plays a distinct role in processing, interpreting, and applying data across research and real‐world contexts [2, 4, 5, 9, 10].

Machine Learning includes “classical” algorithms like Random Forests (RF), Support Vector Machines (SVMs), and Gradient‐Boosted Trees (e.g., XGBoost). These models excel at learning from structured data, such as quantitative structure–activity relationship (QSAR) models, which learn mappings from molecular descriptors (e.g., fingerprints and physicochemical features) to endpoints such as potency, selectivity, toxicity, and absorption, distribution, metabolism, and excretion properties or drug side‐effect profiles [39]. They are often favored for their relative interpretability and effectiveness with smaller datasets [3, 7, 40]. For instance, a Decision‐Tree Regression model was used to prioritize leads against the COVID‐19 main protease (3CLpro) [41], while an RF model helped identify inhibitors for SYK kinase in cancer [3]. A recent innovation involves using a “selective cleaning” algorithm with XGBoost to improve the quality of bioactivity data before model training, significantly boosting predictive accuracy for MDM2‐p53 axis inhibitors [42]. In parallel, ML pipelines for drug‐response prediction and repurposing increasingly leverage recommender‐style learning and feature engineering for precision oncology [25, 28, 29]. These approaches can also perform well when biomedical datasets are incomplete, provided missingness is handled explicitly (e.g., imputation, missingness indicators, or robust tree‐based learning); however, non‐random missingness can encode confounding and bias, so sensitivity checks and external validation remain essential.

Deep Learning utilizes neural networks with many layers to model intricate, non‐linear patterns in data. Graph neural networks (GNNs), a class of neural networks designed to learn from graph‐structured data in which entities are represented as nodes, and their relationships as edges, and transformers have become dominant in this space. In drug repurposing, GNNs are especially useful because molecules, protein interaction networks, and drug‐disease associations can naturally be represented as graphs [6, 13, 26, 42]. Examples include the “eVir” platform, which used Siamese neural networks (twin subnetworks with shared weights that learn a similarity/distance function between two inputs) on graph embeddings to screen for COVID‐19 antivirals (Figure 1) [43], and a deep autoencoder ensemble that identified drugs simultaneously blocking three pro‐angiogenic cancer targets [44]. DL also underpins large‐scale ligand‐based virtual screening and structure‐guided prioritization against SARS‐CoV‐2 proteins (e.g., Mpro), illustrating rapid triage from crystallographic targets to ranked candidates [45, 46, 47]. Beyond small molecules, DL has accelerated new‐to‐nature design and indication mapping (e.g., halicin) and continues to set performance bars in de novo and repurposing pipelines [47, 48]. As with machine learning, these models often require larger labeled datasets (or substantial pretraining), careful selection of hyperparameters, and GPU‐accelerated training and inference, which makes rigorous curation and benchmarking essential. In practice, not all hyperparameters can be exhaustively optimized because the search space is large and computationally expensive. Researchers, therefore, usually tune a limited set of influential hyperparameters, such as learning rate, batch size, network depth, dropout, or regularization strength, using validation sets and strategies such as grid search, random search, or Bayesian optimization. This creates a practical trade‐off—more intensive tuning may improve predictive performance, but it can also increase computational cost, reduce reproducibility across settings, and increase the risk of overfitting to a particular dataset [4, 12, 49].

FIGURE 1.

EVir Platform [38].

Network Pharmacology models the complex interplay between drugs, protein targets, and disease pathways within biological networks. It is particularly effective at elucidating polypharmacology and system‐level effects, supporting repositioning hypotheses that may not surface from single‐target screens. It has been used to reposition fostamatinib for acute myeloid leukemia (AML) by analyzing its impact on the PI3K‐AKT pathway [50] and to find that the antipsychotic aripiprazole could rescue hormone‐therapy resistance in endometrial cancer by modulating the EGFR/PI3K/AKT axis [51]. Methodologically, modern network pharmacology often overlaps with KG construction and embedding, and benefits from representation learning to capture higher‐order interactions [42, 52, 53]. Recent patents and platform descriptions explicitly build large drug‐disease‐target graphs as a foundational step for AI‐driven prediction [54, 55].

Knowledge Graphs (KGs) are large‐scale, structured representations of biomedical knowledge, integrating heterogeneous data from sources like DrugBank, OMIM, EHR‐derived phenotypes, and the scientific literature into a network of entities (e.g., drugs, genes, diseases) and their relationships [52, 56]. AI algorithms then traverse or embed these graphs to infer novel links, effectively generating new repurposing hypotheses. A prominent success story is BenevolentAI's KG, which identified baricitinib's potential for COVID‐19 by linking AAK1 inhibition to viral endocytosis pathways [15, 17, 18, 57]. COVID‐era KGs such as DRKG and CoV‐KGE, as well as causal multi‐modal graphs, demonstrated how curated, computable knowledge accelerates indication discovery [21, 22]. Other platforms (e.g., RepurposeDB; systems from Zhijiang Laboratory) also use KGs as core engines for hypothesis generation and prioritization [52, 55, 58, 59].

Method integration involves practice; high‐performing pipelines blend these approaches,e.g., KG‐derived embeddings feeding GNNs or structure‐aware DL models constrained by network‐level mechanism hypotheses, while large language models (LLMs) increasingly assist with literature mining, hypothesis formulation, and protocol drafting (Figure 2) [34, 36, 37, 47, 60, 61]. In addition, early forms of foundation models (large pretrained models for chemistry, proteins, and biomedical text) are beginning to be used for few‐shot indication ranking and evidence synthesis when paired with retrieval from curated knowledge graphs. For example, emerging clinician‐centered and multimodal foundation‐model frameworks have been explored for prioritizing candidate drugs for new indications by integrating biomedical text, molecular information, and clinical context, thereby illustrating how pretrained models may support repurposing decisions even when task‐specific labeled data are limited.

FIGURE 2.

K‐Paths workflow for multi‐modal reasoning in biomedical knowledge graphs. (1) Starting from a user query about the influence of entity u (e.g., a drug) on entity v (e.g., a disease), (2) the system retrieves multiple reasoning paths connecting u and v within an enriched knowledge graph. (3) A diversity filter is applied to select a representative set of paths. (4a) The chosen paths are converted into natural‐language descriptions and supplied to a large language model for textual inference and hypothesis generation. (4b) Alternatively, the paths are combined to form a compact subgraph, which is then processed by a graph neural network to generate structured predictions.

1.2.2. Performance Metrics

Accuracy, Speed, and Cost Savings.

The integration of AI has yielded quantifiable improvements across key performance indicators compared to traditional computational methods like simple virtual screening or docking. These performance gains arise from multiple paradigms (Table 1) and are reflected in reported model evaluation metrics and workflow‐level impacts across representative studies (Table 2). Modern AI models frequently achieve strong predictive performance, with metrics such as R², AUROC, and classification accuracy often exceeding 80%–90% in retrospective tasks, surpassing classical baselines by 5–15 percentage points in direct comparisons when data quality and splits are controlled [4, 12, 41, 49, 60]. The field is also maturing its evaluation standards, moving toward enrichment‐focused metrics (e.g., BEDROC, early precision, PR‐AUC) and time‐aware/indication‐aware splits that better reflect practical utility and reduce leakage [5, 12, 20, 49, 63].

TABLE 1.

Comparative summary of major AI paradigms for drug repurposing.

Paradigm	Typical inputs	Strengths	Common failure modes/limitations	Best‐fit stages
Machine Learning (QSAR/classical ML)	Fingerprints, descriptors, side‐effect profiles, curated assay features	Strong on structured data; relatively interpretable; works with smaller datasets; fast training	Dataset bias/leakage; limited mechanistic insight; feature engineering dependence; confounding by indication in clinical data	Early screening, property/toxicity prediction, candidate prioritization
Deep Learning (GNNs/Transformers/Autoencoders)	Molecular graphs, sequences, images, multi‐omics; pretrained embeddings	Learns representations; strong performance on large datasets; supports multi‐modal fusion and transfer learning	Data/compute intensive; overfitting/leakage; harder interpretability; distribution shift across assays/institutions	Large‐scale virtual screening, multi‐omics response prediction, and combination design
Network Pharmacology/Network Medicine	Drug‐target networks, pathways, protein–protein interaction graphs, omics signatures	Mechanistic/pathway reasoning; polypharmacology; captures system‐level effects; supports combination hypotheses	Network incompleteness; edge uncertainty; causal ambiguity; pathway overgeneralization	Mechanism discovery, combination hypotheses, target/pathway prioritization
Knowledge Graphs (KGs)	Integrated drug‐gene‐disease knowledge from databases, literature, and clinical signals	Evidence‐linked reasoning paths; integrates heterogeneous sources; supports link prediction and traceable hypotheses	Noisy edges; provenance variability; bias toward well‐studied entities; leakage risks without careful evaluation	Hypothesis generation, evidence triangulation, mechanistic narratives, prioritization
LLM‐assisted workflows (with retrieval)	Biomedical text corpora + retrieved KG paths/documents	Fast literature synthesis; hypothesis articulation; protocol drafting; converts KG paths into explanations	Hallucinations without retrieval; citation/provenance issues; needs guardrails and human review	Evidence synthesis, decision support, research planning, and triage

TABLE 2.

Representative AI methodologies and performance evaluation metrics in drug repurposing.

Paradigm/example	Task context/dataset	Reported model metric(s)	Evaluation notes (as reported)	Operational/workflow impact (if reported)
Classical machine learning/quantitative structure–activity relationship (QSAR) (e.g., Random Forest, Support Vector Machine, XGBoost using fingerprints + physicochemical descriptors)	Bioactivity prediction, quantitative structure–activity relationship modeling, absorption/distribution/metabolism/excretion/toxicity triage; typical tasks include classification/regression on assay/compound datasets (e.g., public benchmarking tasks such as Therapeutics Data Commons).	Often reports R2, AUROC, accuracy, sensitivity/specificity depending on task; strong performance when feature engineering is appropriate, and data is moderately sized.	Performance is sensitive to dataset construction, label noise, and split strategy. Robust conclusions require scaffold/temporal splits and external validation to avoid inflated results from chemical‐series leakage.	Faster training and lower compute requirements than deep learning; rapid baseline models for triage and prioritization.
Deep learning on molecular representations (e.g., multilayer perceptrons, convolutional models, Transformer encoders on SMILES/sequence representations)	Drug‐target interaction prediction, virtual screening classification/regression, and activity prediction.	Frequently reports AUROC/AUPRC; often shows improvements over classical baselines on benchmark tasks (task‐dependent).	Requires careful control of leakage (random splits can overestimate generalization). Benefits are most reliable when using scaffold/temporal splits and external test sets.	Can scale across larger libraries once trained; higher compute and tuning burden than classical machine learning.
Knowledge graph/network medicine (knowledge graphs combining drug‐target‐disease‐pathway relationships; graph embeddings such as Node2Vec)	Link prediction and hypothesis generation on multi‐source biomedical graphs (drug‐disease association, target nomination, mechanism paths).	AUROC ~0.82–0.90 reported for link prediction in representative studies; early enrichment metrics (e.g., BEDROC/early precision) used in some papers; improved ranking quality when well‐curated knowledge graphs are used.	Generalization improves with temporal validation (training on earlier data, testing on later). Results depend heavily on curation quality, missingness, and the confounding structure of the graph.	In silico time reduced by ~60% by avoiding serial single‐target docking; estimated ~6 months saved in target identification in representative reports [50, 64].
Graph neural networks (molecular & biomedical graphs) (graph convolution, message passing neural networks, heterogeneous graph neural networks)	Predicting drug–target interactions, side‐effects, drug combinations, polypharmacy graph learning; often evaluated on benchmark tasks such as Therapeutics Data Commons.	AUROC/AUPRC improvements (~ + 5–15 points) over classical baselines reported in representative benchmarking studies; improved side‐effect/combination prediction in polypharmacy graphs [12, 13, 24, 27, 42, 49].	Performance strongly depends on split design (scaffold/temporal splits are preferred). Interpretability remains challenging without explicit explanation layers or pathway constraints.	Can enable multi‐objective ranking (efficacy + safety signals) and prioritize candidates for orthogonal validation; compute heavier than classical models.
Structure‐aware deep learning/virtual screening (e.g., SARS‐CoV‐2 main protease (Mpro) workflows; docking + deep learning rescoring; molecular dynamics refinement)	Structure‐guided screening when target structures are available; triage of compounds to smaller hit lists.	Reported improvements in top‐k hit rates; enrichment and rescoring gains; narrowing from large candidate libraries to a short list of chemotypes for testing [45, 46, 47].	Often retrospective; true performance depends on experimental confirmation and the presence of reliable negative controls.	Enables rapid narrowing from 103–104 candidates to tens of prioritized chemotypes for experimental follow‐up [45, 46, 47].
Quantum machine learning/Quantum‐Kernel Support Vector Machine (reported as exploratory)	Small‐to‐moderate datasets exploring quantum kernels for classification/regression tasks in chemistry.	Reported higher accuracy in some tasks; sometimes similar wall‐time to classical support vector machine, depending on setup [60].	Still early‐stage; results are highly dependent on dataset choice and experimental control; reproducibility varies.	Achieved higher accuracy with wall‐time similar to classical support vector machine in representative reports [60].
Large language model‐assisted literature extraction + workflow orchestration (retrieval‐augmented synthesis; prompt‐driven extraction; linking literature to knowledge graphs)	Automated evidence mining: Extracting drug‐target‐disease relations, trial outcomes, adverse events; summarization of mechanistic rationale and candidate support.	Reported improvements in recall of mechanistic links; task‐dependent performance on benchmark information extraction/summarization [12, 34, 36, 37, 49].	Requires reproducible prompting, careful citation handling, and guardrails to avoid hallucination. Best used with retrieval and structured constraints.	Analyst time reduced from days to minutes in representative workflow reports; GPU cost ~US$10 per run reported for some setups [63].

Note: Metrics summarized from representative studies; many results are retrospective. Robust conclusions require leakage‐aware splits (scaffold/temporal), external validation, and staged experimental confirmation.

Speed integrates GPU‐accelerated inference and batched screening, enabling throughput of thousands to tens of thousands of compounds in hours to a few days, vs. weeks or months for sequential docking/simulation workflows [19, 41, 46, 47, 66]. Structure‐guided DL for targets like SARS‐CoV‐2 Mpro exemplifies rapid triage from crystal structures to prioritized chemotypes at scale [45, 46]. Cost savings and funnel efficiency arise as AI primarily reduces costs by lowering false‐ positive burden before wet‐lab work. Higher precision upstream cuts the number of compounds requiring synthesis and validation; reports indicate 70%–97% reductions in follow‐up volume when curation and modeling are combined effectively in specific campaigns [40, 41]. At a portfolio level, systematic reviews and industry analysts suggest that AI‐enabled repurposing can reduce overall R&D expenditure (e.g., 40%–60%) and compress cycle times from hypothesis to validated hits (e.g., < 12 months in focused programs), with the exact magnitude contingent on data assets, automation, and assay capacity [4, 5, 36, 67, 68].

Public datasets and benchmarks such as Therapeutics Data Commons (TDC) and curated graph/cheminformatics corpora have been key to standardizing tasks and enabling reproducible comparisons across methods [12, 49].

Practical cautions emphasize that reported gains depend on rigorous dataset construction (de‐duplication, assay harmonization), robust splitting (scaffold/temporal), and external validation to avoid optimistic estimates; explainability and mechanistic plausibility remain important for translational credibility and decision support [5, 9, 11, 12, 30, 49]. Predictive safety (toxicity, DDI/polypharmacy) should be integrated early to prevent costly dead‐ends downstream [27, 31].

These reported gains should be interpreted cautiously, as their magnitude depends strongly on the underlying dataset quality, endpoint definition, validation strategy, comparator choice, and study design; accordingly, cross‐study comparisons are not always directly equivalent.

1.3. Case Analyses

The practical impact of AI in drug repurposing is best illustrated through case studies from the global COVID‐19 response and the ongoing fight against cancer.

1.3.1. Case Insights: COVID‐19

The pandemic served as a global, real‐time test for AI‐driven repurposing. AI platforms were deployed to rapidly scan existing arsenals of approved drugs for activity against SARS‐CoV‐2 or the host response (Table 3). In practice, three complementary streams emerged: (i) network/KG pipelines that mapped host‐virus biology to nominate actionable nodes and approved drugs, (ii) structure‐aware DL/virtual screening guided by crystallographic targets (e.g., Mpro), and (iii) representation learning on real‐world and multi‐omic data to rank candidates for follow‐up. These approaches compressed hypothesis generation and triage from months to weeks while preserving mechanistic visibility where possible [15, 16, 17, 18, 19, 20, 21, 22, 45, 46].

TABLE 3.

Selected AI‐Driven drug repurposing case studies in coronavirus disease 2019.

Drug/intervention	AI/computational technique(s)	Key finding (summary)	Evidence level/validation
Baricitinib	Knowledge graph reasoning; graph embeddings (Node2Vec) [57]	Linked JAK/AAK1 inhibition to viral endocytosis and inflammatory signaling; supported rapid clinical evaluation.	Clinical trials (ACTT‐2, COV‐BARRIER) and Food and Drug Administration approval for hospitalized Coronavirus disease 2019 patients (as cited).
Remdesivir	Siamese neural networks on graph embeddings (eVir platform) [43]	Used as a benchmark/positive control; the AI pipeline prioritized additional oral candidates that outperformed in models.	Comparative preclinical validation reported (cell and animal models, as cited).
Dexamethasone	Automated literature text‐mining; rule‐based Boolean network modeling [67]	Frequently ranked as an anchor drug in optimal anti‐inflammatory combination designs, consistent with its clinical role.	Clinical adoption supported by evidence base; AI result aligns with later practice.
Tocilizumab	Executable host‐virus logic network model [68]	Predicted interleukin‐6 blockade is most effective when paired with corticosteroids; it aligns with adopted treatment strategies.	Supported by later clinical practice and guideline adoption (as cited).
Hydroxychloroquine	Ensemble docking + molecular dynamics + gradient‐boosted regressor [69]	Provided stereochemical rationale for angiotensin‐converting enzyme 2 binding differences; contextualized conflicting/negative trial outcomes.	Mechanistic computational analysis; clinical outcomes ultimately negative in trials (as discussed).
Drug combinations/safety	XGBoost on ion‐channel data for torsades de pointes risk mapping [70]	Mapped arrhythmia risk for drug pairs; informed electrocardiogram monitoring strategy for combinations.	Computational safety prediction; used for risk stratification/monitoring guidance.

The COVID‐19 Experience Demonstrated That AI Could Not Only Generate Novel, Clinically Successful Hypotheses (Baricitinib) but Also Provide Mechanistic Rationale (Hydroxychloroquine), predict Optimal Combinations (Tocilizumab), and Guide Safety Monitoring for Repurposed Drugs [23]. In Particular, KG‐Centric Workflows Linked Baricitinib's JAK/AAK1 Activity to Endocytosis and Inflammatory Signaling, Enabling Rapid Trialing; Structure‐Led Models Leveraged Solved Mpro Structures to Sift Large Libraries Efficiently; and Graph/Embedding Methods Helped Align Disease Signatures With Drug Mechanisms Using Real‐ World Evidence. Together, These Strands Illustrate How Network Medicine, Structure‐Based DL, and Representation Learning Can Be Fused Into End‐To‐End, Crisis‐Ready Pipelines [15, 16, 17, 18, 19, 20, 21, 22, 45, 46].

1.3.2. Case Insights: Oncology

In oncology, artificial intelligence is increasingly being used to support precision medicine by matching existing drugs to specific cancer types, molecular subtypes, and patient‐level profiles based on multi‐omics and clinical data. The selected oncology case studies summarized in Table 4 illustrate how machine learning, deep learning, network pharmacology, and knowledge‐graph‐based approaches are being applied to prioritize repurposing candidates, propose mechanistic rationales, and guide follow‐up validation. Collectively, these examples highlight both the opportunities for faster hypothesis generation and the continuing need for rigorous external validation and safety‐aware translational assessment.

TABLE 4.

Selected AI‐Driven drug repurposing case studies in oncology.

Cancer indication	AI/computational methodology	Repurposed drug(s) prioritized	Key takeaway (summary)	Evidence level/validation
Acute myeloid leukemia	Transcriptomic signature reversal (connectivity mapping) + machine learning prioritization	Itraconazole; thioridazine (representative examples)	Gene‐expression‐driven matching identifies candidates predicted to reverse disease signatures and modulate oncogenic pathways.	In silico + in vitro (cell‐line) validation reported in representative studies.
Breast cancer (subtype stratification)	Multi‐omics integration + supervised learning; pathway‐aware feature selection	Metformin; disulfiram (representative examples)	Integrates genomic and transcriptomic signals to identify subtype‐specific vulnerabilities and prioritize repurposing candidates.	In silico + retrospective clinical association and/or in vitro validation (study‐dependent).
Glioblastoma	Knowledge graph/network pharmacology with mechanistic path scoring	Mebendazole; chlorpromazine (representative examples)	Network/pathway reasoning identifies drugs targeting interconnected signaling nodes; supports mechanistic hypotheses for follow‐up.	In silico + in vitro; occasional small‐scale clinical/observational evidence (study‐dependent).
Colorectal cancer	Drug‐response prediction using recommender‐system style modeling on cell‐line screening data	Auranofin; niclosamide (representative examples)	Learns response patterns from large screening panels to rank repurposed candidates and propose biomarkers of sensitivity.	In silico + in vitro validation on cell lines/organoids (as reported).
Multiple cancers (combination therapy)	Graph neural networks for drug synergy and polypharmacology‐aware combination design	Repurposed agents in combination with standard‐of‐care (examples vary)	Models prioritize combinations with improved predicted efficacy and manageable safety profiles; highlights polypharmacology.	In silico; follow‐up in vitro synergy assays in representative studies.
Patient‐derived organoid validation (precision oncology)	Patient‐level multi‐omics + predictive modeling; organoid response validation	Candidate drugs tailored to the patient's molecular profile (study‐dependent)	Bridges in silico prediction to functional testing, supporting mechanism‐anchored iteration and translational prioritization.	In vitro (patient‐derived organoids); occasional clinical translation (study‐dependent).

Note: Oncology examples are representative of common artificial intelligence repurposing paradigms; specific candidates and evidence levels vary by study. Robust conclusions require external validation and safety‐aware translational assessment.

These oncology cases also highlight a clear trend toward integrating multi‐omics data with sophisticated machine learning models to create predictive biomarkers that stratify patients and help identify tailored therapeutic options [73, 74]. Methodologically, this spans network pharmacology and knowledge‐graph embeddings for mechanistic repurposing, recommender systems for drug‐response prediction, and graph neural networks for combination design and polypharmacology‐aware prioritization [13, 23, 24, 25, 26, 27, 28, 29]. The use of patient‐derived organoids as a validation platform is a particularly powerful development, helping bridge the gap from in silico prediction to personalized clinical action and enabling faster, mechanism‐anchored iteration cycles [73].

Data augmentation and synthetic data generation are increasingly explored in oncology repurposing to address small cohorts and rare subtypes. However, augmentation can amplify biases or introduce distribution shift if synthetic samples do not preserve biological constraints. Best practice is to confine augmentation to training only, evaluate on untouched external cohorts, and report sensitivity analyses demonstrating robustness to augmentation choices.

2. Landscape, Challenges, and Limitations

Despite significant progress, the widespread adoption of AI in drug repurposing faces several systemic hurdles. At the same time, a key tension is reproducibility: Commercial platforms may rely on proprietary datasets and closed models that limit independent benchmarking, detailed error analysis, and transparent validation. This can slow community learning and make it harder to compare claims against open baselines, reinforcing the need for shared benchmarks, clear reporting, and external validation where possible.

2.1. The Overall Landscape

Commercial and academic platforms form an ecosystem that includes open‐access academic tools like CANDO and RepurposeDB [58, 63], as well as sophisticated commercial platforms from companies like BenevolentAI and Recursion. New meta‐platforms such as Europe's REMEDi4ALL aim to orchestrate end‐to‐end repurposing workflows and shared services (clinical, regulatory, IP, and funding pathways) [74]. The market for AI‐driven repurposing tools was estimated at ~$1.2 billion in 2024 and is growing, reflecting a strategic shift in pharma R&D toward data‐driven portfolio management and AI‐first triage [75]. This momentum is mirrored by a surge in patents for AI discovery platforms that explicitly feature deep learning and knowledge‐graph fusion [54, 55, 78], alongside public milestones (e.g., AI‐designed/AI‐discovered programs entering early‐phase trials) that signal increasing industrial maturity [79, 80, 81, 82].

Regulatory posture is evolving in parallel. Both the FDA and the EMA emphasize data quality, human oversight, transparency, lifecycle risk management, and the need for clear documentation when AI contributes to biomedical decision‐making. Regulatory bodies such as the FDA are increasingly receptive to AI‐supported evidence streams, and sponsors report submissions that incorporate AI modules across the development pipeline, although regulatory uncertainty persists, particularly around the validation of dynamic or continuously learning systems [32, 83]. However, the emphases of the FDA and EMA differ somewhat. The FDA has been more operationally explicit in framing AI oversight through a total product lifecycle perspective, with strong attention to version control, change management, and submission‐ready documentation for systems that may evolve over time [32, 84]. By contrast, the EMA has taken a broader medicines‐lifecycle view, placing comparatively greater emphasis on governance, scientific validity, ethical use, and the integration of AI across discovery, development, regulation, and post‐marketing activities [33]. As a result, the current best practice is to “lock” the algorithm version for a given submission and provide detailed Model Inspection Files, including data lineage, training procedures, versioning, performance, and monitoring, to ensure transparency and reproducibility [82]. For AI‐driven drug development, this means that sponsors targeting multiple jurisdictions should not assume full regulatory equivalence; globally relevant evidence packages will need both technical traceability and broader governance justification. Industry outlooks further project continued growth as generative AI, knowledge graphs, and graph neural networks diffuse from target discovery into indication expansion and label‐extension strategies [36, 37].

Scientific machine learning (SciML) provides an additional pathway toward trustworthy repurposing by embedding mechanistic priors and constraints (e.g., physical chemistry, kinetic models, pathway structure) into learning objectives. By restricting predictions to biologically plausible regimes, SciML can improve generalization under dataset shift and make model outputs more interpretable for translational decision‐making, especially when paired with knowledge‐graph provenance and orthogonal simulation.

Academic environments contribute many of the field's methodological innovations, but they also face distinct constraints. Compared with industry, academic groups often rely more heavily on open‐source software, public datasets, shared computing infrastructure, and short‐term grant funding. These conditions can limit access to large proprietary datasets, high‐end GPU resources, long‐term software maintenance, and extensive experimental validation. As a result, academic studies may excel in methodological novelty yet still struggle with scalability, benchmarking against closed industrial systems, or translation beyond proof‐of‐concept. Recognizing these structural differences is important when interpreting apparent performance gaps between academic and commercial platforms.

2.2. Key Challenges

Data Quality and Standardization matter because the “garbage‐in, garbage‐out” principle remains decisive. Performance is bounded by the completeness, harmonization, and accessibility of multi‐omics, chemical, clinical, and literature data; batch effects, assay drift, and heterogeneous curation can inflate headline metrics and impair external validity [6]. Community benchmarks (e.g., Therapeutics Data Commons) and curated graph/cheminformatics corpora are improving comparability via standard tasks, leakage‐aware splits (scaffold/temporal), and clearer documentation, but broad variability persists across modalities and endpoints [4, 5, 12, 42].

Interpretability and trust are affected because the “black‐box” nature of complex DL systems can impede clinical adoption. Clinicians and regulators often require mechanistic plausibility for decisions, motivating explainable AI, pathway/network‐level rationales, and triangulation with physics‐based simulation and KG reasoning [2, 9, 10, 11, 30]. Emerging practices include attention/attribution maps, counterfactuals, subgraph explanations for GNNs, and traceable KG paths that link predictions to biomedical evidence [42, 52, 53].

Validation and “False Theoretical Friends” highlight that in silico results are hypotheses, not conclusions. The eVir study's progression from in silico prediction to in vitro (cell‐based) and in vivo (animal) validation is still relatively rare as a fully connected pipeline [43]. Without rigorous and staged experimental confirmation (orthogonal assays, dose–response, safety/ADME screens), AI can generate plausible but ultimately incorrect leads. Integrating predictive toxicity/DDI models and polypharmacy graphs early helps retire unsafe hypotheses before costly wet‐lab steps [5, 27, 31].

Infrastructure and Talent requirements stem from the fact that Cutting‐edge AI typically requires Ethics, privacy, and consent remain central when using clinical or real‐world data for repurposing. De‐identification does not eliminate re‐identification risk in high‐dimensional datasets, and secondary use may fall outside the original consent scope. Practical mitigations include ethics review, data minimization, secure access controls, federated or privacy‐preserving learning where appropriate, and transparent documentation of data provenance and permitted uses.

GPU‐centric compute, MLOps for versioning/monitoring, and scarce skills spanning ML, cheminformatics, and translational biology. This widens the gap between AI pioneers and organizations with legacy infrastructure, elevating the importance of platformization and shared evaluation assets [4, 36, 68].

Economic and IP Barriers arise because, as seen during the pandemic, repurposing off‐patent drugs can encounter uncertain incentives (limited exclusivity, fragmented IP around combinations/biomarkers), which may deter investment even with a strong scientific rationale [82]. Strategic use of regulatory pathways, data exclusivity, combination IP, and companion diagnostics, combined with portfolio‐level evidence from AI triage, can partially mitigate these frictions [3, 36, 68].

Governance and Regulation require that, beyond model performance, sponsors must manage model risk, dataset shift, drift monitoring, update procedures, and human‐in‐the‐loop review. Current FDA/EMA communications emphasize documentation, change control, and post‐market monitoring; KG/LLM/GNN components should be versioned and auditable with clear boundaries between research and decision‐support use [32, 33]. Clinical translation benefits from aligning AI outputs with study endpoints, fit‐for‐purpose evidence packages, and early interaction with regulators.

3. Future Outlook (2025–2030)

Convergence with Foundation Models is expected as, over the next five years, large‐scale foundation models trained across chemistry, proteomics, functional genomics, and clinical text are poised to generalize beyond narrow tasks. These models, spanning protein‐language models, molecule‐protein co‐embedding, and multimodal large language models (LLMs), may increasingly support “zero‐shot” or “few‐shot” indication ranking and candidate triage, with the potential to shorten parts of early‐stage hypothesis generation and prioritization. However, the extent of such acceleration will depend on data quality, task complexity, validation requirements, and the degree of integration with experimental workflows [5, 11, 12, 26, 34, 42, 49, 85, 86]. In practice, we anticipate hybrid stacks: KG/GNN backbones for structure and causality, with foundation models handling literature synthesis, hypothesis drafting, and protocol generation [10, 12, 34, 42, 52, 87].

Maturing Regulatory Frameworks are expected as regulatory guidance is likely to crystallize around lifecycle controls for AI/ML in medicines, mirroring SaMD's “total product lifecycle” paradigm [83, 88] and aligning with recent FDA and EMA directions on documentation, change control, and human oversight [34, 35, 88, 89]. Sponsors should plan for model version “locking,” evidence packages tied to clinically meaningful endpoints, and auditable data lineage, plus drift monitoring for continuously learning components [10, 32, 33, 88].

Closed‐Loop, Automated Systems will accelerate as AI‐lab integration enables model‐suggested hypotheses executed by robotic platforms, feeding high‐content assays and omics back into models for active learning. Such closed‐loop pipelines, already hinted at in recent intellectual property and platform announcements, should increase hit quality, reduce iteration time, and standardize validation [54, 90, 91], with industrial exemplars from AI‐first discovery pipelines and early‐phase entries providing operational templates [36, 68, 79, 80, 81, 82, 91].

Safety‐by‐Design and Combination Reasoning will advance as Next‐gen workflows will embed predictive safety (toxicity, DDI, TdP risk) and polypharmacy modeling upstream, screening out unsafe hypotheses earlier and informing dose/sequence design [27, 31, 87, 92]. Graph‐based synergy predictors and recommender systems will increasingly propose rational combinations and label extensions, particularly for oncology and immunology [13, 28].

Knowledge Graphs and Real‐World Evidence (RWE) remain central as Biomedical KGs serve as the connective tissue across modalities, linking evidence trails from literature, experiments, and clinics; KG embeddings plus causal annotations will improve mechanistic interpretability and external validity [21, 22, 35, 42, 52, 53, 73, 93]. Benchmarks like TDC will continue to raise the bar for reproducibility (leakage‐aware splits, temporal scaffolds, clinically aligned metrics) [12, 50, 94].

Economics, IP, and Market Trajectory indicate that as platform costs fall and automation rises, AI‐driven repurposing will become a mainstream economic driver of portfolio strategy, lowering per‐program spend, extending patent lifecycles via biomarker‐defined populations and combinations, and expanding incentives for rare diseases andpediatrics. Market projections already anticipate multi‐fold growth of dedicated tools, with estimates of the segment more than tripling to >$4B by 2034 [77, 95, 96], supported by enterprise adoption of generative/graph stacks [36, 66] and broader ecosystem momentum [3, 5, 81, 82, 95, 96, 97, 98]. Strategic use of regulatory/data exclusivity and companion diagnostics can further improve incentives for off‐patent assets [3, 36, 68, 89].

Frontiers: Quantum & High‐Fidelity Simulation suggests that, while early, joint advances in quantum computing and high‐fidelity physics/ML hybrids may unlock harder target classes and conformational problems, improving the “explainability‐to‐actionability” loop for challenging indications [11, 37, 47, 99].

4. Conclusion

The period from 2020 to 2025 has marked a pivotal era in the evolution of drug repurposing, where artificial intelligence has fundamentally shifted the paradigm from sporadic, serendipitous findings to structured, data‐driven, and evidence‐integrated workflows. This transformation is vividly illustrated in the rapid deployment of AI during the COVID‐19 pandemic, where knowledge graphs and deep learning models expedited the identification of candidates like baricitinib, compressing timelines from hypothesis generation to clinical validation and regulatory approval. Similarly, in oncology, AI's integration of multi‐omics data with graph neural networks and machine learning has enabled precision repurposing, tailoring existing drugs to specific cancer subtypes and patient profiles, as seen in cases like fostamatinib for acute myeloid leukemia or ceritinib for hepatocellular carcinoma. These advancements have not only reduced development time—often from over a decade to mere months—but also slashed costs by 40%–60% in targeted programs, minimized uncertainty through predictive modeling of drug‐target interactions, and enhanced mechanistic insights via interpretable network pharmacology and embedding techniques.

Looking forward, the horizon for AI in drug repurposing is expansive and promising, with emerging technologies poised to amplify its impact across the entire pharmaceutical lifecycle. Foundation models, pretrained on vast multimodal datasets encompassing genomics, proteomics, chemistry, and clinical records, will facilitate “zero‐shot” and “few‐shot” predictions, enabling rapid indication ranking and candidate prioritization without extensive retraining. Closed‐loop automated systems, integrating AI with robotic high‐throughput platforms, will create iterative cycles where in silico hypotheses are swiftly tested in wet‐lab environments, refining models through active learning and accelerating validation from silicon to bedside. Moreover, maturing regulatory frameworks from bodies like the FDA and EMA will provide clearer guidelines on lifecycle management, emphasizing locked algorithm versions, audit trails, and human‐in‐the‐loop oversight to ensure safety and reproducibility in AI‐supported submissions.

However, harnessing this potential requires addressing persistent challenges through deliberate and sustained efforts. Investments in data quality—via standardized benchmarks like the Therapeutics Data Commons (TDC) and harmonized multi‐source corpora—will mitigate biases and improve generalizability. Transparent validation protocols, incorporating scaffold‐aware splits and orthogonal experimental assays, must become standard to bridge the gap between computational predictions and real‐world efficacy, avoiding “false theoretical friends.” Safety‐by‐design principles, embedding early toxicity, drug–drug interaction (DDI), and polypharmacy assessments, will preempt downstream failures and foster trust among clinicians and regulators. Ethical considerations, including data privacy, algorithmic fairness, and equitable access to AI tools, should guide development to prevent widening disparities in biomedical innovation.

Ultimately, by prioritizing these elements, AI‐driven drug repurposing can evolve into a cornerstone of pharmaceutical R&D—a reliable, first‐line strategy that not only revitalizes existing therapeutic arsenals but also democratizes access to novel treatments for underserved diseases. This integration holds the promise of a more agile, efficient, and patient‐centric drug development ecosystem, where AI serves as a collaborative force multiplier, empowering researchers to tackle complex health challenges with unprecedented speed and precision.

Funding

The authors have nothing to report.

Disclosure

Use Of Artificial Intelligence : No generative artificial intelligence tools were used to generate scientific content. Any language editing, if performed, was limited to improving readability and did not alter the scientific meaning.

Conflicts of Interest

The authors declare no conflicts of interest.