Artificial Intelligence to Guide Repurposing of Drugs

Abstract

With the pharmacokinetics, dosing, safety, and manufacturing of approved or investigational drugs already well-characterized, drug repurposing/repositioning offers an emerging strategy to rapidly develop effective treatments for various challenging diseases. However, the growing mass of genetic and multi-omics data has not been effectively explored by the drug repurposing community due to a lack of accurate approaches. This Review aims to be an authoritative, critical, and accessible review and discussion of general interest to the drug repurposing community concerning the use of Artificial Intelligence (AI) and Machine Learning (ML) tools. Emerging questions include what is achievable with AI in this domain and what will the impact be, what does AI/ML embrace, and how can we, as geneticists, pharmacologists, and computational scientists, contribute to the discovery of new, cheap, and affordable repurposable medicines. The fast growth of genetics and multi-omics data (Genomics, Transcriptomics, Proteomics, Metabolomics, and Radiomics) and Electronic Health Records (EHR) in diverse populations contribute to answering questions, such as how to rapidly identify effective repurposable medicines, what is a clinically meaningful effect size in trials, and what are the potential implications for precision medicine. This review will discuss AI/ML for drug repurposing in the context of genetics, multi-omics, real-world data collection, and crowdsourcing of knowledge. We finally conclude by considering questions on how AI/ML methodologies can unite the diverse aspects of translational medicine for emerging treatment development in human challenging diseases.

1. Introduction

Although investment in biomedical and pharmaceutical research has increased significantly over the past two decades, specifically in understanding disease genetic risk factors, we still need to develop more effective disease-modifying treatments for multiple challenging diseases, such as Alzheimer’s disease (AD), heart disease, cancer, and Coronavirus disease 2019 (COVID-19)^1–3. This is partly because the growing mass of genetic and multi-omics datasets has not been effectively explored for drug development due to a lack of accurate approaches. Drug repurposing/repositioning reduces the time and cost of drug development^1–4. With the pharmacokinetics, dosing, safety, and manufacturing of approved or investigational drugs already well-characterized, the goal of drug repositioning is to identify new indications for drugs^4–6, e.g., an anti-inflammatory agent for arthritis might be repositioned for treatment of AD or AD-related dementia (ADRD). However, how to prioritize drug targets and candidate repurposable medicines for human complex diseases at drugome-wide and genome-wide scales is challenging.

The recent increase in generation of “multi-omics” data, including genetics, genomics, epigenomics, transcriptomics, proteomics, metabolomics, lipidomics, radiomics, and phenomics, as well as data digitalization in patient care and the pharmaceutical sector, present both challenges and opportunities (Figure 1). For example, the growing availability of big data prompts challenges for personalized clinical diagnosis and treatment of human disease. These challenges have motivated the application of advanced Artificial Intelligence (AI) and Machine Learning (ML) tools to help scientists identify repurposable drugs and accelerate progress toward effective treatments for a variety of challenging complex diseases.

Figure 1.

A proposed novel AI agent-based drug discovery pipeline for personalized medicine. a, An overview pipeline in AI-based drug repurposing. b, Constructing AI-ready datasets. These databases consist of three pivotal datasets, biomedical knowledge graph (KG) dataset, multi-omics dataset, and patient/clinical dataset. The biomedical KG dataset covers relations between drugs, diseases, genes, pathways, and tissues. The multi-omics dataset includes data about genome, transcriptome, phenome, interactome, proteome, and drug, while the patient dataset contains patient health record and genome data of patient. c, AI-assisted drug approaches for both disease-centric and target-centric drug repurposing. d, Experimental validation of AI-prioritized repurposable drugs. The AI-prioritized candidate drugs can be validated using in vitro and in vivo experiments to ensure the efficacy and safety. Subsequently, the candidate drugs with ideal efficacy and safety profiles will be moved to clinical trial on patients with similar diseases, symptoms, and genetic profiles under the precision medicine hypothesis.

2. Target-centric and disease-centric drug repurposing

There are two major ways to do drug repurposing: target-centric drug repurposing and disease-centric drug repurposing. In a recent analysis of repurposed drugs, the majority of repurposed drugs are developed from disease-centric repurposing⁷. The premise of disease-centric drug repurposing is that the same drug can be applied to treat different diseases when these diseases share similar biological pathways, symptoms, or traits⁷. A key step in conducting disease-centric drug repurposing is to identify underlying biological mechanisms of action for the target disease which are homologous to the original disease that a drug treats⁸. Sildenafil is an oral medication to treat pulmonary arterial hypertension and erectile dysfunction^9,10. Sildenafil reduces breakdown of cyclic guanosine monophosphate (cGMP) via inhibiting cGMP specific phosphodiesterase type 5 (PDE5)¹¹. Recent works suggest sildenafil has potential in AD treatment¹². By studying AD endophenotype disease modules within protein-protein interaction (PPI) networks, sildenafil stands out as a novel candidate drug for AD, along with reduced incidence of AD in real-world patient data¹³. In vivo investigations imply sildenafil helps brining back cognitive function of AD mouse by reducing activity of glycogen synthase kinase 3$\mathrm{\beta }$ (GSK3$\mathrm{\beta }$) and cyclin-dependent kinase 5 (CDK5), as well as bringing up the level of brain-derived neurotrophic factor (BDNF)¹⁴. Studies using AD patient induced pluripotent stem cells (iPSC) also indicate that sildenafil targets AD related genes and pathways¹⁵.

On the other hand, target-centric drug repurposing assumes that the same target protein is associated with different diseases, so the drug that inhibits or activates this target protein has the potential to treat these diseases. Metformin is the first line treatment for type 2 diabetes via reducing hepatic glucose production and internal absorption of glucose while improving insulin sensitivity¹⁶. Metformin activates a cellular energy sensor, AMP-activated protein kinase (AMPK) which is considered to be a significant pathway in atrial fibrillation (AF) through transcriptomic-based network analysis¹⁷. Semaglutide is a glucagon-like peptide 1 (GLP-1) receptor agonist (RA) that controls type 2 diabetes, treat obesity, and reduce risk of cardiovascular diseases^18–20. Semaglutide promotes secretion of insulin from pancreatic beta cells and cuts production of glucagon from pancreatic alpha cells, which reduces fasting and postprandial plasma glucose²¹. Administration of GLP-1 receptor agonists is reported to lower the rewarding effect of alcohol and drugs, thus mediating substance abuse²². GLP-1 receptor may also be associated with neurodegenerative diseases, such as AD, via insulin signaling²³. Studies show semaglutide demonstrates neuroprotection in rat model via reducing inflammation and apoptosis²⁴. These successful examples of drug repurposing have offered effective strategies to rapid development of potential treatment for human disease using disease-centric or target-centric approaches (Figure 1).

Yet, existing data, including genomics, transcriptomics, proteomics, longitudinal real-world data, has not yet been fully utilized and integrated for disease-modifying drug repurposing for human disease. Systematic characterization and identification of underlying pathobiology could provide a foundation for identifying disease-modifying targets and repurposable medicine. Integration of the genome, transcriptome, proteome and human interactome from diverse populations is essential for such identification using computational models as discussed below.

3. Artificial Intelligence (AI) for Drug Repurposing

3.1. Artificial Intelligence (AI)-ready datasets

Effective drug discovery pipelines require multi-modal biomedical data, such as genomics, transcriptomics, proteomics, metabolomics, imaging, biofluid markers, real-world patient data (Figure 1), and in-vivo and in vitro validation in both animal and human models. These rich databases offer necessary data for developing AI/ML models or tools for drug repurposing. There are 5 different types of databases (Table 1) which can be leveraged in drug repurposing: (1) chemoinformatic database, where structures, formulas, and other properties of molecules are included; (2) bioinformatics database, where structure, function, and additional information of proteins and genes are provided; (3) systems biology database, where molecular reactions and interactions with diseases are stored; (4) multi-omics database, where details of diseases, and genes and mutations associated with diseases are deposited, and (5) pharmacological database, where interactions and bindings between drugs and targets are present. Commonly used chemoinformatic databases (Table 1) include ChEMBL²⁵, PubChem²⁶, and DrugBank²⁷. Protein and gene databases include UniProt²⁸, PDB²⁹, AlphaFold Protein Structure Database³⁰, GenBank³¹, and Ensembl³². Systems biology and pathway databases include KEGG³³ and Reactome³⁴. Disease-gene databases include DISEASES³⁵ and DisGeNET³⁶. Drug-target interaction databases include BindingDB³⁷ and PDBbind³⁸. These comprehensive databases offer rich AI-ready datasets to build and evaluate various AI/ML models for drug repurposing/repositioning studies.

Table 1.

The list of the AI/ML ready databases and selected tools for drug repurposing.

Database name	Database type	Description	Website
ChEMBL25	Chemoinformatic	A database including chemical and genomic information of bioactive compounds.	https://www.ebi.ac.uk/chembl/
DrugBank27		A database with drugs, drug targets, and clinical information.	https://go.drugbank.com/
PubChem26		A database of molecule properties, structure, and clinical information.	https://pubchem.ncbi.nlm.nih.gov/
BindingDB37	Drug-target	A database containing information of protein-ligand binding affinity.	https://www.bindingdb.org/
PDBbind38		A database of experimentally determined binding affinity data of protein-ligand binding.	http://www.pdbbind.org.cn/
Diseases35	Disease-gene	A database of disease-gene associations curated from literature, mutation data, and genome-wide association studies.	https://diseases.jensenlab.org/
DisGeNet123		A database of genomic and human diseases.	https://disgenet.com/
GeneCards124		A database of genomic, genetic, clinical, and functional information.	https://www.genecards.org/
Open Targets125		A database of identification and prioritization of potential genes related with diseases.	https://platform.opentargets.org/
Epic Cosmos	Electronic health record (EHR)	A database of EHR data from Epic system.	https://cosmos.epic.com/
TriNetX		A database of EHR data from multiple EHR systems, including claims and mortality data.	https://trinetx.com/solutions/real-world-datasets/
AlphaFold Protein Structure Database30	Protein-gene	A database of AlphaFold predicted protein structures.	https://alphafold.ebi.ac.uk/
Ensembl126		A database of annotated genomes information.	https://www.ensembl.org/index.html
GenBank31		A database of annotated genetic sequences.	https://www.ncbi.nlm.nih.gov/genbank/
PDB127		A database of experimentally validated protein crystal 3D structures.	https://www.rcsb.org/
UniProt28		A database of protein sequences and functions.	https://www.uniprot.org/
Interactome INSIDER128	Systems biology and pathway	A protein-protein interaction database with genomic variant information.	http://interactomeinsider.yulab.org/
KEGG33		A database of genes and pathways.	https://www.genome.jp/kegg/
Reactome129		A database of human pathways and genes.	https://reactome.org/
STRING130		A database of functional protein-protein association networks	https://string-db.org/

3.2. Machine Learning (ML) and Deep Learning (DL) techniques

Machine learning (ML) and deep learning (DL) are subfields of AI, which utilize different algorithms to learn pattern from input data³⁹. ML and DL algorithms can be categorized to three major types: (1) supervised learning, where label information is required; (2) unsupervised learning, where no label information is needed; and (3) semi-supervised learning, where labels for only a small portion of the data are necessary. ML algorithms usually use structured or vector data. Molecular fingerprints, which capture chemical features of molecules, are commonly used as input of ML algorithms. Circular fingerprints⁴⁰, molecular access system (MACCS) fingerprints⁴¹, and PubChem fingerprints⁴² are widely adopted to represent structures and chemical properties of molecules. DL methods can be applied to unstructured data, such as simplified molecular input line entry system (SMILES)⁴³ formulas, amino acid sequences, molecular images, 3D structures, and electronic health records (EHRs). Drug-target interactions (DTIs) are primarily used in target-centric drug repurposing (Figure 1). Several early studies combined fingerprints and ML models to predict DTIs, which is a fundamental step in target-centric drug repurposing^44–46. Beyond using molecular fingerprints, similarities between drugs or targets can also be used as input to classical ML algorithms. In this case, similarities between chemical structure, side effect profile, amino acid sequence, and gene expression responses of drugs and targets can be measured using different metrics, followed by generating similarity matrices to recognize potential repurposable drugs and their corresponding targets for diseases⁴⁷. Jacob et al.⁴⁸, Bleakley et al.⁴⁹, and Mordelet et al.⁵⁰ developed several models based on classical ML techniques to leverage similarity matrices target-centric drug repurposing.

Sequence models are user-friendly as they do not require hard work to obtain fingerprints or similarity matrices. Smiles2vec⁵¹, DeepSMILES⁵², CHEM-BERT⁵³, and SMILES-BERT⁵⁴ are several notable examples which leverage sequence data. Sequence models can achieve promising results, yet they lack information about the 3D structure of molecules and proteins which is essential to the function of drugs and targets. Molecular image data provides 2-dimensional (2D) information about molecular structure, leading to better performance. DEEPScreen⁵⁵ and Chemception⁵⁶ are early examples that use molecular images in drug discovery. ImageMol exploits the power of pretraining to improve accuracy and generalizability⁵⁷. There are two primary ways to encode 3D structures: using nodes and edges in geometric data to represent atoms and connections or saving molecules from different angles as a video. A notable example of this is VideoMol⁵⁸, where each molecule is encoded in a video to elucidate the dynamics of molecules. Another trend in DL techniques is going multi-modality, where distinct modalities of data, such as sequence data and structure data, are fused into one DL model to provide comprehensive understanding of drugs and targets⁵⁹. MRL-Mol⁶⁰ and MMELON⁶¹ provide pilot studies in combining diverse modalities of molecules, while CLEAN-Contact⁶² achieves better performance by incorporates different modalities of proteins.

3.3. Network-based approaches

Network-based techniques are largely used in disease-centric drug repurposing, as they use information derived from disease-related networks to find repurposable drugs for a specific disease⁶³. Network-based techniques use graph algorithms to analyze graph data which consists of nodes and edges. In particular, nodes can represent either drugs, targets, patients, genes, or pathways, while edges represent relationships between different entities. Widely used networks include the protein-protein interactome (PPI) network, which helps understanding how protein functions within cells, and the drug-drug-disease network, which offers insight into how different drugs can interact with each other and cure diseases. Network-based techniques stand out at learning undiscovered relations between drugs and their potential targets, and diseases with their target proteins. A systems pharmacology-based network approach was developed in order identify hydroxychloroquine, an anti-malarial and anti-rheumatic drug, as a potential repurposable drug to reduce risk of cardiovascular diseases (CVDs) ⁶⁴. More recently, a genome-wide positioning systems network (GPSnet)⁶⁵ was developed to aim at human PPI networks with patients’ DNA and RNA sequences mapped. Ouabain, a cardiac arrhythmia and heart failure drug, was identified with anti-tumor activity in lung adenocarcinoma. DeepDTnet⁶⁶ achieves high accuracy by integrating 15 different types of networks and predicts topotecan, a topoisomerase inhibitor, to be a potential therapy against multiple sclerosis (MS) by inhibiting human retinoic-acid-receptor-related orphan receptor-gamma t (ROR-γt). By concentrating on endophenotype, an intermediate characteristic on the pathway between the genotype and disorder, a medicine network found sildenafil as a potential treatment to AD¹³.

3.4. Clinical trial emulation from real-world patient databases

Randomized controlled trials (RCT) are the gold standard for drug development⁶⁷. However, due to the stringent definition of eligibility criteria, the number of eligible trial participants is typically small. This makes the trial cohort “ideal” rather than representative of real-world patient populations. Real-world data (RWD) are comprised of practice-based observations from real-world patients and thus provide a valuable resource for population-based validation of drug repurposing⁶⁸. However, RWD is challenged by confounding factors, like sex, race, and lack of detailed clinical, biomarker, genetic information, socioeconomic, and other unknown factors. Recently there have been attempts to replicate the treatment effects obtained from RCT in RWD with the help of AI techniques. RWD cohorts with features like those of the trial participants were identified, and if individual data were available for RCT participants then (weighted) propensity score matching approaches can be applied. Due to complicated confounding factors in RWD (e.g., high-dimensionality and temporality), classical ML approaches are unable to estimate the propensity scores with high accuracy. Deep learning approaches can resolve this problem through the target trial emulation method^69,70. This method can be applied in large-scale insurance claims and found zolpidem (a FDA-approved anti-insomnia medicine) as a potential drug for slowing progression of Parkinson’s Dementia⁷¹. Using a deep learning framework on emulating clinical trials from RWD⁷², they identified 14 drug candidates which reduced risk of for AD in patient subgroups with specific clinical features. Under high-throughput target trial emulation, the team identified five top-ranked drugs (pantoprazole, gabapentin, atorvastatin, fluticasone, and omeprazole) originally intended for other indications with potential benefits for AD patients⁷³.

Due to the complex confounding situations in RWD, conventional propensity score estimation approaches based on logistic regression may not be able to accurately estimate the probability of treatment and censoring. Advanced AI approaches may achieve this goal. Recent studies have evaluated three types of strategies for the representation of EHR data for predictions: (1) convolutional neural network (CNN)-based methods, which represents EHR of each patient as an event by time matrix and perform a series of one-side convolution, activation plus pooling, followed by a final fully connected layer to perform the predictions⁷⁴; (2) recurrent neural network (RNN)-based methods, which represents each individual’s EHR as an event sequence and can be used to train a RNN-based prediction model⁷⁵; and (3) graph-based methods, which represents the EHR as a heterogeneous information network with patients and clinical events as nodes and event co-occurrence as edges, applying graph neural network based approaches to perform the predictions⁷⁶. Yet, a recent study showed that the deep learning-based propensity score model did not necessarily outperform logistic regression-based methods in confounding factor justification⁷³. In addition, data missingness is another feature of noisy RWD. To handle missingness not at random, researchers can consider selection model-based methods (e.g., outcome-dependent sampling in longitudinal outcomes⁷⁷). In addition, we can also utilize DL-based imputation methods such as the last observation carry-forward⁷⁸ and generative-adversarial nets (GANs)⁷⁹.

3.5. Biomedical Knowledge Graph for drug repurposing hypothesis generation

With the rich, relevant, and high-quality knowledge found in the literature, we can both validate the drug repurposing hypotheses generated from RWD and generate knowledge based repurposing hypotheses on their own. A recent study first generated embeddings for the entities (drugs, diseases, or genes/proteins) and relations (i.e., drug-disease associations) using knowledge graph embedding methods such as DGL-KE⁸⁰, suitable for large-scale data. These embedding methods can generate vector-based representations for both entities and relations in the same embedded semantic space, such that the repurposing hypotheses can be generated according to the ranking scores of <drug, relation, disease> triples evaluated with vector-based similarities. Using a graph foundation model for zero-shot drug repurposing from a large biomedical knowledge graph, a graph neural network and metric learning module with high accuracy were developed to rank drugs as potential indications and contraindications for 17,080 diseases, including a large number of rare diseases⁸¹.

Another work developed a comprehensive biomedical knowledge graph concerning COVID-19⁸². This COVID-19 knowledge graph consists of 15 million edges and 39 different types of relations, including connections between drugs, diseases, genes, anatomies, pharmacologic classes, and gene expression. Similar to the previous work, an embedding model was developed to extract representation vectors for <head entity, relation type, tail entity>. A deep learning model trained on this knowledge graph identified more than 40 potential repurposable drugs for COVID-19. Additional enrichment analysis was conducted to validate the predicted repurposable drugs which shows high confidence in treating COVID-19. Such biomedical knowledge graphs accompanied with highly accurate deep learning models are important AI methods in advancing drug repurposing.

3.6. Clinical and experimental validation of AI-based drug repurposing

Experimental and clinical validation are crucial steps after repurposing drugs using AI models, as such validations could confirm their real-world accuracy and reliability. Primary approaches to conduct experimental validation include using induced pluripotent stem cells (iPSCs) derived models and animal models⁶⁸, while clinical validation may use electronic health record (EHR) data or health insurance claim data (Figure 1). In a recent work where sildenafil was repurposed as a potential AD treatment, AD patient iPSC-derived neurons treated with sildenafil showed significant reduction in phosphorylated-tau181 (p-tau 181), an early biomarker of AD pathology¹³. Another work conducted additional experiments to validate the effectiveness of sildenafil on the reduction of both p-tau 181 and p-tau 205¹⁵. Additionally, enrichment analysis of differentially expressed genes (DEGs) between control and sildenafil treated groups displayed a neuroprotective effect. Another work which repurposed metformin as a therapeutic treatment for atrial fibrillation (AF) leveraged human iPSC-derived atrial-like cardiomyocytes (a-iCMs) to validate in silico predictions¹⁷. Critical cardiovascular-related markers were significantly upregulated accompanied by an increase in expression of several known markers by metformin associated with low expression in AF. In another study, AD patient iPSC-derived microglia were treated with ketorolac, a moderate to severe pain treatment by downregulating the Type-I interferon signaling, mechanistically supporting potential beneficial of ketorolac in reducing incidence of AD via targeting disease-relevant microglia⁸³.

4. Applications of AI Approaches to Drug Repurposing

Critically, drug repurposing depends on efficient searching of the vast drug space, for which the optimal approach is rapidly evolving. As drug repurposing is a complex process involving many steps, multimodal machine learning tools can significantly reduce the time and cost of drug development. For instance, multimodal machine learning approaches^84,85 improve accuracy of patient subphenotyping during clinical trial design by assembling neuroimaging, genetic, and multi-omics profiling data. With the help of deep learning, effective representations can be learned for different data modalities^86,87, which can then be fused by simple concatenation or more complicated nonlinear transformation⁸⁸ to perform downstream tasks such as molecular design, pharmacokinetics property evaluation and optimization. AI/ML tools has become a leading technique in expediting and reducing the cost of the drug development. We next turned to utilize four types of human challenging complex diseases (including AD, cancer, COVID-19, and cardiovascular diseases) to illustrate how AI accelerates the finding of treatments by repurposing approved drugs.

There are 50 repurposable drug trials (40 unique repurposed agents) based on the latest Alzheimer’s disease drug development pipeline 2024⁸⁹. To expand drug repurposing efforts, a recent machine learning-based framework, named DRIAD (Drug Repurposing In AD), has been developed to quantify the potential associations between AD biological processes and linked genetic datasets, thereby prioritizing drug candidates for repurposing⁹⁰. DRIAD prioritized baricitinib as a candidate AD drug and baricitinib is being tested in an open-label, biomarker-driven basket trial (ClinicalTrials.gov Identifier: NCT05189106) in people with AD and Amyotrophic lateral sclerosis (ALS).

Another recently developed tool, AlzGPS, is a systems biology platform with over 100 multi-omics datasets that capture molecular profiles underlying AD pathobiology⁹¹. This tool enables network-based prioritization of potential targets for AD drug repurposing⁹¹. NETTAG is a network topology-based deep learning framework to identify disease-associated genes for AD and prioritize candidate drugs and targets. Using NETTAG⁹², the team successfully identified gemfibrozil (an approved lipid regulator) is significantly associated with reduced risk of AD compared to simvastatin using an active-comparator design from real-world patient data. A deep learning methodology (deepDTnet) was developed for new target identification and drug repurposing in a heterogeneous drug-gene-disease network embedded by 15 types of chemical, genomic, phenotypic, and cellular network profiles⁶⁶ (Table 1). Trained on 732 U.S. FDA-approved small molecule drugs, deepDTnet shows high accuracy in identifying novel molecular targets for known drugs, outperforming previously published state-of-the-art methodologies. Importantly, the authors experimentally validated deepDTnet-predicted topotecan as a candidate repurposable drug for multiple sclerosis by targeting human retinoic-acid-receptor-related orphan receptor-gamma t (ROR-γt)⁶⁶.

Using multimodal analysis of single-cell/nucleus RNA-sequencing data from AD patient brains, two approved asthma drugs (fluticasone and mometasone) were found to be significantly associated with reduced likelihood of AD by targeting AD-associated microglia⁹³. Via analysis of real-world electronic insurance record data from 7.2 million patients from the IBM^® MarketScan^® Medicare Supplemental Database, two FDA-approved p300/CBP inhibitors, salsalate or diflunisal, were found to be associated with decreased incidence of AD, and neuroprotective efficacy was also validated in mice^94,95. Using an endophenotype-based in silico network medicine approach¹³, a team showed that sildenafil usage was significantly associated with reduced likelihood of AD, and further validated the findings using an AD patient induced Pluripotent Stem Cell (iPSC)-derived neuron model¹³. Another team demonstrated that bumetanide (an FDA-approved oral diuretic) provides a potential treatment for apolipoprotein APOE4-related AD using in silico approaches combining experimental and real-world evidence⁹⁶. These prototypical examples illustrate how AI-based multi-omics approaches combined with real-world patient databases and experimental approaches can rapidly identify potential treatments for AD (Figure 1). The details of computational methods for AD drug repurposing can be found in recent reviews^4–6.

Cancer indicates challenging diseases that are caused by malignant tumors involving uncontrolled growth of cells in the body⁹⁷. Cancer is the second biggest contributor to the death in the U.S. with lung cancer as the deadliest cancer⁹⁸. A major method to apply AI to repurpose drugs for cancer is leveraging cancer cell line models and patient-derived primary cells⁹⁹. A ML algorithm based on drug response to human breast cancer cell line finds 16 from 28 drugs have significantly better performance in treating human breast cancer. Additionally, electronic health records (EHRs) of patients can be used for predicting drug effectiveness on patients with cancer. For example, investigations have been made to discover potential drugs that can lower cancer mortality by using EHRs from Mayo Clinic and Vanderbilt University Medical Center¹⁰⁰.

COVID-19 is an infectious disease caused by SARS-CoV-2 virus, which became a world pandemic¹⁰¹. In the early stage of the pandemic, AI is a key technique in discovering drugs repurposable to the treatment of COVID-19¹⁰². Studies has been focused on inhibiting SARS-CoV-2 spike proteins, a key essential in the infection of virus, and reducing inflammation and immune response caused by the disease, which plays a vital role in the mortality of patients. A network-based approach identified 16 drugs and 3 drug combinations with the potential to treat COVID-19¹⁰³. ImageMol has been used to find inhibitors of SARS-CoV-2 spike proteins⁵⁷. Remarkably, melatonin, a hormone that regulates sleep, and dexamethasone, a glucocorticoid agonist, were proposed to reduce inflammation and immune response, and toremifene was presented as a potential SARS-CoV-2 spike protein inhibitor^82,104,105.

Cardiovascular disease (CVD) refers to a group of diseases and disorders concerning heart and blood vessels¹⁰⁶. CVD is the primary reason of death around the globe. Main symptoms of CVD include chest pain, palpitations, and shortness of breath. By studying human protein-protein interactome, carbamazepine, a drug to control seizures, is believed to increase the risk of CVD, while hydroxychloroquine, a malaria treatment, is considered to lower the risk of CVD⁶⁴. By incorporating information about adverse effects of drugs, approved drugs have been repurposed for CVD treatment while lowering potential adverse effects¹⁰⁷.

5. Discussion, Perspective, and Future Directions

In this review, we briefly summarizied and discussed existing AI techniques and their applications to repurposing drugs in several challenging diseases. Despite substantial advancements in AI recently, the research circles of AI and biomedicine still face significant challenges. One formidable challenge is that the selection of hyperparameters, such as learning rate and size of hidden layers, can substantially affect the model’s performance. Early strategies to find optimal hyperparameters mainly use random search and grid search, which can be computationally expensive for complex AI models with large hyperparameter space¹⁰⁸. Recent development of automated machine learning (AutoML) incorporating neural architecture search (NAS) provides a viable solution for this issue¹⁰⁹. Even though an accurate AI model is trained with high-quality labeled datasets, experimental validation of drug discovery results is still required to confirm.

Recent growth in large language models (LLMs) provides heartening solutions for advancing drug repurposing through the integration of heterogeneous data sources. These AI models can have hundreds of billions to trillions of parameters while trained on innumerable data and documents¹¹⁰. Such training equips LLMs with the ability to find relations between drugs, diseases, targets, and pathways. Yet LLMs trained on generic data are not immune to hallucinations and biased responses¹¹¹. Retrieval-augmented generation (RAG) is a method where a specialized dataset is used to supplement LLMs to reduce hallucinations and biased responses. Moreover, AI agents, emerging AI systems designed to automatically carry out tasks and complete goals with limited human interaction, are another promising direction to drug discovery^112,113. A future AI agentic system based on LLMs could execute different kinds of drug discovery inquiries by leveraging multiple state-of-the-art drug repurposing tools, such as AlphaFold3¹¹⁴ and NETTAG⁹² (Figure 1). By integrating biomedical knowledge graphs (KGs) into the LLMs of AI agents, patients, physicians, and researchers can ask any questions regarding drugs and diseases. For example, physicians can ask the AI agent to find a repurposable drug for a patient, as the patient may display severe adverse effects on existing drugs for the disease.

AI-based drug repurposing communities are leading big data and open science. Data used for training AI models must be annotated by domain experts to ensure the quality of AI models before effectively advancing the drug development process for human complex diseases. The application of AI solutions to drug repurposing is possible via cross-disciplinary teamwork and cooperation to address current gaps and challenges, such as analysis of clinical trial data and application of trial outcomes to patients who could benefit greatly from repurposable drugs. Altogether, AI is an indispensable aspect of the future of drug repurposing and precision medicine for human challenging diseases.

There are several intimidating challenges in building the current AI-ready datasets. First, multi-omics and clinical data are generated from heterogeneous patient samples across different laboratories and health care systems. Harmonization of clinical and multi-omics data plays a crucial role in securing the quality of AI models in real-world drug repurposing and data harmonization is challenging in for most basic and clinical scientists. Limited data sharing is another hurdle in AI-based drug repurposing. Bio-pharmaceutical companies have generated massive data compared to academic institutions, which cannot be shared due to various intellectual property issues. Because of the complexity of human diseases, heterogeneous data sets covering genomic, cellular, clinical, and behavioral aspects are required to advance disease understanding^115,116. Collaborations across different entities and institutions therefore need to incorporate diverse patient populations, as well as participants with diverse disease characteristics¹¹⁷, using AI technologies.

Data security is another key to inclusiveness and maintaining the trust of AI technologies. For example, Fast Healthcare Interoperability Resource (FHIR)¹¹⁸ for clinical data transmission, are crucial to data credibility and confidentiality. Another potential solution is federated learning¹¹⁹, which protects individual patient data through collective learning from multiple local sites without transferring the original raw data. Another important direction is to improve transparency and interpretability of AI models so that scientists can understand the decision-making process during drug repurposing in order to insure proper assessments of usability and potential failure cases, such as Explainable AI technologies⁹². Thus, model interpretability can be contextualized¹²⁰, with different levels of model transparency required for different applications. Model interpretation methods have various potential uses¹²¹. For example, knowledge distillation is a popular interpretation technique that aims at advancing a secondary interpretable model to approximate the prediction results of the first model¹²². These models could be vulnerable to adversarial attacks, which could manipulate the model explanations deliberately and make them unreliable. Deriving robust, reliable, and secure model interpretations is therefore an important future research direction.

Table 2.

Commonly used data resources to develop AI drug repurposing tools.

Name	Description	Link
CHEM-BERT53	A molecular sequence model that predicts molecular properties and drug targets.	https://github.com/HyunSeobKim/CHEM-BERT
MMELON61	A foundation multi-modal model that integrates different modalities of small molecules to predict molecular properties and drug targets.	https://github.com/BiomedSciAI/biomed-multi-view
ImageMol57	A foundation model that trained on molecule images to predict drug targets and molecule properties.	https://github.com/ChengF-Lab/ImageMol
VideoMol58	A foundation model that converts molecule images to videos to study molecule properties and activities.	https://github.com/ChengF-Lab/VideoMol
deepDR131	A network-based model combining 10 types of networks to repurpose drugs.	https://github.com/ChengF-Lab/deepDR
deepDTnet66	A network-based model combining 15 types of networks to identify drug targets.	https://github.com/ChengF-Lab/deepDTnet
LISA-CPI132	A molecular image model combined with target 3D structure information to predict drug target binding activity.	https://github.com/ChengF-Lab/LISA-CPI
Chemception56	A molecular image model predicts activity on drug targets.	https://github.com/Abdulk084/Chemception
MolCLR133	A molecular structure-based model that can predict molecular properties and drug targets.	https://github.com/yuyangw/MolCLR