Novel graph attention autoencoder framework with multilayered validation identifies drug repurposing candidates for COVID-19 treatment

doi:10.21203/rs.3.rs-8274293/v1

Download PDF

Research Article

Novel graph attention autoencoder framework with multilayered validation identifies drug repurposing candidates for COVID-19 treatment

https://doi.org/10.21203/rs.3.rs-8274293/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Background

The COVID-19 pandemic highlighted the critical need for drug repurposing to rapidly identify therapeutic options. While computational graph-based approaches show promise, conventional single analytical methods often fail to capture the complex pathological mechanisms needed for clinical translation.

Methods

We developed a Graph Attention Autoencoder (GATE) framework with multilayered validation integrating computational prediction, real-world data, and phenome analyses. Network-based embeddings identified drug candidates through latent space similarity analysis within biomedical knowledge graphs. Clinical potential was assessed via disproportionality analysis (DPA) using adverse event reporting system, and biological plausibility was evaluated through gene expression profiling and pathway analyses.

Results

Our GATE framework identified 17 drug candidates, including 13 with known clinical effects on COVID-19 symptoms, validating model performance. Four novel candidates (cilastatin, megestrol, drotrecogin alfa, and ethacrynic acid) with minimal prior COVID-19 associations were identified. DPA further narrowed these to two promising candidates with significant protective signals: cilastatin (reporting odds ratio [ROR]: 0.19, 95% confidence interval [CI]: 0.09–0.43) and megestrol (ROR: 0.63, 95% CI: 0.42–0.94). Pathway profiling confirmed that both drugs share molecular signatures with drugs investigated in COVID-19 clinical trials. Gene expression analyses suggest that cilastatin may have anti-viral and anti-inflammatory effects via suppression of splicing, ribosomal function and mitochondrial pathway, through DPEP1 inhibition.

Conclusions

This study presents the first systematic integration of GATE-based computational prediction, real-world evidence from adverse event databases, and phenome-level pathway profiling for COVID-19 drug repurposing. This multilayered approach enabled multidimensional candidate validation unattainable by single analytical methods alone. The identification of clinically approved drugs with established safety profiles may facilitate accelerated clinical evaluation. Furthermore, the pathway and gene expression analyses provide mechanistic working hypotheses that can streamline subsequent preclinical validation. Although preclinical and clinical validation remain essential, this framework offers a generalizable strategy for rapid candidate identification during pandemics and for diseases with unmet therapeutic needs.

drug repurposing

drug repositioning

COVID-19

SARS-CoV-2

graph neural networks

real world evidence

gene expression

hypothesis generation

The spread of infectious diseases and the emergence of new pathologies have particularly increased the demand for treatments targeting rare diseases. The COVID-19 pandemic highlighted critical gaps in rapid therapeutic development, demonstrating the urgent need for drug repurposing strategies [1]. While traditional drug development requires over 10 years and billions of dollars and enormous time commitments [2–4], with approximately 10% of success rates in clinical development [5], drug repurposing offers accelerated timelines and higher success probabilities due to established safety and pharmacological profiles [6, 7]. Indeed, several computational approaches have successfully identified COVID-19 therapeutics, including Baricitinib through network-based repositioning [8], validating the potential of systematic computational drug discovery.

Computational drug repurposing leverages public databases and bioinformatics to propose novel therapeutic candidates [6]. Graph structure-based approaches – which model relationships between diseases, drugs, and genes as interconnected networks – have emerged as powerful tools for capturing this biological complexity. Ritonavir (a component of Paxlovid), and Atorvastatin were successfully identified as potential COVID-19 treatments through network-derived relationships analysis, highlighting how graph-based representations can serve as powerful foundation for systematically uncovering repurposing opportunities [9].

In graph structure-based drug repurposing, it is common to convert graph nodes and edges into vector representations through embedding. Graph embedding methods have been applied to recommendation systems in web stores and other applications [10]. Building upon these foundations, Graph Neural Networks (GNNs), which employ neural networks for graph embedding, have been proposed [11] and have become widely used.

While various GNN methods exist, Graph Attention Networks (GATs) that utilize attention mechanisms have demonstrated superiority [12]. The attention mechanism is a technology used in Transformer Architecture, which forms the foundation of generative AI using large language models such as ChatGPT [13], and supports today's revolutionary AI evolution. Recently, these GATs have increasingly been applied to drug repurposing research [14]. However, existing GNN approaches face critical limitations when confronting complex pathological mechanisms that constrain their therapeutic discovery potential.

In fact, COVID-19 pathogenesis involves complex, multifaceted mechanisms encompassing viral replication, host immune responses, and cellular metabolic processes. Approximately 200 existing drugs may target SARS-CoV-2-associated pathways, spanning diverse targets from viral proteins (spike protein, papain-like protease (PLpro), 3-chymotrypsin-like protease (3CLpro), RNA-dependent RNA polymerase (RdRp), and helicase [15]) and host cellular machinery (mTOR-PI3K pathway [16], autophagy [17], vesicular transport systems [18], and immune response [19]). This molecular complexity presents a fundamental challenge; single analytical approaches may be insufficient for adequate exploration of therapeutic candidates.

To address this challenge, we developed a comprehensive multilayered analytical framework integrating Graph Attention Autoencoder (GATE) with validation using real-world evidence and phenome data. This represents the first systematic combination of these approaches for drug repurposing, enabling identification of candidate drugs such as cilastatin and megestrol that would remain invisible to conventional computational discovery methods while providing actionable mechanistic insights for therapeutic development.

[Graph Attention Autoencoder (GATE)]

We constructed a knowledge graph comprising three types of nodes: disease, drug, and gene (target). Data were primarily obtained from Open Targets (version 24.06) [20]. Open Targets aggregates target and disease association data from multiple public sources. Edges were defined using the following four types of relationship information: (i) "Molecular interactions" for target-target relationships, (ii) "Associations - direct (overall score)" for target-disease relationships, (iii) "Drug - mechanism of action" for drug-target relationships, (iv) "Drug - indications" for drug-disease relationships.

Initial attribute vectors for nodes were generated using Large Language Models (LLMs) to create vector representations reflecting the characteristics of each node. Using this graph and a Graph Attention Autoencoder (GATE), we computed vector representations for each node and identified drug nodes proximal to the COVID-19 node by calculating cosine distances. The implementation was performed using the Python programming language with Spektral [21] utilized for GATE construction.

[Disproportionality Analysis with Spontaneous Reporting Systems]

We utilized a spontaneous adverse event (AE) reporting database constructed by U.S. Food and Drug Administration (FDA). The FDA Adverse Event Reporting System (FAERS) contains reports of spontaneous AEs reported by consumers, pharmaceutical companies, and medical professionals worldwide. FAERS consists of seven datasets: DEMO (patient demographic and administrative information), DRUG (drug information), REAC (AEs), INDI (indications for the use of reported drugs), THER (start and end dates for the reported drugs), OUTC (outcomes for the event) and RPSR (sources for the event). The present study used FAERS data from January 2020 through March 2022, representing 3,617,885 reports. This data was obtained through JAPIC-AERS (Japan Pharmaceutical Information Center (JAPIC), Tokyo, Japan), a commercially processed FAERS database.

AEs in the FAERS database are coded using Medical Dictionary for Regulatory Activities (MedDRA) Preferred Terms (PTs), which are grouped according to defined medical conditions or areas of interest. We used the Standardized MedDRA Queries to identify PTs related to COVID-19. The search terms used for “COVID-19” are described in Table S1.

AE risk was evaluated by calculating the reporting odds ratio (ROR) and corresponding 95% confidence intervals (CI) .

The ROR and 95% CI were defined as follows:

ROR = (a/c) / (b/d)

95% CI = exp [ln(ROR) ± 1.96√(1/a + 1/b + 1/c + 1/d)]

where a, b, c and d refer to the number of individuals in the FAERS database as divided into the following four groups: (a) individuals who received the drug of interest (i.e., cilastatin) and exhibited COVID-19-related AEs; (b) individuals who received the drug of interest but did not exhibit COVID-19-related AEs; (c) individuals who did not receive the drug of interest and exhibited COVID-19-related AEs; and (d) individuals who did not receive the drug of interest and did not exhibit COVID-19-related AEs. If the upper limit of the 95% CI was < 1, a significant association was assumed between use of the drug of interest and the decreased occurrence of COVID-19-related AEs.

[Pathway Profiling]

To explore potential therapeutic agents at the phenome level, we conducted analyses using pathway alteration profiles following drug treatment. We utilized gene expression profile data from the Expanded CMAP (Connectivity Map) LINCS (Library of Integrated Network-Based Cellular Signatures) Resource 2020 [22], which contains transcriptome data generated from cell lines treated with various compounds.

To mitigate noise inherent in high-throughput methodologies, we employed a quality control process based on gene expression distribution in duplicate samples, retaining only high-confidence data. Plate-to-plate normalization was performed to eliminate assay bias. Furthermore, we applied distribution analysis to determine gene expression alteration levels, selecting only samples with gene alterations at a Q-value threshold of < 0.2.

Gene Set Enrichment Analysis (GSEA) [23, 24] was conducted on filtered samples to identify significantly altered pathways for each sample, using a false discovery rate (FDR) cutoff of < 0.20. For the GSEA analysis, we utilized ontology gene sets (C5) [25] from the Molecular Signatures Database (MSigDB) [26, 27]. Correlations between pathway profiles across samples were evaluated using Fisher's exact test to identify compounds with similar phenotypic effects.

Statistical analysis was implemented in C and Ruby, ensuring computational efficiency for large-scale comparisons. Multiple testing correction was applied using the Benjamini-Hochberg method, with a FDR threshold of 0.2 to determine statistical significance.

Pathway enrichment visualization was performed using Enrichment Map (v3.4.0) [28], a plugin for Cytoscape software (Version 3.10.1) [29]. Significantly enriched pathways with FDR < 0.20 were included in the analysis. The pathway network was constructed with each node representing an enriched gene set, and edges connecting pathways with shared genes, where the edge weight corresponds to the overlap coefficient between the gene sets.

To improve interpretability of the resulting network, we utilized the AutoAnnotate plugin (v1.5.1) [30] to cluster and label groups of functionally related pathways based on their shared gene content. This approach allowed for the identification of major biological themes among the differentially regulated pathways. Clusters were manually reviewed and annotated to ensure biological relevance.

[Use of Generative Artificial Intelligence Tools]

During the preparation of this manuscript, we used large language models (ChatGPT, OpenAI; Claude, Anthropic) to assist with English translation, improve readability and clarity, and support literature search strategies. All AI-assisted outputs were carefully reviewed and edited by the authors, who take full responsibility for the accuracy and integrity of the final manuscript.

[GATE Approach Reveals Novel Candidate Drugs for COVID-19 Treatment]

To systematically identify and validate novel drug repurposing candidates we developed a comprehensive analytical framework that addresses the limitations of standalone computational approaches. Our methodology consists of three integrated phases: (1) computational identification of candidate drugs through state-of-the-art GATE approaches utilizing attention mechanisms and data obtained from Open Targets [20], (2) concept validation of clinical potential using real-world evidence, and (3) molecular biological assessment for mechanism of action (MOA) hypotheses generation using phenome data (Fig. 1). This multilayered evaluation strategy enables us to move beyond simple similarity predictions to provide actionable insights for therapeutic development. To demonstrate the utility of this system, we applied it to identify and validate novel drug repurposing candidates for COVID-19.

AI network analysis (GATE) selects candidate drugs for COVID-19, real-world data analysis (signal detection) narrows them to drugs with protective signals, and phenome analysis (pathway profiling) evaluates their molecular plausibility and mechanisms of action.

Based on the latent space representation of GATE, we identified drug nodes showing high similarity to the COVID-19 node (Table 1). In interpreting the results, we considered different salt or derivative forms of the same compound (e.g., cilastatin and cilastatin sodium) as representing a single drug entity. Among the 17 drugs identified, 13 drugs were already known to have therapeutic effects against COVID-19 symptoms, indicating that our model identifies clinically meaningful drug-disease associations. The remaining 4 drugs (cilastatin, megestrol, drotrecogin alfa, ethacrynic acid) showed minimal prior associations with COVID-19 in the literature, revealing hidden therapeutic opportunities that would have remained invisible to conventional computational discovery approaches. Importantly, these candidate drugs and COVID-19 were not directly connected in the original knowledge graph (Fig. 2, Figure S1-3), demonstrating our GATE’s ability to discover latent associations implicit in network relationships.

Table 1

Top-ranked drug candidates based on embedding similarity to COVID-19 in the latent space learned by GATE model
ID	Name	Description	Cosine Similarity	OpenTargets Association Score	OpenTargets Rank of Score
CHEMBL1201057	CILASTATIN SODIUM	Antibiotic, Dipeptidyl Peptidase I Inhibitor	0.86067	0.01	2672
CHEMBL766	CILASTATIN	Antibiotic, Dipeptidyl Peptidase I Inhibitor	0.86006	0.01	2672
CHEMBL1201014	PREDNISOLONE SODIUM PHOSPHATE	Steroid Drug	0.85693	0.609	4
CHEMBL1201231	PREDNISOLONE PHOSPHORIC ACID	Steroid Drug	0.85526	0.609	4
CHEMBL679	EPINEPHRINE	Adrenergic Receptor Agonist	0.85111	0.372	169
CHEMBL2134724	IPRATROPIUM BROMIDE	Anticholinergic Drug, COPD and Bronchial Asthma Treatment	0.84696	0.076	737
CHEMBL1621597	IPRATROPIUM	Anticholinergic Drug, COPD and Bronchial Asthma Treatment	0.84621	0.076	737
CHEMBL1200689	NITRIC OXIDE	Nitric Oxide, Pulmonary Hypertension Treatment	0.84535	0.345	397
CHEMBL1201335	GLYCOPYRRONIUM	Anticholinergic Drug, COPD	0.84439	0.076	737
CHEMBL1437	NOREPINEPHRINE	Adrenergic Receptor Agonist	0.84266	0.372	169
CHEMBL1215	PHENYLEPHRINE	Adrenergic Receptor Agonist	0.84121	0.372	169
CHEMBL2108429	MEPOLIZUMAB	Anti-IL-5 Antibody	0.83772	0.068	799
CHEMBL1201139	MEGESTROL ACETATE	Antineoplastic Drug, Contraceptive, Progesterone Receptor Agonist	0.83667	0.086	636
CHEMBL1370	BUDESONIDE	Steroid Drug	0.83639	0.609	4
CHEMBL1364144	METHYLPREDNISOLONE ACETATE	Steroid Drug	0.83551	0.609	4
CHEMBL1201027	GLYCOPYRROLATE	Anticholinergic Drug	0.83493	0.076	737
CHEMBL2109065	DROTRECOGIN ALFA (ACTIVATED)	Antithrombotic, Anti-inflammatory, Fibrinolysis-promoting Human Activated Protein C	0.83448	0.026	1339
CHEMBL52440	DEXTROMETHORPHAN	Cough Suppressant and Expectorant (Medicon)	0.83335	0.365	334
CHEMBL1473	FLUTICASONE PROPIONATE	Steroid Drug	0.83256	0.609	4
CHEMBL3707243	GLYCOPYRRONIUM TOSYLATE	Anticholinergic Drug, Primary Axillary Hyperhidrosis Treatment	0.83239	0.076	737
CHEMBL456	ETHACRYNIC ACID	Loop Diuretic Drug	0.83171	0.004	4822

Drug candidates are ranked by cosine similarity scores in GATE latent space. Rows shown in bold indicate drugs not previously used for COVID-19 treatment, representing novel repositioning candidates. The OpenTargets association scores range from 0 to 1, with higher values indicating stronger relevance.

Subnetwork extracted from a multimodal knowledge graph consisting of drugs, diseases, and genes. Nodes are colored by type: orange for diseases, blue for drugs, and green for genes.

To further evaluate the findings, we investigated the association scores between COVID-19 and the target genes of each candidate drug using the Open Targets Platform, an integrated drug discovery database. Association scores are comprehensive relatedness indicators based on literature information, genetic evidence, experimental data, and other sources. This analysis revealed that all four candidate drugs showed low association scores with COVID-19 (Table 1), suggesting that these drugs would not be visible through existing knowledge bases. In particular, our top candidate, cilastatin, had an association score of 0.01, which is significantly lower than scores (0.068–0.609) of drugs that are known to have effects on COVID-19.

To benchmark against other approaches, we investigated whether these four drugs had been reported as COVID-19 repurposing candidates in existing studies. Cilastatin was reported as one of 78 candidates for SARS-CoV-2 3CLpro inhibition through docking simulation approaches, but detailed validation was not performed [31]. Megestrol was proposed as one of 34 COVID-19 therapeutic candidates through network-based approaches, though detailed investigation was not conducted [32]. Drotrecogin alfa has no existing reports, representing the first identification of this drug as a COVID-19 candidate. Ethacrynic acid was reported as a promising candidate for SARS-CoV-2 main protease (Mpro) inhibition through docking simulation approaches, with experimental validation also performed [33].

Although some candidate drugs identified by the GATE approach have been previously reported by other computational methods, these studies provided only preliminary identification without comprehensive evaluation of therapeutic potential. Importantly, three drugs (cilastatin, drotrecogin alfa, ethacrynic acid) excluding megestrol were not identified by conventional network-based approaches and were first detected by the GATE approach. This indicates that GATE can identify complex associations that remain undetected by conventional network analysis, exemplified by novel identification of drotrecogin alfa.

These results suggest that the GATE approach is complementary to conventional methods and has the potential to efficiently discover novel relationships between unknown drug-disease pairs. Furthermore, the identification of candidates such as ethacrynic acid, for which experimental validation has been conducted, provides additional evidence supporting the reliability of our approach. However, these remain computational predictions that require experimental validation. To provide mechanistic insights that could guide future preclinical studies, we next analyzed the potential mechanisms of action of these candidate drugs.

[Evaluation of Clinical Potential of Candidate Drugs through FAERS Disproportionality Analysis (DPA)]

We further analyzed the four drugs identified by the GATE approach using real-world adverse events data from FAERS to examine potential clinical associations with COVID-19 by DPA. In DPA, ROR and 95% CI are used to evaluate associations between drugs and diseases. An inverse association (ROR < 1) is considered protective and a 95% CI range with the upper limit < 1 indicates statistical significance. This analysis revealed striking protective signals for cilastatin and megestrol, with ROR for COVID-19 infection of 0.19 (95% CI: 0.09–0.43) and 0.63 (95% CI: 0.42–0.94), respectively. (Fig. 3). Notably, cilastatin demonstrated an exceptionally strong protective association, with patients taking this medication showing approximately 5-fold lower odds of COVID-19 infection compared to non-cilastatin users. Megestrol also exhibited a significant protective effect, with approximately 40% reduction in infection odds. These findings suggest substantial clinical potential for both drugs in COVID-19 infection, with cilastatin showing one of the most pronounced protective signals observed in real-world data.

(A) Reporting odds ratios (ROR) and 95% confidential intervals (CI) for candidate drugs through GATE-based analysis. “NA” indicates values that could not be calculated due to insufficient sample size. (B) Forest plot visualizing RORs for the same set of drugs. The horizontal bars represent 95% CIs. RORs for drotrecogin alfa and ethacrynic acid were not calculable and are indicated with asterisks (*)

To examine whether the observed inverse associations were specific to each drug, we conducted DPA for drugs in the same class as each candidate drug or for drugs that are often co-administered with the candidate drug. Cilastatin is often co-administered with imipenem, a β-lactamase inhibitor. DPA for β-lactamase inhibitors showed ROR values around 1.0, and the remarkable inverse associations observed with cilastatin were not observed (Figure S4). This result suggests that the observed COVID-19 infection suppression effect may be specific to cilastatin. Megestrol is a progesterone receptor agonist. Progesterone preparations classified as progesterone receptor agonists showed RORs around 0.5 with 95% CI upper limit less than 1, indicating inverse association trends similar to megestrol (Figure S5). This result suggests that the observed effect represents a pharmacological class effect mediated through progesterone receptors rather than a mechanism that is specific to megestrol. However, more detailed analysis is needed to clearly distinguish the influence of class effects.

The remarkable inverse association observed for cilastatin is consistent with previous studies demonstrating that this drug may attenuate COVID-19 pathology by reducing SARS-CoV-2 viral replication and providing protection from the cytokine storm observed in severe COVID-19 [34]. Based on these findings, cilastatin warrants further investigation as potential therapeutic candidate for COVID-19.

Regarding megestrol, no literature has reported direct associations with COVID-19. However, progesterone, a drug in the same class that our DPA has indicated to be protective against COVID-19, has anti-inflammatory properties, suggesting potential as a COVID-19 treatment [35–37]. Clinical studies on COVID-19 treatment using progesterone have been conducted (NCT04365127, NCT04865029) and in the NCT04865029 trial, progesterone administration in moderate to severe COVID-19 patients resulted in significant improvement in clinical status (a score of 1.5 points, corresponding to a 3-day reduction in oxygen supplementation and 2.5-day reduction in hospitalization duration) [38]. Given the observed class effect of progesterone receptor agonists and the clinical evidence supporting progesterone, megestrol merits evaluation to determine whether it confers similar protective effects.

Note that drotrecogin alfa and ethacrynic acid could not be evaluated using this DPA due to insufficient co-occurrence reports with COVID-19-related AEs in the FAERS database during the analysis period.

[Evaluation of Molecular Potential of Cilastatin and Megestrol for COVID-19 Treatment]

To further explore these findings at the phenome level, we conducted correlation analysis using pathway profiling with LINCS 2020 for cilastatin and megestrol, the two drugs that showed statistically significant inverse associations in DPA. LINCS 2020 is the successor database to the Connectivity Map (CMAP), an established platform for comparing phenomic signatures of any given drug with those associated with specific diseases or disease therapeutics [22]. For each drug, we explored compounds with correlated pathway profiles from 17,170 compounds (241,597 profiles) using Fisher's exact test.

The compounds showing the strongest correlations with cilastatin (Table 2) included artesunate and sirolimus, which are drugs that have undergone clinical trials for COVID-19. The analysis also revealed ipidacrine, ursolic acid, wortmannin, and thapsigargin, which have been reported in the literature as COVID-19 therapeutic candidates. These results suggest that cilastatin has potential as a therapeutic candidate at the phenome level.

Table 2

Top-ranked compounds positively correlated with the pathway profile of cilastatin
PertID	Name	MOA	Cell	Conc	Time	Adj. P-value	COVID-19 relation
BRD-K79759031	artesunate		HT29	10uM	6h	4.4E-42	10 clinical studies
BRD-K66896231	ipidacrine	Acetylcholinesterase inhibitor	HT29	10uM	6h	1.0E-40	1 report
BRD-K60067222	BRD-K60067222		PC3	10uM	24h	8.7E-40
BRD-K68185022	ursolic-acid	11-beta hydroxysteroid dehydrogenase inhibitor	PC3	70uM	24h	4.2E-39	4 reports
BRD-A19037878	BRD-A19037878		K562	2.5uM	24h	8.0E-39
BRD-A75409952	wortmannin	PI3K inhibitor	PC3	10uM	6h	2.6E-38	2 reports
BRD-K81855038	roxatidine		HT29	10uM	6h	4.9E-38
BRD-A79768653	sirolimus	MTOR inhibitor	PC3	3.33uM	24h	1.5E-37	9 clinical studies
BRD-K69023402	thapsigargin	ATPase inhibitor	JURKAT	0.37uM	24h	1.5E-37	3 reports
BRD-K56032964	AP-26113		HT29	0.37uM	24h	1.6E-37

The "COVID-19 relation" column indicates whether each compound has been previously associated with COVID-19 through publications or clinical trials. Details of the corresponding reports and studies are provided in Supporting Information 1 and 2.

Similarly, the compounds showing the strongest correlations with megestrol (Table 3) included tianeptine, fedratinib (TG-101348), AZD-6482 and vorinostat. Tianeptine and fedratinib (TG-101348) have undergone clinical trials for COVID-19. AZD-6482 and vorinostat have been reported as COVID-19 therapeutic candidates. These results suggest that megestrol has potential as a COVID-19 therapeutic candidate at the phenome level.

Table 3

Top-ranked compounds positively correlated with the pathway profile of megestrol
PertID	Name	MOA	Cell	Conc	Time	Adj. P-value	COVID-19 relation
BRD-K37142460	MI-2		OCILY3	10uM	24h	1.3E-08
BRD-A81370665	BI-D1870	Ribosomal protein inhibitor	HBL1	10uM	24h	1.3E-08
BRD-K01436366	XMD-1150	Leucine rich repeat kinase inhibitor	BJAB	10uM	24h	3.4E-08
BRD-K36363294	I-BET-151		HIMG002	2.5uM	24h	3.4E-08
BRD-A53077924	tianeptine	Selective serotonin reuptake enhancer	HIMG001	15uM	24h	4.1E-08	1 clinical study
BRD-K12502280	TG-101348	FLT3 inhibitor	OCILY3	10uM	4h	2.5E-07	1 clinical study
BRD-K58772419	AZD-6482	PI3K inhibitor	MCF7	1.11uM	24h	3.0E-07	1 report
BRD-K12502280	TG-101348	FLT3 inhibitor	TMD8	2.5uM	4h	5.7E-07	1 clinical study
BRD-K81418486	vorinostat	HDAC inhibitor	JURKAT	10uM	24h	5.7E-07	3 reports
BRD-K54606188	BRD-K54606188		TMD8	0.66uM	4h	1.9E-06

[Hypothesized Mechanisms of Action of Cilastatin for COVID-19 Treatment]

The correlation with established COVID-19 therapeutic candidates provided substantial evidence for the intrinsic potential of cilastatin for clinical application. To further strengthen this rationale and evaluate whether cilastatin possesses sufficient mechanistic basis to incentivize clinical development, we sought to elucidate the underlying biological mechanisms based on pathway profiling data. We argued that because pathway-level information represents a functional layer closer to biological responses than individual genes, a detailed analysis of the LINCS 2020 data can facilitate hypothesis generation regarding mechanisms of action.

In the LINCS 2020 data, cilastatin treatment significantly decreased the expression of 42 pathways in the HT29 cell line (Table S2). Enrichment Map analysis revealed that gene sets related to splicing, ribosome, mitochondria, telomere, and ubiquitin pathways were altered (Fig. 4). Based on these pathway alterations, we constructed a mechanistic hypothesis for how cilastatin might intervene in the pathophysiological processes of COVID-19 induced by SARS-CoV-2 infection (Fig. 5).

This network visualizes gene ontology (GO) biological process terms significantly down-regulated by cilastatin treatment. Each node represents a gene set corresponding to a GO term, with node size proportional to the number of genes in the set and node color reflecting statistical significance (log₁₀-transformed p-value). Edges indicate gene overlap between gene sets. Functionally related clusters are annotated.

This schematic illustrates potential therapeutic mechanisms by which cilastatin may alleviate COVID-19 pathology, inferred from pathway-level gene expression changes. Cilastatin-mediated inhibition of DPEP1 is proposed to influence multiple downstream biological processes, including ribosome biogenesis, RNA splicing, and mitochondrial metabolism. These changes converge on antiviral and anti-inflammatory outcomes such as suppression of viral replication and inflammation. Pathways shaded in blue indicate those with down-regulated gene expression following cilastatin treatment.

Cilastatin is clinically used as a dipeptidase-1 (DPEP1) inhibitor in combination with imipenem to prevent renal degradation of the antibiotic by DPEP1. Beyond this established role in antibiotic protection, our analysis of LINCS2020 data revealed that cilastatin modulatesmultiple cellular pathways related to splicing, ribosomal function, and mitochondrial processes in cells. DPEP1, in addition to its primary function of dipeptide hydrolysis in the kidney, also functions as a neutrophil adhesion receptor on vascular endothelial cells and is known to promote neutrophil recruitment during inflammation [39]. DPEP1 may restrict host ribosomes utilized by SARS-CoV-2 by controlling the PI3K/Akt/mTOR pathway [40], which regulates cellular protein synthesis and metabolism [41]. This may suppress viral replication through a mechanism similar to that of mTOR inhibitors such as sirolimus (rapamycin) [42].

Of particular interest, it is known that the nonstructural protein 16 (NSP16) of SARS-CoV-2 binds to U1/U2 snRNA and interferes with the host splicing mechanism [43]. Suppression of splicing-related pathways may function as a cellular stress response that limits host gene expression and potentially suppresses viral replication by restricting protein synthesis. Moreover, splicing abnormalities may activate antiviral responses via RIG-I-like receptors [44].

Beyond the splicing mechanism, cilastatin also affects mitochondrial pathways. Decreased expression of mitochondria-related genes suggests changes in cellular metabolic activity and immune signaling. These changes are expected to achieve balanced immune regulation by limiting energy available for viral replication while maintaining type I interferon responses through mitochondrial antiviral signaling protein (MAVS) [45].

[Study Limitations]

While our multilayered computational approach successfully identified several promising therapeutic candidates and elucidated potential mechanisms of action, there are several important limitations that must be addressed.

Limitations of the GATE approach

The predictive performance of our knowledge graph-based method heavily depends on the comprehensiveness and quality of the base knowledge graph being used. In biomedical knowledge graphs, data bias tends to occur based on research progress and attention levels, with particularly insufficient information regarding rare diseases and drugs in early development stages. This deficiency may limit the discovery of novel drug-disease associations and reduce predictability. Therefore, developing comprehensive knowledge graphs that systematically incorporate data on rare diseases and emerging therapeutics will be essential for improving comprehensiveness and reducing bias in future studies.

Limitations of real-world data analysis

Since FAERS is a spontaneous AE reporting system, it inherently contains problems such as reporting bias, underreporting, and selective reporting. A critical limitation emerged from our analysis of drotrecogin alfa and ethacrynic acid, which revealed that the method cannot analyze drugs with insufficient data in the database. This particularly affects newer drugs that lack extensive real-world usage data, limiting our ability to detect potential associations. Conversely, this suggests that classical, widely used drugs may have higher detection sensitivity due to larger volumes of available data. Furthermore, inferring causal relationships from observed associations remains challenging, and the influence of confounding factors cannot be completely eliminated.

Limitations of phenome analysis

The LINCS database used is primarily based on drug responses in cancer cell lines, which exhibit different characteristics from alveolar epithelial cells, vascular endothelial cells, and immune cells, which are important in COVID-19 pathogenesis. Additionally, drug responses under the inflammatory environment in patients during SARS-CoV-2 infection may differ significantly from normal cell culture conditions. Furthermore, it is difficult to capture host-virus interactions and other complex biological processes that can only be detected in in vivo conditions using cell line-based phenome data. While extensive animal data have been accumulated and could potentially be curated and integrated into such databases, comprehensive incorporation remains challenging due to species differences in drug target molecules, including variations in binding site homology and pharmacological responses between species.

Need for experimental validation

This study primarily focused on computational predictions and analysis of existing databases to systematically identify therapeutic candidates and elucidate potential mechanisms. While this approach provided valuable insights into drug repurposing opportunities for COVID-19 treatment, future studies should include direct validation using SARS-CoV-2-infected models and/or clinical evaluation.

In this study, we systematically evaluated the applicability of existing drugs for COVID-19 treatment using the GATE-based latent space analysis with real-world evidence and phenome-level validation. This represents the first systematic combination of these approaches for drug repurposing. We identified two candidate drugs for further investigation: cilastatin and megestrol, both clinically approved medications with established safety profiles, which may be amenable to accelerated clinical evaluation.

Our multilayered validation approach achieved high-confidence candidate drug selection by comprehensively validating clinical and molecular biological plausibility. Integration of these heterogeneous data sources – static knowledge graphs, real-world clinical-observations, and dynamic gene expression data – enabled multi-dimensional insights unattainable by single analytical method based approaches. Real-world data analysis confirmed clinical potential, while phenome data analysis established molecular biological validity and mechanistic hypotheses for both drugs.

This integrated framework establishes a new paradigm for computational drug discovery that bridges the gap between in silico prediction and clinical translation. This multimodal methodology has applications for rapid candidate identification during pandemics and for diseases and conditions with limited therapeutic options. Future developments in large language models (LLMs) and agentic AI may further enhance this GATE approach by improving knowledge graph comprehensiveness.

Future research should focus on in vivo validation and clinical trials of these repurposed drug candidates. It should also explore the applicability of this methodology to other diseases and establish selection criteria for optimal computational methods based on data scale and intended use.

Data Availability

The datasets used and analyzed in this study were obtained from publicly accessible resources, including the Open Targets Platform datasets (https://platform.opentargets.org/downloads), the CMAP LINCS 2020 collection on the CLUE platform (https://clue.io/data/CMap2020#LINCS2020), and the MSigDB C5 collection (https://www.gsea-msigdb.org/gsea/msigdb/).

Spontaneous adverse event data were obtained from JAPIC-AERS, a licensed database derived from FAERS. Due to licensing and commercial restrictions, the JAPIC-AERS data and other processed datasets used in this study are not publicly available and cannot be shared.

adverse event

confidence interval

CMAP

Connectivity Map

DPA

disproportionality analysis

DPEP1

dipeptidase-1

FAERS

FDA Adverse Event Reporting System

FDA

Food and Drug Administration

GAT

Graph Attention Network

GATE

Graph Attention Autoencoder

GNN

Graph Neural Network

LINCS

Library of Integrated Network-Based Cellular Signatures

LLM

Large Language Model

MOA

mechanism of action

ROR

reporting odds ratio

Data Availability

Acknowledgments

The authors gratefully acknowledge the valuable discussions and technical advice provided by K. Matsubara, T. Morimoto, N. Hirashima, Dr. K. Shimada, and T. Naito. We would like to thank K. Hashizume for collaboration on this work. We also thank Dr. J. Urano for insightful comments and critical suggestions on the manuscript.

In addition, a preliminary version of this work was presented as an oral presentation at the 145th Annual Meeting of the Pharmaceutical Society of Japan (Fukuoka, March 26–29, 2025), and the present manuscript has been substantially expanded to include further analyses and more in-depth mechanistic discussion.

Funding

This work and its publication were supported by GEXVal Inc. through its internal research and development budget; and the Business Restructuring Subsidy Program of the Ministry of Economy, Trade and Industry (METI), Japan [grant number R2148U00245-000]. The funders provided financial support in the form of salaries and research expenses but had no additional role in study design, data collection and analysis, manuscript preparation, or the decision to publish.

Author Information

Authors and Affiliations

GEXVal Inc., Shonan iPark 2-26-1 Muraoka Higashi, Fujisawa-shi, Kanagawa, 251-8555, Japan

Yusuke Nakayama & Juran Kato-Suzuki

Research Center for Advanced Science and Technology (RCAST), The University of Tokyo, 4-6-1 Komaba, Meguro-ku, Tokyo, 153-8904, Japan.

Shingo Tsuji

Independent Researcher, Ibaraki, Japan

Koji Yamamoto

Faculty of Pharmacy, Kindai University, 3-4-1, Kowakae, Higashi-osaka, Osaka 577-8502, Japan.

Kouichi Hosomi

Contributions

Conceptualization, Y.N., S.T., J.K-S. and K.H.; Methodology, S.T., K.H. and K.Y.; Data curation, S.T. and K.Y.; Formal analysis, Y.N. and K.H.; Investigation, Y.N. and K.H.; Writing—original draft preparation, Y.N.; Writing—review and editing, S.T., J.K-S. and K.H.; Supervision, Y.N., S.T., J.K-S. and K.H. All authors have read and agreed to the published version of the manuscript.

Corresponding author

Correspondence to Yusuke Nakayama

Ethics declarations

Ethics approval and consent to participate

This study used only secondary, de-identified data from publicly available databases and did not involve human participants, animals, or human tissue. Ethics approval and consent to participate were therefore not required. All research was conducted in accordance with ICMJE, MEXT, and JSPS ethical guidelines.

Consent for publication

All authors have agreed to the publication.

Competing interests

J.K. is the founder representative director of GEXVal Inc. Y.N. is an employee of GEXVal Inc. S.T., K.Y. and K.H. are external consultants providing scientific and technical advice under service agreements with GEXVal Inc. for this academic study. The algorithms and methodologies reported in this study will be utilized by GEXVal Inc. for internal research and development activities as well as external collaborations. All authors have reviewed and approved this statement.

Serafin MB, et al. Drug repositioning is an alternative for the treatment of coronavirus COVID-19. Int J Antimicrob Agents. 2020;55(6):105969.
Thomas D et al. Clinical Development Success Rates and Contributing Factors 2011–2020 (Biotechnology Innovation Organization, 2021). 2021; Available from: https://go.bio.org/rs/490-EHZ-999/images/ClinicalDevelopmentSuccessRates2011_2020.pdf. Accessed 27 November 2025.
Simoens S, Huys I. R&D Costs of New Medicines: A Landscape Analysis. Front Med (Lausanne). 2021;8:760762.
Sertkaya A, et al. Costs of Drug Development and Research and Development Intensity in the US, 2000–2018. JAMA Netw Open. 2024;7(6):e2415445.
Wong CH, Siah KW, Lo AW. Estimation of clinical trial success rates and related parameters. Biostatistics. 2018;20(2):273–86.
Pushpakom S, et al. Drug repurposing: progress, challenges and recommendations. Nat Rev Drug Discov. 2019;18(1):41–58.
Nosengo N. Can you teach old drugs new tricks? Nature. 2016;534(7607):314–6.
Saber-Ayad M et al. Current Status of Baricitinib as a Repurposed Therapy for COVID-19. Pharmaceuticals (Basel), 2021. 14(7).
Morselli Gysi D et al. Network medicine framework for identifying drug-repurposing opportunities for COVID-19. Proc Natl Acad Sci U S A, 2021. 118(19).
Cai H, Zheng VW, Chang KC-C. A Comprehensive Survey of Graph Embedding: Problems, Techniques, and Applications. IEEE Trans Knowl Data Eng. 2018;30(9):1616–37.
Cao S, Lu W, Xu Q. Deep Neural Networks for Learning Graph Representations. Proceedings of the AAAI Conference on Artificial Intelligence, 2016. 30(1).
Salehi A, Davulcu H. Graph Attention Auto-Encoders. IEEE.
Vaswani AS, Parmar N, Uszkoreit N, Jones J, Gomez L, Kaiser AN, Polosukhin Łukasz. Illia, Attention Is All You Need. Advances in Neural Information Processing Systems. Curran Associates, Inc.; 2017. p. 30.
Li W et al. Drug repurposing based on the DTD-GNN graph neural network: revealing the relationships among drugs, targets and diseases. BMC Genomics, 2024. 25(1).
Muhammed Y. Molecular targets for COVID-19 drug development: Enlightening Nigerians about the pandemic and future treatment. Biosaf Health. 2020;2(4):210–6.
Kindrachuk J, et al. Antiviral potential of ERK/MAPK and PI3K/AKT/mTOR signaling modulation for Middle East respiratory syndrome coronavirus infection as identified by temporal kinome analysis. Antimicrob Agents Chemother. 2015;59(2):1088–99.
Gassen NC, et al. SKP2 attenuates autophagy through Beclin1-ubiquitination and its inhibition reduces MERS-Coronavirus infection. Nat Commun. 2019;10(1):5770.
Sureda A, et al. Endoplasmic reticulum as a potential therapeutic target for covid-19 infection management? Eur J Pharmacol. 2020;882:173288.
Han N et al. Identification of SARS-CoV-2-induced pathways reveals drug repurposing strategies. Sci Adv, 2021. 7(27).
Open Targets Platform Datasets. Available from: https://platform.opentargets.org/downloads. Accessed 27 November 2025.
Grattarola D, Alippi C. Graph Neural Networks in TensorFlow and Keras with Spektral. 2020. arXiv:2006.12138 DOI: 10.48550/arXiv.2006.12138.
CMAP LINCS. 2020. Available from: https://clue.io/data/CMap2020#LINCS2020. Accessed 27 November 2025.
Subramanian A et al. Gene set enrichment analysis: A knowledge-based approach for interpreting genome-wide expression profiles. Proceedings of the National Academy of Sciences, 2005. 102(43): pp. 15545–15550.
Mootha VK, et al. PGC-1α-responsive genes involved in oxidative phosphorylation are coordinately downregulated in human diabetes. Nat Genet. 2003;34(3):267–73.
MSigDB. : c5. Available from: https://www.gsea-msigdb.org/gsea/msigdb/human/genesets.jsp?collection=C5. Accessed 27 November 2025.
MSigDB. Available from: https://www.gsea-msigdb.org/gsea/msigdb/. Accessed 27 November 2025.
Liberzon A, et al. Molecular signatures database (MSigDB) 3.0. Bioinformatics. 2011;27(12):1739–40.
Merico D, et al. Enrichment Map: A Network-Based Method for Gene-Set Enrichment Visualization and Interpretation. PLoS ONE. 2010;5(11):e13984.
Shannon P, et al. Cytoscape: a software environment for integrated models of biomolecular interaction networks. Genome Res. 2003;13(11):2498–504.
Kucera M et al. AutoAnnotate: A Cytoscape app for summarizing networks with semantic annotations. F1000Res, 2016. 5: p. 1717.
Wu C, et al. Analysis of therapeutic targets for SARS-CoV-2 and discovery of potential drugs by computational methods. Acta Pharm Sinica B. 2020;10(5):766–88.
Zhou Y, et al. A network medicine approach to investigation and population-based validation of disease manifestations and drug repurposing for COVID-19. PLoS Biol. 2020;18(11):e3000970.
Isgrò C, Sardanelli AM, Palese LL. Systematic Search for SARS-CoV-2 Main Protease Inhibitors for Drug Repurposing: Ethacrynic Acid as a Potential Drug. Viruses. 2021;13(1):106.
González-Nicolás M, et al. Cilastatin: a potential treatment strategy against COVID-19 that may decrease viral replication and protect from the cytokine storm. Clin Kidney J. 2020;13(5):903–5.
Shilpa. COVID-19 and Progesterone: Part 1. SARS-CoV-2, Progesterone and its potential clinical use. Endocr Metabolic Sci. 2021;5:100109.
Su S et al. Modulation of innate immune response to viruses including SARS-CoV-2 by progesterone. Signal Transduct Target Therapy, 2022. 7(1).
Al-Kuraishy HM et al. New insights on the potential effect of progesterone in Covid‐19: Anti‐inflammatory and immunosuppressive effects. Immunity, Inflammation and Disease, 2023. 11(11).
Ghandehari S, et al. Progesterone in Addition to Standard of Care vs Standard of Care Alone in the Treatment of Men Hospitalized With Moderate to Severe COVID-19: A Randomized, Controlled Pilot Trial. Chest. 2021;160(1):74–84.
Choudhury SR, et al. Dipeptidase-1 Is an Adhesion Receptor for Neutrophil Recruitment in Lungs and Liver. Cell. 2019;178(5):1205–e122117.
Cui X, et al. DPEP1 is a direct target of miR-193a-5p and promotes hepatoblastoma progression by PI3K/Akt/mTOR pathway. Cell Death Dis. 2019;10(10):701.
Appelberg S, et al. Dysregulation in Akt/mTOR/HIF-1 signaling identified by proteo-transcriptomics of SARS-CoV-2 infected cells. Emerg Microbes Infect. 2020;9(1):1748–60.
Khan N. mTOR: A possible therapeutic target against SARS-CoV-2 infection. Arch Stem Cell Ther. 2021;2(1):5–7.
Banerjee AK, et al. SARS-CoV-2 Disrupts Splicing, Translation, and Protein Trafficking to Suppress Host Defenses. Cell. 2020;183(5):1325–e133921.
Chang AY, et al. Modulation of SF3B1 in the pre-mRNA spliceosome induces a RIG-I-dependent type I IFN response. J Biol Chem. 2021;297(5):101277.
Hanada Y, et al. MAVS is energized by Mff which senses mitochondrial metabolism via AMPK for acute antiviral immunity. Nat Commun. 2020;11(1):5711.

Competing interest reported. J.K. is the founder representative director of GEXVal Inc. Y.N. is an employee of GEXVal Inc. S.T., K.Y. and K.H. are external consultants providing scientific and technical advice under service agreements with GEXVal Inc. for this academic study. The algorithms and methodologies reported in this study will be utilized by GEXVal Inc. for internal research and development activities as well as external collaborations. All authors have reviewed and approved this statement.

Download PDF

Reviewers invited by journal
07 Jan, 2026
Editor invited by journal
10 Dec, 2025
Editor assigned by journal
04 Dec, 2025
Submission checks completed at journal
04 Dec, 2025
First submitted to journal
03 Dec, 2025

You are reading this latest preprint version

Novel graph attention autoencoder framework with multilayered validation identifies drug repurposing candidates for COVID-19 treatment

Status:

Version 1

Abstract

Background

Methods

Results

Conclusions

Figures

Introduction

Materials and Methods

[Graph Attention Autoencoder (GATE)]

[Disproportionality Analysis with Spontaneous Reporting Systems]

[Pathway Profiling]

[Use of Generative Artificial Intelligence Tools]

Results and Discussion

[GATE Approach Reveals Novel Candidate Drugs for COVID-19 Treatment]

[Evaluation of Clinical Potential of Candidate Drugs through FAERS Disproportionality Analysis (DPA)]

[Evaluation of Molecular Potential of Cilastatin and Megestrol for COVID-19 Treatment]

[Hypothesized Mechanisms of Action of Cilastatin for COVID-19 Treatment]

[Study Limitations]

Conclusions

Data Availability

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1