Data statisticsThe human interactome information was obtained from the CODA (Context-Oriented Directed Association) database, resulting in a human gene network with 22,119 nodes and 530,761 edges11. CODA database is relational database aggregating biological relationship information within the human body by extracting and standardizing data from various sources, including international public databases (including BioGrid, CTD, DiseaseConnect, EndoNet, PhenoGO, RegNetwork and so on), literature databases, and experimental data. It is the world’s largest biological relationship information database and enables high-level analysis that considers all levels from molecules to cellular functions and diseases. Biological pathway information was obtained from the KEGG (Kyoto Encyclopedia of Genes and Genomes) database, and 307 biological pathways were utilized in the research12. The gene expression data from Parkinson’s disease patients were collected from the GEO (Gene Expression Omnibus) database, comprising three series of data: GSE20292 (26 samples: 15 controls, 11 Parkinson’s disease patients), GSE20291 (35 samples: 20 controls, 15 Parkinson’s disease patients), and GSE20168 (29 samples: 15 controls, 14 Parkinson’s disease patients). Based on the biological pathway information, 34,702 edge weights were modified among the 530,709 edges (6.1%). Drug, disease and their relationship data were collected from the DrugBank and OMIM databases4. A total of 238 drugs were found to have more than 20 target genes, and among them, 11 were used for PD treatment. Additionally, 105 genes were identified as PD-related genes. PD-related genes were referenced from Menche et al.13. This paper collected relationships between diseases and genes from the OMIM database14 and UniProtKB/Swiss-Prot15. Significant relationships between diseases and genes were filtered using GWAS data from PheGenI16.Proximity between drug targets and disease genesAccording to the proximity-based drug repurposing approach, known drug-disease interactions tend to be proximal within the human gene network compared to unknown drug-disease interactions7. We performed calculations to determine the proximity of PD-associated genes with both known and unknown PD drugs using different versions of the gene network: the network with no edge weight modification, the network with modified edge weights considering the number of pathways (STEP 1) and the network with modified edge weights considering both the number of pathways and the correlation of the pathways with PD (STEP 1, 2) (Fig. 2). GSEA(Gene Set Enrichment Analysis) scores were used to quantify the correlation of the pathways with PD. Statistical analysis revealed no significant proximity difference between known and unknown Parkinson’s disease drugs within the non-weighted human gene network (p-value> 0.05). In contrast, a statistically significant difference was observed using the weighted network considering the number of pathways (STEP 1; p-value = 0.05). Finally, the weighted network considering both the number of pathways and the correlation of the pathways with PD (STEP 1, 2) demonstrated the most statistically significant difference (p-value = 0.03). Consequently, it was observed that the proximity-based PD drug interaction inference was not appropriate using the non-weighted human gene network, as there was no statistical distinction between known and unknown PD drug proximity. In contrast, the human gene network with edge weight modification based on the biological pathway information showed statistical distinction between known and unknown PD drug proximity and the network yielded more accurate inference of PD drug interactions.Fig. 2Proximity difference between known and unknown Parkinson’s disease drugs using (a) the non edge-weighted network(p-value = 0.13), (b) the edge-weighted network considering the number of pathways(p-value = 0.05), (c) the edge-weighted network considering the number of pathways and the pathways’ correlation with the Parkinson’s disease(p-value = 0.03).Performance measurement using the biological pathway based edge-weighted human gene networkThe statistical significance of the biological pathway based edge-weighted human gene network in inferring drug-disease interactions was evaluated using Receiver Operating Characteristic (ROC) curves. The tendency of known drug-disease interactions to exhibit higher proximity rankings than unknown drug-disease interactions was assessed by comparing the Area Under the ROC Curve (AUROC) values (Fig.3(a)). The non-weighted path length analysis, calculated using the non edge-weighted human gene network, resulted in an AUROC value of 0.66. The weighted path length analysis considering only the number of pathways (STEP 1) showed a higher AUROC value (AUROC=0.68). Finally, the weighted path length analysis utilizing both the number of pathways and the pathway relationship with PD (STEP 1, 2) demonstrated the highest performance, with an AUROC value of 0.69. In conclusion, when employing the proximity-based method, incorporating biological pathway knowledge into the human gene network yielded superior performance in inferring Parkinson’s disease-drug relationships. To demonstrate the statistically significant difference in AUROC between proximity calculations using non-weighted and weighted path length, we employed a permutation test. We calculated AUROC based on proximity values after randomly extracting the same number of drugs for PD drugs from the remaining drug set. Fig.3(b) demonstrates that the AUROC obtained from proximity calculations using weighted path length is significantly improved compared to calculations using non-weighted path length (p-value < 0.05).Fig. 3(a) ROC curve using the non edge-weighted network, the edge-weighted network considering the number of pathways (STEP 1) and the edge-weighted network considering the number of pathways and the relationship of the pathways with the Parkinson’s disease (STEP 1, 2). (b) Permutation test showing significantly improved AUROC using weighted path length.In order to demonstrate the statistical reliability of drug-disease interaction inference using the biological pathway based edge-weighted human gene network, we utilized the recall measure among the top 50 drugs with high proximity values. That is, we calculated the ratio of known PD drugs among these top 50 drugs. Using the non edge-weighted human gene network, the recall value was 0.27, and only three PD drugs were identified among the top 50 drugs. However, when the number of pathways was incorporated into the network edge weights, the recall improved to 0.45, with five PD drugs appearing in the top 50 drugs. Finally, when the pathway relationship with PD was included, the recall further increased to 0.64, with seven PD drugs found among the top 50 drugs. These findings indicate that the inclusion of pathway information and its relationship with PD in the biological pathway based edge-weighted human gene network increases the number of known PD drugs within the top 50 drugs with high proximity values (Fig. 4).Fig. 4Recall with top 50 highest proximity drugs using the non edge-weighted network, the edge-weighted network considering the number of pathways (STEP 1) and the edge-weighted network considering the number of pathways and the relationship of the pathways with the Parkinson’s disease (STEP 1, 2).To demonstrate the superiority of the relationship prediction performance between PD and drugs using the biological pathway based edge-weighted network construction method, we compared it with disease-drug relationship prediction models based on gene expression and biological network. The comparative models were SAveRUNNER, MNBDR, and OCTAD. SAveRUNNER utilizes proximity and similarity between drugs and diseases module on biological networks, MNBDR leverages gene connectivity in the network and gene expression in the context of diseases and drugs, and OCTAD predicts relationships between drugs and diseases by calculating the rank of gene expression for each gene in the context of diseases and drugs, thus identifying drugs that can reverse gene expression changes caused by diseases. For the construction of MNBDR and OCTAD models using gene expression, changes in gene expression due to PD were based on the GEO datasets (GSE20292, GSE20291, GSE20168) already utilized in the research, while changes in gene expression due to drugs were obtained from CMap LINCS 2020 Data an expansion upon the previous 2017 data (GSE92742) and contains \(\tilde{3}\)M gene expression profiles. As a result, in both AUROC and recall evaluation metrics, the biological pathway based edge-weighted network construction method demonstrated superior performance in predicting relationships between PD and drugs. Fig.5Fig. 5AUROC and recall performance between PD and drugs using the biological pathway based edge-weighted network construction method(Weighted path length) and other prediction methods. The method demonstrated superior performance in predicting relationships between PD and drugs.Parkinson’s disease drugs MoA analysis with biological pathway based edge-weighted human gene networkDespite the importance of understanding the mechanism of action(MoA), the proximity-based drug MoA analysis using the non edge-weighted human gene interaction network fails to explain how drugs affect diseases. In order to address this limitation, we conducted an analysis of PD drugs that exhibited the most significant increase in proximity compared to the non edge-weighted human gene network. We focused on the three drugs with the most significant increase in proximity and examined the shortest paths between the drug targets and PD-associated genes using the biological pathway based edge-weighted human gene network. Remarkably, the biological pathway based edge-weighted human gene network provided logical explanations for the mechanism of actions of the three drugs. We observed a remarkable coincidence between the improved proximity of these drugs and the shortest paths derived from the biological pathway based edge-weighted human gene network. These findings suggest that by utilizing the biological pathway based edge-weighted human gene network, we gain valuable insights into the mechanisms of action of drugs, providing a more comprehensive understanding of how the drugs impact PD.Trihexyphenidyl exhibits the most improved proximity when utilizing the biological pathway based edge-weighted human gene network. The mechanism of action of trihexyphenidyl involves the inhibition of efferent impulses by affecting dopamine and M1 muscarinic (acetylcholine) receptors17. In contrast to the non edge-weighted human gene network, which suggested a path between the drug targets and disease genes (CHRM4 – RREB1), the biological pathway based edge-weighted human gene network provides a more accurate representation. The biological pathway based edge-weighted network suggests the path; CHRM4 – GNG12 – PLCB3 – PRKN(Fig. 6). The interaction between CHRM4 and GNG12 is found in the cholinergic synapse, while the interaction between GNG12 and PLCB3 is found in the dopaminergic pathway. Finally, the interaction between PLCB3 and PRKN is found in the neurodegenerative pathway. These interactions align with the mechanism of action of trihexyphenidyl, providing a more comprehensive and accurate depiction of its mode of action.Benzatropine exhibits the second most improved proximity when using the biological pathway based edge-weighted human gene network. Benzatropine acts as an antagonist of acetylcholine and muscarinic receptors, effectively prolonging the action of dopamine18. In contrast to the non edge-weighted human gene network, which suggested the path CHRM1 – CTCF – CDH8, these interactions were not found in the KEGG pathways, thus failing to accurately describe the mechanism of action for benzatropine (Fig. 6). However, with the biological pathway based edge-weighted human gene network, the suggested path CHRM1 – GNAQ – PLA2G6 partially explains the drug’s mechanism of action; the interaction between CHRM1 and GNAQ is found in the cholinergic synapse pathway, providing a more comprehensive understanding of benzatropine’s effects in PD.Bromocriptine exhibits the third most improved proximity when utilizing the biological pathway based edge-weighted human gene network. Bromocriptine functions as a dopamine receptor agonist10. However, the non edge-weighted human gene network suggests a shortest path consisting of DRD4 – ARNT – RREB1. None of these interactions were found in the pathway database, therefore this path could not sufficiently describe the mechanism of action for bromocriptine. Conversely, the biological pathway based edge-weighted human gene network proposes an alternative path: DRD4 – GNG13 – PLCB3 – PRKN. Notably, the interaction between DRD4 and GNG13, as well as the interaction between GNG13 and PLCB3, were identified within the dopaminergic pathway. Additionally, the interaction between PLCB3 and PRKN was found within the neurodegenerative pathway (Fig. 6). These pathways are directly relevant to the mechanism of action of bromocriptine. Thus, the biological pathway based edge-weighted human gene network suggests a more accurate path between the drug and the disease, effectively describing the drug’s biological functions.Fig. 6Shortest path between (a) trihexyphenidyl, (b) benzatropine, (c) bromocriptine and the Parkinson’s disease using the pathway-weighted network corresponded to the each drug mechanism of actions.Drug repurposing with biological pathway based edge-weighted human gene networkWe utilized the distance between drug targets and PD-associated genes to identify unknown drug-disease interactions with significantly high proximity. Through this approach, we filtered and identified unknown drug-PD interactions with improved and high proximity, indicating their potential for repurposing. We suggested 10 drugs with the most improved proximity and 10 drugs with the highest proximity as potential repurposable drugs for PD, along with supporting literature evidence (Table. 1). Improved proximity means the difference between the distance between drug and PD-associated genes measured from non-weighted and biological pathway based edge-weighted human gene network. For example, if proximity between drug and PD-associated genes in the non-weighted human gene network is 1 and in the biological pathway based edge-weighted human gene network is 0.4, improved proximity is 0.6. We analyzed highly improved 10 drugs and among them, 9 drugs showed statistically significant improvement (p-value < 0.05); Baclofen showed not significant improvement (p-value=0.05). First, previous evidence supports the use of drugs with highly improved proximity when utilizing the biological pathway based edge-weighted network for PD treatment. Losartan, originally an hypertension drug, was found to reduce neurodegeneration and behavioral symptoms in a PD rat model19. Tetracycline, an antibiotic used for susceptible infections, was proposed as a treatment for PD20. Gabapentin, primarily used for peripheral neuropathic pains, postherpetic neuralgia, and partial-onset seizures, showed improvements in rigidity, bradykinesia, and tremor associated with the PD21. Baclofen, a GABA-ergic agonist used for spasticity treatment, was suggested for PD management when used in conjunction with acamprosate, which is utilized for alcohol abstinence22.Second, drugs with the highest proximity based on the biological pathway based edge-weighted network were linked to PD (Table. 2). Colchicine, used to relieve gout pain, was shown to protect dopaminergic neurons in a rat model, suggesting its potential efficacy as a PD therapeutic23. Vincristine, used to treat various conditions such as acute leukemia and malignant lymphoma, was inversely associated with PD. When administered with adriamycin to acute leukemia patients, vincristine was found to induce parkinsonian-like symptoms as a side effect24. It is worth noting that the biological pathway based edge-weighted network did not consider the improvement or aggravation of PD symptoms, hence suggesting drugs with PD as a side effect. Finally, felbamate, used for epilepsy treatment, exhibited anti-parkinsonian potential in a rat model25. In summary, the proximity-based drug repurposing approach utilizing the biological pathway based edge-weighted network proposed potential PD drugs supported by previous findings offering reliable options for further investigation and potential treatment options.Table 1 Non-Parkinson’s disease drugs with highly improved proximity.Table 2 Non-Parkinson’s disease drugs with highest proximity.
