DREG： A large-scale biological knowledge graph for Drug Repurposing

doi:10.21203/rs.3.rs-2734221/v1

Background: Developing new drugs involves significant costs. However, repurposing existing drugs can identify potential application areas for treating new diseases. Current drug repurposing methods often require strong domain knowledge and numerous biological experiments. The knowledge graph approach which leverages existing drug research and development knowledge has emerged as a promising alternative.

Results: This paper introduces the DREG (Drug Repurposing Entity Knowledge graph) model, which is based on a large-scale knowledge graph. By aggregating multiple related knowledge networks, the model extracts association information between known drug and disease entities. It then calculates entity similarity to predict unknown drug-disease relationships. The DREG model's performance, with an MRR index of 0.308 and Hits@10 of 0.628, surpasses that of the best-performing model by 4.7% (MRR) and 18.1% (Hits@10). The DREG model's effectiveness rate in quantitative experiments can reach 80%. The recommended results have clinical value, as some are already undergoing clinical trials on ClinicalTrials.gov and the China Clinical Trial Center.

Conclusion: The DREG model can predict candidate drugs for repurposing, reducing the scope of human expert evaluation and treating new diseases.

Drug repurposing

Knowledge graph

Deep learning algorithm

DREG

Currently, global drugs are generally facing problems such as lengthy R&D^[1] and production cycle and low success rate, the research and development of new drugs is a vast and complex system that projects the development of a new compound into a new drug with a particular treatment effect requires a long process. In the research and development process for new medicines, even if many funds and practices are invested, the success rate of developing new drugs is low^[2]. In recent years, with the continuous development of pharmaceutical sciences and research and developers' exploration and practice of drugs, drug repurposing is to discover the existing drugs for new applications^[3]. Since existing drugs have undergone many clinical trials, and researchers have a more thorough understanding of various indicators of existing drugs, the primary conditions for drug safety evaluation and metabolism can be quickly met, making existing drugs safer. On the other hand, the development cycle of new drugs is long, it requires a lot of human and financial resources, and the risk factors in the R&D process still need to be apparent. The successful development of a new drug and its final clinical application requires a long period of drug efficacy testing and evaluation. It can be seen that the research on drug repurposing can reduce the time and risk of new drug development. Therefore, researching drug repurposing in the computer field is a new force to promote the development of medical undertakings. It is worthy of further research by more researchers.

Over the past few years, various web-based approaches have been developed to repurpose existing drugs. For example, Zhang R et al.^[4] adopted a PubMedBERT-based precision classifier and knowledge graph technology to make it possible for the model to identify other drug candidates and generate plausible hypotheses about the link between drug candidates and covid-19. Sun P et al.^[5] developed the first three-cluster editing method, applied to drug-disease gene triangulation. Combining it with a literature mining pipeline and applying it to several databases calculates drug repurposing yields. It utilizes the n-CluE algorithm to solve the long-standing computer science problem of weighted n-cluster graph editing. Moon C et al.^[6] proposed and verified a new biomedical knowledge graph embedding algorithm (DDTE). In the analysis of the DTINet dataset, new links between drugs and diseases or between drugs and side effects were discovered, demonstrating the potential use of the algorithm in drug repurposing or side effect prediction research activities. Although predecessors have done much work on drug reusability and achieved good results, however, this field is still faced with the most significant disadvantage of the existing drugs and algorithms, which is that the model verification effect is not strong, the physical connection ability is unstable, the experimental results guide fewer questions in biological research.

To address the aforementioned issues, this paper proposes a drug repurposing model framework called DREG based on a knowledge graph. The framework leverages different similarity measurement methods and alignment and splicing principles to strengthen connections between entities. The DREG model is trained on various biological data to capture information between drug entities and disease entities, thereby inferring interactions between new drugs and diseases. We also compare the DREG model with several advanced models and use literature knowledge and clinical trials to enhance the effectiveness of the model verification.

This article is structured as follows: Section 2 introduces the data processing, similarity methods, dictionary encoding, and knowledge graphs that are crucial in building and validating models. Section 3 presents the framework and core principal technology of DREG, demonstrating that the combination of the similarity measurement fusion mechanism and the alignment and splicing principle significantly improves entity information capture and connection ability. Section 4 develops various validation methods to assess the model's effectiveness and compares its performance with existing literature methods. The experimental results demonstrate that the DREG model framework can enhance the degree of entity connection, model verification effect, and entity information capture. Finally, Section 5 concludes that the DREG model proposed in this paper can be applied to most biomedical entity datasets to study unspecified relationships between reusable drugs and diseases and achieve better-expected results.

2.1 Data preprocessing

The entity data is from multiple medical and biological websites such as Drug-central, Bio-GRID, KEGG, and Reactome^[7–10]. Due to duplication and semantically identical entities in the data, data preprocessing is required, and the purpose of data processing is to extract valuable and meaningful information from incomprehensible data. The dataset used in this paper contains related entity data such as drugs, diseases, targets, and information pathways. The duplicate and wrong items in the entity data are deleted through entity data integration. The PubTator platform^[11] verifies the entities' association. Drug-Disease, Drug-Target, Disease-Target, Target-Pathway associations and the number of datasets are shown in Table 1. The model training uses the Drug-Disease entity dataset; the rest are used to build the knowledge graph semantic network.

Table.1 Number of datasets

Dataset name	Number
Drug-Disease	90655
Drug-Target	29833
Disease-Target	1048575
Target-Pathway	538913
ALL	1707976

2.2 Similarity measurement methods

Similarity measurement can provide a stronger entity connection. Various similarity measurement methods are used in the research to evaluate the similarity between entities^[12]. The closer the entities are, the larger the similarity measurement. The farther the two entities are, the more similarity measurement is petite. Common similarity measurement methods include cosine similarity, Euclidean distance, etc. Cosine similarity evaluates the similarity between entities by calculating the cosine of the angle between two vectors^[13]. As shown in Eq. 1, A and B are n-dimensional vectors, and ||A|| and ||B|| denote the magnitude of the vector. The cosine similarity between two vectors is calculated as follows.

Similarity = cos(θ)= $\frac{{A}·{B}}{\left|\right|{A}\left|\right|\left|{B}\right||}$ = $\frac{\sum _{{i}=1}^{{n}}{A}{i}· {B}{i}}{\sqrt{\sum _{{i}=1}^{{n}}{\left({A}{i}\right)}^{2}·\sqrt{\sum _{{i}=1}^{{n}}{\left({B}{i}\right)}^{2}}}}$ (1)

Euclidean distance judges the similarity between vectors by calculating the absolute distance of each vector in multidimensional space^[14], as shown in Eq. 2, $ai$ and $ri$ are two points, where n means multidimensional. The absolute distance between $ai$ and $ri$ is calculated by Euclidean distance as follows.

Di$(ai,ri)$=$\sqrt{\sum _{{i}=1}^{{n}}{\left({a}{i}-{r}{i}\right)}^{2}}$ (2)

In this study, once the entities are connected, their data is fused and calculated using both cosine similarity and Euclidean distance measurement methods. The resulting outputs are then processed in the subsequent steps to provide valuable reference information for constructing the knowledge graph.

2.3 Dictionary Encoding

Text data cannot directly train the model. Establishing a dictionary mapping of the entity data in the dataset is necessary, and then putting the compiled text sequence data into the model for training^[15]. To clean the data of the Drug-Disease entity library and establish a dictionary mapping, we map entity names to corresponding entity codes using a one-to-one mapping and serialize the Drug-Disease entity library in text form to generate entity code sequences. These sequences are encoded in the form of [[Drug-Entity-Code_1], [Drug-Entity-Code_2], ...[Drug-Entity-Code_n]]. The detailed process is illustrated in Fig. 1.

3.1 Description of the DREG method

The DREG model is an entity connectivity model that aims to predict hidden entity relationships from existing entity connectivity relationships. DREG combines the principles of alignment and collocation with similarity measures to enable the model to predict entities accurately. The DREG model framework is divided into an embedding layer, a linear fully connected layer, and a parameter fusion layer. The role of the embedding layer is to construct entity data information, improve computational efficiency, and continuously update entity data information. The Linear fully connected layer realizes the feature fusion of entity data. To make the backward propagation speed during training faster, a hyperbolic tangent function is added for processing. Finally, the Parameter fusion layer combines the cosine similarity calculation method with the Euclidean distance calculation function. DREG model that combines these three layers can explore the similarity between various entities in a multidimensional space and finally realize the reusable drug candidate recommendation function. The overall flow of data processing is shown in Fig. 2.

3.2 Parameter similarity measurement fusion mechanism

The Parameter similarity measure fusion mechanism used in this paper can convert untrainable tensors into trainable parameter types and bind them to the list of model-trainable parameters. This ensures that when the model parameters are updated, the converted tensors are also updated accordingly^[16]. The similarity between entities is calculated by combining the cosine similarity function and the Euclidean distance through a parameter similarity measure fusion mechanism, which is then incorporated into the model. Figure 3 shows a schematic diagram of the parametric similarity measure fusion mechanism.

3.3 Drug-Disease relationship entity alignment splicing model principle

To enhance the effectiveness of the DREG model's recommendations, this paper incorporates the alignment and splicing of Drug-Disease relationships by integrating the relationships between drug and disease entities and mapping them one-to-one. From a mathematical perspective, this principle calculates a score by measuring the similarity of two vectors ^[17]. This method aims to establish a strong connection between two entities, capture relevant entity information, and assist in predicting unknown entities based on known ones. The DREG model utilizes alignment stitching to learn the entity connections, uses a fully connected layer to combine entity information, and implements prediction functions through entity associations.

The DREG drug repurposing model is based on the similarity algorithm and utilizes the idea of alignment and stitching of input and output in Attention^[18]. The Drug-Disease association was calculated as shown in Eq. 3. In this formula, ht represents the remote state of the output word, hs represents the hidden state of processing the word's context, and Va and Wa are the weight matrices to be learned by the DREG model. Use tanh as the nonlinear activation function.

Score(ht,hs) = v^T_a·tanh(Wa[ht;hs]) (3)

The role of the tanh activation function is to increase the nonlinearity of the neural network model, improve the model convergence speed, and reduce the number of iterations^[19], as shown in Eq. 4. In the formula, x represents a vector, and the hyperbolic tangent is obtained by sinhx/coshx.

tanhx =$\frac{{s}{i}{n}{h}{x}}{{c}{o}{s}{h}{x}}$=$\frac{{{e}}^{{x}}-{{e}}^{-{x}}}{{{e}}^{{x}}+{{e}}^{-{x}}}$ (4)

During the training process of the DREG model, the drug and disease entities that have been processed are aligned and spliced together to generate a new feature tensor through the embedding layer. The Parameter similarity measurement fusion mechanism is then used to obtain the similarity between entities in order to predict unknown entities. The structure of the DREG model is illustrated in Fig. 4.

4.1 Quantitative evaluation and analysis of the DREG model

To assess the validity of the model, it is usually quantified using the internal coverage method to measure the indicators. The coverage method evaluates the model by ranking the relationship between drugs and diseases. In this paper, the quantitative evaluation is carried out by learning the remaining entities, covering a single entity, and then predicting the single covered entity. This approach results in an overall quantitative evaluation of all predicted entities. The results of the quantitative evaluation experiments are presented in Table 2.

Table 2 presents a sample of randomly selected drugs from the database for coverage verification, where Randomly drawn percentage (RDP) indicates the percentage of randomly selected drugs, Effective ratio (ER) reflects the proportion of successfully verified drugs at a particular rank, and the sum of boost scores (SBS) is the product of the reciprocal of the rank of successful verification and the number of successful verifications. The calculation formula is the number of successful verifications * (1/rank before sorting). Draw proportional probability (DPP) represents the proportion of drawn drugs to all drugs in the database. Model effect multiplier (MEM) is a multiple that reflects the DREG model's effectiveness compared to untrained entity data, where MEM = SBS/DPP. For example, if 50 drugs are randomly selected from the database, the random sampling percentage is 0.88%, and two drugs are ranked before the 10th position, the effective ratio is 4%. The reciprocal is 1/10 before ranking 10 and 1/20 before ranking 20, and the sum of the improved scores is 2*(1/10) = 0.2. After calculation, the extraction probability ratio is 0.0088, and the model effect multiplier is 22.56. This indicates that the model effect multiplier is 22.56 times higher than that of the unused model effect multiplier. By quantifying the experimental results, it is possible to estimate the likelihood of reusing all drugs in the database.

**Table.2 Model Validation Quantitative Metrics**
DREG
RDP	ER	SBS	DPP	MEM
RDP	Position 10
0.08%	40%	0.2	0.0008	225.6
0.17%	20%	0.2	0.0017	112.8
0.88%	4%	0.2	0.0088	22.56
1.61%	5.49%	0.2	0.01613	12.3956
	Position 20
0.08%	40%	0.1	0.0008	112.8
0.17%	20%	0.1	0.0017	56.4
0.88%	8%	0.2	0.0088	22.56
1.61%	8.79%	0.4	0.01613	24.7912
	Position 50
0.08%	80%	0.08	0.0008	90.24
0.17%	40%	0.08	0.0017	45.12
0.88%	18%	0.18	0.0088	20.304
1.61%	19.78%	0.36	0.01613	22.312
	Position 100
0.08%	80%	0.04	0.0008	45.12
0.17%	60%	0.06	0.0017	33.84
0.88%	28%	0.14	0.0088	15.792
1.61%	32.97%	0.3	0.01613	18.5934
	Position 150
0.08%	80%	0.026666	0.0008	30.079248
0.17%	60%	0.04	0.0017	22.56
0.88%	30%	0.1	0.0088	11.28
1.61%	21.97%	0.13333	0.01613	8.2635

The DREG model has been evaluated quantitatively and analyzed to show that the top 100 drug candidates perform well in terms of coverage and verification. To determine their reusability, a case study was conducted using Advanced breast cancer as an example, where the top 100 drugs recommended by the DREG model were investigated. The results showed that among the top 100 recommended drugs, 20% belonged to the "other antineoplastic drugs" category, while the other 80% were categorized as "pain relief" and "antiviral drugs". Additionally, a search of ClinicalTrials.gov related results showed that the top 50 recommended drugs had a higher proportion of research results compared to the last 50 recommended drugs. Therefore, the top 50 recommended reusable drug candidates have been selected.

4.2 Override method for model validation results

The study suggests that the validity of the DREG model's recommended effects should be verified through the coverage method for a single drug. This involves randomly selecting a certain percentage of drugs for coverage experiments, covering one disease relationship among all related diseases of a drug, and training the model on the uncovered relationships. After training, the recommended disease entity data is obtained, and the associated covered diseases are found through the original dataset. If there are covered disease entities in the recommended disease entity database, it confirms the effectiveness of the drug repurposing model's recommendation results.

The data in Table 3 shows that some drugs perform coverage verification on diseases with existing relationships and verify whether there are diseases that cover the relationship in the recommended results. The recommended entities are ranked according to their relevance scores. The higher the drug-disease relevance, the higher the Top ranking. Verify the coverage relationship is the best Top (VRT) represents the sorted position of coverage of the disease. For example, the hypertensive disease is covered, and fish oils recommend treatable disease through training, and the recommended disease hypertensive disease is ranked 24th after the correlation score. The number of existing diseases related to the drug in the database (DRD) in the table is the number of each drug and all related diseases, and the database uses the Drug-Disease database. For example, fish oils have 31 known relationships with different diseases in the Drug-Disease database. The relevance Score is the correlation score between the drug and the disease calculated by the DREG model. For example, the correlation score between fish oils and hypertensive disease is 0.9995.

Table.3 Coverage verification example

Drug name	Override Disease name	VRT	DRD	Relevance Score
fish oils	hypertensive disease	24	31	0.9995
alteplase	myocardial infarction	13	20	0.9843
zotepine	hyperactive behavior	27	20	0.9891
reteplase	cardiovascular diseases	59	26	0.8897
levorphanol	substance withdrawal syndrome	21	18	0.9900

4.3 Knowledge Graph Visualization

This study utilizes a knowledge graph to illustrate the connections between entities, providing a more practical and intuitive approach to drug repurposing. The knowledge graph, from an R&D perspective, facilitates a semantic understanding of entity information and captures the relationships between entities. To obtain more valuable drug recommendation information, the study focuses on constructing a knowledge graph around diseases and recommends reusable drug candidates through the DREG model, achieving an effective and intuitive display of reference information. For instance, data on related drug entities of advanced breast cancer, related targets, information pathways, and other data information are integrated into a knowledge graph using Neo4j visualization software, as depicted in Fig. 5^[20].

4.4 Literature Knowledge and Results of Clinical Trials

To prove the rationality of the recommended drug results of the DREG model proposed in this paper, the rationality was demonstrated by querying the published literature and clinical trial results. First, check whether the relevant literature on the reusable drug candidates and diseases recommended by the DREG model exists through the PubMed literature website. If there is, it proves that the drug candidates recommended by the proposed model are compelling. The greater the number of literature recommending reusable drug candidates and diseases, the more meaningful the research represents the drug candidates and the diseases they recommend.

Table 4 presents the suggested drug candidates, correlation scores, and the number of PubMed documents for select diseases, as determined by the model. For instance, the model recommends cetrimonium as a drug candidate for advanced cancer with a correlation score of 0.9152, despite the lack of an association between the two in the original database. The PubMed website features five relevant literature mentions for both, and the Fitted value is derived by the sigmoid fitting of the Relevance Score and Number of related literature.

Table.4
Literature verification example
Drug name	Disease name	Relevance Score	Is it known	Number of related literature	Fitted value
diphenhydramine	advanced cancer	0.9140	recommend	74	1.0
diazepam	advanced cancer	0.9186	recommend	25	1.0
cetrimonium	advanced cancer	0.9152	recommend	5	0.997
benfotiamine	advanced cancer	0.9161	recommend	2	0.949
ajmaline	advanced cancer	0.9260	recommend	2	0.949

To reduce the amount of research required when studying drug diseases, we have implemented a rare recommendation function that utilizes a resounding validation method for literature knowledge. The process involves separating recommended reusable drug candidates from their respective diseases to obtain their literary quantity. Next, the number of relevant literature on the co-occurrence of the recommended drugs and diseases is obtained. Finally, the rarity value calculation is performed, and the rarity values are sorted from small to large, with a smaller value indicating a higher rare value.

The rare value formula is determined using the following equation: Score = (number of articles in common between drugs and diseases/number of articles about drugs + number of articles in common between drugs and diseases/number of articles about diseases). For example, if the number of documents on the drug cisplatin is 84,337 on the PubMed website, and the number of documents on advanced cancer is 365,556, with 18,848 related documents appearing together, the rare value can be calculated using the formula: 18,848 / 84,337 + 18,848 / 365,556 = 0.0970, resulting in a rare value of 0.097002. If the recommended number of articles is 0, the rare value is classified as "UnKnown". Table 5 presents some rare examples of Drug-Disease validation.

Table.5

Examples of rare recommendations

Drug name	Disease name	Relevance Score	Is it known	Number of common literature	Fitted value	Rare value
temozolomide	advanced cancer	0.9874	known	899	1.0	0.0970
cisplatin	advanced cancer	0.9855	known	18848	1.0	0.2763
sunitinib	advanced cancer	0.9852	known	1803	1.0	0.2611
cabergoline	advanced cancer	0.9276	recommend	21	1.0	0.1079
diazepam	advanced cancer	0.9186	recommend	25	1.0	0.0010
cetrimonium	advanced cancer	0.9152	recommend	5	0.997	0.0018
amoxapine	advanced cancer	0.9548	recommend	0	0.722	UnKnow

To determine the validity of the experimental results of the DREG model in clinical trials, the clinical trial progress on the ClinicalTrials.gov website was verified. ClinicalTrials.gov^[21] is a clinical trial database developed by the US National Library of Medicine (NML) and the US Food and Drug Administration (FDA) in 1997 and officially launched in February 2002. To confirm the validity of the model's recommendation, you can check the ClinicalTrials.gov website for the "drug-disease" clinical trials that correspond to the recommendation. If such trials exist, it verifies the effectiveness of the recommended drug and confirms the accuracy of the model's experimental results in clinical trials. Table 6 provides some examples of the clinical trials of drug-disease entities recommended by the model on ClinicalTrials.gov. For instance, a Phase I clinical trial for the use of paclitaxel in treating advanced breast cancer is scheduled to begin in 2022.

Table.6

Clinical Trials Clinical Validation Examples

Disease name	Drug name	Mode	Start Time	Phase	Relevance Score
advanced breast cancer	Paclitaxel	Active	2022	1	0.8378
Biliary Tract Cancer	leucovorin	Active	2023	1	0.8799
Prostate Cancer	Niraparib	Active	2020	1	0.8981
Prostate Cancer	Cyclophosphamide	Active	2023	1	0.8722
non-small cell lung cancer metastatic	Cyclophosphamide	Completed	2018	1	0.8721
non-small cell lung cancer stage iii	Carboplatin	Recruiting	2022	1	0.9305
non-small cell lung cancer	Cyclophosphamide	Recruiting	2024	1	0.8769
Colorectal Cancer Metastatic	Tipiracil	Recruiting	2024	1	0.8113
Colorectal Cancer Metastatic	celecoxib	Recruiting	2023	1	0.7924
Colorectal Cancer	Gevokizumab	Recruiting	2019	1	0.8676
Colorectal Cancer	Oxaliplatin	Recruiting	2023	1	0.8435
hormone refractory prostate cancer	Nivolumab	Recruiting	2021	1	0.8930
Prostate Cancer Metastatic	Olaparib	Recruiting	2022	1	0.8645
Prostate Cancer Metastatic	Enzalutamide	Recruiting	2023	1	0.8604
ovarian epithelial cancer recurrent	Paclitaxel	Recruiting	2021	1	0.8521
nervous system disorder	cyclophosphamide	Recruiting	2022	1	0.8882
hypertensive disease	Nitroglycerin	Recruiting	2020	1	0.8273
hypertensive disease	Trimethaphan	Recruiting	2022	1	0.8212
Seizures	levetiracetam	Recruiting	2021	1	0.7924

To strengthen the credibility of the experimental findings of the DREG model in clinical trials, a clinical trial case inquiry was conducted on the China Clinical Trials Registry website. The China Clinical Trials Registry is the primary registry of the International Clinical Trials Registration Platform of the World Health Organization, ensuring the reliability and validity of clinical trial data^[22]. To verify the efficacy of drugs recommended by the model and validate the experimental results in clinical trials, we suggest searching for relevant clinical trial cases on the China Clinical Trials Registry website. If such cases exist, they can demonstrate the potential reuse value of the drugs recommended by the model. Table 7 presents the clinical validation cases of the candidate drugs and their related diseases, as identified by the Chinese Clinical Trial Registry. For instance, in 2020, the Cancer Hospital of the Chinese Academy of Medical Sciences will conduct Phase I and Phase II clinical trials of fluorouracil for the treatment of metastatic colorectal cancer.

Table.7

Examples of Clinical Validation by China Clinical Trials Registry

Drug name	Disease name	Time	Study Phase	Location	Relevance Score
fluorouracil	Colorectal cancer metastatic	2020	I + II	Chinese Academy of Medical Sciences Cancer Hospital	0.8780
Paclitaxel	Colorectal Cancer	2020	I	Ruijin Hospital Affiliated to Shanghai Jiaotong University School of Medicine	0.8283
Oxaliplatin	Colorectal Cancer	2021	II	Department of Medical Oncology, Union Hospital Affiliated to Fujian Medical University	0.8435
Paclitaxel	advanced breast cancer	2021	II	Liaoning Cancer Hospital	0.8378
capecitabine	advanced breast cancer	2020	Exploratory studies/pre-trials	Fudan University Cancer Hospital	0.8146
doxorubicin	advanced breast cancer	2019	Post-marketing drugs	The Fourth Hospital of Hebei Medical University	0.7716
Eribulin	advanced breast cancer	2021	Post-marketing drugs	Shandong Cancer Hospital	0.8409
Olaparib	Prostate Cancer Metastatic	2021	retrospective study	Renji Hospital, Shanghai Jiaotong University School of Medicine	0.8645
Enzalutamide	Prostate Cancer Metastatic	2022	Clinical trials of new therapeutic technologies	Beijing Hospital	0.8604
Carboplatin	non-small cell lung cancer iii	2021	Exploratory studies/pre-trials	The Fifth Affiliated Hospital of Sun Yat-Sen University	0.9305
levetiracetam	Seizures	2022	Exploratory studies/pre-trials	The Seventh Affiliated Hospital of Sun Yat-sen University	0.7924

4.5 Comparison with other literature methods

Through extensive research on the theme of drug repurposing, the MRR and Hits@N indicators are evaluated for the experimental results of the proposed DREG model, and the DREG model is compared with several other current advanced models for MRR and Hits@N evaluation metrics.

The evaluation calculation of MRR is shown in Eq. 5, where S is the triplet set, |S| is the number of triplet sets, and ${rank}_{i}$refers to the link prediction rank (distance score) of the i-th triplet. The larger the MRR value, the better, indicating that the ranking conforms to the facts and that the embedding effect is good^[23].

$$MRR=\frac{1}{\left|S\right|}\sum _{i=1}^{S}\frac{1}{{rank}_{i}}$$

5

The evaluation calculation of Hits@N is shown in Eq. 6, the symbol is the same as above, and II is the indicator function (if the condition is proper, the function value is 1; otherwise, it is 0). Generally, take n as 1, 3, or 10. The larger the Hits@N index, the better^[24].

$$Hits@N=\frac{1}{\left|S\right|}\sum _{i=1}^{S}\text{Ⅱ}({rank}_{i}\le n)$$

6

Table.8

Overall predictive performance of drug-disease associations

Dataset	Model	MRR	Hits@1	Hits@3	Hits@10
GP-KG^[29]	DistMult^[25]	0.191	0.103	0.207	0.379
	TransE^[26]	0.209	0.116	0.226	0.399
	ConvE^[27]	0.216	0.126	0.232	0.399
	RotatE^[28]	0.212	0.119	0.231	0.403
	KG-Predict^[29]	0.261	0.174	0.266	0.447
	DREG	0.308	0.171	0.314	0.628

The statistics for Hits@N and MRR scores of DREG and other models are presented in Table 8. Among the five knowledge graph embedding techniques, DREG exhibits the best MRR performance. For instance, DREG outperforms the best-performing model by 4.7% (MRR), 4.8% (Hits@3), and 18.1% (Hits@10) while outperforming ConvE by 4.5% (Hits@1). As illustrated in Fig. 6, the experimental results demonstrate that the DREG model outperforms other literature methods in terms of MRR and Hits@10. Thus, the DREG model is proven to have superior prediction performance compared to other models.

This paper proposes a drug repurposing model called DREG, which is based on multiple similarity measures and alignment splicing principles. This model can effectively recommend drug candidates for reuse, reducing the need for human experts to evaluate reusable drugs. The DREG model has an effective ratio of up to 80%, according to quantitative experiments. The MRR index is 0.308, and Hits@10 is 0.628. Compared with the other five models, the DREG model performs 4.7% better in terms of MRR and 18.1% better in terms of Hits@10, making it the best model available. The recommended results of the model were consulted on ClinicalTrials.gov and China ClinicalTrial Center. The DREG model overcomes the stability problems between entities and the model verification effect and can capture entity information to enhance prediction consistency. This model is suitable for recommending drug and disease entity data and can predict the undefined relationship between drugs and diseases through biomedical entity data sets.

While the DREG model has proven effective in recommending drug candidates for reuse, some errors in entity prediction still exist. In the future, we aim to improve the model's performance and drug repurposing methods by utilizing internal knowledge from the literature. We believe that the proposed DREG drug repurposing model provides researchers with promising drug candidates and serves as a valuable starting point for future drug repurposing clinical trials.

Ethics approval and consent to participate

Not applicable.

Consent to publication

All authors give consent to publish.

Availability of Data and Material

The data that support the findings of this study are available from DrugCentral but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are however available from the corresponding author upon reasonable request and with permission of DrugCentral.

Competing interests

The authors declare that we have no known competing financial interests or personal relationships that could influence the work reported in this article.

Funding

This work is supported by the Natural Science Foundation of Jilin Province(YDZJ202301ZYTS401,YDZJ202301ZYTS288),the Science and Technology Project of the Education Department of Jilin Province (JJKH20220245KJ, JJKH20220226SK), the National Natural Science Foundation of China Joint Fund Project (U19A200496).

Authors’ contributions

Xin Feng: Conceived the study, Designed and performed the experiment, Writing – original draft. Zhansen Ma, Yongxian Xin: Provided inputs on method design, Evaluation, Case studies, Manuscript. Cuinan Yu: Evaluation, Case studies, Manuscript. Ruihao Xin: Conceived the study, Edited the manuscript.

Acknowledgments

No acknowledgement to report.

Vrontis, D. & Christofi, M. R&D internationalization and innovation: A systematic review, integrative framework and future research directions. Journal of Business Research 128, 812-823 (2021).
Schlander, M., Hernandez-Villafuerte, K., Cheng, C.-Y., Mestre-Ferrandiz, J. & Baumann, M. How much does it cost to research and develop a new drug? A systematic review and assessment. PharmacoEconomics 39, 1243-1269 (2021).
Brasch, A., Eierle, B. & Jarvis, R. Research and development investments, development costs capitalization, and credit ratings: Exploratory evidence from UK R&D-active private firms. Journal of International Accounting, Auditing and Taxation 48, 100485 (2022).
Zhang, R. et al. Drug repurposing for COVID-19 via knowledge graph completion. Journal of biomedical informatics 115, 103696 (2021).
Sun, P., Guo, J., Winnenburg, R. & Baumbach, J. Drug repurposing by integrated literature mining and drug–gene–disease triangulation. Drug discovery today 22, 615-619 (2017).
Moon, C. et al. Learning Drug-Disease-Target Embedding (DDTE) from knowledge graphs to inform drug repurposing hypotheses. Journal of Biomedical Informatics 119, 103838 (2021).
Ursu, O. et al. DrugCentral 2018: an update. Nucleic acids research 47, D963-D970 (2019).
Breitkreutz, B.-J. et al. The BioGRID interaction database: 2008 update. Nucleic acids research 36, D637-D640 (2007).
Ogata, H. et al. KEGG: Kyoto encyclopedia of genes and genomes. Nucleic acids research 27, 29-34 (1999).
Orlic-Milacic, M., Williams, M. & Wundenberg, T. Inositol phosphate metabolism.
Wei, C.-H., Allot, A., Leaman, R. & Lu, Z. PubTator central: automated concept annotation for biomedical full text articles. Nucleic acids research 47, W587-W593 (2019).
Xue, X. & Jiang, C. Matching sensor ontologies with multi-context similarity measure and parallel compact differential evolution algorithm. IEEE Sensors Journal 21, 24570-24578 (2021).
Singh, R. & Singh, S. Text similarity measures in news articles by vector space model using NLP. Journal of The Institution of Engineers (India): Series B 102, 329-338 (2021).
Wang, S., Tian, Z., Dong, K. & Xie, Q. Inconsistency of neighborhood based on Voronoi tessellation and Euclidean distance. Journal of Alloys and Compounds 854, 156983 (2021).
Nundloll, V., Smail, R., Stevens, C. & Blair, G. Automating the extraction of information from a historical text and building a linked data model for the domain of ecology and conservation science. Heliyon, e10710 (2022).
Le, T.-M., Nguyen-Tat, B.-T. & Ngo, V. M. Automated evaluation of Tuberculosis using Deep Neural Networks. EAI Endorsed Transactions on Industrial Networks and Intelligent Systems 9, e4-e4 (2022).
Wold, S., Esbensen, K. & Geladi, P. Principal component analysis. Chemometrics and intelligent laboratory systems 2, 37-52 (1987).
Ayodeji, A., Wang, W., Su, J., Yuan, J. & Liu, X. An empirical evaluation of attention-based multi-head models for improved turbofan engine remaining useful life prediction. arXiv preprint arXiv:2109.01761 (2021).
Shen, S.-L., Zhang, N., Zhou, A. & Yin, Z.-Y. Enhancement of neural networks with an alternative activation function tanhLU. Expert Systems with Applications 199, 117181 (2022).
Webber, J. in Proceedings of the 3rd annual conference on Systems, programming, and applications: software for humanity. 217-218.
Zarin, D. A., Tse, T., Williams, R. J., Califf, R. M. & Ide, N. C. The ClinicalTrials. gov results database—update and key issues. New England Journal of Medicine 364, 852-860 (2011).
Wu, T. et al. Chinese clinical trial registry: mission, responsibility and operation. Journal of evidence-based medicine 4, 165-167 (2011).
Zhang, Y., Yao, Q., Shao, Y. & Chen, L. in 2019 IEEE 35th International Conference on Data Engineering (ICDE). 614-625 (IEEE).
Zhu, R., Janowicz, K., Cai, L. & Mai, G. Reasoning over higher-order qualitative spatial relations via spatially explicit neural networks. International Journal of Geographical Information Science, 1-32 (2022).
Yang, B., Yih, W.-t., He, X., Gao, J. & Deng, L. Embedding entities and relations for learning and inference in knowledge bases. arXiv preprint arXiv:1412.6575 (2014).
Bordes, A., Usunier, N., Garcia-Duran, A., Weston, J. & Yakhnenko, O. Translating embeddings for modeling multi-relational data. Advances in neural information processing systems 26 (2013).
Dettmers, T., Minervini, P., Stenetorp, P. & Riedel, S. in Proceedings of the AAAI conference on artificial intelligence.
Sun, Z., Deng, Z.-H., Nie, J.-Y. & Tang, J. Rotate: Knowledge graph embedding by relational rotation in complex space. arXiv preprint arXiv:1902.10197 (2019).
Gao, Z., Ding, P. & Xu, R. KG-Predict: A knowledge graph computational framework for drug repurposing. Journal of biomedical informatics 132, 104133 (2022).

No competing interests reported.

Data.zip

DREG： A large-scale biological knowledge graph for Drug Repurposing

Status:

Version 1

Abstract

Figures

1 Background

2 Methods