Abstract
Indonesia has the largest medicinal plant species in the world and these plants are used as Jamu medicines. Jamu medicines are popular traditional medicines from Indonesia and we need to systemize the formulation of Jamu and develop basic scientific principles of Jamu to meet the requirement of Indonesian Healthcare System. We propose a new approach to predict the relation between plant and disease using network analysis and supervised clustering. At the preliminary step, we assigned 3138 Jamu formulas to 116 diseases of International Classification of Diseases (ver. 10) which belong to 18 classes of disease from National Center for Biotechnology Information. The correlation measures between Jamu pairs were determined based on their ingredient similarity. Networks are constructed and analyzed by selecting highly correlated Jamu pairs. Clusters were then generated by using the network clustering algorithm DPClusO. By using matching score of a cluster, the dominant disease and high frequency plant associated to the cluster are determined. The plant to disease relations predicted by our method were evaluated in the context of previously published results and were found to produce around 90% successful predictions.
1. Introduction
Big data biology, which is a discipline of data-intensive science, has emerged because of the rapid increasing of data in omics fields such as genomics, transcriptomics, proteomics, and metabolomics as well as in several other fields such as ethnomedicinal survey. The number of medicinal plants is estimated to be 40,000 to 70,000 around the world [1] and many countries utilize these plants as blended herbal medicines, for example, China (traditional Chinese medicine), Japan (Kampo medicine), India (Ayurveda, Siddha, and Unani), and Indonesia (Jamu). Nowadays, the use of traditional medicines is rapidly increasing [2, 3]. These medicines consist of ingredients made from plants, animals, minerals, or combination of them. The traditional medicines have been used for generations for treatments of diseases or maintaining health of people and the most popular form of traditional medicine is herbal medicine. Blended herbal medicines as well as single herb medicines include a large number of constituent substances which exert effects on human physiology through a variety of biological pathways. The KNApSAcK Family database systems can be used to comprehensively understand the medicinal usage of plants based upon traditional and modern knowledge [4, 5]. This database has information about the selected herbal ingredients, that is, the formulas of Kampo and Jamu, omics information of plants and humans, and physiological activities in humans. Jamu is generally composed based on the experience of the users for decades or even hundreds of years. However, versatile scientific analyses are needed to support their efficacy and their safety. Attaining this objective is in accordance with the 2010 policy of the Ministry of Health of Indonesian Government about scientification of Jamu. Thus, it is required to systemize the formulations and develop basic scientific principles of Jamu to meet the requirement of Indonesian Healthcare System. Afendi et al. initiated and conducted scientific analysis of Jamu for finding the correlation between plants, Jamu, and their efficacy using statistical methods [6–8]. They used Biplot, partial least squares (PLS), and bootstrapping methods to summarize the data and also focused on prediction of Jamu formulations. These methods give a good understanding about relationship between plants, Jamu, and their efficacy. Among 465 plants used in 3138 Jamu, 190 plants were shown to be effective for at least one efficacy and these plants were considered to be the main ingredients of Jamu. The other 275 plants are considered to be supporting ingredients in Jamu because their efficacy has not been established yet.
Network biology can be defined as the study of the network representations of molecular interactions, both to analyze such networks and to use them as a tool to make biological predictions [9]. This study includes modelling, analysis, and visualizations, which holds important task in life science today [10]. Network analysis has been increasingly utilized in interpreting high throughput data on omics information, including transcriptional regulatory networks [11], coexpression networks [12], and protein-protein interactions [13]. We can easily describe relationship between entities in the network and also concentrate on part of the network consisting of important nodes or edges. These advantages can be adopted for analyzing medicinal usage of plants in Jamu and diseases. Network analysis provides information about groups of Jamu that are closely related to each other in terms of ingredient similarity and thus allows precise investigation to relate plants to diseases. On the other hand, multivariate statistical methods such as PLS can assign plants to efficacy by global linear modeling of the Jamu ingredients and efficacy. However, there is still lack of appropriate network based methods to learn how and why many plants are grouped in certain Jamu formula and the combination rule embedding numerous Jamu formulas.
It is needed to explore the relationship between Indonesian herbal plants used in Jamu medicines and the diseases which are treated using Jamu medicines. When effectiveness of a plant against a disease is firmly established, then further analysis about that plant can be proceeded to molecular level to pinpoint the drug targets. The present study developed a network based approach for prediction of plant-disease relations. We utilized the Jamu data from the KNApSAcK database. A Jamu network was constructed based on the similarity of their ingredients and then Jamu clusters were generated using the network clustering algorithm DPClusO [14, 15]. Plant-disease relations were then predicted by determining the dominant diseases and plants associated with selected Jamu clusters.
2. Methods
2.1. Concept of the Methodology
Jamu medicines consist of combination of medicinal plants and are used to treat versatile diseases. In this work we exploit the ingredient similarity between Jamu medicines to predict plant-disease relations. The concept of the proposed method is depicted in Figure 1. In step 1 a network is constructed where a node is a Jamu medicine and an edge represents high ingredient similarity between the corresponding Jamu pair. In Figure 1, the nodes of the same color indicate the Jamu medicines used for the same disease. The similarity is represented by Pearson correlation coefficient [16, 17]; that is, where is the weight of plant- in Jamu , is the weight of plant- in Jamu , is mean of Jamu , and is mean of Jamu . The higher similarity between Jamu pairs the higher the correlation value. In the present study, and are assigned as 1 or 0 in cases the th plant is, respectively, included or not included in the formula. Under such condition, Pearson correlation corresponds to fourfold point correlation coefficient; that is, where , , , and represent the numbers of plants included in both and , in only , in only , and in neither nor , respectively.
In step 2 the Jamu clusters are generated using network clustering algorithm DPClusO. DPClusO can generate clusters characterized by high density and identified by periphery; that is, the Jamu medicines belonging to a cluster are highly cohesive and separated by a natural boundary. Such clusters contain potential information about plant-disease relations.
In step 3 we assess disease-dominant clusters based on matching score represented by the following equation: Matching score of a cluster is the ratio of the highest number of Jamu associated with a single disease to the total number of Jamu in the cluster. We assign a disease to a cluster for which the matching score is greater than a threshold value. In step 4, we determine the frequency of plants associated with a cluster if and only if a disease is assigned to it in the previous step. The highest frequency plant associated to a cluster is considered to be related to the disease assigned to that cluster. True positive rates (TPR) or sensitivity was used to evaluate resulting plants. TPR is the proportion of the true positive predictions out of all the true predictions, defined by the following formula [18]: where true positive is the number of correctly classified and false negative is the number of incorrectly rejected entities. We refer to the proposed method as supervised clustering because after generation of the clusters we narrow down the candidate clusters for further analysis based on supervised learning and thus improve the accuracy of prediction of the proposed method.
3. Result and Discussion
3.1. Construction and Comparison of Jamu and Random Networks
We used the same number of Jamu formulas from previous research [6], 3138 Jamu formulas, and the set union of all formulas consists of 465 plants. We assigned 3138 Jamu formulas to 116 diseases of International Classification of Diseases (ICD) version 10 from World Health Organization (WHO, Table 1) [19]. Those 116 diseases are mapped to 18 classes of disease, which contains 16 classes of disease from National Center for Biotechnology Information (NCBI) [20] and 2 additional classes. Table 2 shows distribution of 3138 Jamu into 18 classes of disease. According to this classification, most Jamu formulas are useful for relieving muscle and bone, nutritional and metabolic diseases, and the digestive system. Furthermore, there is no Jamu formula classified into glands and hormones and neonatal disease classes. We excluded 4 Jamu formulas which are used to treat fever in the evaluation process because this symptom is very general and almost appeared in all disease classes. Jamu-plant-disease relations can be represented using 2 matrices: first matrix is Jamu-plant relation with dimension and the second matrix is Jamu-disease relation with dimension .
After completion of data acquisition process, we calculated the similarity between Jamu pairs using correlation measure. The similarity measures between Jamu pairs were determined based on their ingredients. Corresponding to (3138 in present case) Jamu formulas, there can be maximum = = 4,921,953 Jamu pairs. We sorted the Jamu pairs based on correlation value using descending order and selected top- (0.7%, 0.5%, and 0.3%) pairs of Jamu formula to create 3 sets of Jamu pairs. The number of Jamu pairs for 0.7%, 0.5%, and 0.3% datasets is 34,454 pairs, 24,610 pairs, and 14,766 pairs and the corresponding minimum correlation values are 0.596, 0.665, and 0.718, respectively. The three datasets of Jamu pairs can be regarded as three undirected networks (step 1 in Figure 1) consisting of 2779, 2496, and 2085 Jamu formulas, respectively (Table 3). Figure 2 shows visualization of 0.7% Jamu networks using Cytoscape Spring Embedded layout. We verified that the degree distributions of the Jamu networks are somehow close to those of scale-free networks, that is, roughly are of power law type. However, in the high-degree region the power law structure is broken (Figure 3). Nearly accurate relation of power laws between medicinal herbs and the number of formulas utilizing them was observed in Jamu system but not in Kampo (Japanese crude drug system) [4]. The difference of formulas between Jamu and Kampo can be explained by herb selection by medicinal researchers based on the optimization process of selection [4]. Thus, the broken structure of power law corresponding to Jamu networks is associated with the fact that selection of Jamu pairs based on ingredient correlation leads to nonrandom selection. We also constructed random networks according to Erdős-Rényi (ER) model [21], Barabási-Albert (BA) model [22], and Vazquez’s Connecting Nearest Neighbor (CNN) model [23] of the same size corresponding to each of the real Jamu network. We used Cytoscape Network Analyzer plugin [24] and R software for analyzing the characteristics of both the Jamu and the random networks.
We determined five statistical indexes, that is, average degree, clustering coefficient, number of connected component, network diameter, and network density of each Jamu network and also of each random network. The clustering coefficient of a node is defined as , where is the number of neighbors of and is the number of connected pairs between all neighbors of . The network diameter is the largest distance between any two nodes. If a network is disconnected, its diameter is the maximum of all diameters of its connected components. A network’s density is the ratio of the number of edges in the network over the total number of possible edges between all pairs of nodes (which is , where is the number of vertices, for an undirected graph). The average number of neighbors and the network density are the same for the real and random networks of the same size as it is shown in Table 3. In case of 0.7% and 0.5% real networks, the clustering coefficient is roughly the same and in case of 0.3% the clustering coefficient is somewhat larger. The number of connected components and the diameter of the Jamu networks gradually decrease as the network grows bigger by addition of more nodes and edges.
Very different values corresponding to clustering coefficient, connected component, and network diameter imply that the Jamu networks are quite different from all 3 types of random networks. The differences between Jamu networks and ER random networks are the largest. Random networks constructed based on other two models are also substantially different from Jamu networks. Based on the fact that the random networks constructed based on all three types of models are different from the Jamu networks, it can be concluded that structure of Jamu networks is reasonably biased and thus might contain certain information about plant-disease relations. Specially, much higher value corresponding to clustering coefficient indicates that there are clusters in the networks worthy to be investigated. To extract clusters from the Jamu networks (step 2 in Figure 1) we applied DPClusO network clustering algorithm [14] to generate overlapping clusters based on density and periphery tracking.
3.2. Supervised Clustering Based on DPClusO
DPClusO is a general-purpose clustering algorithm and useful for finding overlapping cohesive groups in an undirected simple graph for any type of application. It ensures coverage and performs robustly in case of random addition, removal, and rearrangement of edges in protein-protein interaction (PPI) networks [14]. While applying DPClusO, the parameter values of density and cluster property that we used in this experiment are 0.9 and 0.5, respectively [15]. Table 3 shows the summary of clustering result by DPClusO. Because clusters consisting of two Jamu formulas are trivial clusters, for the next steps we only use clusters each of which consists of 3 or more Jamu formulas. The number of total clusters increases along with the larger dataset, although the threshold correlation between Jamu pairs decreases. We evaluated the clustering result using matching score to determine dominant disease for every cluster (step 3 in Figure 1). Matching score of a cluster is the ratio of the highest number of Jamu associated with the same disease to the total number of Jamu in the cluster. Thus matching score is a measure to indicate how strongly a disease is associated to a cluster. Figure 4 shows the distribution of the clusters with respect to matching score from three datasets. All datasets have the highest frequency of clusters at matching score >0.9 and overall most of the clusters have higher matching score, which means most of the DPClusO generated clusters can be confidently related to a dominant disease. Furthermore the number of clusters with matching score >0.9 is remarkably larger compared to the same in other ranges of matching score in case of the 0.3% dataset (Figure 4(c)). If we compare the ratio of frequency of clusters at matching score >0.9 for every dataset, the 0.3% dataset has the highest ratio with 40.84% (of 453), compared to 29.67% (of 873) and 21.91% (of 1296), in case of 0.5% and 0.7% datasets, respectively. Thus, the most reliable species to disease relations can be predicted at matching score >0.9 corresponding to the clusters generated from 0.3% dataset.
(a) 0.7%
(b) 0.5%
(c) 0.3%
Figure 5(a) shows the success rate for all 3 datasets with respect to threshold matching scores. Success rate is defined as the ratio of the number of clusters with matching score larger than the threshold to the total number of clusters. As expected it tends to produce lower success rate if we decrease correlation value to create the datasets. However more clusters are generated and more information can be extracted when we lower the threshold correlation value. The success rate increases rapidly as the matching score decreases from 0.9 to 0.6 and after that the slope of increase of success rate decreases. Therefore in this study we empirically decide 0.6 as the threshold matching score to predict plant-disease relations.
(a)
(b)
3.3. Assignment of Plants to Disease
By using DPClusO resulting clusters, we assigned plants to classes of disease. Based on a threshold matching score we assigned dominant disease to a cluster. Then we assign a plant to a cluster by way of analyzing the ingredients of the Jamu formulas belonging to that cluster and determining the highest frequency plant, that is, the plant that is used for maximum number Jamu belonging to that cluster (step 4 in Figure 1). Thus we assign a disease and a plant to each cluster having matching score greater than a threshold. Our hypothesis is that the disease and the plant assigned to the same cluster are related.
The total number of assigned plants depends on matching score value. Figure 5(b) shows the number of predicted plants that can be assigned to diseases in the context of matching score. With higher matching score value, the number of predicted plants assigned to classes of disease is supposed to remain similar or decrease but the reliability of prediction increases. In Figure 5(b) a sudden change in the number of predicted plants is seen at matching score 0.6 which we consider as empirical threshold in this work. Based on the 0.7% dataset, the largest number of plants (135 plants, Table 4) was assigned to diseases. There are 63 plants assigned to only one class of disease, whereas the other 72 plants are assigned to at least two or more classes of disease (Figure 6).
3.4. Evaluation of the Supervised Clustering Based on DPClusO
We used previously published results [6] as gold standard to evaluate our results. The previous study assigned plants to 9 kinds of efficacy whereas we assigned the plants to 18 disease classes (16 from NCBI and 2 additional classes). For the sake of evaluation we got done a mapping of the 18 disease classes to 9 efficacy classes by a professional doctor, which is shown in Table 5. Table 6 shows the prediction result of plant-disease relations for all 3 datasets, corresponding to clusters with matching score greater than 0.6. Table 6 also shows corresponding efficacy, the number of assigned plants, number of correctly predicted plants, and true positive rates (TPR), respectively.
We determined TPR corresponding to a disease/efficacy class by calculating the ratio of the number of correct prediction to the number of all predictions. When a disease corresponds to more than one kind of efficacy, the highest TPR can be considered the TPR for the corresponding disease. For all 3 datasets the TPR corresponding to each disease is roughly 90% or more. The 0.3% dataset consists of Jamu pairs with higher correlation values and based on this dataset 117 plants are assigned to 14 disease classes. The 0.7% dataset contains more Jamu pairs and assigned plants to 11 disease classes, one less disease class compared to 0.5% dataset. The two disease classes covered by 0.3% dataset but not covered by 0.5% and 0.7% datasets are the nervous system (D13) and disease of the immune system (D9). The only disease class covered by 0.3% and 0.5% datasets but not covered by 0.7% dataset is mental and behavioural disorders (D18). The larger dataset network tends to have lower coverage of disease classes. The number of Jamu pairs, that is, the number of edges in the network, affect the number of DPClusO resulting clusters and number of Jamu formulas per cluster. As a consequence, for the larger dataset networks, the success rate becomes lower and the coverage of disease classes is lower but prediction of more plant-disease relations can be achieved.
4. Conclusions
This paper introduces a novel method called supervised clustering for analyzing big biological data by integrating network clustering and selection of clusters based on supervised learning. In the present work we applied the method for data mining of Jamu formulas accumulated in KNApSAcK database. Jamu networks were constructed based on correlation similarities between Jamu formulas and then network clustering algorithm DPClusO was applied to generate high density Jamu modules. For the analysis of the next steps potential clusters were selected by supervised learning. The successful clusters containing several Jamu related to the same disease might be useful for finding main ingredient plant for that disease and the lower matching score value clusters will be associated with varying plants which might be supporting ingredients. By applying the proposed method important plants from Jamu formulas for every classes of disease were determined. The plant to disease relations predicted by proposed network based method were evaluated in the context of previously published results and were found to produce a TPR of 90%. For the larger dataset networks, success rate and the coverage of disease classes become lower but prediction of more plant-disease relations can be achieved.
Conflict of Interests
The authors declare that there is no financial interest or conflict of interests regarding the publication of this paper.
Acknowledgments
This work was supported by the National Bioscience Database Center in Japan and the Ministry of Education, Culture, Sports, Science, and Technology of Japan (Grant-in-Aid for Scientific Research on Innovation Areas “Biosynthetic Machinery. Deciphering and Regulating the System for Creating Structural Diversity of Bioactivity Metabolites (2007)”).