Data Mining Methods for Omics and Knowledge of Crude Medicinal Plants toward Big Data Biology

Transcription

1 , CSBJ Data Mining Methods for Omis and Knowledge of Crude Mediinal Plants toward Big Data Biology Farit M. Afendi a,b, Naoaki Ono a, Yukiko Nakamura a, Kensuke Nakamura d, Latifah K. Darusman, Nelson Kibinge a, Aki Hirai Morita a, Ken Tanaka e, Hisayuki Horai f, Md. Altaf-Ul-Amin a, Shigehiko Kanaya a,* Abstrat: Moleular biologial data has rapidly inreased with the reent progress of the Omis fields, e.g., genomis, transriptomis, proteomis and metabolomis that neessitates the development of databases and methods for effiient storage, retrieval, integration and analysis of massive data. The present study reviews the usage of KNApSAK Family DB in metabolomis and related area, disusses several statistial methods for handling multivariate data and shows their appliation on Indonesian blended herbal mediines (Jamu) as a ase study. Exploration using Biplot reveals many plants are rarely utilized while some plants are highly utilized toward speifi effiay. Furthermore, the ingredients of Jamu formulas are modeled using Partial Least Squares Disriminant Analysis (PLS-DA) in order to predit their effiay. The plants used in eah Jamu mediine served as the preditors, whereas the effiay of eah Jamu provided the responses. This model produes 71.6% orret lassifiation in prediting effiay. Permutation test then is used to determine plants that serve as main ingredients in Jamu formula by evaluating the signifiane of the PLS-DA oeffiients. Next, in order to explain the role of plants that serve as main ingredients in Jamu mediines, information of pharmaologial ativity of the plants is added to the preditor blok. Then N-PLS-DA model, multiway version of PLS-DA, is utilized to handle the three-dimensional array of the preditor blok. The resulting N-PLS-DA model reveals that the effets of some pharmaologial ativities are speifi for ertain effiay and the other ativities are diverse toward many effiaies. Mathematial modeling introdued in the present study an be utilized in global analysis of big data targeting to reveal the underlying biology. 1. Introdution 1 Data-intensive sienes have progressed in modern astronomy [1], biology [2-8], omputational materials siene [9], eology [10-11] and soial siene [12] beause open-aess data has inreased drastially. Data-intensive or -driven disovery in biology requires a large open pool of data aross the full breadth of the life sienes and the aess to the pool will invite New logi, strategies and tools to disover new trends, assoiations, disontinuities, and exeptions that reveal aspets of the underlying biology [2, 5, 6]. Big data biology, whih is a disipline of data-intensive siene, was proposed based on agraduate Shool of Information Siene, Nara Institute of Siene and Tehnology, Nara , Ikoma, Japan bdepartment of Statistis, Bogor Agriultural University, Jln. Meranti, Kampus IPB Darmaga, Bogor 16680, Indonesia Biopharmaa Researh Center, Bogor Agriultural University, Kampas IPB Taman Kenana, Jln. Taman Kenana No. 3 Bogor 16151, Indonesia dmaebashi Institute of tehnology, Kamisadori, Maebashi-shi, Gunma, Japan edepartment of Mediinal Resoures, Institute of Natural Mediine, University of Toyama, 2630 Toyama, , Japan fdepartment of Eletroni and Computer Engineering, Ibaraki National College of Tehnology, 866 Nakane, Hitahinaka, Ibaraki , Japan * Corresponding author. address: skanaya@gt.naist.jp (Shigehiko Kanaya) the rapid inreasing of omis data produed by genomis, transriptomis, proteomis and metabolomis [2-8]. This situation is also a feature of the ethnomediinal survey and the number of mediinal plants is estimated to be 40,000 to 70,000 around the world [13] and many ountries utilize these plants as blended herbal mediines, e.g., China (traditional Chinese mediine), Japan (Kampo mediine), India (Ayruveda, Siddha and Unani) and Indonesia (Jamu). Blended herbal mediines as well as single herb mediines inlude a large number of onstituent substanes whih exert effets on human physiology through a variety of biologial pathways. To omprehensively understand the mediinal usage of plants based upon traditional and modern knowledge, we add to KNApSAK Family database systems the seleted herbal ingredients i.e., the formulas of Kampo and Jamu, omis information in plants and humans, and physiologial ativities in humans [14-16]. These information need to be onneted in a way that enables sientists to make preditions based on general priniples. In this mini-review, we disuss the usage of KNApSAK Family DB in metabolomis, explain mining tehniques suh as prinipal omponent analysis (PCA), partial least square regression (PLSR) and multiway model, and show their appliation on Indonesian blended herbal mediines (Jamu) as a ase study. 2. KNApSAK Family Database Omis biology, like most sientifi disiplines, is in an era of aelerated inrease of data, so alled big data biology [2-8]. Largesale sequening enters, high-throughput analytial failities and

2 2 individual laboratories produe vast amounts of data suh as nuleotide and protein sequenes, gene expression measurements, protein and geneti interations, mass spetra of metabolites and phenotype studies. The goal of investigating the interations between mediinal/edible plants and humans is to omprehensively understand the moleular mehanism of mediinal plants on human physiology based on urrent and traditional knowledge. Optimization of blended herbal formulas should be developing using information derived from plant and human omis. To reah this goal we need to develop databases based on the platform shown in Fig. 1A. KNApSAK family DBs have been developed for this purpose [14-16]. Relations among individual DBs are illustrated in Fig. 1A and main page of KNApSAK Family DB is shown in Fig. 1B. A B Figure 1. Integrated platform of knowledge of mediinal plants and plant and human omis and KNApSaK Family databases. (A) The relations of attributes among individual DBs. (B) Main window of KNApSAK Family DB, indexes from a to i in panel A orrespond to those in panel B. Four DBs (Lunh Box DB, DietNavi DB, Food Proessor DB and DietDish DB, a-d in Fig. 1) are about Food & Health related with Japanese foods and ingredients explained in Japanese language beause initially we developed them targeting the Japanese people, but we are Data Mining Methods for Omis planning to translate them into English as early as possible. Lunh Box DB omprises information on 800 edible speies whih inlude the speies introdued to Japan from outside or originally grown in Japan, general information of the rops and the effet of them on human health. Nonommuniable diseases suh as heart disease, metaboli disease, aner and respiratory disease, whih superseded the infetious diseases beause of the development and widespread distribution of vaines and antimirobial drugs, aount for 60% of all deaths worldwide and 80% of deaths in low- and middle-inome ountries [17]. Food and ingredients in sanative diet and more effetive ombination of foods benefiial against those nonommuniable diseases are aumulated in DietNavi and DietDish DBs, respetively (b and d in Fig. 1). FoodProessor DB omprises 309 retortable pouh foods enompassed by 261 food ingredients produed in Japan, and onneted with DietNavi and KNApSAK ore by speies names of foods. To systematize rude drugs by multifaeted view points, we have developed four DBs (WorldMap, KAMPO, JAMU and TeaPot DBs as shown in e-h of Fig. 1). The KNApSAK WorldMap DB omprises 46,256 geographi zone-plant pair entries in 217 geographial zones exept mini-states suh as the Prinipalities of Liehtenstein, Monao and Andorra, and the Vatian City. Presriptions orresponding to Japanese and Indonesian herbal mediines have been aumulated in KAMPO and JAMU DBs, respetively. KAMPO DB is omprised of 1,581 primary formulas lassified in to 336 formula names enompassed by 278 mediinal plants whih are approved by the National health insurane authority in Japan. JAMU DB is omprised of 5,310 formulas enompassed by 550 mediinal plants and 12 anatomial regions whih are approved by the National Ageny of Drug and Food Control (NA-DFC) of Indonesia. Mediinal/edible plants reported in the sientifi literature have been lassified into geographi zones using the International Organization for Standardization (ISO3166), whih defines geographi zones based on the borders between nations and small islands. Herbs are defined as any plants with leaves, seeds, and flowers used for flavoring, food, mediine, perfume and parts of suh a plant as used in ooking. Those are aumulated in TeaPot DB. Two types of biologial ativities, that is, ativities of natural resoures and metabolites to other speies inluding human, i.e., antibioti, antianer and so on are aumulated in Natural Ativity and Metabolite Ativity DBs (Fig. 1B), respetively. The former and the latter omprised 33,703 and 6,677 entries, respetively. For extension of speies-metabolite relationship DB to metaboli pathways, it is needed to design seondary metaboli pathway DB for detetion of metaboli pathways based on enzyme reations and predition of reations by peptide sequenes. So we have developed Motoeryle DB ontaining 2,421 entries. The metabolomis of plants is developing rapidly [18-20 and referenes in Table 1], and it will be an important topi in the systems-biologial studies of interations between plants and humans, whih is inluded in the topis of big data biology [2-8], with the goal of ahieving a holisti understanding of plant funtion and healthare, inluding the ativity of mediinal plants as well as interation between plants and their environment [14-16, 21, 22]. To failitate aess to metabolite information obtained from analytial tehniques, we have developed speies-metabolite relationship DB (KNApSAK Core DB) whih ontains 106,418 speies-metabolite relationships enompassing 21,705 speies and 50,897 metabolites. Nine databases of KNApSAK family (exept DietDish) are onneted with KNApSAK Core DB to easily obtain andidates of seondary metabolites in speies utilized in several purposes [23]. The KNApSAK Core DB was utilized in very

3 3

4 4 diverged purposes of metabolomis studies inluding identifiation of metabolites ( Exp in Table 1), onstrution of integrated databases ( DB ), bioinformatis and systems biology ( Bioinfo ), and ited in at least 110 papers listed in Table 1, that is, in 29 papers in the period of , 25 papers in the period of 2009, 20 papers in 2010, 18 papers in 2011, 18 papers in In addition, it was applied in diverged speies from bateria to plants and animals, in total 28 speies, that is, Angelia autiloba [74], Arabidopsis lyrata ssp. petraea [56], Arabidopsis thaliana[25, 30, 33, 35, 37, 46, 47, 62, 70, 86, 99, 103, 104, 108, 109, 121, 122], Atriplex halimus [127], Baillus subtilis [113], Brassia oleraeae var apitata [60], Brufelsia alyina [81], Capsium sp. [123], Citrus sinensis [131], Curuma longa [77], Ephedra sp. [67], Esherihia oli [51], Fragaria x ananassa [40, 43, 44], Fragaria vesa [105], Glyine max [53], Glyyrrhiza uralensis [94], Hordeum vulgare [80, 102], Homo sapiens [63, 101], Jatropha uras [124, 125], Malx x domestia [126], Ophiorrhiza pumila [117], Oryza sativa [49, 61], Papaver somniferum [42], Rattus norvegius [39, 97], Rizotania solani [79], Solanum lyopersium [45, 48], Solanum tuberosum [98] and Zea mays [120]. In the period of , many review papers [ Review in Table 1] foused on metabolomis platforms integrated by massspetrometry and metabolite databases inluding KNApSAK Core [29, 31, 34, 38, 42, 49, 52] and on linking hemistry with biology [24], and on metabolome researhes targeting the model plant Arabidopsis thaliana [30, 33, 35, 37]. In 2009, metabolome studies were extended to diverged speies suh as rops and mediinal plants [53, 60, 61, 67, 68, 73, 74, 78] and to engineering studies suh as quality assessment based on metabolomis [73, 74]. Thus metabolomis was applied from model speies to rops and mediinal herbs. In the period of , metabolomis was further extended to genetis suh as QTL [80, 98, 126], and to explanation of speies by metabolites, i.e., eologial subjets [85] phytoalexins [119], herbivore-indued metabolites [120] and defense against pathogens [131], and to stress responses [115, 116, 127]. In addition, metabolomis has also been tried in imaging studies [112, 129]. Speies-metabolite relation database KNApSAK Core has been utilized in the extended fields of metabolomis researhes and the horizon of metabolomis researhes ould be reognized by reviewing the works that utilized and/or ited the KNApSAK DB. Methodologies for multivariate analysis to statistially proess the massive amount of metabolome data were reviewed in [16] and to systematize blended herbal mediines in Kampo [15]. In the following setion, we fous on the mining studies of blended herbal mediines for systematially understanding the omposition of mediinal herbs to effiaies on humans, that is, prinipal omponent analysis (PCA) that makes it possible to systematize the ingredient in individual blending systems, partial least squares (PLS) that an relate the ingredients of mediinal herbs to the effiaies and N-PLS that an onnet multi-fators to the effiaies. We initially explain individual tehniques in Setion 3 and then disuss their appliation in datamining of blended types of herbal mediines in Setion Mathematial Methods of Data Mining PCA is a linear transformation of a large number of interrelated variables into a new set of variables, alled as the prinipal omponents (PCs), whih are unorrelated and ordered so that the first few retain most of the variation present in all the original variables [132]. Consider a data matrix A = (a1 a2 ap) with n observations and let V (p x p) be the variane-ovariane matrix of A. The prinipal omponents of A, Z = (z1 z2 zp), are alulated as z j = A j ( j = 1, 2,, p) where j is the j-th eigenvetor of V whih orrespond to the j-th eigenvalue of V ( j). The properties of PCs are: (1) Var(zj) = j; (2) Cov(zj,zj ) = 0, j j ; (3) Var(z1) Var(z2) Var(zp). The umulative proportion of variane of the original variables explained by the first J prinipal omponents an be obtained as Pr( J j 1 z j ) p j 1 j j PLSR is a regression method, whih assumes underlying fators among the preditors aount for most of the response variation [133, 134]. These underlying fators of X-variate T = XW are obtained by maximizing their ovariane with the orresponding underlying fators of Y-variate where X is an n m matrix of preditors, Y is an n p matrix of responses, T is an n matrix of X-sore fators, and W is m matrix of weight. Note that n is the number of observations, m is the number of preditors, p is the number of responses, and is the number of omponents. The X-sore fators, i.e. matrix T, have the following properties [133]. a. When multiplied by loadings P, they are good summaries of X, i.e. the X-residuals E are small X = TP t + E b. The X-sore fators are good preditors of Y, i.e. Y = TQ t + F The Y-residuals F express the deviations between the observed and modeled responses. Based on Eq. (3), Eq. (5) an be rewritten as a multiple regression model Y = XWQ t + F = XB + F Thus, PLSR oeffiients B an be written as B = WQ t whereas predition of the responses an be obtained from t Yˆ XWQ Data Mining Methods for Omis Although PLSR is not speifially designed to disriminate among groups, Barker and Rayens [135] have demonstrated that PLSR an be used for suh purposes by onneting PLSR and Linear

5 Disriminant Analysis (LDA); this ombined method is alled as Partial Least Square Disriminant Analysis (PLS-DA). In PLS-DA, group membership is transformed into a dummy matrix, and this dummy matrix provides the response variables for PLSR. = 1, 2,, K). The deomposition of both the preditor and the response blok based on N-PLS model are as follows X ijk C TiW 1 J j W K k E ijk Y il C V 1 i V l F il The array X is deomposed into a tri-linear model onsisting of one sore vetor for observation alled t (I x 1), and two weight vetors, one for type I variable alled w (J x 1) and one for type II K variable alled w (K x 1). Similarly, a bi-linear model is used in deomposing the matrix Y into one sore vetor v (I x 1) and one weight vetor u (L x 1). The deomposition is onduted suh that the ovariane among the sore of preditor t and the orresponding sore of the response v is maximized. All sores and weights are indexed with showing that they orrespond to th multiway omponent, while C represents the total number of multiway omponents used in N-PLS model. Moreover, E and F are the residuals of the deomposition of the three-dimensional array X and matrix Y, respetively. J 5 Figure 2. Shemati diagram of the deomposition of both preditor and response bloks for: (a) PLS and (b) N-PLS model. Figure 3. Illustration of matriizing three-dimensional array X (I x J x K) into matrix X (I x JK). An extension of PLSR to deal with multidimensional data known as Multiway Partial Least Squares has been developed by Bro [136] and is alled as N-PLS. In this model, the same priniple of PLSR for two dimensional data is utilized, that is, both preditor and response bloks are deomposed suessively into multi-linear model suh that the pairwise sores have maximal ovariane. The sore of the preditor is then regressed to the response variable. Fig. 2 illustrates the deomposition of N-PLS model. Moreover, N-PLS model an also be used for disrimination purpose, whih is alled as N-PLS- DA, that is the multiway version of PLS-DA, by utilizing the dummy matrix of group membership as the response variable. Consider the three-dimensional array X indexed by observation (i = 1, 2,, I), type I variable (j = 1, 2,, J) and type II variable (k Furthermore, let Xk (I x J) be the kth slie of X (I x J x K) for the orresponding kth of type II variable, then matriizing threedimensional array X into matrix X (I x JK) is performed as follows [137] X = [X 1 X 2 X K] Fig. 3 depits this unfolding proess of array X into matrix X. Using this notation, the sore t of the th omponent an be alulated as [138] t or X( w K w J )

6 i J K t x j 1 k 1 ijk w J j w K k From Eq. (12), the weight orresponding to th omponent, w (JK x 1), an be defined as w ( w K w J ) Smilde [140] also desribed that, due to the deflation in X during the deomposition, the weight matrix W (JK x C) an be applied diretly to the original unfolded matrix X is defined as pharmaologial ativities: A2 and A4. Plant P2 also has two pharmaologial ativities: A1 and A2, while plant P3 has three ativities: A3, A4, and AK. The other onnetions an be desribed similarly. From the onept of integrated platform of knowledge of mediinal plants and plant and human-omis depited in Fig. 1, the effiay layer in Fig. 4 represents the physiologial ativity layer in human-omis attribute, the herbal mediine and plant layer represent the presription and mediinal herb layer, respetively, in knowledge of mediinal plants attribute, while the pharmaologial ativity layer represents the metabolomis layer in plant-omis attribute. On the following setion we will illustrate the data mining tehniques on herbal mediine database analyzing relationship among entities for two, and more than two attributes. t t t t 2 ) wq W [ w ( I JK w1w1 ) w2... ( I JK w1w1 )( I JK w2w2 )...( I JK wq 1wQ 1 ] Hene, the sores in T (I x C) expressed diretly in terms of the X- olumns is T = XW After the deomposition proedure, the next step is to regress Y on the omponent sores T Yˆ TB with B = (T t T) -1 T t Y 6 From Eq. (15) and (16) we have Ŷ XWB Therefore, the regression oeffiients BNPLS (JK x L) needed to predit Y from X are obtained as B NPLS = WB 4. Illustration of Data Mining Tehniques Indonesia, the mega-biodiversity enter like Brazil, has at least 9,600 speies of plants with pharmaologial ativity [110] and has developed blended herbal mediines alled Jamu taking modern and traditional knowledge of herbs into onsideration. To prepare Jamu, several plants are seleted and mixed suh that the onotion has the desired effiay. Traditionally, plants are hosen based on prior experiene whih is passed down from generation to generation. In uring a partiular disease, eah ethni group in Indonesia may have its own formulas, whose speifi nature depends strongly on the loal plant resoures in the region where a given population lives and the effiaies of Jamu mediines have been empirially demonstrated [ ]. Data mining tehniques with the blended herbal mediine databases suh as KAMPO and JAMU (Fig. 1) makes it possible to omprehensively and mathematially understand those blended herbal systems. Fig. 4 illustrates a network onneting effiay, herbal mediine, plant, and pharmaologial ativity of plant. The network showing that rude mediines M1, whih is useful for effiay E1, use three plants in its ingredients: plant P1, P3, and P4. Plant P1 has two Figure 4. A typial network illustrating onnetions between effiay, herbal mediine, plant, and pharmaologial ativity of plant. As an illustration for data mining of herbal mediine database whih rely on relationship between two attributes, the relationship between the effiay of Jamu and mediinal plants used in Jamu is explored using PCA [ ]. The effiaies of 3,138 Jamu are lassified into one of nine ategories, namely: (1) disorders of appetite (DOA), (2) disorders of mood and behavior (DMB), (3) female reprodutive organ problems (FML), (4) gastrointestinal disorders (GST), (5) musuloskeletal and onnetive tissue disorders (MSC), (6) pain/inflammation (PIN), (7) respiratory disease (RSP), (8) urinary related problems (URI), and (9) wounds and skin infetions (WND). In total, those 3,138 Jamu use 465 plants in their ingredients. The distribution of Jamu and plant utilized in Jamu for eah effiay is shown in Table 2. Note that, one plant may be used in many Jamu with varying effiaies. Hene, it is interesting to find out the most signifiant effets of speifi plants by analyzing their usage in Jamu, and onsidering that the more useful a given plant in having ertain effet, the more frequently the plant will be used in Jamu when that effet is desired. Biplot, a multivariate exploration tool, is suitable for this purpose beause it provides simultaneous plot of prinipal omponent sores and loadings, as representation of observations and variables, respetively [145]. Considering plants as observations and effiay groups as variables, the relationship between them an be explored using a biplot.

7 7 Following the explanation of PCA in previous setion, the data matrix A as an input for PCA is generated by putting plant as observation and effiay as variables. So, A onsists of 465 rows and 9 olumns. Eah ell aij shows the number of Jamu that use plant i and useful for effiay j. Figure 5. Biplot onfiguration based on PCA analysis of Jamu data. Plants and Jamu effiaies are represented as red points and blue lines, respetively. Biplot onfiguration using the first two omponents is shown in Fig. 5. In the figure, plants are represented as red points while Jamu effiaies as blue lines, i.e. vetors based on loadings. The length of a given effiay line showing the variability of plant usage for the orresponding effiay, that is, the longer the effiay line the larger the variability of plant usage for that effiay. From Fig. 6, it is obvious that effiay MSC has the largest variability of plant usage, followed by effiay GST and FML. On the other hand, effiay DMB has the smallest variability of plant usage, followed by effiay URI and RSP. This finding an be addressed due to two fators, that is, the number of Jamu as well as the number of plant utilized in the orresponding effiay (see Table 2). Effiaies with large variability of plants usage (MSC, GST, and FML) have large values for both fators; in ontrast, effiaies with small variability of plants usage (effiay DMB, URI, and RSP) have small values for both fators. In the onfigurations, many plants are lustered in the enter. Note that, the projetion value of plants' point on a given effiay line is the predition of the frequeny of plants usage on that effiay. So, these lustered plants are basially plants whose frequenies of usage in Jamu are very low. In ontrast to the lustered plants, some plants are spread out and loated near the effiay for whih the plants are highly utilized. For example, Ginger (Zingiber offiinale) is loated near the effiay MSC. Ginger is well known for its funtion of refreshing body, and for this reason many Jamu use Ginger for effiay MSC whih an easily be identified from biplot onfiguration. Another example is Turmeri (Curuma longa) whih loated near the effiay FML. Due to its analgesi and antimirobial ativity, this plant is well known and highly utilized in Indonesia as ingredient of Jamu formula for women during menstruation, whih is a problem that lassified into effiay FML. Thus, the biplot onfiguration exhibits useful information in exploring the relationship between plants and the effiay of Jamu. Another illustration for relationship between two attributes on data mining of herbal mediine database is the modeling of Jamu ingredients (representation of knowledge of mediinal plants) to predit the effiay (representation of human omis). This analysis is performed beause of the fat that Jamu is prepared from a mixture of several plants. The plants are hosen so that the Jamu has the desired effiay. As a result, the omposition of the plants used in Jamu formula determines the effiay. Thus, it is interesting to model the ingredients of Jamu, i.e. the onstituent plants, and use this model to predit effiay. PLS-DA, a statistial model for lassifiation and disrimination based on Partial Least Square Regression (PLSR), is suitable for this analysis beause a large number of plants are used in Jamu, whereas Jamu effiaies an be grouped into a few ategories or lasses. In this method, the plants used in eah Jamu mediine served as the preditors, whereas the effiay of eah Jamu provided the responses. The data struture used for PLS-DA is as follows. The data matrix X in X-blok ontains plant usage status. The dimension of matrix X is (I x J), where I is the number of Jamu (in this ase, 3,138), and J is the number of plants (in this ase, 465). Beause of the availability of information about Jamu produts, whih generally do not state in detail the mixing ratio of the plants used, the preditors X is onstruted only in binary data. Eah ell xij (i = 1, 2,, I; j = 1, 2,, J) is set to 1 if Jamu i uses plant j, and is set to 0 otherwise. In the present study, nine indiator variables, whih orrespond to the 9 effiaies listed in Table 2 perform as the Y- blok in PLS-DA modeling. Thus, the dimension of data matrix Y is (I x 9). Eah ell yil (l = 1, 2,, 9) is set to 1 if Jamu i is lassified into effiay group l, and is set to 0 otherwise. Note that 9 l 1 il Data Mining Methods for Omis y 1 beause eah Jamu is lassified to one effiay only. Using the derived PLS-DA model, we an then use it to predit the effiay of Jamu given information of the ingredients. In this analysis, among the 3,138 Jamu mediines, the effiaies of 2,248 Jamu mediines (71.6%) an be assigned to an individual effiay reported. Hene, the effiay in most Jamu mediines an be predited on the basis of mediinal plants used. The perentages of orret predition for eah effiay (see Table 3) vary from 22.7% for effiay DMB to 89.8% for effiay GST. The low perentage of orret predition for effiay DMB an be addressed due to the small number of Jamu for this effiay, whih is only 22 out of 3,138 Jamu (see Table 2).

8 Furthermore, plants in the ingredients of Jamu are used as main ingredients, whih ontribute primarily to the mediines' effiaies; other plants are used as supporting ingredients [146, 147]. Investigating whih plants are main ingredients and whih are supporting is important in order to omprehensively understand the mehanisms by whih speifi plants ahieve desired effiaies. The regression oeffiients of previous PLS-DA model, whih relates plants usage in Jamu as preditors and Jamu effiay as response, an be helpful in this attempt beause they summarize the effet of plant on effiay. Plants that at as main ingredients will have signifiant effet on the model developed. Furthermore, due to the absene of parametri testing for the PLS-DA oeffiients, the evaluation for signifiane is performed using permutation testing, in whih the distribution of oeffiients under the null hypothesis is generated via resampling of the existing data [149]. the PLS-DA oeffiients obtained from this proess generates a distribution, against whih a p-value an be alulated and subsequently evaluated for signifiane [150]. The results of the signifiane testing of all plants used in eah 9 effiaies are shown in Table 4. Note that one plant may be used for more than one effiay. From the testing, we observed 234 plants (50.3% among all 465 plants) showing no signifiant status for all 9 effiaies; whereas the other 231 plants have signifiant status whih omprise of 189 plants (40.6%) are signifiant only for 1 effiay, 38 plants (8.2%) are signifiant for 2 effiaies, and the other 4 plants (0.9%) are signifiant for 3 effiaies. Besides testing the plants usage statistially, furthermore, we also heked from sientifi papers the usage of signifiant plants in their orresponding effiay. Many of the results we obtained by our analysis are supported by sientifi papers. Note that in prediting Jamu effiay based on the information of its ingredients we an also use other methods suh as disrimination analysis, nominal logisti regression, and support vetor mahine. However, in the present study we fous on PLS-DA in lassifying Jamu effiay by taking into onsideration that we also intend to evaluate the signifiane of plant usage in Jamu to ahieve speifi effiay as well as extending the analysis into three-way model by adding the plant pharmaologial ativity into preditors blok. 8 Figure 6. Clustergram of pharmaologial ativity against Jamu effiay. The red and blak ells indiate that the pharmaologial ativity is signifiant or non-signifiant, respetively, to the orresponding effiay. The resampling is performed by permuting the order of the responses (in this ase, Jamu effiaies) while maintaining the order of the preditors (in this ase, plant utilization as Jamu ingredients) so that the existing relationship between the preditors and the response is destroyed and a new data set is generated under the null hypothesis, i.e., plant utilization in Jamu does not affet Jamu effiay. If we perform suh resampling many times and apply the PLS-DA model on the new data generated from the resampling, the aumulation of

9 9 During the modeling proess of PLS-DA in the previous setion, the ingredients of Jamu provide the preditor while the Jamu effiay serves as the response. In order to identify the funtion of the plants in Jamu to ahieve speifi effiay, the reported pharmaologial ativities of the plants are added to the preditors blok. Thus, the preditors blok an be represented as a three-dimensional array X (I x J x K) indexed by Jamu mediine (i), plant (j), and pharmaologial ativity (k) as depited in Fig. 2 with Jamu mediine, plant, and pharmaologial ativity serve as observation, type I and type II variables, respetively. Furthermore, the response blok is represented as matrix Y (I x 9). This analysis then onnets three attributes: (1) knowledge of mediinal plants (represented by Jamu and plants orresponding to JAMU DB in Fig 1); (2) plant omis (represented by pharmaologial ativity orresponding to Biologial ativity (Nat) in Fig 1); and (3) human omis (represented by effiay). The detail about the elements of array X and matrix Y is as the following. Let xijk (k = 1, 2,, K; K = 46 where K is the number of reported pharmaologial ativity; see previous setion on definition of i, j, I, and J) denotes the usage status of plant j with pharmaologial ativity k in Jamu i, where xijk = 1 if the plant j with pharmaologial ativity k is used in Jamu i, and xijk = 0 otherwise. On the other hand, let yil represents the status of Jamu i on effiay l, where yil = 1 if Jamu i is lassified into effiay l, and yil = 0 otherwise. In order to identify the pharmaologial ativity that is signifiantly related with the effiay, we adopt the guidelines from Hair et al. [150] that all weights w K (in absolute values) of 0.3 or above are signifiant for sample sizes of 350 or greater. Figure 6 depits the 2-dimensional dendrogram of Jamu effiay and the pharmaologial ativity signifiantly related with the effiay. The luster of Jamu effiay and the pharmaologial ativity was performed using Ward Linkage based on the Eulidean distane among the entities. The lustering of the pharmaologial ativity side learly exhibits two groups. The first group onsists of ativities useful for one or two effiaies only. This group an be regarded as a group of speifi ativity beause the effets of the ativities are speifi for ertain effiay. For example the diureti ativity is useful for effiay URI and DOA. Diureti is an agent that inreases the seretion and elimination of urine from the body [151]. Obviously, this ativity is benefiial for the effiay URI. Diureti also help the body eliminate waste and support the whole proess of inner leansing, whih is an ation that is useful for effiay DOA espeially related with a slimming purpose. The five ativities (antihaemorrhoidal, arminative, hypoglyaemi, depurative, and anthelminti) are speifially related with effiay GST. Antihaemorrhoidal means an ativity that treats haemorrhoids (piles), while the arminative is defined as an ativity that eases disomfort aused by flatulene. Hypoglyaemi ativity helps redue the levels of sugar in the blood, whereas the depurative eliminates toxins and purifies the system espeially the blood, and the anthelminti helpful in expelling parasites from the gut. Thus, all of these ativities are helpful for the problem related with the digestive system, i.e. the effiay GST. Furthermore, the seond group of ativity revealed by the dendrogram onsists of ativities useful for at least four effiaies. In ontrast to the first group, this group an be regarded as the general ativities beause of the diverse effiaies related to this group. Among all ativities lustered to this group, antimirobial ativity is signifiantly related with all 8 effiaies. We an interpret this result as follows. Due to the environmental onditions, hygiene, and its loation as a tropial ountry whih led to many mirobes that are harmful to health, then it is reasonable that antimirobial ativity is important and should be available in many Jamu formulas in Indonesia. It should be noted that many popular mediinal plants in Indonesia suh as Temulawak (Curuma xanthorriza), Ginger (Zingiber offiinale), Turmeri (Curuma longa) or Kenur (Kaempferia galanga) have ontent of this ativity [152]. Anti-inflammation, antispasmodi, analgesi, sedative, and stimulant are also lustered into this general ativity group. Sine many health problems or diseases are often aompanied with inflammation or spasm, then the plants with anti-inflammation and/or antispasmodi ativity are hosen in many Jamu formulas. Those health problems/diseases often ause pain or other disomforts, thus plants with ertain ativities suh as analgesi or sedative effets are hosen in many Jamu mediines. Finally, stimulant ativity, whih exites or quikens ativity of the physiologial proesses, is important for the reovery reason after one experiening those health problems or diseases. From the previous explanation regarding the grouping of pharmaologial ativity, it an be onluded that in formulating Jamu the plants are seleted so that, beside uring the targeted diseases or health problems as indiated by the speifi ativities, the plants also should overome the other disomforts aused by the targeted diseases or health problems as indiated by the general ativities. It is in aordane with the proess of making the Jamu mediines that involving whole part of plant and not only the speifi ative omponents. Hene speifi or general pharmaologial ativities of omponents are involved during the uring proess of Jamu mediines towards targeted diseases or health problems. 5. Conluding Remarks Data Mining Methods for Omis Biology, like most sientifi disiplines, is in an era of aelerated information gathering and sientists inreasingly depend on the availability of amounts of data suh as nuleotide and protein sequenes, protein and gene expression, dynamis of metabolites et. The nature of urrent systemati understanding of big data biology towards health, nutrition, and other soietal issues have reently beome the fous of sholar in soietal studies of siene and information studies. The rise of ommunity databases, i.e., KNApSAK family DB introdued in the present review, has been strongly assoiated with the urrent emphasis on data-intensive siene. The entral question is whether sientists an dedue how systems and whole organisms work from this torrent of moleular data. To progress this situation, data-intensive approah is needed for understanding intra- and inter-relations in individual layers represented in Fig. 1. The former an be solved based on a type of multivariate analyses suh as luster analysis and prinipal omponent analysis. Though the latter is more ompliated, several approahes inluding PLS and N-PLS make it possible to larify and understand those relations. The big data biology has beome an inevitable part of biology, and the laws of nature ould be larified based on global analysis of big data biology the era of whih has appeared. For enturies biologial researh mainly depended on experiments and for a deade or two omputational analysis has usually followed experimentation but future it might be the opposite i.e., omputational analysis is done first to guide the experimental design failitated by versatile and freely available omis data at various databases.

10 10

11 11

12 12

13 13

14 Competing Interests: The authors have delared that no ompeting interests exist Afendi et al. Liensee: Computational and Strutural Biotehnology Journal. This is an open-aess artile distributed under the terms of the Creative Commons Attribution Liense, whih permits unrestrited use, distribution, and reprodution in any medium, provided the original author and soure are properly ited. What is the advantage to you of publishing in Computational and Strutural Biotehnology Journal (CSBJ)? 14 Easy 5 step online submission system & online manusript traking Fastest turnaround time with thorough peer review Inlusion in sholarly databases Low Artile Proessing Charges Author Copyright Open aess, available to anyone in the world to download for free