Research Article For reprint orders, please contact: [email protected]
Predicting drug–target interaction networks of human diseases based on multiple feature information Aim: Drug–target interaction is crucial in the drug design process. Predicting the drug–target interaction networks of important human diseases can provide valuable clues for the characterization of the mechanism of action of diseases. Materials & methods: A new graph-based semisupervised learning (GBSSL) method is proposed to predict the drug–target interaction networks involved in 13 types of diseases. According to the method, each drug–target pair is initially described with different biological features including sequence, structure, function and network topology information. Then, the optimal feature selection procedures based on the relief and minimum redundancy maximum relevance are executed, respectively. Finally, unknown drug–target interactions can be predicted by the GBSSL method effectively. Results: The proposed method can effectively predict drug–target interactions (with a receiver operating characteristic score of 94.8% and a precision-recall score of 76.5%). Conclusion: Compared with the existing methods, the GBSSL method provides an efficient means of generating optimal features obtained from the combination of multiple sources of feature information. Original submitted 22 April 2013; Revision submitted 14 August 2013 KEYWORDS: disease network n drug–target interaction n semisupervised learning
Typically, a drug target is a key molecule involved in a particular metabolic or signaling pathway that is specific to a disease condition. Identification of new drug targets is the pre liminary step for drug discovery. Recent studies have shown that there are thousands of targets for drug action, whereas the number of cur rent identified drug targets is less than 500 [1,2]. It means that there are still a large number of targets undetected, especially for some impor tant human diseases. Much research has dem onstrated that drug–target interactions have been the basis of drug discovery and design in recent decades. There is no doubt that identify ing interactions between drug compounds and target proteins can provide valuable clues for discovering new targets and understanding the mechanism of drug activity [3]. As the detection of drug–target interactions by experimental methods is often laborious and costly, there is a strong incentive to develop new methods that are able to detect these potential drug–target interactions effectively [4]. With the development of bioinformatics tools and systems biology, a large amount of data regarding pharmaceutically targeted pro teins and drug–target interactions are publicly available for research and educational use. The availability of these public data motivated the development of machine learning methods for
drug–target interactions prediction. To date, a variety of computational methods including molecular docking [5], text mining [6], semi supervised learning (Laplacian regularized least square [NetLapRLS]) [7,8], kernel regres sion-based [9,10], bipartite local model (BLM) [11,12] and network-based inference [13,14] have been proposed to predict drug–target interac tions. Most of the methods attempted to infer unknown drug–target interactions by integrat ing both chemical (drug chemical structure) and genomic (protein structure) space information into a unified space. However, the drug–target interaction networks they predicted only utilized the chemical similarity information between drug compounds and genomic similarity infor mation between target proteins [10], which are limited and often difficult to describe for each drug–target pair. More importantly, little research has been devoted to studying the dis ease networks that are involved in drug–target interactions, especially for some important human tumor diseases. Therefore, in order to predict the unknown drug–target interaction pairs underlying disease networks, it is reason able to utilize the increased linkage between biological sequences and molecular informa tion by an effective learning method. Recently, semisupervised learning methods have attracted substantial research interest for their ability to
10.2217/PGS.13.162 © 2013 Future Medicine Ltd
Pharmacogenomics (2013) 14(14), 1701–1707
Weiming Yu1, Yan Yan1, Qing Liu1, Junxiang Wang1 & Zhenran Jiang*1 Department of Computer Science & Technology, East China Normal University, 200241, Shanghai, China *Author for correspondence: Tel.: +86 21 54345188 [email protected] 1
part of
ISSN 1462-2416
1701
1702
Pharmacogenomics (2013) 14(14)
H
Tertiary amine
P… P…
…
…… ……
……
Allylic group
HO
H H
CH3
H3C H
HO H
Protein feature space
Feature selection
H H
CH3
H3C H
x
Optimal feature subset
H
CH3
CH3
CH3
Protein sequence similarity
H
CH3
Protein sequence features Original features F1 F2 F3 F4 …… F... F... …… P1 …… P2 …… P 3
N
GO semantic similarity
Allylic alcohol
HO
O
Phenol
Chemical structure similarity
CH3 CH3
Representation of interaction vectors {(d1,p1),(d2,p2),…,(dn,pn)}
Semisupervised learning classifier
New interaction
Known drug
Known target
Predicted drug–target network
Figure 1. Predicting drug–target pairs in known drug–target interaction networks. Multiple feature information including protein homologous similarity, protein sequence features, protein function information, chemical structure similarity, functional groups and drug–target network topology information are integrated for drug–target interaction network construction. The graph-based semisupervised learning method is developed for interaction network prediction.
Known interaction
Known drug
Known target
Known drug–target network
Ether
HO
Chemical functional group
Chemical feature space
f(x)
Research Article Yu, Yan, Liu, Wang & Jiang
future science group
Predicting drug–target interaction networks of human diseases
maximize the generalization ability from chemi cal and genomic spaces by utilizing the small amount of existing label data and the abundant unlabeled data together [7,8]. In this paper, a graph-based semisupervised learning (GBSSL) method is developed to pre dict the unknown drug–target interactions based on multiple biological features including protein homologous similarity, protein sequence features, protein function information, chemi cal structure similarity, functional groups and drug–target network topology information. As a result, the method we proposed is shown to give better performance than two related meth ods, which can achieve an area under receiver operating characteristic (ROC) curve (AUC) score of over 94% and an area under precisionrecall (AUPR) score of up to 76%, respectively. In addition, several new drug–target interaction pairs are obtained for further function research through a comprehensive analysis of drug–target interaction networks.
Materials & methods In order to predict the drug–target interaction networks involved in 13 types of diseases, the proposed method utilized the functional groups and the chemical structure similarity to repre sent the chemical features of each FDA approved drug. Similarly, the features including average Gene Ontology semantic similarity, average pro tein homology similarity and the optimal pro tein sequence features obtained by the feature selection method, were used to describe each target protein. Meanwhile, the drug–target network topology features were obtained from the known drug–target interaction networks. Finally, 174-D features of each drug–target pair was regarded as feature vectors for the GBSSL classifier. Experiment results demonstrated that the GBSSL method can predict new drug– target interactions effectively (see Figur e 1 & Table 1) [8,9,15–40]. For a full method please see S upplementary M ethods at www.futuremedicine. com/doi/suppl/10.2217/pgs.13.162. Results & discussion Results of different kernel functions In this article, we applied two feature selection methods including relief and minimum redun dancy maximum relevance (mRMR) to gain the optimal protein sequence features with the pack age of Libsvm-3.1 [41]. Because the consequences of erroneous overoptimism may be estimated if both training and testing sets are used for super vised feature selection, the training set is used for future science group
Research Article
Table 1. Protein sequence features of each protein. Dimension
Properties
20
Composition of the 20 amino acid residues
400
Dipeptide composition
240
M-B autocorrelation group
240
Moran autocorrelation group
240
Geary autocorrelation group
147
Composition, transition and distribution groups
160
Sequence order group
80
Amphiphilic pseudo amino acid composition group
supervised feature selection and evaluated by an independent test set [42]. As different support vec tor machine (SVM) kernel functions can affect the performance of the feature selection methods, we tested four common kernels including linear, polynomial, radial and sigmoid basis function with the default parameters of C and g in order to obtain the best performance of the SVM model. The performance of each kernel function is evalu ated by tenfold cross-validation tests. The former 250-dimension features of each protein ranked by relief and mRMR are used as the SVM input. The results of the four kernel functions on the data set is shown in Table 2. As listed in Table 2 , the best performance of radial basis function is approximately 73 and 77% in overall accuracy for relief and mRMR, respectively. Therefore, we select the SVM model with radial basis function for the feature selection method in further study. Results of feature selection methods It is clear that the original protein sequence fea tures may include a large amount of irrelevant and redundant information. Removal of this information is a key step for drug–target inter action prediction. After the relief and mRMR algorithms list the ranks of the original feature set, incremental feature selection and forward feature selection (FFS) methods are used to Table 2. Tenfold cross-validation results for different kernel functions (accuracy). Kernel function
Relief (%)
mRMR (%)
SVM-linear
65.94
72.66
SVM-polynomial
61.98
66.25
SVM-RBF
73.27
77.33
SVM-sigmoid
70.01
75.52
mRMR: Minimum redundancy maximum relevance; RBF: Radial basis function; SVM: Support vector machine.
www.futuremedicine.com
1703
Research Article
Yu, Yan, Liu, Wang & Jiang
The IFS and FFS curves (mRMR)
Accuracy (%)
Accuracy (%)
The IFS and FFS curves (relief) 100 90 80 70 60 50 40 30 20 10 0
FFS IFS 0
50
100
150
200
250
100 90 80 70 60 50 40 30 20 10 0
FFS IFS 0
Number of feature
50
100
150
200
250
Number of feature
Figure 2. The curves of incremental feature selection and forward feature selection based on relief and minimum redundancy maximum relevance for 250-dimension features. FFS: Forward feature selection; IFS: Incremental feature selection; mRMR: Minimum redundancy maximum relevance.
generate the optimal feature subset in this study. The incremental feature selection curves based on relief and mRMR with their corresponding FFS curves are obtained (Figure 2). By compari son, the mRMR algorithm has better perfor mance than the single relief algorithm, and the peaks of FFS based on mRMR can reach the overall accuracy rate of 83.4% with 87 features. Comparison with other methods BLM [10] is a statistical method that is used to predict unknown drug–target interactions from chemical structure and genomic sequence information simultaneously. NetLapRLS [7] can reduce labeling costs and improve the perfor mance of learning, especially suitable for very small-labeled data and large unlabeled data. Fur thermore, we compared the performance with the two classical methods, we not only tested
our methods on this data set, but also tested it on four human drug–target interaction data sets involving enzymes, ion channels, G-proteincoupled receptors and nuclear receptors, which are publicly available at [101] (see also [9]). The three methods: GBSSL, BLM and NetlapRLS were further evaluated on the bench mark data set, respectively. We executed ten tri als of a tenfold cross-validation procedure: the data set of drug–target pairs was divided into ten subsets, each subset was then taken in turn as a test set and the remaining nine were used as training sets. For example, there are 103 US FDA approved drugs and 143 target proteins with 312 known interactions and 14,417 unknown drug– protein pairs in our study. In each cross-valida tion, the 281 drug–target interactions are used for training classification while the remaining 1441 unknown drug–protein pairs and 31 drug–target
NetLapRLS BLM GBSSL Random
Precision
True positive rate (sensitivity)
ROC curve 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
Precison-recall curve NetLapRLS BLM GBSSL
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
False positive rate (1-specificity)
Recall
Figure 3. Comparison of the receiver operating characteristic curves and precision-recall curves for the three different methods. BLM: Bipartite local model; GBSSL: Graph-based semisupervised learning; ROC: Receiver operating characteristic.
1704
Pharmacogenomics (2013) 14(14)
future science group
Predicting drug–target interaction networks of human diseases
interactions are designated as the testing data set. The performances of the methods are evaluated by two quality measures called AUC and AUPR. Figure 3 shows the ROC curves and precision-recall curves of three different methods. As demon strated in Figure 3, the GBSSL method can achieve a better precision than the two other methods. The proposed method increased by 5 and 6% on the AUC scores, and improved by 11 and 35% on the AUPR scores for BLM and NetlapRLS meth ods, respectively. We can obtain the best results – an AUC score of 94.8% and AUPR score of 76.5% – with the GBSSL method. Why does the GBSSL method shows a higher AUC and AUPR than both the other methods for this data set? The reasons are that the method used in the current research possesses several mer its. First, many drugs are associated with two or more diseases of the 13 diseases in the data set. These 13 diseases are all associated with cancer and have some similarities in disease phenotype. Our method can utilize the disease information that is involved in drug–target interaction to the greatest extent. Second, the GBSSL method incor porates several critical similarity measures, includ ing Gene Ontology similarities, protein sequence similarities and chemical similarities, to build reli able drug–target networks. Third, we performed the feature selection processes to obtain a subset of relevant features (variables) for constructing more robust learning models, which is effective in capturing the effects of more interacting features. Meanwhile, we compare the performance of the three methods on four classes of drug–target interactions. The AUC scores and AUPR scores were used to evaluate the performance of these methods [43]. The prediction results for GBSSL,
Research Article
Table 3. Performance of four classes of drug target families. Dataset
Method
AUC
AUPR
Enzyme
GBSSL
98.2
84.5
BLM
97.3
85.1
NetLapRLS
95.6
75.4
GBSSL
96.7
84.6
BLM
97.2
83.5
NetLapRLS
94.4
71.3
GBSSL
94.2
54.7
BLM
94.5
66.3
NetLapRLS
84.8
45.2
GBSSL
93.6
67.1
BLM
84.7
59.3
NetLapRLS
82.5
37.8
Ion channel
GPCR
Nuclear receptor
AUC: Area under receiver operating characteristic curve; AUPR: Area under precision-recall; BLM: Bipartite local model; GBSSL: Graph-based semisupervised learning; GPCR: G-protein-coupled receptors.
BLM and NetlapRLS methods after tenfold cross-validation procedures are shown in Table 3. Because we only consider the information from drug structure similarity, target sequence simi larity and network topology information instead of more biological features for predicting the drug–target interaction networks, we have not achieved a greater advantage than the latter two methods. However, we believe that accuracy will be improved if more biological features are combined into our method in future work.
Conclusion & future perspective In conclusion, we proposed a GBSSL method to predict unknown drug–target interactions in 13 types of diseases from multiple biological
Executive summary Background Predicting the drug–target interaction networks of human diseases can provide valuable clues for the characterization of the mechanism of action of diseases. The availability of the public drug–target data motivated the development of machine learning methods for the prediction of drug–target interactions. However, little research has been devoted to studying the disease networks that are involved in drug–target interactions, especially for some important human tumor diseases. It is necessary to develop a method to predict the drug–target interaction networks of human diseases based on multiple feature information. Results Compared with the existing methods, the graph-based semisupervised learning method provides an efficient means of generating optimal features obtained from the combination of multiple sources of information. The proposed method can effectively predict drug–target interactions (with a receiver operating characteristic score of 94.8% and a precision-recall score of 76.5%). Limitations Each drug–target pair can be encoded with more chemical and biological features. The biological significance of the networks can be utilized more effectively. The predicted drug–target pairs need to be validated experimentally.
future science group
www.futuremedicine.com
1705
Research Article
Yu, Yan, Liu, Wang & Jiang
information sources. The proposed method is a complementary method that can be used for studying the function of disease interaction net works. The merit of the study is that we not only consider the drug structure similarity and tar get sequence similarity, but also integrate novel features such as protein function information, sequence features, functional groups and net work topology information into the classifier. The experiment demonstrates that our method has a superior performance than BLM and NetLapRLS, especially for the sample imbal ance problem that is involved in drug–target interaction network prediction. It should be noted that although network pharmacology was regarded as the next paradigm in drug discovery [44], the predicted pairs by our methods should still be validated by experi ment methods. Furthermore, as the process of projecting a bipartite drug–target network into the corresponding monopartite drug and target networks is intrinsically associated with a certain loss of original interaction information [45], we urgently need to develop effective computational methods to integrate more valuable information for drug–target interaction network construc tion. With the continuous improvement of
Drews J. Drug discovery: a historical perspective. Science 287, 1960–1964 (2000).
2
Imming P, Sinning C, Meyer A. Drugs, their targets and the nature and number of drug targets. Nat. Rev. Drug Discov. 5, 821–834 (2006).
3
4
5
6
7
Haggarty SJ, Koeller KM, Wong JC, Butcher RA, Schreiber SL. Multidimensional chemical genetic analysis of diversity oriented synthesis-derived deacetylase inhibitors using cell-based assays. Chem. Biol. 10, 383–396 (2003). Kuruvilla FG, Shamji AF, Sternson SM, Hergenrother PJ, Schreiber SL. Dissecting glucose signalling with diversity-oriented synthesis and small-molecule microarrays. Nature 416, 653–657 (2002). Cheng AC, Coleman RG, Smyth KT et al. Structure-based maximal affinity model predicts small molecule druggability. Nat. Biotechnol. 25, 71–75 (2007). Yang Y, Adelstein SJ, Kassis AI. Target discovery from data mining approaches. Drug Discov. Today 14, 147–154 (2009). Xia Z, Wu LY, Zhou X, Wong ST. Semisupervised drug–protein interaction
1706
Acknowledgements The authors appreciate the constructive criticism and valuable suggestions from the anonymous reviewers.
Financial & competing interests disclosure This work was supported by National Basic Research Program of China (Grants No. 2012CB910400) and the Fundamental Research Funds for the Central Universities (Grants No. 78260026). The authors have no other rel evant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed. No writing assistance was utilized in the production of this manuscript.
Ethical conduct of research The authors state that they have obtained appropriate insti tutional review board approval or have followed the princi ples outlined in the Declaration of Helsinki for all human or animal experimental investigations. In addition, for investigations involving human subjects, informed consent has been obtained from the participants involved.
prediction from heterogeneous biological spaces. BMC Syst. Biol. 4, S6 (2010).
References 1
these technologies, we believe that this proposed method will offer valuable insights for the iden tification of novel disease-relevant targets.
8
9
Chen H, Zhang Z. A semi-supervised method for drug–target interaction prediction with consistency in networks. PLoS ONE 8(5), e62975 (2013). Yamanishi Y, Araki M, Cutteridge A, Honda W, Kanehisa M. Prediction of drug–target interaction networks from the integration of chemical and genomic spaces. Bioinformatics 24, i232–i240 (2008).
10 Gönen M. Predicting drug–target
interactions from chemical and genomic kernels using Bayesian matrix factorization. Bioinformatics 28(18), 2304–2310 (2012). 11 Bleakley K, Yamanishi Y. Supervised
prediction of drug–target interactions using bipartite local models. Bioinformatics 25, 2397–2403 (2009). 12 Mei JP, Kwoh CK, Yang P, Li XL, Zheng J.
Drug–target interaction prediction by learning from local information and neighbors. Bioinformatics 29(2), 238–245 (2013). 13 Alaimo S, Pulvirenti A, Giugno R, Ferro A.
Drug–target interaction prediction through domain-tuned network-based inference. Bioinformatics 29(16), 2004–2008 (2013).
Pharmacogenomics (2013) 14(14)
14 Cheng F, Liu C, Jiang J et al. Prediction of
drug–target interactions and drug repositioning via network-based inference. PLoS Comput. Biol. 8(5), e1002503 (2012). 15 Wishart DS, Knox C, Guo AC et al.
DrugBank: a knowledgebase for drugs, drug actions and drug targets. Nucleic Acids Res. 36, D901–D906 (2008). 16 Hamosh A, Scott AF, Amberger JS, Bocchini
CA, McKusick VA. Online Mendelian Inheritance in Man (OMIM), a knowledgebase of human genes and genetic disorders. Nucleic Acids Res. 33, D514–D517 (2005). 17 Jain E, Bairoch A, Duvaud S et al.
Infrastructure for the life sciences: design and implementation of the UniProt website. BMC Bioinformatics 10, 136 (2009). 18 Kanehisa M, Goto S, Hattori M et al. From
genomics to chemical genomics: new developments in KEGG. Nucleic Acids Res. 34, D354–D357 (2006). 19 Altschul SF, Gish W, Miller W, Myers EW,
Lipman DJ. Basic local alignment search tool. J. Mol. Biol. 215, 403–410 (1990). 20 Hattori M, Okuno Y, Goto S, Kanehisa M.
Development of a chemical structure comparison method for integrated analysis of
future science group
Predicting drug–target interaction networks of human diseases
chemical and genomic information in the metabolic pathways. J. Am. Chem. Soc. 125, 11853–11865 (2003). 21 Hattori M, Okuno Y, Goto S, Kanehisa M.
Heuristics for chemical compound matching. Genome Inform. 14, 144–155 (2003). 22 He Z, Zhang J, Shi XH et al. Predicting
drug–target interaction networks based on functional groups and biological features. PLoS ONE 5, e9603 (2010). 23 Peng ZL, Yang JY, Chen X. An improved
classification of G-protein-coupled receptors using sequence-derived features. BMC Bioinformatics 11, 420 (2010). 24 Li ZR, Lin HH, Han LY, Jiang L, Chen X,
Chen YZ. PROFEAT: a web server for computing structural and physicochemical features of proteins and peptides from amino acid sequence. Nucleic Acids Res. 34, W32–W37 (2006). 25 Tetko IV, Gasteiger J, Todeschini R et al.
Virtual computational chemistry laboratorydesign and description. J. Comput. Aid Mol. Des. 19, 453–463 (2005). 26 Ashburner M, Ball CA, Blake JA et al. Gene
ontology: tool for the unification of biology. The Gene Ontology Consortium. Nat. Genet. 25(1), 25–29 (2000). 27 Resnik P. Semantic similarity in a taxonomy:
an information-based measure and its application to problems of ambiguity in natural language. J. Artif. Intell. Res. 11, 95–130 (1999). 28 Ovaska K, Laakso M, Hautaniemi S. Fast
gene ontology based clustering for microarray experiments. BioData Min. 1, 11 (2008).
future science group
29 Zhang Y, Ding D, Li T. Gene selection
algorithm by combining reliefF and mRMR. BMC Genomics 9, S27 (2008). 30 Robnik-Sikonja M, Kononenko I. Theoretical
and empirical analysis of ReliefF and RReliefF. Mach. Learn. 53, 23–69 (2003). 31 Saeys Y, Inza I, Larranaga P. A review of
feature selection techniques in bioinformatics. Bioinformatics 23, 2507–2517 (2007). 32 Peng H, Long F, Ding C. Feature selection
based on mutual information: criteria of maxdependency, max-relevance, and minredundancy. IEEE Trans. Pattern Anal. Mach. Intell. 27, 1226–1238 (2005). 33 Lee MC. Using support vector machine with
a hybrid feature selection method to the stock trend prediction. Expert Syst. Appl. 36(8), 10896–10904 (2009). 34 Widodo A, Yang BS. Application of nonlinear
feature extraction and support vector machines for fault diagnosis of induction motors. Expert Syst. Appl. 36(7), 10266–10273 (2007). 35 Zhou D, Bousquet O, Lal TN, Weston J,
Schölkopf B. Learning with local and global consistency. Advances in Neural Information Processing System 16, 321–328 (2004). 36 Xu Q, Hu QX, Xue H, Yu W, Yang Q. Semi-
supervised protein subcellular localization. BMC Bioinformatics 10, S47 (2009). 37 You ZH, Yin Z, Han K, Huang DS, Zhou X.
A semi-supervised learning approach to predict synthetic genetic interactions by combining functional and topological properties of functional gene network. BMC Bioinformatics 11, 343 (2010).
www.futuremedicine.com
Research Article
38 Gestel TV, Suykens JAK, Baesens B et al.
Benchmarking least squares support vector machine classifiers. Mach. Learn. 54(1), 5–32 (2004). 39 Yang Y, Wang YP, Li KB. MiRTif: a support
vector machine-based microRNA target interaction filter. BMC Bioinformatics 9, S4 (2008). 40 Zhu X, Ghahramani Z, Lafferty J. Semi-
supervised learning using Gaussian fields and harmonic functions. Proceedings of the 20th International Conference (ICML 2003), 912–919 (2003). 41 Chang CC, Lin CJ. LIBSVM: a library for
support vector machines. ACM Trans. Intell. Syst. Technol. 2(3), 27 (2011). 42 Smialowski P, Frishman D, Kramer S. Pitfalls
of supervised feature selection. Bioinformatics 26, 440–443 (2010). 43 Park SH, Goo JM, Jo CH. Receiver
Operating Characteristic (ROC) curve practical review for radiologists. Korean J. Radiol. 5, 11–18 (2004). 44 Hopkins AL. Network pharmacology: the
next paradigm in drug discovery. Nat. Chem. Biol. 4, 682–690 (2008). 45 Vogt I, Mestres J. Drug-target networks. Mol.
Inform. 29, 10–14 (2010).
Website 101 Prediction of drug–target interaction
networks from the integration of chemical and genomic spaces. http://web.kuicr.kyoto-u.ac.jp//supp/yoshi/ drugtarget
1707
Predicting drug–target interaction networks of human diseases based on multiple feature information Aim: Drug–target interaction is crucial in the drug design process. Predicting the drug–target interaction networks of important human diseases can provide valuable clues for the characterization of the mechanism of action of diseases. Materials & methods: A new graph-based semisupervised learning (GBSSL) method is proposed to predict the drug–target interaction networks involved in 13 types of diseases. According to the method, each drug–target pair is initially described with different biological features including sequence, structure, function and network topology information. Then, the optimal feature selection procedures based on the relief and minimum redundancy maximum relevance are executed, respectively. Finally, unknown drug–target interactions can be predicted by the GBSSL method effectively. Results: The proposed method can effectively predict drug–target interactions (with a receiver operating characteristic score of 94.8% and a precision-recall score of 76.5%). Conclusion: Compared with the existing methods, the GBSSL method provides an efficient means of generating optimal features obtained from the combination of multiple sources of feature information. Original submitted 22 April 2013; Revision submitted 14 August 2013 KEYWORDS: disease network n drug–target interaction n semisupervised learning
Typically, a drug target is a key molecule involved in a particular metabolic or signaling pathway that is specific to a disease condition. Identification of new drug targets is the pre liminary step for drug discovery. Recent studies have shown that there are thousands of targets for drug action, whereas the number of cur rent identified drug targets is less than 500 [1,2]. It means that there are still a large number of targets undetected, especially for some impor tant human diseases. Much research has dem onstrated that drug–target interactions have been the basis of drug discovery and design in recent decades. There is no doubt that identify ing interactions between drug compounds and target proteins can provide valuable clues for discovering new targets and understanding the mechanism of drug activity [3]. As the detection of drug–target interactions by experimental methods is often laborious and costly, there is a strong incentive to develop new methods that are able to detect these potential drug–target interactions effectively [4]. With the development of bioinformatics tools and systems biology, a large amount of data regarding pharmaceutically targeted pro teins and drug–target interactions are publicly available for research and educational use. The availability of these public data motivated the development of machine learning methods for
drug–target interactions prediction. To date, a variety of computational methods including molecular docking [5], text mining [6], semi supervised learning (Laplacian regularized least square [NetLapRLS]) [7,8], kernel regres sion-based [9,10], bipartite local model (BLM) [11,12] and network-based inference [13,14] have been proposed to predict drug–target interac tions. Most of the methods attempted to infer unknown drug–target interactions by integrat ing both chemical (drug chemical structure) and genomic (protein structure) space information into a unified space. However, the drug–target interaction networks they predicted only utilized the chemical similarity information between drug compounds and genomic similarity infor mation between target proteins [10], which are limited and often difficult to describe for each drug–target pair. More importantly, little research has been devoted to studying the dis ease networks that are involved in drug–target interactions, especially for some important human tumor diseases. Therefore, in order to predict the unknown drug–target interaction pairs underlying disease networks, it is reason able to utilize the increased linkage between biological sequences and molecular informa tion by an effective learning method. Recently, semisupervised learning methods have attracted substantial research interest for their ability to
10.2217/PGS.13.162 © 2013 Future Medicine Ltd
Pharmacogenomics (2013) 14(14), 1701–1707
Weiming Yu1, Yan Yan1, Qing Liu1, Junxiang Wang1 & Zhenran Jiang*1 Department of Computer Science & Technology, East China Normal University, 200241, Shanghai, China *Author for correspondence: Tel.: +86 21 54345188 [email protected] 1
part of
ISSN 1462-2416
1701
1702
Pharmacogenomics (2013) 14(14)
H
Tertiary amine
P… P…
…
…… ……
……
Allylic group
HO
H H
CH3
H3C H
HO H
Protein feature space
Feature selection
H H
CH3
H3C H
x
Optimal feature subset
H
CH3
CH3
CH3
Protein sequence similarity
H
CH3
Protein sequence features Original features F1 F2 F3 F4 …… F... F... …… P1 …… P2 …… P 3
N
GO semantic similarity
Allylic alcohol
HO
O
Phenol
Chemical structure similarity
CH3 CH3
Representation of interaction vectors {(d1,p1),(d2,p2),…,(dn,pn)}
Semisupervised learning classifier
New interaction
Known drug
Known target
Predicted drug–target network
Figure 1. Predicting drug–target pairs in known drug–target interaction networks. Multiple feature information including protein homologous similarity, protein sequence features, protein function information, chemical structure similarity, functional groups and drug–target network topology information are integrated for drug–target interaction network construction. The graph-based semisupervised learning method is developed for interaction network prediction.
Known interaction
Known drug
Known target
Known drug–target network
Ether
HO
Chemical functional group
Chemical feature space
f(x)
Research Article Yu, Yan, Liu, Wang & Jiang
future science group
Predicting drug–target interaction networks of human diseases
maximize the generalization ability from chemi cal and genomic spaces by utilizing the small amount of existing label data and the abundant unlabeled data together [7,8]. In this paper, a graph-based semisupervised learning (GBSSL) method is developed to pre dict the unknown drug–target interactions based on multiple biological features including protein homologous similarity, protein sequence features, protein function information, chemi cal structure similarity, functional groups and drug–target network topology information. As a result, the method we proposed is shown to give better performance than two related meth ods, which can achieve an area under receiver operating characteristic (ROC) curve (AUC) score of over 94% and an area under precisionrecall (AUPR) score of up to 76%, respectively. In addition, several new drug–target interaction pairs are obtained for further function research through a comprehensive analysis of drug–target interaction networks.
Materials & methods In order to predict the drug–target interaction networks involved in 13 types of diseases, the proposed method utilized the functional groups and the chemical structure similarity to repre sent the chemical features of each FDA approved drug. Similarly, the features including average Gene Ontology semantic similarity, average pro tein homology similarity and the optimal pro tein sequence features obtained by the feature selection method, were used to describe each target protein. Meanwhile, the drug–target network topology features were obtained from the known drug–target interaction networks. Finally, 174-D features of each drug–target pair was regarded as feature vectors for the GBSSL classifier. Experiment results demonstrated that the GBSSL method can predict new drug– target interactions effectively (see Figur e 1 & Table 1) [8,9,15–40]. For a full method please see S upplementary M ethods at www.futuremedicine. com/doi/suppl/10.2217/pgs.13.162. Results & discussion Results of different kernel functions In this article, we applied two feature selection methods including relief and minimum redun dancy maximum relevance (mRMR) to gain the optimal protein sequence features with the pack age of Libsvm-3.1 [41]. Because the consequences of erroneous overoptimism may be estimated if both training and testing sets are used for super vised feature selection, the training set is used for future science group
Research Article
Table 1. Protein sequence features of each protein. Dimension
Properties
20
Composition of the 20 amino acid residues
400
Dipeptide composition
240
M-B autocorrelation group
240
Moran autocorrelation group
240
Geary autocorrelation group
147
Composition, transition and distribution groups
160
Sequence order group
80
Amphiphilic pseudo amino acid composition group
supervised feature selection and evaluated by an independent test set [42]. As different support vec tor machine (SVM) kernel functions can affect the performance of the feature selection methods, we tested four common kernels including linear, polynomial, radial and sigmoid basis function with the default parameters of C and g in order to obtain the best performance of the SVM model. The performance of each kernel function is evalu ated by tenfold cross-validation tests. The former 250-dimension features of each protein ranked by relief and mRMR are used as the SVM input. The results of the four kernel functions on the data set is shown in Table 2. As listed in Table 2 , the best performance of radial basis function is approximately 73 and 77% in overall accuracy for relief and mRMR, respectively. Therefore, we select the SVM model with radial basis function for the feature selection method in further study. Results of feature selection methods It is clear that the original protein sequence fea tures may include a large amount of irrelevant and redundant information. Removal of this information is a key step for drug–target inter action prediction. After the relief and mRMR algorithms list the ranks of the original feature set, incremental feature selection and forward feature selection (FFS) methods are used to Table 2. Tenfold cross-validation results for different kernel functions (accuracy). Kernel function
Relief (%)
mRMR (%)
SVM-linear
65.94
72.66
SVM-polynomial
61.98
66.25
SVM-RBF
73.27
77.33
SVM-sigmoid
70.01
75.52
mRMR: Minimum redundancy maximum relevance; RBF: Radial basis function; SVM: Support vector machine.
www.futuremedicine.com
1703
Research Article
Yu, Yan, Liu, Wang & Jiang
The IFS and FFS curves (mRMR)
Accuracy (%)
Accuracy (%)
The IFS and FFS curves (relief) 100 90 80 70 60 50 40 30 20 10 0
FFS IFS 0
50
100
150
200
250
100 90 80 70 60 50 40 30 20 10 0
FFS IFS 0
Number of feature
50
100
150
200
250
Number of feature
Figure 2. The curves of incremental feature selection and forward feature selection based on relief and minimum redundancy maximum relevance for 250-dimension features. FFS: Forward feature selection; IFS: Incremental feature selection; mRMR: Minimum redundancy maximum relevance.
generate the optimal feature subset in this study. The incremental feature selection curves based on relief and mRMR with their corresponding FFS curves are obtained (Figure 2). By compari son, the mRMR algorithm has better perfor mance than the single relief algorithm, and the peaks of FFS based on mRMR can reach the overall accuracy rate of 83.4% with 87 features. Comparison with other methods BLM [10] is a statistical method that is used to predict unknown drug–target interactions from chemical structure and genomic sequence information simultaneously. NetLapRLS [7] can reduce labeling costs and improve the perfor mance of learning, especially suitable for very small-labeled data and large unlabeled data. Fur thermore, we compared the performance with the two classical methods, we not only tested
our methods on this data set, but also tested it on four human drug–target interaction data sets involving enzymes, ion channels, G-proteincoupled receptors and nuclear receptors, which are publicly available at [101] (see also [9]). The three methods: GBSSL, BLM and NetlapRLS were further evaluated on the bench mark data set, respectively. We executed ten tri als of a tenfold cross-validation procedure: the data set of drug–target pairs was divided into ten subsets, each subset was then taken in turn as a test set and the remaining nine were used as training sets. For example, there are 103 US FDA approved drugs and 143 target proteins with 312 known interactions and 14,417 unknown drug– protein pairs in our study. In each cross-valida tion, the 281 drug–target interactions are used for training classification while the remaining 1441 unknown drug–protein pairs and 31 drug–target
NetLapRLS BLM GBSSL Random
Precision
True positive rate (sensitivity)
ROC curve 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
Precison-recall curve NetLapRLS BLM GBSSL
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
False positive rate (1-specificity)
Recall
Figure 3. Comparison of the receiver operating characteristic curves and precision-recall curves for the three different methods. BLM: Bipartite local model; GBSSL: Graph-based semisupervised learning; ROC: Receiver operating characteristic.
1704
Pharmacogenomics (2013) 14(14)
future science group
Predicting drug–target interaction networks of human diseases
interactions are designated as the testing data set. The performances of the methods are evaluated by two quality measures called AUC and AUPR. Figure 3 shows the ROC curves and precision-recall curves of three different methods. As demon strated in Figure 3, the GBSSL method can achieve a better precision than the two other methods. The proposed method increased by 5 and 6% on the AUC scores, and improved by 11 and 35% on the AUPR scores for BLM and NetlapRLS meth ods, respectively. We can obtain the best results – an AUC score of 94.8% and AUPR score of 76.5% – with the GBSSL method. Why does the GBSSL method shows a higher AUC and AUPR than both the other methods for this data set? The reasons are that the method used in the current research possesses several mer its. First, many drugs are associated with two or more diseases of the 13 diseases in the data set. These 13 diseases are all associated with cancer and have some similarities in disease phenotype. Our method can utilize the disease information that is involved in drug–target interaction to the greatest extent. Second, the GBSSL method incor porates several critical similarity measures, includ ing Gene Ontology similarities, protein sequence similarities and chemical similarities, to build reli able drug–target networks. Third, we performed the feature selection processes to obtain a subset of relevant features (variables) for constructing more robust learning models, which is effective in capturing the effects of more interacting features. Meanwhile, we compare the performance of the three methods on four classes of drug–target interactions. The AUC scores and AUPR scores were used to evaluate the performance of these methods [43]. The prediction results for GBSSL,
Research Article
Table 3. Performance of four classes of drug target families. Dataset
Method
AUC
AUPR
Enzyme
GBSSL
98.2
84.5
BLM
97.3
85.1
NetLapRLS
95.6
75.4
GBSSL
96.7
84.6
BLM
97.2
83.5
NetLapRLS
94.4
71.3
GBSSL
94.2
54.7
BLM
94.5
66.3
NetLapRLS
84.8
45.2
GBSSL
93.6
67.1
BLM
84.7
59.3
NetLapRLS
82.5
37.8
Ion channel
GPCR
Nuclear receptor
AUC: Area under receiver operating characteristic curve; AUPR: Area under precision-recall; BLM: Bipartite local model; GBSSL: Graph-based semisupervised learning; GPCR: G-protein-coupled receptors.
BLM and NetlapRLS methods after tenfold cross-validation procedures are shown in Table 3. Because we only consider the information from drug structure similarity, target sequence simi larity and network topology information instead of more biological features for predicting the drug–target interaction networks, we have not achieved a greater advantage than the latter two methods. However, we believe that accuracy will be improved if more biological features are combined into our method in future work.
Conclusion & future perspective In conclusion, we proposed a GBSSL method to predict unknown drug–target interactions in 13 types of diseases from multiple biological
Executive summary Background Predicting the drug–target interaction networks of human diseases can provide valuable clues for the characterization of the mechanism of action of diseases. The availability of the public drug–target data motivated the development of machine learning methods for the prediction of drug–target interactions. However, little research has been devoted to studying the disease networks that are involved in drug–target interactions, especially for some important human tumor diseases. It is necessary to develop a method to predict the drug–target interaction networks of human diseases based on multiple feature information. Results Compared with the existing methods, the graph-based semisupervised learning method provides an efficient means of generating optimal features obtained from the combination of multiple sources of information. The proposed method can effectively predict drug–target interactions (with a receiver operating characteristic score of 94.8% and a precision-recall score of 76.5%). Limitations Each drug–target pair can be encoded with more chemical and biological features. The biological significance of the networks can be utilized more effectively. The predicted drug–target pairs need to be validated experimentally.
future science group
www.futuremedicine.com
1705
Research Article
Yu, Yan, Liu, Wang & Jiang
information sources. The proposed method is a complementary method that can be used for studying the function of disease interaction net works. The merit of the study is that we not only consider the drug structure similarity and tar get sequence similarity, but also integrate novel features such as protein function information, sequence features, functional groups and net work topology information into the classifier. The experiment demonstrates that our method has a superior performance than BLM and NetLapRLS, especially for the sample imbal ance problem that is involved in drug–target interaction network prediction. It should be noted that although network pharmacology was regarded as the next paradigm in drug discovery [44], the predicted pairs by our methods should still be validated by experi ment methods. Furthermore, as the process of projecting a bipartite drug–target network into the corresponding monopartite drug and target networks is intrinsically associated with a certain loss of original interaction information [45], we urgently need to develop effective computational methods to integrate more valuable information for drug–target interaction network construc tion. With the continuous improvement of
Drews J. Drug discovery: a historical perspective. Science 287, 1960–1964 (2000).
2
Imming P, Sinning C, Meyer A. Drugs, their targets and the nature and number of drug targets. Nat. Rev. Drug Discov. 5, 821–834 (2006).
3
4
5
6
7
Haggarty SJ, Koeller KM, Wong JC, Butcher RA, Schreiber SL. Multidimensional chemical genetic analysis of diversity oriented synthesis-derived deacetylase inhibitors using cell-based assays. Chem. Biol. 10, 383–396 (2003). Kuruvilla FG, Shamji AF, Sternson SM, Hergenrother PJ, Schreiber SL. Dissecting glucose signalling with diversity-oriented synthesis and small-molecule microarrays. Nature 416, 653–657 (2002). Cheng AC, Coleman RG, Smyth KT et al. Structure-based maximal affinity model predicts small molecule druggability. Nat. Biotechnol. 25, 71–75 (2007). Yang Y, Adelstein SJ, Kassis AI. Target discovery from data mining approaches. Drug Discov. Today 14, 147–154 (2009). Xia Z, Wu LY, Zhou X, Wong ST. Semisupervised drug–protein interaction
1706
Acknowledgements The authors appreciate the constructive criticism and valuable suggestions from the anonymous reviewers.
Financial & competing interests disclosure This work was supported by National Basic Research Program of China (Grants No. 2012CB910400) and the Fundamental Research Funds for the Central Universities (Grants No. 78260026). The authors have no other rel evant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed. No writing assistance was utilized in the production of this manuscript.
Ethical conduct of research The authors state that they have obtained appropriate insti tutional review board approval or have followed the princi ples outlined in the Declaration of Helsinki for all human or animal experimental investigations. In addition, for investigations involving human subjects, informed consent has been obtained from the participants involved.
prediction from heterogeneous biological spaces. BMC Syst. Biol. 4, S6 (2010).
References 1
these technologies, we believe that this proposed method will offer valuable insights for the iden tification of novel disease-relevant targets.
8
9
Chen H, Zhang Z. A semi-supervised method for drug–target interaction prediction with consistency in networks. PLoS ONE 8(5), e62975 (2013). Yamanishi Y, Araki M, Cutteridge A, Honda W, Kanehisa M. Prediction of drug–target interaction networks from the integration of chemical and genomic spaces. Bioinformatics 24, i232–i240 (2008).
10 Gönen M. Predicting drug–target
interactions from chemical and genomic kernels using Bayesian matrix factorization. Bioinformatics 28(18), 2304–2310 (2012). 11 Bleakley K, Yamanishi Y. Supervised
prediction of drug–target interactions using bipartite local models. Bioinformatics 25, 2397–2403 (2009). 12 Mei JP, Kwoh CK, Yang P, Li XL, Zheng J.
Drug–target interaction prediction by learning from local information and neighbors. Bioinformatics 29(2), 238–245 (2013). 13 Alaimo S, Pulvirenti A, Giugno R, Ferro A.
Drug–target interaction prediction through domain-tuned network-based inference. Bioinformatics 29(16), 2004–2008 (2013).
Pharmacogenomics (2013) 14(14)
14 Cheng F, Liu C, Jiang J et al. Prediction of
drug–target interactions and drug repositioning via network-based inference. PLoS Comput. Biol. 8(5), e1002503 (2012). 15 Wishart DS, Knox C, Guo AC et al.
DrugBank: a knowledgebase for drugs, drug actions and drug targets. Nucleic Acids Res. 36, D901–D906 (2008). 16 Hamosh A, Scott AF, Amberger JS, Bocchini
CA, McKusick VA. Online Mendelian Inheritance in Man (OMIM), a knowledgebase of human genes and genetic disorders. Nucleic Acids Res. 33, D514–D517 (2005). 17 Jain E, Bairoch A, Duvaud S et al.
Infrastructure for the life sciences: design and implementation of the UniProt website. BMC Bioinformatics 10, 136 (2009). 18 Kanehisa M, Goto S, Hattori M et al. From
genomics to chemical genomics: new developments in KEGG. Nucleic Acids Res. 34, D354–D357 (2006). 19 Altschul SF, Gish W, Miller W, Myers EW,
Lipman DJ. Basic local alignment search tool. J. Mol. Biol. 215, 403–410 (1990). 20 Hattori M, Okuno Y, Goto S, Kanehisa M.
Development of a chemical structure comparison method for integrated analysis of
future science group
Predicting drug–target interaction networks of human diseases
chemical and genomic information in the metabolic pathways. J. Am. Chem. Soc. 125, 11853–11865 (2003). 21 Hattori M, Okuno Y, Goto S, Kanehisa M.
Heuristics for chemical compound matching. Genome Inform. 14, 144–155 (2003). 22 He Z, Zhang J, Shi XH et al. Predicting
drug–target interaction networks based on functional groups and biological features. PLoS ONE 5, e9603 (2010). 23 Peng ZL, Yang JY, Chen X. An improved
classification of G-protein-coupled receptors using sequence-derived features. BMC Bioinformatics 11, 420 (2010). 24 Li ZR, Lin HH, Han LY, Jiang L, Chen X,
Chen YZ. PROFEAT: a web server for computing structural and physicochemical features of proteins and peptides from amino acid sequence. Nucleic Acids Res. 34, W32–W37 (2006). 25 Tetko IV, Gasteiger J, Todeschini R et al.
Virtual computational chemistry laboratorydesign and description. J. Comput. Aid Mol. Des. 19, 453–463 (2005). 26 Ashburner M, Ball CA, Blake JA et al. Gene
ontology: tool for the unification of biology. The Gene Ontology Consortium. Nat. Genet. 25(1), 25–29 (2000). 27 Resnik P. Semantic similarity in a taxonomy:
an information-based measure and its application to problems of ambiguity in natural language. J. Artif. Intell. Res. 11, 95–130 (1999). 28 Ovaska K, Laakso M, Hautaniemi S. Fast
gene ontology based clustering for microarray experiments. BioData Min. 1, 11 (2008).
future science group
29 Zhang Y, Ding D, Li T. Gene selection
algorithm by combining reliefF and mRMR. BMC Genomics 9, S27 (2008). 30 Robnik-Sikonja M, Kononenko I. Theoretical
and empirical analysis of ReliefF and RReliefF. Mach. Learn. 53, 23–69 (2003). 31 Saeys Y, Inza I, Larranaga P. A review of
feature selection techniques in bioinformatics. Bioinformatics 23, 2507–2517 (2007). 32 Peng H, Long F, Ding C. Feature selection
based on mutual information: criteria of maxdependency, max-relevance, and minredundancy. IEEE Trans. Pattern Anal. Mach. Intell. 27, 1226–1238 (2005). 33 Lee MC. Using support vector machine with
a hybrid feature selection method to the stock trend prediction. Expert Syst. Appl. 36(8), 10896–10904 (2009). 34 Widodo A, Yang BS. Application of nonlinear
feature extraction and support vector machines for fault diagnosis of induction motors. Expert Syst. Appl. 36(7), 10266–10273 (2007). 35 Zhou D, Bousquet O, Lal TN, Weston J,
Schölkopf B. Learning with local and global consistency. Advances in Neural Information Processing System 16, 321–328 (2004). 36 Xu Q, Hu QX, Xue H, Yu W, Yang Q. Semi-
supervised protein subcellular localization. BMC Bioinformatics 10, S47 (2009). 37 You ZH, Yin Z, Han K, Huang DS, Zhou X.
A semi-supervised learning approach to predict synthetic genetic interactions by combining functional and topological properties of functional gene network. BMC Bioinformatics 11, 343 (2010).
www.futuremedicine.com
Research Article
38 Gestel TV, Suykens JAK, Baesens B et al.
Benchmarking least squares support vector machine classifiers. Mach. Learn. 54(1), 5–32 (2004). 39 Yang Y, Wang YP, Li KB. MiRTif: a support
vector machine-based microRNA target interaction filter. BMC Bioinformatics 9, S4 (2008). 40 Zhu X, Ghahramani Z, Lafferty J. Semi-
supervised learning using Gaussian fields and harmonic functions. Proceedings of the 20th International Conference (ICML 2003), 912–919 (2003). 41 Chang CC, Lin CJ. LIBSVM: a library for
support vector machines. ACM Trans. Intell. Syst. Technol. 2(3), 27 (2011). 42 Smialowski P, Frishman D, Kramer S. Pitfalls
of supervised feature selection. Bioinformatics 26, 440–443 (2010). 43 Park SH, Goo JM, Jo CH. Receiver
Operating Characteristic (ROC) curve practical review for radiologists. Korean J. Radiol. 5, 11–18 (2004). 44 Hopkins AL. Network pharmacology: the
next paradigm in drug discovery. Nat. Chem. Biol. 4, 682–690 (2008). 45 Vogt I, Mestres J. Drug-target networks. Mol.
Inform. 29, 10–14 (2010).
Website 101 Prediction of drug–target interaction
networks from the integration of chemical and genomic spaces. http://web.kuicr.kyoto-u.ac.jp//supp/yoshi/ drugtarget
1707
