Ensemble Learning-based Prediction of Drug-pathway Interactions

based on Features Integration

Mingyuan Xin

, Jun Fan

and Zhenran Jiang

Shanghai Key Laboratory of Regulatory Biology, Institute of Biomedical Sciences and School of Life Sciences,

East China Normal University, Shanghai 200241, China

Shanghai Key Laboratory of Multidimensional Information Processing, Department of Computer Science and Technology,

East China Normal University, Shanghai 200241, China

Keywords: Drug-pathway Interaction, Ensemble Learning, AdaBoost, Bagging, Random SubSpace.

Abstract: Recently, developing computational methods to explore drug-pathway interaction relationships has attracted

attention for their potentiality in discovering unknown targets and mechanisms of drug actions. However,

mining suitable features of drugs and pathways is challenging for available prediction methods. This paper

performed an ensemble learning-based method to predict potential drug-pathway interactions by integrating

different drug-based and pathway-based features. The main characteristic of our method lies in using the

Relief algorithm for feature selection and regarding three ensemble methods (AdaBoost, Bagging and

Random Subspace) for classifiers. Cross validation results showed the AdaBoost algorithm that based on the

Decision Tree classifier can obtain a higher prediction accuracy, which indicated the effectiveness of

ensemble learning. Moreover, some new predicted interactions were validated by database searching, which

demonstrated its potentiality for further biological experiment investigation.

1 INTRODUCTION

Traditional drug discovery primarily tries to seek the

specific drug molecule to act on individual target

(Hopkins, 2008). However, it is well recognized that

many drugs are far beyond targeting individual

proteins, but rather influencing the complex

interactions among the relevant biological pathways.

Therefore, the inferences of drug-pathway

associations are critical for identifying unknown

targeted pathways and drug action mechanisms (Ma

and Zhao, 2012).

Increasing effort has been devoted to detecting

these potential associations and several drug-

pathway interactions prediction methods have been

proposed from different aspects (Subramanian et al.,

2005; Ma and Zhao, 2012). Generally, most of the

methods attempted to analyze the drug-pathway

interactions mainly based on gene expression data.

For instance, ‘iFad’ mainly combined the gene

expression and drug sensitivity datasets to analyze

the drug-pathway interactions (Ma and Zhao, 2012),

but it is always difficult to obtain adequate drug-

pathway information merely on the gene expression

data. To tackle the problem, some methods attempt

to utilize different machine learning algorithms by

integrating more chemical and biological

information (Silberberg et al., 2012; Pratanwanich

and Lio, 2014; Song et al., 2014). For instance,

protein-protein interaction networks (PPI)

(Silberberg et al., 2012), other target structure

information have been utilized effectively recently.

However, the extraction and fusion of the drug-

pathway association information is still challenging

for drug-pathway interactions prediction (Song et

al., 2014).

Inspired by the challenges, we attempted to use

the ensemble learning methods to predict potential

drug-pathway associations. As similar drugs often

act similar target proteins, we assume that similar

drugs also act on similar pathways. Based on the fact

that the drug mode of actions (MoA) is a central

concept linking drug structures to a set of biological

activities, we used drug structure and MoA

similarity to represent drug feature information.

Further, we used the ‘RNA: AffyHG-U133 (A, B)’

gene expression data of NCI-60 cell lines (Reinhold

et al., 2012) to obtain related genes which covered

by these pathways, then these genes ontology

semantic similarity and sequence similarity are

Xin M., Fan J. and Jiang Z.

Ensemble Learning-based Prediction of Drug-pathway Interactions based on Features Integration.

DOI: 10.5220/0006096701170124

In Proceedings of the 10th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2017), pages 117-124

ISBN: 978-989-758-214-1

117

calculated to represent pathway information.

Further, the drug-pathway network topology

information was merged into the drug and pathway

feature profiles, respectively.

It is known that ensemble learning methods

usually exhibits better generalization performance

than a single classifier. In this study, we used three

well-established methods: AdaBoost (Freund and

Schapire, 1997), Bagging (Breiman, 1996) and

Random SubSpace (Ho, 1998) to achieve a good

ensemble result. Meanwhile, three widely used

learning methods: Support Vector Machine (SVM)

(Cortes and Vapnik, 1995), Navie Bayesian (NB)

(Rish, 2001) and Decision Tree (DT) (Friedl and

Brodley, 1997) are chosen as the base classifier.

Compared of these method combinations, the

AdaBoost algorithm that based on the DT classifier

is selected as the final model to predict the drug-

pathway interactions.

2 MATERIAL AND METHOD

2.1 Dataset

This study focus on 58 pathways that have been

proved to be related to cancers (Ahmed et al., 2011)

and 362 drugs obtained from KEGG database

(Kanehisa et al., 2012), which contains most of

pathways and molecular information in genomics,

transcriptomics, proteomics and metabolomics. In

addition, these drugs have complete drug

information and most of them are proved to be

related to these pathways.

2.1.1 Features Construction

(1) Drug features

Drug structure-based feature

S : Drug structure

similarity is calculated based on their molecular

fingerprints which include 881 chemical

substructures defined by the PubChem database

(Chen et al., 2009). PaDel-Descriptor (Yap, 2011)

was used to convert each drug Mol file into 881

dimensional binary vectors. Then the corresponding

fingerprints are used to compute the similarity scores

between two drugs by Tanimoto scores (Lipkus,

1999).

MoA based feature

: Since drugs which share a

similar MoA are likely to target same pathways, thus

the drug MoA similarity can be utilized to predict

associations between drugs and pathways. Here we

retrieved MoA information from DrugBank database

(Wishart et al., 2006) and calculate the similarities

based on 341 MoAs. We consider drugs as samples

and each MoA as a label and take known drug-MoA

association matrix M as local correlations.

According to the local correlations between labels of

samples in drug-MoA interaction network, we

calculate the cosine similarity of each two drug

vectors in M:

(, ) cos( , )

|| |||| ||

dij

Sij mm





(1)

(2)Pathway features

This study mainly concentrated on 1863 genes

covered by the 58 pathways for the pathway features

construction.

Gene ontology Semantic feature

: The Gene

Ontology terms of 1863 genes were retrieved from

Quick GO database (Binns et al., 2009), and

semantic similarity scores between these pathway-

related genes were calculated by the csbl.go R

package (Ovaska et al., 2008). What’s more, the

similarity scores between the pathways from gene

semantic similarity scores were computed in

accordance with the reference (Song et al., 2014).

Gene sequence similarity

: Sequence similarity

between the corresponding pathway-related genes

was calculated based on a Smith-Waterman

sequence alignment score (Smith et al., 1985), and

the similarity between two pathways can be

calculated as the sum of similarity between all the

gene sequences related to the two pathways.

(3)Drug-pathway network topology feature

The drug-pathway network topology information

was calculated based on the literature (Van et al.,

2013). In the drug-pathway network, the average

shortest path of each node and the number of shared

drugs or pathways are denoted as

,,,

dpdp

DKK

respectively.

As showed in Table 1, the drug features

Sim

include drug structure information

S , drug mode of

actions

, network topology information

and the pathway features

Sim

are combined by

gene ontology semantic similarity

, gene

sequence similarity

and ,

DK. Construction of

the drug-pathway feature is followed the theory: for

drug i and pathway j, their features can be

constructed by combining row i in

Sim and row j

Sim

, namely.

Fea<drug(i),pathway(j)>=

Sim (i)+

Sim

(j). (2)

BIOINFORMATICS 2017 - 8th International Conference on Bioinformatics Models, Methods and Algorithms

118

Table 1: The construction of drug-pathway features.

Drug-Pathway Features

Drug

Features

Sim

Drug structure similarity

Drug mode of actions

similarity

Drug-pathway interaction

topology information

Pathway

Features

Sim

Pathway-related gene

ontology semantic feature

Pathway-related gene

sequence similarity

Drug-pathway interaction

topology information

2.1.2 Features Selection

Existing facts demonstrate that irrelevant and

redundant features can lead the model to overfit.

Here we perform the Relief method (Sun et al.,

2011) to avoid redundancy of feature variables. At

each iteration, the algorithm picks randomly a

sample K, then picks at random the feature sample

of the instance closest to K from each class, the

same class instance is called ‘near-hit’ and the

different class instance is called ‘near-miss’. Then

the weight vector is updated as:

()( )

iii i i i

W W x nearHit x nearMiss   

(3)

Thus, the weight of any given feature increases if

the distance between K and near-hit is shorter than

the distance between K and near-miss for the feature,

and decreases otherwise. After n iterations, the

relevance vector is updated by dividing each element

of the weight vector by n, then feature are selected if

their relevance is greater than a threshold k. In this

study, we set the threshold as zero and finally

selected 551 features with positive weight from 764

features.

2.2 Ensemble Learning Method

Ensemble learning is a machine learning paradigm

which constructs a set of classifiers and then

combines them for classifying data by taking a vote

of their predictions (Schwenker, 2013). Here we take

three well-established methods in practice to achieve

a good ensemble. AdaBoost and Bagging are two

instance partitioning methods and Random Subspace

is a feature partitioning method (Van et al., 2013).

2.2.1 AdaBoost

AdaBoost is an iterative algorithm where the

conjuncture of many weak classifiers is employed to

construct a ‘strong’ classifier (Ho, 1998). It works

by choosing a base algorithm and iteratively

improving it by accounting for the incorrectly

classified examples in the training set. The final

predictions are retrieved from a weighted vote. The

AdaBoost algorithm’s pseudo code is shown as

followed:

2.2.2 Bagging

Bagging is an ensemble meta-estimator where each

base classifier is trained on random subsets of the

original dataset and then aggregated their individual

predictions to form a final prediction (Breiman,

1996). It improves the stability and reduces variance,

and avoids overfitting of learning algorithms. The

base classifiers’ combination strategy for Bagging is

majority vote. The Bagging algorithm pseudo code

is shown as followed:

2.2.3 Random Subspace

Random Subspace is a combination model that

consists of several classifiers and each are trained on

randomly chosen subspaces of the original feature

space (Ho, 1998). The outputs of the models are

usually combined by majority vote. The Random

Subspace algorithm’s pseudo codes are shown as

followed:

Ensemble Learning-based Prediction of Drug-pathway Interactions based on Features Integration

119

2.3 Procedure

In this model, we choose four widely used base

classifier for implementing the three ensemble

methods: SVM, NB and

DT. SVM algorithm has

been used for a variety of application and it performs

structural risk minimization on a nested set structure

of separating hyperplanes (Cortes and Vapnik, 1995).

Navie Bayesian algorithm is a simple classification

based on the Bayes theory for conditional

probability. Decision Tree algorithm is an easily

understandable and transparent sequential model but

it has relatively low prediction accuracy compared to

other methods. In this study, we chose the widely

used method C4.5. Here we use the toolkit WEKA,

which includes a collection of machine learning

algorithms for solving data mining problems (Hall et

al., 2009). The AdaBoost, Bagging and Random

SubSpace are selected to implement the ensemble

algorithms. The drug-pathway associations we used

include 643 positive samples and 17390 negative

samples, and the positive sample density of the

dataset is 0.036.

In order to evaluate the performances of different

models, 10-fold cross validation tests are executed

on the models. For the datasets, all of the drug-

pathway samples are randomly spilt into ten subsets

with equal size, and nine subsets are combined as

the training set and the remaining one subset is taken

as the testing set each time. The overview procedure

of the model is shown in the Fig.1.

Figure 1: Figure summarizes the overview of this model.

The model is mainly composed of three sections: (a) the

process of feature construction. (b) ensemble methods

operation. (c) the comparison of these methods.

3 RESULTS

3.1 Performance Evaluation

Here several metrics, i.e., precision, recall, accuracy

(ACC), area under ROC curve (AUC) and the area

under the precision-recall curve (AUPR), F-measure

(F), are used to evaluate the performances of the

models. Among the metrics, accuracy represents the

overall accuracy of the classification, precision

represents the measure of the reliability of positive

instances prediction and recall represents the

probability of correct prediction. F-measure is a

score from 0 to 1 as a measure of test accuracy. The

metrics were calculated in a 10-fold cross-validation

procedure by using the equations as followed:

()

precision

TP FP

recall

TP FN

TP TN

ACC

TP TN FP FN

recision recall



















(4)

where TP, FP, TN and FN represent the number

of true positive, false positive, true negative and

false negative samples, respectively.

BIOINFORMATICS 2017 - 8th International Conference on Bioinformatics Models, Methods and Algorithms

120

3.2 Performance of Features

Integration

To quantitatively assess the efficiency of all the

features and each single feature in predicting the

drug-pathway interactions, we performed a 10-fold

cross validation with the AdaBoost algorithm based

on DT classifier, respectively. As a result, the model

that integrated features exhibits a better performance

than those with single feature (See in Fig. 2).

Further, the Relief method is performed to avoid

the redundancy of feature variables. In our study, we

get 551 features with positive weight from 764

features after feature selection. By comparison, the

selected features have better classification

performances than the original features (See in

Fig.3).

3.3 Performance of Ensemble Methods

In this model, we compared the performance of 12

methods, including SVM, NB and DT, and their

corresponding ensemble methods of AdaBoost,

Bagging and Random Subspace. The performance of

base classifiers and ensemble methods based on

three base classifiers is shown in Table 2. As

demonstrated in Table 2, we find all the three base

classifiers have a poorer performance than the

ensemble methods, and AdaBoost method has the

best performance in every base classifier. The

possible reason for this situation is that AdaBoost

more fully account the weight of each classifier

relative to other algorithms.

Next, we compared the three base classifiers in

the case of AdaBoost ensemble models. The ROC

and PR curves of the three approaches are shown in

the Fig. 4, we can see that the AdaBoost ensemble

algorithm based on the DT classifier can achieve the

best performance.

3.4 New Predictions

Here we used the Comparative Toxicogenomics

Database (CTD) (Davis et al., 2015) as reference to

validate the predicted interactions. The CTD

database integrates chemical, gene, disease and their

interactions from curated literatures. There are 502

new predicted interactions and 241 associations have

been proved existence by searching the CTD

database. For instance, the interaction between the

drug ‘Theophylline’ and the pathway ‘Neuroactive

ligand-receptor interaction’ can be found in both

KEGG database and CTD database. Some predicted

samples that have been confirmed in CTD database

are listed in Table 3.

In addition, we focused on the pathway:

Kegg05223 and associated predicted drugs. We

found there are 15 predicted drugs related with the

pathway ‘Non-small cell lung cancer’. Meanwhile,

we confirmed that eleven drugs have associations

with the pathway in CTD database. Among the other

four drugs, we cannot find the interactions between

the drug and the pathway ‘Non-small cell lung

cancer’, but from the aspect of disease we find that

the three drug ‘Aminoglutethimide’, ‘Sunitinib

malate’ and ‘sunitinib’ have been tested in clinical

trials for lung cancer in the literatures (Xiao et al.,

2010; Chen et al., 2011; Shin et al., 2013; Xue et al.,

2014), which have been laterally validated that the

drugs have associations with this pathway.

Figure 2: The comparison between the integrated features and each single feature.

Ensemble Learning-based Prediction of Drug-pathway Interactions based on Features Integration

121

Figure 3: The comparison between selected features and original features.

Table 2: Performance comparisons of different learning methods.

Method AUC AUPR Recall Precision ACC F

SVM 0.827 0.770 0.827 0.827 0.653 0.827

AdaBoost SVM 0.922 0.925 0.834 0.835 0.669 0.834

Bagging SVM 0.856 0.811 0.826 0.826 0.652 0.826

RS SVM 0.859 0.821 0.804 0.804 0.608 0.804

NB 0.772 0.732 0.715 0.715 0.429 0.715

AdaBoost NB 0.851 0.845 0.779 0.779 0.558 0.778

Bagging NB 0.813 0.792 0.725 0.726 0.451 0.725

RS NB 0.804 0.790 0.708 0.709 0.418 0.708

DT 0.891 0.857 0.882 0.883 0.765 0.882

AdaBoost DT 0.975 0.976 0.925 0.925 0.850 0.925

Bagging DT 0.965 0.966 0.901 0.902 0.803 0.900

RS DT 0.974 0.972 0.916 0.916 0.833 0.916

Figure 4: The evaluation of the four methods: DT, AdaBoost DT, Bagging DT and Random SubSpace DT.

BIOINFORMATICS 2017 - 8th International Conference on Bioinformatics Models, Methods and Algorithms

122

Table 3: The top 20 confirmed drug-pathway interactions.

DrugID (Kegg) Drug Name Pathway Name Validated Database

D00371 Theophylline Neuroactive ligand-receptor interaction Kegg; CTD

D04197 Floxuridine Natural killer cell mediated cytotoxicity CTD

D04023 Erlotinib hydrochloride Chronic myeloid leukemia CTD

D03881 Dobutamine tartrate Vascular smooth muscle contraction CTD

D03879 Dobutamine Vascular smooth muscle contraction CTD

D00371 Theophylline Vascular smooth muscle contraction Kegg; CTD

D00632 Dobutamine hydrochloride Vascular smooth muscle contraction CTD

D08111 Lercanidipine Vascular smooth muscle contraction Kegg; CTD

D01849 Lercanidipine hydrochloride Vascular smooth muscle contraction Kegg; CTD

D00126 Ibuprofen Insulin signaling pathway CTD

D01366 Bezafibrate Insulin signaling pathway CTD

D00341 Hydroxycarbamide Natural killer cell mediated cytotoxicity CTD

D00330 Flurbiprofen Insulin signaling pathway CTD

D00565 Fenofibrate Insulin signaling pathway CTD

D00586 Flutamide Non-small cell lung cancer CTD

D04023 Erlotinib hydrochloride Pancreatic cancer CTD

D00562 Propylthiouracil Natural killer cell mediated cytotoxicity CTD

D02368 Gemcitabine Cytokine-cytokine receptor interaction CTD

D01441 Imatinib mesilate Non-small cell lung cancer CTD

D01155 Gemcitabine hydrochloride Jak-STAT signaling pathway CTD

4 CONCLUSIONS

In this article, we evaluated the ensemble

algorithms: AdaBoost, Bagging and Random

SubSpace, for predicting drug-pathway interactions

based on three base classifiers: SVM, NB and DT.

Our results show that ensemble methods have the

advantage over the individual classifier on drug-

pathway interactions prediction. The merit of this

study lied in selecting the effective features obtained

from drug chemical structure information, drug

mode of actions and pathway-related gene

information. Some validated results to some extent

demonstrated the reliability of the models.

Although our method has utilized different types

of drug-based and pathway-based information, more

useful drug-pathway information can be further

mined. Therefore, our future study will focus on

fusing more biological prior information to improve

the prediction reliability.

ACKNOWLEDGEMENTS

This work was partly supported by National Natural

Science Foundation of China (Grants No.

81330049), National Major Scientific and

Technological Special Project for "Significant New

Drugs Development" (2013ZX09507001) and the

Science and Technology Commission of Shanghai

Municipality (14DZ2270100).

REFERENCES

Ahmed, J., Meinel, T., Dunkel, M., Murgueitio, M.S.,

Adams, R., Blasse, C., et al. 2011. CancerResource: a

comprehensive database of cancer-relevant proteins

and compound interactions supported by experimental

knowledge. Nucleic Acids Res., 39, D960-D967.

Binns, D., Dimmer, E., Huntley, R., Barrell, D.,

O'Donovan, C. and Apweiler, R., 2009. QuickGO: a

web-based tool for Gene Ontology searching.

Bioinformatics, 25(22), 3045-6.

Breiman, L., 1996. Bagging predictors. Machine Learning,

24(2), 123-140.

Chen, B., Wild, D. and Guha, R., 2009. PubChem as a

source of polypharmacology. J. Chem. Inf. Model., 49

(9), 2044-2055.

Chen, C., Fu, X., Zhang, D., Li, Y., Xie, Y., Li, Y. and

Huang Y., 2011. Varied pathways of stage IA lung

adenocarcinomas discovered by integrated gene

expression analysis. Int. J. Biol. Sci., 7(5), 551-66.

Ensemble Learning-based Prediction of Drug-pathway Interactions based on Features Integration

123

Cortes, C. and Vapnik, V., 1995. Support vector machine.

Machine Learning, 20(3), 273-297.

Davis, A.P., Grondin, C.J., Lennon-Hopkins, K., Saraceni-

Richards, C., Sciaky, D., King, B.L., et al. 2015. The

Comparative Toxicogenomics Database's 10th year

anniversary: update 2015. Nucleic Acids Res., 43,

D914-20.

Dietterich, T.G., 2000. Ensemble methods in machine

learning. In: Multiple classifier systems. Springer

Berlin Heidelberg, 1-15.

Freund, Y. and Schapire, R. E., 1997. A decision-theoretic

generalization of on-line learning and an application to

boosting. J Comput. Syst. Sci., 55(1), 119-139.

Friedl, M.A. and Brodley, C.E., 1997. Decision tree

classification of land cover from remotely sensed data.

Remote Sens. Environ., 61(3), 399-409.

Hall, M., Frank, E., Holmes, G., Pfahringer, B.,

Reutemann, P. and Witten, I. H., 2009. The WEKA

data mining software: an update. ACM SIGKDD

explorations newsletter, 11(1), 10-18.

Ho, T.K., 1998. The random subspace method for

constructing decision forests. IEEE Transactions on

Pattern Analysis and Machine Intelligence, 20(8),

832-844.

Hopkins, A.L., 2008. Network pharmacology: the next

paradigm in drug discovery. Nat Chem Biol, 4(11),

682-90.

Kanehisa, M., Goto, S., Sato, Y., Furumichi, M. and

Tanabe, M., 2012. KEGG for integration and

interpretation of large-scale molecular data sets.

Nucleic Acids Res., 40, D109-D114.

Lipkus, A.H., 1999. A proof of the triangle inequality for

the Tanimoto distance. J Math Chem., 26, 263-265.

Ma, H. and Zhao, H., 2012. iFad: an integrative factor

analysis model for drug-pathway association

inference. Bioinformatics, 28(14), 1911-1918.

Ma, H. and Zhao, H., 2012. FacPad: Bayesian sparse

factor modeling for the inference of pathways

responsive to drug treatment. Bioinformatics, 28(20),

2662-70.

Ovaska, K., Laakso, M. and Hautaniemi, S., 2008. Fast

Gene Ontology based clustering for microarray

experiments. BioData Min., 1(1), 11.

Pratanwanich, N. and Lio, P., 2014. Exploring the

complexity of pathway–drug relationships using latent

Dirichlet allocation. Comput. Biol. Chem., 53,144-152.

Reinhold, W.C., Sunshine, M., Liu, H., Varma, S., Kohn,

K.W., Morris, J., et al. 2012. CellMiner: a web-based

suite of genomic and pharmacologic tools to explore

transcript and drug patterns in the NCI-60 cell lineset.

Cancer Res., 72(14), 3499-3511.

Rish, I., 2001. An empirical study of the naive Bayes

classifier. In: IJCAI 2001 workshop on empirical

methods in artificial intelligence, Hoos, H.H. Ed., IBM

New York, 3(22), pp. 41-46.

Schwenker, F., 2013. Ensemble methods: Foundations and

algorithms. Computational Intelligence Magazine,

IEEE, 8(1), 77-79.

Shin, J.Y., Hong, S.H., Kang, B., Minai-Tehrani, A. and

Cho, M.H., 2013. Overexpression of beclin1 induced

autophagy and apoptosis in lungs of K-rasLA1 mice.

Lung Cancer, 81(3), 362-70.

Silberberg, Y., Gottlieb, A., Kupiec, M., Ruppin, E. and

Sharan, R., 2012 Large-scale elucidation of drug

response pathways in humans. J. Comput. Biol., 19(2),

163-74.

Smith, T.F., Waterman, M.S. and Burks, C., 1985. The

statistical distribution of nucleic acid similarities.

Nucleic Acids Res., 13, 645-656.

Song, M., Yan, Y. and Jiang, Z., 2014. Drug-pathway

interaction prediction via multiple feature fusion. Mol.

Biosyst., 10(11), 2907-2913.

Subramanian, A., Tamayo, P., Mootha, V.K., Mukherjee,

S., Ebert, B.L., Gillette, M.A., et al. 2005. Gene set

enrichment analysis: a knowledge-based approach for

interpreting genome-wide expression profiles. Proc.

Natl. Acad. Sci. U. S. A., 102(43), 15545-50.

Sun, Y., Lou, X. and Bao, B., 2011. A novel relief feature

selection algorithm based on mean-variance model. J

Inf Comput Sci., 8, 3921-3929.

Van, L.T., Nabuurs, S.B. and Marchiori, E., 2013.

Predicting drug-target interaction networks of human

diseases based on multiple feature information.

Pharmacogenomics, 14(14), 1701-7.

Wishart, D.S., Knox, C., Guo, A.C., Shrivastava, S.,

Hassanali, M., Stothard, P., et al. 2006. DrugBank: a

comprehensive resource for in silico drug discovery

and exploration. Nucleic Acids Res., 34, D668-72.

Xiao, G., Lu, Q., Li, C., Wang, W., Chen, Y. and Xiao, Z.,

2010. Comparative proteome analysis of human

adenocarcinoma. Med Oncol., 27(2), 346-56.

Xue, D., Lu, M., Gao, B., Qiao, X. and Zhang, Y., 2014.

Screening for transcription factors and their regulatory

small molecules involved in regulating the functions

of CL1-5 cancer cells under the effects of

macrophage-conditioned medium. Oncol. Rep., 31(3),

23-33.

Yap, C.W., 2011. PaDEL-descriptor: An open source

software to calculate molecular descriptors and

fingerprints. J. Comput. Chem., 32(7), 1466-1474.

BIOINFORMATICS 2017 - 8th International Conference on Bioinformatics Models, Methods and Algorithms

124