Feature Selection for MicroRNA Target Prediction

Comparison of One-Class Feature Selection Methodologies

Malik Yousef

1,2

, Jens Allmer

3,4

and Waleed Khalifa

1,2

Computer Science, The College of Sakhnin, Sakhnin, 30810, Israel

The Institute of Applied Research, The Galilee Society, P.O. Box 437 ShefaAmr,20200, Israel

Molecular Biology and Genetics, Izmir Institute of Technology, Urla, Izmir, 35430, Turkey

Bionia Incorporated, IZTEKGEB A8, Urla, Izmir, 35430, Turkey

Keywords: MicroRNA Targets, One-Class, Two-Classes, Machine Learning, Feature Selection.

Abstract: Traditionally, machine learning algorithms build classification models from positive and negative examples.

Recently, one-class classification (OCC) receives increasing attention in machine learning for problems where

the negative class cannot be defined unambiguously. This is specifically problematic in bioinformatics since

for some important biological problems the target class (positive class) is easy to obtain while the negative

one cannot be measured. Artificially generating the negative class data can be based on unreliable

assumptions. Several studies have applied two-class machine learning to predict microRNAs (miRNAs) and

their target. Different approaches for the generation of an artificial negative class have been applied, but may

lead to a biased performance estimate. Feature selection has been well studied for the two–class classification

problem, while fewer methods are available for feature selection in respect to OCC. In this study, we present

a feature selection approach for applying one-class classification to the prediction of miRNA targets. A

comparison between one-class and two-class approaches is presented to highlight that their performance are

similar while one-class classification is not based on questionable artificial data for training and performance

evaluation. We further show that the feature selection method we tried works to a degree, but needs

improvement in the future. Perhaps it could be combined with other approaches.

1 INTRODUCTION

MicroRNAs (miRNAs) are short (~21 nt) nucleotide

sequences that are either co-transcribed during

transcription or are organized in intergenic regions

with their own promoters. One or more mature

miRNAs are split from ~70-100 nucleotide long pre-

miRNAs (hairpins) which consist of double-stranded

region (stem) containing one or more loops and

bulges. Interaction of a miRNA with its target

messenger RNAs (mRNAs) leads to repressing or

translation and causes mRNA degradation (Bartel et

al., 2004). It has been shown that this process depends

on binding of the miRNA to the 3’UTR of the target,

which is a region searched by most target programs.

However, recent findings (Lytle et al., 2007) suggest

that microRNAs may affect gene expression by also

binding to 5’ UTRs of mRNA.

1.1 Target Identification

Numerous computational approaches have been

proposed for the prediction of miRNA’s targets

(Yousef et al., 2009). All of these methods depend on

the parameterization of the miRNA:mRNA duplex

and related information. Currently, sequence

complementarity, thermodynamic calculations, and

evolutionary conservation between species are used

to predict the miRNA-target structure (Bartel et al.,

2004; Lai, 2004). MiRanda (John et al., 2004), for

example, uses dynamic programming to find the

optimum alignment between mature miRNAs and

their target genes. Another tool, RNAhybrid (Bartel

et al., 2004; Lai, 2004), employs the prediction of

RNA secondary structure (like the Mfold algorithm

(Zuker, 2003)) to evaluate target propensity.

TargetScanS (Lewis et al., 2003), scores target sites

based on their evolutionary conservation using

multiple genomes (e.g.: human, mouse, rat, dog, and

chicken). Similarly, PicTar (Krek et al., 2005) is

216

Yousef, M., Allmer, J. and Khalifa, W.

Feature Selection for MicroRNA Target Prediction - Comparison of One-Class Feature Selection Methodologies.

DOI: 10.5220/0005701602160225

In Proceedings of the 9th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2016) - Volume 3: BIOINFORMATICS, pages 216-225

ISBN: 978-989-758-170-0

based on a statistical method using genome-wide

alignments of related species. TargetBoost (Saetrom

et al., 2005) uses machine learning based on sequence

information to create weighted sequence motifs that

extract a profile describing binding characteristics

between miRNAs and their targets. Likewise, Sung-

Kyu et al. (Kim et al., 2005) and Yan, X., et al (Yan

et al., 2007), used machine learning algorithms (SVM

and ensemble learning, respectively) to predict

miRNA-mRNA duplexes. MicroTar (Thadani and

Tammi, 2006) is a statistical computational tool,

which in contrast to many competitors does not use

sequence homology for the prediction of miRNA’s

targets. RNA22 (Miranda et al., 2006) is a program

based on pattern discovery, scanning UTR sequences

for targets. Yousef et al. also employed machine

learning, to develop the NBmiRTar program (Yousef

et al., 2007) which does not require sequence

conservation, but generates a model from sequence

and structure information. For more information

please refer to Table 1 and Xiao et al. (Fan and

Kurgan, 2014) who conducted a comprehensive

review and assessment of existing computational

tools of microRNA targets prediction in animals.

Recently, we compared one-class and two-class

approaches (Yousef et al., 2010) and concluded that

the advantage of one class methods is that they don’t

require the generation of an arbitrary negative class.

Table 1: Summary of the computational tools used for the

prediction of microRNA targets.

Tool

Name

URL/ Reference

TargetScanS

http://genes.mit.edu/targetscan

(Lewis et al., 2005)

miRanda

http://www.microrna.org

(John et al., 2004)

PicTar

http://pictar.mdc-berlin.de/

(Krek et al., 2005)

RNAhybrid

http://bibiserv.techfak.uni-

bielefeld.de/rnahybrid

(Krüger and Rehmsmeier, 2006)

Diana-

microT

http://diana.imis.athena-

innovation.gr/DianaTools/index.php

(Kiriakidou et al., 2004)

Target

Boost

http://www.interagon.com/demo

(Saetrom et al., 2005)

Rna22

https://cm.jefferson.edu/rna22/

(Miranda et al., 2006)

MicroTar

http://tiger.dbs.nus.edu.sg/microtar/

(Thadani and Tammi, 2006)

NBmiRTar

http://wotan.wistar.upenn.edu/NBmiRTar

(Yousef et al., 2007)

miRecords

http://mirecords.umn.edu/miRecords/

(Xiao et al., 2009)

1.2 One Class Classification and

Feature Selection

Supervised learning approaches for miRNA detection

generally consider both positive and negative

examples during training, testing, and application of

the learned models. This binary (two-class) learning

approach depends on the a priori knowledge of both

classes in the examples. In contrast to binary learning

strategies, one-class classification (OCC) uses only

one class during training of the model. Anything not

belonging to the class is rejected as an outlier by the

trained model. For problems where the negative class

cannot be unambiguously defined (e.g.: miRNA

detection) one-class classification has received

increasing attention (Crammer and Chechik, 2004;

Gupta and Ghosh, 2005; Kowalczyk and Raskutti,

2002; Spinosa and Carvalho, 2005; Yousef et al.,

2010) and has been successfully applied for example

in text mining (Manevitz and Yousef, 2002),

functional Magnetic Resonance Imaging (Thirion and

Faugeras, 2004), signature verification (Koppel and

Schler, 2004), and miRNA gene and target discovery

(Yousef et al., 2008; Yousef et al., 2010). Recently,

Khan and Madden, 2014 discussed OCC and

presented a taxonomy based on the availability of

training data, OCC algorithms, and the application

domain.

Parameterization of biological information is of

prime importance for proper learning, but often an

abundance of features is derived. Therefore, feature

selection, to determine the smallest subset of

meaningful features, has to be performed. Feature

selection for classification is well studied for two–

class classification. Unfortunately, few methods are

available for feature selection for OCC. Moreover,

existing two-class feature selection methods may not

be applicable to the OCC problem because they use

two classes during feature ranking. Recently different

studies suggest new or updated methods (Jeong et al.,

2012; Lian, 2012; Lorena et al., 2014), for OCC

feature selection such as the SVDD-radius-recursive

feature elimination (Jeong et al., 2012).

Here, we present our feature selection approach

for OCC for miRNA genes and target discovery and

compare the results with two-class classification. Our

feature selection approach leads to a good

improvement of the OCC and for some cases it

reached the performance of the two-class approach.

Feature Selection for MicroRNA Target Prediction - Comparison of One-Class Feature Selection Methodologies

217

2 MATERIALS AND METHODS

2.1 MicroRNA Target Data

A collection of 326 confirmed MicroRNA targets

(human, mouse, fruit fly, worm, and virus) were

downloaded from the TarBase (Sethupathy et al.,

2006) (TarBase_V4, Tarbase flat file data as of

04/2007) web-site to serve as positive examples and

1,000 negative examples were chosen at random from

the negative class pool generated for our previous

study (Yousef et al., 2007).

2.2 Structure and Sequence Features

Target

Feature extraction was done according to (Yousef et

al., 2007). The miRNA:mRNA duplex was

partitioned into two, the seed (5’ 8 nt of the miRNA)

and out-seed (3’ remainder). 57 structural features

were extracted for these parts. Here sequence features

(words) are defined as short sequences having lengths

equal to or less than 3 which leads to 84 features in

total. The complete length of the feature vector was

thus 141 (141 = 57 + 84). Supplementary Table 1

categorizes the features and Supplementary Table 5

presents the complete list of features and their ranking

using 4 different methods (http://

bioinformatics.iyte.edu.tr/supplements/binfo2016)

2.3 Feature Selection for One-Class

Classification

Feature selection has been well studied for the two–

class classification problem, while few methods are

available for feature selection in respect to OCC.

Unfortunately, existing two-class feature selection

methods cannot be applied to feature selection for

OCC because they also use two classes for the

ranking of the features. Recently different studies

suggest novel or updated methods for feature

selection under the premise of OCC (Bailey and

Elkan, 1994; Goymer, 2006; Hall et al., 2009; Lorena

et al., 2014; Novak, 2006; Xuan et al., 2011). We

considered these methods in this study but only

compare to the Pearson method (Lorena et al., 2014)

since no significant performance difference was seen

among the suggested feature selection methods

(Supplementary File 2 in http://

bioinformatics.iyte.edu.tr/supplements/ binfo2016).

The Pearson correlation measure allows detection of

linear relation among features. The pair-wise

distances among all feature were calculated using

Pearson correlation. Features with lower correlation

were preferred during feature selection.

2.4 Zero-Norm Feature Selection

We define for each feature’s vector a zero-norm to be

the non-zero values over all positive examples. Our

approach to performing feature selection based on

zero-norm is to remove a feature whose vector values

are all zero. Moreover, we have defined a #(v) as the

number of values with non-zero value. For example

if v = (0.4, 0, 0.6, 0, 0, 0.8, 0, 1, 1.4) the value of #(v)

is 5. Furthermore, we define different levels for

thresholds of #(v) to determine the relevance of a

feature in order to remove it from the set of features

(valid for both, the positive and the negative dataset).

For example given a threshold of 3 for #(v) the feature

is not selected if #(v) is less than the threshold. We

considered the following thresholds: 3, 5, 10, 15, 20,

25, 30, and 35. Obviously, only positive data was

used for feature selection.

2.5 One-Class Classifiers

Two-class classification depends on properly

assigned examples from both the positive (miRNA)

and negative (non-miRNA) classes in order to build a

classifier that can effectively discriminate between

them. One-class classification employs only the

information of one class (target) during training of the

model which then is able to recognize the examples

belonging to that class and rejecting others as outliers.

Many one-class classification algorithms are

available and we chose three one-class algorithms for

comparison. In the following the algorithms will be

briefly described, but more information is available in

(Schölkopf et al., 2001; Tax, 2001). The LIBSVM

library (Chang and Lin, 2011) was used for the

implementation of the SVM-based (two-class)

classifiers and DDtools (Tax, 2015) were used to

implement all other selected OCCs. WEKA software

(Witten et al., 2011) was used as implementation of

the two-class classifiers enabling comparison with

existing tools like the popular SVM method

(Schölkopf et al., 1999; Vapnik, 1995). In the

following OC-Gaussian (2.5.1), OC-kMeans (2.5.2),

OC-kNN (2.5.3), and for comparison the two-class

classifiers NB (2.6.1), SVM (2.6.2), random forest

(2.6.3), and C4.5 (2.6.4) are briefly described.

2.5.1 One-Class Gaussian

This OCC algorithm (OC-Gaussian) uses a density

estimation model which is under the assumption of a

BIOINFORMATICS 2016 - 7th International Conference on Bioinformatics Models, Methods and Algorithms

218

multivariate normal distribution and that the

probability density function can be calculated for a

given test sample in n-dimensional space and

compared to the training sample distribution (Yousef

et al., 2007). Here we use g to depict the density.

2.5.2 One-Class kMeans

Kmeans is a well-known clustering algorithm which

can partition data into k clusters. Using OC-kMeans

we divide the data into k clusters. For an unknown

sample z the distance d(z) is calculated to all k

clusters. Generally, the class is assigned by returning

the label of the closest cluster. In this case learned

clusters are from the target class and thus if the

unknown example is closer to the clusters than a

threshold, they are assigned the target class or

otherwise receive the label ‘unknown’.

2.5.3 One-Class k-Nearest Neighbour

As a modification of the two-class nearest neighbour

classifier the one-class nearest neighbour classifier

(OC-kNN) learns from positive examples only. OCC-

kNN stores all positive training examples as its

model. When classifying an unknown example z, the

distance to its nearest neighbour y (y = NN(z)) is

calculated as d(z,y). In case the distance to y is smaller

than to any of y’s nearest neighbours, the example is

classified as y. Here we consider the average distance

of the k nearest neighbours in the OC-kNN

implementation.

2.6 Two Class Methods

Two-class classification methods were selected based

on their popularity in bioinformatics. Currently, we

see a rise of the use of random forest-based machine

learning can be seen while a decrease for the use of

simple decision trees (e.g.: C4.5) is apparent.

2.6.1 Naïve Bayes

Naïve Bayes is a classification algorithm based on

posterior probabilities (Mitchell, 1997) and can,

therefore, provide a probability for the membership

of an unknown example. Important is the assumption

that the features are conditionally independent given

the class which may not hold in this case.

We used the Rainbow program (McCallum,

1996) to train the naïve Bayes classifier. To combine

the numeric features identified in the miRNA-target

duplex with the sequence features (words) in the

target candidate sequence, a dictionary of all the

unique words was generated and the frequency of

each word in the sequence was used.

2.6.2 Support Vector Machines (SVMs)

Support Vector Machines (SVMs) have been

employed in bioinformatics (Donaldson et al., 2003;

Haussler, 1999; Pavlidis et al., 2001). Linear SVMs

are usually defined as SVMs with linear kernel. The

training data for linear SVMs could be linear non-

separable and then soft-margin SVM could be

applied. Linear SVM separates the two classes in the

training data by producing the optimal separating

hyper-plane with a maximal margin between the class

1 and class 2 samples; given a proper training set.

2.6.3 Random Forest

Random forests are an ensemble of tree predictors.

Each tree depends on the values of a random vector

sampled independently for all trees in the forest

assuring same distribution for all trees (Breiman,

2001). The improvement in the classification

accuracy is due to the growing or an ensemble of tress

that vote for the most popular class. Random forests

are becoming increasingly popular because of their

ability to deal with small sample size and high-

dimensional space.

2.6.4 C4.5

C4.5 is a decision tree algorithm, developed by

(Quinlan, 1993). A decision tree is a simple structure

where non-terminal nodes represent tests on one or

more attributes and terminal nodes (leaves) reflect

decision outcomes.

2.7 Classification Performance

Evaluation

To evaluate classification performance, we used the

data generated from the positive class and 1,000

negative examples. The negative class is not used for

training of the one-class classifiers, but merely for

estimating the classification specificity.

Each one-class algorithm was trained using 90%

of the positive class and the remaining 10% were used

for sensitivity evaluation. The randomly selected

1,000 negative examples were used for the evaluation

of specificity. The whole process was repeated 100

times in order to evaluate the stability of the methods.

The procedure is depicted as a flowchart in Figure 1.

Feature Selection for MicroRNA Target Prediction - Comparison of One-Class Feature Selection Methodologies

219

Figure 1: Training and testing procedure for the one-class

classifiers. Competing two-class classifiers also received

negative data during training.

3 RESULTS AND DISCUSSION

3.1 Zero-Norm Feature Selection

Since the current standard approach in miRNA

prediction is using two-class classification, the OCCs

are compared to two-class classifiers.

Feature selection effectiveness is only presented

for OCC since feature selection is well established for

two-class classification. Table 2 shows the

effectiveness of OC-kMeans, OC-kNN, and OC-

Gaussian for different number of selected features

using the zero-norm feature selection method. Nine

different feature set sizes were tested with their

associated zero-norm thresholds. OC-kMeans and

OC-kNN achieved similar maximum accuracy (96.65

and 96.8, respectively) while OC-Gaussian was

somewhat less accurate (94.4). In general, accuracy

rises to the maximum and then decreases with the

number of features although some outliers can be

seen (Table 2). OC-kNN shows best performance for

the unfiltered feature set at a k of four (Table 2). The

number of clusters is generally at four, but for some

feature sets it increases to six and even ten. The

related method OC-kMeans interestingly performs

best in a range from 18-35 for k inversely related with

the number of features. The lowest accuracy for OC-

kMeans (95.68) and OC-kNN (94.8) was still better

than the best accuracy for OC-Gaussian. Perhaps, the

density estimation for OC-Gaussian is not as effective

as the clustering methods in OC-kMeans and OC-

kNN. This is also seen by the range of accuracies

achieved by these different methods.

The accuracy spread for OC-kMeans (~1) and

OC-kNN (2) is lower than OC-Gaussian (8). Given

this data and feature selection method, OC-Gaussian

doesn’t seem to be performing well. On the other

hand, OC-kMeans and OC-kNN perform well and

feature selection was most effective for OC-kMeans.

Figure 2: Classifier performance in respect to number of

selected features using Pearson feature selection.

It is always possible that several features that

individually have not much discriminative power

together are very potent at discriminating among

classes. This problem cannot be captured with the

feature selection method employed here.

Unfortunately, feature selection is NP-hard (Amaldi

and Kann, 1998) and it is not possible to test all

combination of features at all k or g. This may explain

the outliers that can be seen in Table 2.

3.2 Pearson-based Feature Selection

For comparison with our zero-norm method, we

performed feature selection results using the Pearson

approach (Figure 2). OC-kNN achieves highest

accuracy (96.82) followed by OC-kMeans (95.8) and

OC-Gaussian (92.39). The accuracy spread for the

selected features between 60 and 141 is 6 for OC-

kMeans, ~13 for OC-kNN, and ~19 for OC-Gaussian

(details in Supplementary Table 2 in

http://bioinformatics.

iyte.edu.tr/supplements/binfo2016. Most

interestingly, the Pearson method shows best

performance for OC-kNN and OC-Gaussian with all

features (141) selected. This indicates that for our

data the Pearson-base feature selection method was

not successful. For OC-kMeans some feature

selection was achieved, however, which shows that in

principle the Pearson-based method was correctly

applied. Whether this is a problem of the features, the

data, or a weakness of the feature selection

100

141

130

120

110

100

Accuracy

Number of Features

OC-Gaussaian Oc-kNN OC-kMeans

BIOINFORMATICS 2016 - 7th International Conference on Bioinformatics Models, Methods and Algorithms

220

Table 2: The accuracy performance of one-class classifiers in respect to selected features at different zero-norm thresholds.

Number of clusters (k) was tested between 1 and 150 and density (g) between 0.01 and 1. Highest accuracy is highlighted in

grey. ACC: accuracy, SE: sensitivity, SP: specificity.

Features

Zero-

norm

Thres

-hold

OC-kMeans OC-kNN OC-Gaussian

ACC SE SP

141 0 18 96.2 82.8 96.6 4 96.8 90.3 97.0 0.6 92.5 82.5 92.8

119 3 19 96.1 80.7 96.6 4 96.7 90.6 96.9 0.4 94.4 80.0 94.8

113 5 20

96.12

80.5 96.62 6 96.0 90.1 96.2 0.5 92.7 83.1 93.0

101 10 25 96.65 80.93 97.15 4 96.3 89.6 96.5 0.5 90.5 83.6 90.7

90 15 20 95.1 78.15 96.9 4 95.8 90.8 96.0 0.3 92.6 81.4 93.0

81 20 25 96.08 79.18 96.62 4 95.7 89.7 95.9 0.3 91.5 81.8 91.8

75 25 25 95.95 79.72 96.46 4 95.4 90.8 95.6 0.2 91.9 80.3 92.2

66 30 25 95.75 79.93 96.25 6 94.8 90.7 95.0 0.2 89.4 83 89.6

58 35 25 95.68 79.38 96.2 10 95.2 91.1 95.3 0.2 86.4 84.1 86.4

methodology cannot be deduced from our study, but

in the future we will investigate this issue further. For

OC-kMeans the accuracy doesn’t seem to have a

correlation with the number of features while for OC-

kNN and OC-Gaussian there is a minimum between

the two maxima with steady decrease and increase

correlated with the number of features.

While the results that this method achieves are

somewhat comparable to our zero-norm feature

selection method, OC-kMeans achieves about 1%

less accuracy for Pearson feature selection when

compared to zero-norm feature selection. The

performance of OC-kNN is the same for both feature

selection algorithms and OC-Gaussian performs

about 2% worse when using Pearson-based feature

selection.

In summary, feature selection methods can be

effective but work differently. A combination of

methods may, therefore, be more successful than the

methods that were compared here. Compared to the

Pearson method, the zero-norm approach appeared to

be more stable.

The separation of positive versus ‘unknown’ class

is better when fewer features are used in training

(Figure 3). The linear projection in Figure 3 visually

confirms that the classes are better separated in fewer

dimensions, however, the separation could be better

and selection of proper features may improve the

situation.

Figure 3: Linear projection of high dimensional data into

two dimensions with red representing the positive and blue

representing the ‘unknown’ class. Top pane shows all

features (141), middle pane shows projection when features

are selected using a zero-norm threshold of 3 (119 features),

and the last pane shows the distribution after filtering using

a threshold of 5 (113 features).

3.3 Comparison to Two-Class

Classification

Since two-class classification is the de facto standard

for miRNA target prediction, there is a need to

compare the accuracy achieved using OCC to the one

Feature Selection for MicroRNA Target Prediction - Comparison of One-Class Feature Selection Methodologies

221

that can be reached using two-class classification. We

chose a few representative two-class classification

algorithms which are also popular in bioinformatics.

Unfortunately, any comparison between one-

class and two class classification’s effectiveness must

remain biased unless perfectly known examples are

available for both classes. For miRNA target

prediction, this is not possible since only when a

miRNA is co-expressed with its target can the effect

be measured on the protein level. This in turn means

that any genomic sequence can be a target until it was

shown that it is not via co-expression with all known

miRNAs. This is extremely difficult for any organism

and futile for any higher organisms.

Table 3: Accuracy performance of two-class classification

algorithms in respect to selected features using the zero-

norm feature selection approach. ACC: accuracy, #:

number (Further details in Supplementary Table 4,

http://bioinformatics.iyte.edu.tr/supplements/binfo2016/).

Feat

ures

Zero-

Norm

Thres-

hold

Random

Forest

LIB-

SVM kNN C4.5

# ACC ACC ACC ACC

119 3 98.53 99.30 93.33 96.38

113 5 98.51 99.27 93.19 96.4

101 10 98.63 99.36 93.38 96.31

90 15 98.61 99.46 93.96 96.49

81 20 98.78 99.45 93.98 96.44

75 25 98.71 99.48 93.99 96.13

66 30 98.71 99.48 94.83 96.83

58 35 98.79 99.52 94.77 96.4

Since truly negative data cannot be determined, it

is necessary to create an artificial negative dataset for

training and testing of the two-class classifiers. This

would pose no problem if targeting would be

understood in detail which would make machine

learning unnecessary. Thus, for all artificial negative

datasets the content of false negative data is unknown.

This causes a bias when comparing to one class

classification where the positive class is comparably

well defined.

Nonetheless, a comparison may be informative

and, therefore, we performed two-class classification

on the same data, using the same features, as we

employed for OCC. Additionally, we employ the

same zero-norm feature selection method (Table 3).

Table 3 shows the performance of the two-class

classification algorithms when feature selection is

only based on the information from the positive class

(exactly as done for OCC, above). There doesn’t

seem to be a clear influence of the feature selection

on the performance of the two-class classification

results and down to 58 features they keep their

accuracy more or less constant within a range of less

than 1% accuracy.

The best accuracy for the two-class classification

methods are achieved with fewer features, something

we would have expected to see in the OCC analysis

as well. Overall, the performance of the two-class

classification seems to be better than the OCC results.

C4.5 and kNN perform worse than OC-kNN and

equal to OC-kMEans. Random forest and SVM are up

to ~2.5% more accurate than the OCC models. This

view is likely biased since the OCC accuracy

measures were established using artificial negative

data. This will overestimate the accuracy of the two-

class classification and underestimate the accuracy of

the one-class models.

4 CONCLUSIONS

We intended to show that one-class classification can

be used for miRNA target prediction. For this we first

attempted feature selection and were partially

successful. From our data it seems clear that proper

selection of features is important and has a positive

influence on classification accuracy (Tables 2 and 3).

A number of problems complicate feature selection,

however:

1) not all features are known

2) negative data is artificial and of unknown

quality

3) feature selection is NP-hard and many

features have already been proposed.

Since feature selection methods perform differently

(Table 2 and Figure 2), we are optimistic that a

combination of feature selection methods may in the

future propose a minimal feature set with maximum

accuracy. We believe that problems 1) and 2) can be

solved, but have no hope for the third issue.

It was our aim to point out that relying on artificial

negative data may be dangerous and that OCC can

perform at a similar accuracy as two-class

classification despite biased accuracy estimates. The

performance difference among methods is only up to

~2.5% (Tables 2 and 3).

The current results show that it is possible to train

a classifier based only on positive examples yielding

a competitive performance. Moreover, using zero-

norm feature selection with the one-class approaches

is able to improve the performance and approach to

two-class performance levels. Clearly OCC is more

BIOINFORMATICS 2016 - 7th International Conference on Bioinformatics Models, Methods and Algorithms

222

sensitive to non-relevant features than two-class

classification. However, since the process of

obtaining reliable biological data that defines the

negative class is a time consuming, if not an

impossible, endeavour a successful application of

OCC can reduce this cost and provide important tools

for classification of biological data and prediction of

unknown data.

ACKNOWLEDGEMENTS

The work was supported by the Scientific and

Technological Research Council of Turkey [grant

number 113E326] to JA.

REFERENCES

Amaldi, E., and Kann, V. (1998). On the approximability

of minimizing nonzero variables or unsatisfied relations

in linear systems. Theoretical Computer Science,

209(1-2), 237–260. doi:10.1016/S0304-

3975(97)00115-1.

Bailey, T. L., and Elkan, C. (1994). Fitting a mixture model

by expectation maximization to discover motifs in

biopolymers. Proceedings / ... International

Conference on Intelligent Systems for Molecular

Biology ; ISMB. International Conference on

Intelligent Systems for Molecular Biology, 2, 28–36.

Retrieved from http://www.ncbi.nlm.nih.gov/

pubmed/7584402.

Bartel, D. P., Lee, R., and Feinbaum, R. (2004).

MicroRNAs : Genomics , Biogenesis , Mechanism , and

Function Genomics : The miRNA Genes, 116, 281–

297.

Breiman, L. (2001). Random Forests. Machine Learning,

45(1), 5–32. doi:10.1023/A:1010933404324.

Chang, C.-C., and Lin, C.-J. (2011). LIBSVM. ACM

Transactions on Intelligent Systems and Technology,

2(3), 1–27. doi:10.1145/1961189.1961199.

Crammer, K., and Chechik, G. (2004). A needle in a

haystack: local one-class optimization. In R. Greiner

and D. Schuurmans (Eds.), Proceedings of the 21st

International Conference on Machine Learning (ICML-

04). Retrieved from http://www.machinelearning.org/

proceedings/icml2004/papers/239.ps.

Donaldson, I., Martin, J., de Bruijn, B., Wolting, C., Lay,

V., Tuekam, B., … Hogue, C. W. V. (2003). PreBIND

and Textomy--mining the biomedical literature for

protein-protein interactions using a support vector

machine. BMC Bioinformatics, 4, 11.

Fan, X., and Kurgan, L. (2014). Comprehensive overview

and assessment of computational prediction of

microRNA targets in animals. Briefings in

Bioinformatics. doi:10.1093/bib/bbu044.

Goymer, P. (2006). Different treatment. Nature Reviews

Cancer, 6(2), 94–95. doi:10.1038/nrc1808.

Gupta, G., and Ghosh, J. (2005). Robust one-class

clustering using hybrid global and local search. In

Proceedings of the 22nd international conference on

Machine learning - ICML ’05 (pp. 273–280). New

York, New York, USA: ACM Press.

doi:10.1145/1102351.1102386.

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

P., and Witten, I. H. (2009). The WEKA data mining

software. ACM SIGKDD Explorations Newsletter,

11(1), 10. doi:10.1145/1656274.1656278.

Haussler, D. (1999). Convolution Kernels on Discrete

Structures. In Technical Report UCSCRL9910 UC,

23(1), 1–38. Retrieved from http://

eprints.kfupm.edu.sa/32597/

Jeong, Y.-S., Kang, I.-H., Jeong, M.-K., and Kong, D.

(2012). A New Feature Selection Method for One-Class

Classification Problems. IEEE Transactions on

Systems, Man, and Cybernetics, Part C (Applications

and Reviews), 42(6), 1500–1509.

doi:10.1109/TSMCC.2012.2196794.

John, B., Enright, A. J., Aravin, A., Tuschl, T., Sander, C.,

and Marks, D. S. (2004). Human MicroRNA targets.

PLoS Biology, 2(11), e363. doi:10.1371/

journal.pbio.0020363.

Khan, S. S., and Madden, M. G. (2014). One-class

classification: taxonomy of study and review of

techniques. The Knowledge Engineering Review,

29(03), 345–374. doi:10.1017/S026988891300043X.

Kim, S.-K., Nam, J.-W., Lee, W.-J., and Zhang, B.-T.

(2005). A Kernel Method for MicroRNA Target

Prediction Using Sensible Data and Position-Based

Features. In 2005 IEEE Symposium on Computational

Intelligence in Bioinformatics and Computational

Biology (pp. 1–7). IEEE.

doi:10.1109/CIBCB.2005.1594897.

Kiriakidou, M., Nelson, P. T., Kouranov, A., Fitziev, P.,

Bouyioukos, C., Mourelatos, Z., and Hatzigeorgiou, A.

(2004). A combined computational-experimental

approach predicts human microRNA targets. Genes and

Development, 18(10), 1165–1178. doi:10.1101/

gad.1184704.

Koppel, M., and Schler, J. (2004). Authorship verification

as a one-class classification problem. In Twenty-first

international conference on Machine learning - ICML

’04 (p. 62). New York, New York, USA, Alberta,

Canada: ACM Press. doi:10.1145/1015330.1015448.

Kowalczyk, A., and Raskutti, B. (2002). One Class SVM

for Yeast Regulation Prediction. SIGKDD

Explorations, 4(2), 99–100.

Krek, A., Grün, D., Poy, M. N., Wolf, R., Rosenberg, L.,

Epstein, E. J., … Rajewsky, N. (2005). Combinatorial

microRNA target predictions. Nature Genetics, 37(5),

495–500. doi:10.1038/ng1536.

Krüger, J., and Rehmsmeier, M. (2006). RNAhybrid:

microRNA target prediction easy, fast and flexible.

Nucleic Acids Research, 34(Web Server issue), W451–

454. doi:10.1093/nar/gkl243.

Feature Selection for MicroRNA Target Prediction - Comparison of One-Class Feature Selection Methodologies

223

Lai, E. C. (2004). Predicting and validating microRNA

targets. Genome Biology, 5(9), 115. doi:10.1186/gb-

2004-5-9-115.

Lewis, B. P., Burge, C. B., and Bartel, D. P. (2005).

Conserved seed pairing, often flanked by adenosines,

indicates that thousands of human genes are microRNA

targets. Cell, 120(1), 15–20. doi:10.1016/

j.cell.2004.12.035.

Lewis, B. P., Shih, I., Jones-Rhoades, M. W., Bartel, D. P.,

and Burge, C. B. (2003). Prediction of mammalian

microRNA targets. Cell, 115(7), 787–798. Retrieved

from http://www.ncbi.nlm.nih.gov/pubmed/14697198.

Lian, H. (2012). On feature selection with principal

component analysis for one-class SVM. Pattern

Recognition Letters, 33(9), 1027–1031.

doi:10.1016/j.patrec.2012.01.019.

Lorena, L. H. N., Carvalho, A. C. P. L. F., and Lorena, A.

C. (2014). Filter Feature Selection for One-Class

Classification. Journal of Intelligent and Robotic

Systems, 1–17. doi:10.1007/s10846-014-0101-2.

Lytle, J. R., Yario, T. A., and Steitz, J. A. (2007). Target

mRNAs are repressed as efficiently by microRNA-

binding sites in the 5’ UTR as in the 3' UTR.

Proceedings of the National Academy of Sciences of the

United States of America, 104(23), 9667–9672.

doi:10.1073/pnas.0703820104.

Manevitz, L. M., and Yousef, M. (2002). One-Class SVMs

for Document Classification. The Journal of Machine

Learning Research, 2, 139–154. Retrieved from

http://dl.acm.org/citation.cfm?id=944808.

McCallum, A. K. (1996). Bow: A toolkit for statistical

language modeling, text retrieval, classification and

clustering. Retrieved from http://www.cs.cmu.edu/

~mccallum/bow.

Miranda, K. C., Huynh, T., Tay, Y., Ang, Y.-S., Tam, W.-

L., Thomson, A. M., … Rigoutsos, I. (2006). A pattern-

based method for the identification of MicroRNA

binding sites and their corresponding heteroduplexes.

Cell, 126(6), 1203–17. doi:10.1016/j.cell.2006.07.031.

Mitchell, T. (1997). Machine Learning.

Novak, K. (2006). Taking out the trash. Nature Reviews

Cancer, 6(2), 92–92. doi:10.1038/nrc1807.

Pavlidis, P., Weston, J., Jinsong, C., and Grundy, W. N.

(2001). Gene functional classification from

heterogeneous data. In Proceedings of the Fifth

International Conference on Computational Molecular

Biology (pp. 242–248). Retrieved from

https://noble.gs.washington.edu/papers/exp-phylo.pdf.

Quinlan, J. R. (1993). C4.5: programs for machine

learning. San Francisco, CA, USA: Morgan Kaufmann

Publishers Inc.

Saetrom, O., Snøve, O., and Saetrom, P. (2005). Weighted

sequence motifs as an improved seeding step in

microRNA target prediction algorithms. RNA, 11(7),

995–1003. doi:10.1261/rna.7290705.

Schölkopf, B., Burges, C. J. C., and Smola, A. J. (1999).

Advances in Kernel Methods. Cambridge, MA: MIT

Press.

Schölkopf, B., Platt, J. C., Shawe-Taylor, J., Smola, A. J.,

and Williamson, R. C. (2001). Estimating the Support

of a High-Dimensional Distribution. Neural Comp.,

13(7), 1443–1471.

Sethupathy, P., Corda, B., and Hatzigeorgiou, A. G. (2006).

TarBase: A comprehensive database of experimentally

supported animal microRNA targets. RNA, 12(2), 192–

7. doi:10.1261/rna.2239606.

Spinosa, E. J., and Carvalho, A. C. P. L. F. de. (2005).

Support vector machines for novel class detection in

Bioinformatics. Genetics and Molecular Research

[electronic Resource] : GMR., 4(3), 608–615.

Tax, D. M. J. (2001). One-class classification. Technical

University Delft. Retrieved from ISBN: 90-75691-05-x.

Tax, D. M. J. (2015). DDtools, the Data Description

Toolbox for Matlab.

Thadani, R., and Tammi, M. T. (2006). MicroTar:

predicting microRNA targets from RNA duplexes.

BMC Bioinformatics, 7 Suppl 5, S20.

doi:10.1186/1471-2105-7-S5-S20.

Thirion, B., and Faugeras, O. (2004). Feature

characterization in fMRI data: The Information

Bottleneck approach. Medical Image Analysis, 8(4),

403–419. doi:10.1016/j.media.2004.09.001.

Vapnik, V. N. (1995). The nature of statistical learning

theory. New York, New York, USA: Springer-Verlag.

Retrieved from http://dl.acm.org/

citation.cfm?id=211359.

Witten, I. H., Frank, E., and Hall, M. A. (2011).

Introduction to Weka. In Data Mining: Practical

Machine Learning Tools and Techniques (pp. 403–

406). Elsevier. doi:10.1016/B978-0-12-374856-

0.00010-9.

Xiao, F., Zuo, Z., Cai, G., Kang, S., Gao, X., and Li, T.

(2009). miRecords: an integrated resource for

microRNA-target interactions. Nucleic Acids Research,

37(Database issue), D105–10. doi:10.1093/nar/gkn851.

Xuan, P., Guo, M., Liu, X., Huang, Y., Li, W., and Huang,

Y. (2011). PlantMiRNAPred: efficient classification of

real and pseudo plant pre-miRNAs. Bioinformatics

(Oxford, England), 27(10), 1368–76.

doi:10.1093/bioinformatics/btr153.

Yan, X., Chao, T., Tu, K., Zhang, Y., Xie, L., Gong, Y., …

Peng, X. (2007). Improving the prediction of human

microRNA target genes by using ensemble algorithm.

FEBS Letters, 581(8), 1587–93. doi:10.1016/

j.febslet.2007.03.022.

Yousef, M., Jung, S., Kossenkov, A. V, Showe, L. C., and

Showe, M. K. (2007). Naïve Bayes for microRNA

target predictions--machine learning for microRNA

targets. Bioinformatics (Oxford, England), 23(22),

2987–92. doi:10.1093/bioinformatics/btm484.

Yousef, M., Jung, S., Showe, L. C., and Showe, M. K.

(2008). Learning from positive examples when the

negative class is undetermined--microRNA gene

identification. Algorithms for Molecular Biology, 3, 2.

doi:10.1186/1748-7188-3-2.

Yousef, M., Najami, N., and Khalifa, W. (2010). A

Comparison Study Between One-Class and Two-Class

Machine Learning for MicroRNA Target Detection.

Journal of Biomedical Science and Engineering.

BIOINFORMATICS 2016 - 7th International Conference on Bioinformatics Models, Methods and Algorithms

224

Yousef, M., Showe, L., and Showe, M. (2009). A study of

microRNAs in silico and in vivo: Bioinformatics

approaches to microRNA discovery and target

identification. FEBS Journal. doi:10.1111/j.1742-

4658.2009.06933.x.

Zuker, M. (2003). Mfold web server for nucleic acid folding

and hybridization prediction. Nucleic Acids Research,

31(13), 3406–3415. doi:10.1093/nar/gkg595.

Feature Selection for MicroRNA Target Prediction - Comparison of One-Class Feature Selection Methodologies

225