Investigation into the use of support vector machine for omics applications

ABSTRACT Machine learning methods have frequently been used in early stage diagnosis at the proteomic level, such as the MHC binding peptides prediction and biomarkers selection for meta

INVESTIGATION INTO THE USE OF SUPPORT VECTOR

MACHINE FOR –OMICS APPLICATIONS

GUO YANGFAN

(B.Sc, DUT, China)

A THESIS SUBMITTED FOR THE DEGREE OF MASTERS IN SCIENCE

DEPARTMENT OF PHARMACY NATIONAL UNIVERSITY OF SINGAPORE

2011

First and foremost, I would like to express my sincere and deepest gratitude to my

supervisors, Assistant Professor Yap Chun Wei and Professor Chen Yu Zong Their

excellent guidance and invaluable advices and suggestions helped and enlightened me in

last two years studies in National University of Singapore

I am grateful to my labmates and friends for their insight suggestions and collaborations

in my research work: Ms Liew Chin Yee, Ms He Yuye, Mr Woo Sze Kwang, Mr

Bhaskaran David Prakash, and Mr Nitin Sharma from PaDEL group, Dr Zhu Feng, Dr Jia

Jia, Ms Liu Xin and Mr Zhang Jingxian from BIDD group and Dr Pasikanti Kishore

Kumar from MPRG group

Lastly, I would like to thank my parents and friends for their encouragement and

understanding It would have been impossible for me to finish this work without them

The financial support from NUS research scholarship is gratefully acknowledged

TABLE OF CONTENTS

ACKNOWLEDGMENT II TABLE OF CONTENTS III ABSTRACT V LIST OF TABLES VI LIST OF FIGURES VII LIST OF ABBREVIATIONS VIII

1 INTRODUCTION 1

1.1 Applications of SVM in bioinformatics 1

1.1.1 Applications of SVM in genomics 1

1.1.2 Applications of SVM in proteomics 3

1.1.3 Applications of SVM in metabonomics 6

1.2 Underlying difficulties in using SVM 7

1.3 Objectives and organization of this thesis 9

1.3.1 Objectives of this thesis 9

1.3.2 Organization of this thesis 13

2 METHODOLOGY 14

2.1 Support vector machines (SVMs) method 14

2.1.1 Linear SVM 14

2.1.2 Nonlinear SVM 20

2.2 Performance evaluation 22

3 MHC BINDING PREDCITION 24

3.1 Data Preparation 24

3.2 Descriptor Generation 27

3.3 Overview of SVM modeling procedure 31

3.4 Results and Performance evaluation 32

3.4.1 Self consistency testing accuracy of dataset without generated non-binders 32

3.4.2 Self consistency testing accuracy of dataset with generated non-binders 32

3.5 Summary and Discussion 36

4 METABOLITES SELECTION IN METABONOMICS 37

4.1 Data collection and normalization 37

4.2 Overview of SVM-RFE selection procedure 38

4.3 Results and Discussion 42

4.3.1 Comparison of prediction performance of multiple machine learning methods 42

4.3.2 The predictive performance of identified metabolites biomarkers 44

4.3.3 The list of selected metabolite biomarkers 49

4.3.4 Performance evaluation with multiple classifiers 58

5 CONCLUSION AND FUTURE WORK 60

BIBLIOGRAPHY 63

ABSTRACT

Machine learning methods have frequently been used in early stage diagnosis at the

proteomic level, such as the MHC binding peptides prediction and biomarkers selection

for metabonomics Although many computational methods have been designed for such

studies, it is necessary to develop more stable and smart system to improve predictive

performance Support vector machine, an artificial intelligence technique, demonstrates

remarkable generalization performance Two groups of MHC binding peptides and two

bladder cancer metabonomics datasets with different number of metabolites has been

investigated by support vector machine and other machine learning methods Recursive

feature elimination, an effective feature selection algorithm, has also been applied to

investigate the metabonomics data The results of MHC binding peptide study showed

that the prediction system can achieve satisfactory performance by constructing the

model with sufficient generated non-binding peptides The second study on

metabonomics prediction suggested that metabolites biomarkers can be effectively

selected from the metabonomics dataset by support vector machine-recursive feature

elimination method

LIST OF TABLES

Table 1 Division of amino acids for different physicochemical properties 29

Table 2 Prediction performance of MHC binding peptides without generated

non-binders 33

Table 3 Datasets and the binder and non-binder prediction accuracies for HLA alleles I

……… 34 Table 4 Prediction performance with metabolites selection for 75 BC samples with 189

metabolites by multiple machine learning methods 43

Table 5 Overall prediction accuracies of 20 times SVM-RFE selection for 75 BC

samples with 189 metabolites 45

Table 6 Selected metabolites list for 75 BC samples with 189 metabolites 46

Table 7 Overall prediction accuracies of 20 times SVM-RFE selection for 75 BC

samples with 398 metabolites 47

Table 8 Selected metabolites list for 75 BC samples with 398 metabolites 48

Table 9 List of 31 Selected metabolites (repeated rate > 80%) for 75 BC samples with

398 metabolites 50

Table 10 List of structures of the 31 Selected metabolites (repeated rate > 80%) 52

Table 11 List of evaluation performance of the 31 Selected metabolites (repeated rate >

LIST OF FIGURES

Figure 1 General pipeline of data mining and knowledge discovery in metabonomics

analysis ……… 12

Figure 2 Diagrams of the process for training and predicting targets 15

Figure 3 Architecture of support vector machines 16

Figure 4 Different hyper planes could be used to separate examples 16

Figure 5 Mapping input space to feature space 20

Figure 6 Workflow of SVM-RFE metabolites selection procedure 40

1 INTRODUCTION

Support vector machines (SVMs) are a group of supervised learning methods that can be

applied to classification or regression problems The support vector (SV) algorithm is a

nonlinear generalization of the Generalized Portrait algorithm developed in the early

60’s.1,2

In the past few decades, SVM showed excellent performance in many real-world

applications such text categorization, hand-written character recognition, image

classification and etc With the advent of the genomic, proteomic and metabonomics era,

the availability of human genome provides an opportunity to elucidate the genetic basis

of biological processes and human diseases However, the huge amount of data requires

the development of high-throughput analysis tools and powerful computational capacity

to facilitate the data analysis Facing these challenges, bioinformatics has created many

techniques, of which SVM as one of them In the following sections, the increasing

applications of SVM in bioinformatics, specifically genomics, proteomics and

metabonomics, are reviewed

1.1 Applications of SVM in bioinformatics

1.1.1 Applications of SVM in genomics

The Human Genome Project (HGP) was launched in 1989 with the initial goal of

producing a draft sequence of the human genome A working draft of genome was

announced in 2000 and completed version in 2003 But knowledge of the genomic

sequence is just the first step towards the understanding of the development and functions

of organisms The next key landmark will be an overview of the characteristics and

activities of the proteins encoded in the genes Since not all genes are expressed at the

same time, a further question is which genes are active under which circumstances One

of the immediate goals of comparative genomics is the understanding of the evolutionary

trajectories of genes and integrating them into plausible evolutionary scenarios for entire

genomes A prerequisite for this process is a phylogenetic classification of genes

The fast progress in genome sequencing projects calls for rapid, reliable and accurate

functional assignments of gene products Genome annotation3 enables the structural and

functional understanding of genome Computational analysis has been extensively

explored to perform automatic annotation to co-exist with and complement mutual

annotation The basic level of annotation is annotating genomes based on BLAST based

similarities Nowadays a lot more additional information is added to the annotation

platform including genome context information, similarity scores, experimental data and

integrations of other resources and a variety of software tools have been developed to

annotate sequences on a large scale In recent years, the application of SVMs in genome

annotation was aroused.4-8 These automated annotation systems develop binary classifiers

based on sequence data and assign these sequences to certain Gene Oncology (GO)

terms.4-8 Compared to other existing genome annotation systems, these SVMs based

annotation tools outperform to some extent with more stable prediction results and better

generalization capacity.5

With the accomplishment of HGP, genome-wide association studies (GWAS) are largely

launched to derive gene signatures to determine common and complex diseases such as

age-related macular degeneration (ARMD)9 and diabetes.10 In 2005, a GWAS found an

association between ARMD and a variation in the gene of complement factor H (CFH)

Together with four other variants, these genes can predict half the risk of ARMD between

siblings and make it the earliest and most successful example of GWAS.9 In 2007, a

GWAS found an association between type 2 diabetes (T2B) and a variation in several

single nucleotide polymorphisms (SNPs) in the genes TCF7L2, SLC30A8 and others.10

In recent years, SVMs have been applied to detect the variations associated with various

diseases Listgarten et al explored combinations of SNPs from 45 genes and detected

their potential relevance to breast cancer etiology in 174 patients and accuracy of 69%

was obtained by using SVMs as the learning algorithm.11 They concluded that multiple

SNPs from different genes over distant parts of the genome are better at identifying breast

cancer patients than any single SNP alone Waddell et al have applied SVMs to predict

the susceptibility to multiple myeloma.12 Their work had 71% accuracy on a dataset

containing 40 cases and 40 controls.12 In 2009, by using several machine learning

techniques including SVM, Uhmn et al predicted patients' susceptibility to chronic

hepatitis from SNPs.13 More recently, Ban et al investigated 408 SNPs in 87 genes

involved in major T2D related pathways in 462 T2D patients and 456 healthy controls

using SVM and achieved a 65.3% prediction rate with a combination of 14 SNPs in 12

genes.14 As the high-throughput technology for genome-wide SNPs improves, it is likely

that a much higher prediction rate with biologically more interesting combination of

SNPs can be acquired and this will further benefit future drug discovery efforts and

choosing of proper treatment strategies

1.1.2 Applications of SVM in proteomics

After genomics, proteomics is considered the next step in the study of biological systems

It is much more complicated than genomics mostly because while an organism's genome

is more or less constant, the proteome differs from cell to cell and from time to time This

is because distinct genes are expressed in distinct cell types This means that even the

basic set of proteins which are produced in a cell needs to be determined In the past, this

was done by mRNA analysis but it was found not to correlate with protein content.15,16 It

is now known that mRNA is not always translated into protein, and the amount of protein

produced for a given amount of mRNA depends on the gene it is transcribed from and on

the current physiological state of the cell Besides, not only does the translation from

mRNA cause differences, many proteins are also subjected to a wide variety of chemical

modifications after translation Many of these post-translational modifications, such as

phosphorylation, ubiquitination, methylation, acetylation, glycosylation, oxidation,

nitrosylation and etc., are critical to the protein's function

Despite the difficulties in proteomic studies, scientists are still interested in proteomics

because it gives a much better understanding of the functions of an organism than

genomics Functional clues contained in the amino acid sequence of proteins and

peptides17-20 have been extensively explored for computer prediction of protein function

and functional peptides A particular challenge is to derive functional properties from

sequences that show low or no homology to proteins of known function

Recently, SVMs have been explored for functional study of proteins and peptides by

determining whether their amino acid sequence derived properties conform to those of

known proteins of a specific functional class21-25 The advantage of this approach is that

more generalized sequence-independent characteristics can be extracted from the

sequence derived structural and physicochemical properties of the multiple samples that

share common functional profiles irrespective of sequence similarity These properties

can be used to derive classifiers19-30 for predicting other proteins that have the same

functional or interaction profiles

The task of predicting the functional class of a protein or peptide can be considered as a

two-class (positive class and negative class) classification problem for separating

members (positive class) and non-members (negative class) of a functional or interaction

class SVM and other well established two-class classification-based machine learning

methods can then be applied for developing an artificial intelligence system to classify a

new protein or peptide into the member or non-member class, which is predicted to have

a functional or interaction profile if it is classified as a member

The reported prediction accuracies for class members (P+) and non-members (P–) of

SVM for predicting protein functional classes are in the range of 25.0%~100.0% and

69.0%~100.0%, with the majority concentrated in the range of 75%~95% and

80%~99.9% respectively21-24,31-45 Based on these reported results, SVM generally shows

a certain level of capability for predicting the functional class of proteins and

protein-protein interactions In many of these reported studies, the prediction accuracy for

the non-members appears to be better than that for the members The higher prediction

accuracy for non-members likely results from the availability of more diverse set of

non-members than that of members, which enables SVM to perform a better statistical

learning for recognition of non-members

Prediction of protein-binding peptides have primarily been focused on MHC-binding

peptides,27 the reported P+ and P– values for MHC binding peptides are in the range of

75.0%~99.2% and 97.5%~99.9%, with the majority concentrated in the range of

93.3%~95.0% and 99.7%~99.9% respectively.46-48 These studies have demonstrated that,

apart from the prediction of protein functional classes, SVM is equally useful for

predicting protein-binding peptides and small molecules

From the above reported results, it can be easily concluded that SVM shows promising

potential for a wide spectrum of protein and peptide classes including some of the low-

and non-homologous proteins This method can thus be explored as a potential tool to

complement alignment-based, clustering-based, and structure-based methods for

predicting protein function and interactions

1.1.3 Applications of SVM in metabonomics

Metabonomics is the comprehensive and quantitative assessment of low molecular

weight analytes (<1500Da) that define the metabolic status of an organism under a given

condition.49 In complementation with genomics and proteomics, the direct measurement

of metabolite expression is essential in the systematic understanding of biological process

Metabolomics is increasingly enjoying widespread applications in areas such as

functional genomics, identification of the onset and progression of disease,

pharmacogenomics, nutrigenomics, and system biology.50-53

Because of its sensitivity and coverage, mass spectrometry (MS) is a favorable

technology for metabolomics study One major bottleneck for current MS-based

metabolomics is the identification of metabolites To identify the correct metabolite from

a large volume of MS/MS spectra, a proper comparison or scoring scheme is needed In

machine learning, SVMs are widely considered to represent the state of the art in

classification accuracy Recently, SVMs have been applied to the supervised

classification of cancer versus control sample sets from MS data.54-63 Xue et al

investigated the serum metabolic difference between hepatocellular carcinoma (HCC)

male patients and normal male subjects by stepwise discriminant analysis (SDA) and

SVM based on gas chromatography (GC)/MS data.61 The resultant diagnostic model

could discriminate between HCC patients and normal subjects with 20-fold cross

validation classifying accuracy of 75% and error count estimate for each group of 0%.61

Henneges et al constructed breast cancer predictive models by profiling of urinary RNA

metabolites using SVM-based feature selection from data obtained from liquid

chromatography ion trap (LC-IT) MS, and had classification sensitivity and specificity of

83.5% and 90.6% respectively.63 The performance of SVM for the classification of liquid

chromatography/time-of-flight (LC/TOF) MS metabolomics data focusing on

recognizing combinations of potential metabolic ovarian cancer diagnostic biomarkers

was evaluated by Guan et al.54 The classification of the serum sample test set was 90%

accurate, which suggests that the developed approach might lead to the development of

an accurate and reliable metabolomics-based approach for detecting ovarian cancer.54

More recently, Zhou et al collected MS/MS spectra for 21 metabolites from both

in-house data and publicly available data from the Human Metabolite Database (HMDB)

and utilized SVM to incorporate both peak and profile similarity measures for spectral

matching The models had accuracies and F-measure ranging from 94.6%~96.3% and

80.7%~85.1% respectively.64 By comparing the identification performance with other

algorithms (NIST, MassBank and SpectraST) and the correlation method, it was observed

that SVM can achieve 7% to 10% improvement on identification performance.64

1.2 Underlying difficulties in using SVM

The performance of SVM critically depends on the diversity of samples in a training

dataset and the appropriate representation of these samples The datasets used in many of

the reported studies are not expected to be fully representative of all of the proteins,

peptides and small molecules with and without a particular functional and interaction

profile Various degrees of inadequate sampling representation likely affect, to a certain

extent, the prediction accuracy of the developed statistical learning models SVM is not

applicable for proteins, peptides and small molecules with insufficient knowledge about

their specific functional and interaction profile Searching of the information about

proteins, peptides and small molecules known to possess a particular profile and those

that do not possess the profile is key to more extensive exploration of statistical learning

methods for facilitating the study of functional and interaction profiles

In the datasets of some of the reported studies, there appears to be an imbalance between

the number of samples having a profile and those without the profile SVM method tends

to produce feature vectors that push the hyper-plane towards the side with smaller

number of data,65 which often lead to a reduced prediction accuracy for the class with a

smaller number of samples or less diversity (usually members) than those of the other

class (usually non-members) It is however inappropriate to simply reduce the size of

non-members to artificially match that of members, since this compromises the diversity

needed to fully represent all non-members Computational methods for re-adjusting

biased shift of hyper-plane are being explored.66 Application of these methods may help

improving the prediction accuracy of SVM in the cases involving imbalanced data

While a number of descriptors have been introduced for representing proteins and

peptides,19,31,67,68 most reported studies typically use only a portion of these descriptors It

has been found that, in some cases, selection of a proper subset of descriptors is useful for

improving the performance of SVM.69-71 Therefore, there is a need to explore different

combination of descriptors and to select an optimum set of descriptors using feature

selection methods.69-71 Efforts have also been directed at the improvement of the

efficiency and speed of feature selection methods,72 which will enable a more extensive

application of feature selection methods Moreover, indiscriminate use of the existing

descriptors, particularly those of overlapping and redundant descriptors, may introduce

noise as well as extending the coverage of some aspects of these special features Thus, it

may be necessary to introduce new descriptors for the systems that have been described

by overlapping and redundant descriptors Investigations of cases of incorrectly predicted

samples have also suggested that the currently-used descriptors may not always be

sufficient for fully representing the structural and physicochemical properties of proteins,

peptides and small molecules.30,55,73 These have prompted works for developing new

descriptors.42

1.3 Objectives and organization of this thesis

1.3.1 Objectives of this thesis

The main objective of this thesis is to investigate and develop novel systems of support

vector machine for –omics application Two types of studies were included in this

investigation These are MHC binding prediction for proteomics level, and metabolites

selection for metabonomics level

The first study is to explore an improved flexible prediction system for MHC binding

prediction Generally, there are several inevitable limitations of the current prediction

systems First of all, most prediction systems were particularly designed for peptides with

fixed lengths Secondly, the dataset size of the existing systems, especially the training

dataset of non-binders are not adequate for building a reliable prediction model Thirdly,

some of the prediction systems represented peptides not by the structural and

physicochemical properties, but by sequence of peptides directly Last but not least, most

MHC binding prediction systems only cover a limited number of MHC alleles, which

leads to a lack of statistically significant number of known peptides in the commonly

studied length ranges

There are several feasible ways to alleviate the above problems These include choosing a

prediction algorithm which works for peptides with flexible lengths; representing the

peptides with sequence-derived structural and physicochemical properties; and

conducting the training data with sufficiently diverse set of non-binders All of these

improvements can be achieved in the studies by using support vector machine According

to previous studies, SVM has shown promising capability for prediction of specific

functional group of flexible lengths with sequence-derived structural and

physicochemical properties Moreover, peptides in same specific functional group are

generally diverse but share similar structural and physicochemical features To some

extents, the MHC binding peptides in specific alleles share similar characteristics, which

mean they have similar structural and physicochemical features Therefore, SVM is

expected to be a potential eligible algorithm to be applied for predicting MHC binding

and non-binding peptides

The second part of this thesis is to investigate a new approach of metabolites selection by

using support vector machine feature selection system The development of a new

approach of metabolites selection is one of the major topics in the area of data mining in

metabonomics studies It is important to find the marker metabolites responsible for

disease reaction This may help in early diagnosis and correct prediction of disease The

general workflow of data mining in metabonomics analysis can be found in Figure 1

There are two major sub-objectives for the second part of study (1) Discovery of marker

metabolites responsible for the distinction between groups of samples related to the

specific interests (2) Development the better metabolites selection methods by advanced

machine learning algorithm Compared with the traditional methods of metabolites

selection, the new approach will be derived from the strategies of gene selection in

microarray data Several feature selection methods and algorithms (e.g.: SVM recursive

feature elimination, forward/backward weighting methods based on Decision tree, Nạve

Bayes kernel function and other traditional weighting methods) will be compared to

determine their performance and usability for metabolite selection

Figure 1 General pipeline of data mining and knowledge discovery in

metabonomics analysis

1.3.2 Organization of this thesis

Chapter 1 introduces the history of SVMs and reviews their increasing applications in

bioinformatics especially in genomics, proteomics and metabolomics

Chapter 2 describes in detail the mathematical theory of SVM as a combination of two

main concepts: Maximal Margin Hyperplanes (also called Optimal Separating

Hyperplanes) and kernel functions The general criteria for evaluating the classifying

performance are also introduced

Chapter 3 elucidated the real application of SVM in MHC binding prediction Several

SVM prediction systems were developed and evaluated for the multiple MHC alleles

The accuracies of these prediction systems were validated using fivefold cross validation

Chapter 4 elaborated the application of SVM for metabolites selection in metabonomics

Urine samples of 75 subjects of bladder cancers were investigated with the methods of

metabonomics The advances of SVM system in metabolites selection were demonstrated

by comparison with several feature selection algorithms

Chapter 5 concludes the achievement and limitation of current work Future works are

also introduced in this chapter

2 METHODOLOGY

2.1 Support vector machines (SVMs) method

The process of training and using a SVM model for screening peptides based on their

physicochemical property descriptors is schematically illustrated in Figure 2 SVM is

based on the structural risk minimization principle of statistical learning theory,74-79

which consistently shows outstanding classification performance, is less penalized by

sample redundancy, and has lower risk for over-fitting.80-82

2.1.1 Linear SVM

In two-class problems, SVM aims to separate examples of two classes with the maximum

hyper plane (Figure 3) Mathematically, the data is composed of n examples of two

classes, denoted as {( ,x y1 2), , ( ,x y n n)}, where x iR N is a vector in feature space

of one class (positive examples) from those of the other one (negative examples) The

hyper plane is represented byw x b  0, where w is slope and bis bias Thus the

objective function of SVM changes to minimize Euclidean norm w 2 with following

Figure 2 Diagrams of the process for training and predicting targets

Figure 3 Architecture of support vector machines

Figure 4 Different hyper planes could be used to separate examples

According to which side those new instances locate, we can easily determine which class

they belong to So the decision function becomes f w b, ( )x sign(w x,  b)

Geometrically, all the points are divided into two regions by a hyper plane H As shown

in Figure 4, there are numerous ways through which a hyper plane can separate these

examples The objective of SVM is to choose the “optimal” hyper plane As all new

examples are supposed to be located under similar distribution as training examples, the

hyper plane should be chosen such that small shifts of data do not result in fluctuations in

prediction result Therefore, the hyper plane that separates examples of two classes

should have the largest margin, which is expected to possess the best generalization

performance Such hyper plane is called the Optimal Separating Hyper plane (OSH).83

Examples locating on the margins are called support vectors, whose presentation

determines the location of the hyper plane OSH could be thus represented by a linear

combination of support vectors The margin i( , )w b of a training point x iis defined as the distance betweenHandx i:

This optimization problem could be efficiently solved by the Lagrange method With the

introduction of Lagrangian multipliersi 0(i1, 2, , )n , one for each of the inequality constraints, we obtain the Lagrangian:

This is a Quadratic Programming (QP) problem We would have to minimize L w b P( , , )

with respect to w , band simultaneously require that the derivatives of L w b P( , , ) with

respect to the multipliers ivanish, L w b P( , , ) 0

By substituting these two equations into equation (11), the QP problem becomes the

Wolfe dual of the optimization problem:

This QP problem could be efficiently solved through several standard algorithms like

Sequential Minimization Optimization86 or decomposition algorithms.87

Oncew0and b0 are determined, the hyper plane is readily drawn The points for which

0

i

  are called support vectors, which lie on the margin88

2.1.2 Nonlinear SVM

Many real-world problems are usually too complicated to be solved with linear classifiers

With the introduction of kernel techniques, input data could be mapped to a

higher-dimension space, where a new linear classifier can be used to classify these

examples (Figure 5)

Figure 5 Mapping input space to feature space

Let  denotes an implicit mapping function from input space to feature spaceF Then all the previous equations are transformed by substituting input vector x i and inner

product ( , )x x i with ( )x i and kernel K x x( , )i respectively, where

( , )i ( )i ( )

K x x   x  x (15)

Equation (13) is then replaced by

may be infinitely dimensional, such as in the case of Gaussian kernel,89 where mapping

function cannot be explicitly represented A function could be used as a kernel function if

and only if it satisfies Merce’s condition.90

Followings are well-known kernel functions:

Polynomial k x z( , ) ( x z,  1)p

Sigmoid k x z( , )tanh( x z,  )

Radial basis function (RBF) k x z( , )exp( x z 2/ 22)

In this work, RBF kernel is used due to its many advantages demonstrated in previous

studies Different SVM models could be developed by using different  values It is thus necessary to scan a number of  values to find the best model, which is evaluated

by their performance on classification tasks Figure 1 illustrates the schematic diagrams

of the process of training and prediction of drug targets by SVM Sequence-derived

feature hi, pi, vi… represents such structural and physicochemical properties as hydrophobicity, polarizability, and volume The calculation of the structural and

physicochemical properties used for representing MHC binding peptides is described in

Chapter 3 and the Recursive Feature Elimination (RFE) method used for metabolites

prediction is introduced in Chapter 4

2.2 Performance evaluation

The performance evaluation aims to find out whether an algorithm is able to be applied to

novel data that have not been used to develop the prediction model, or measure the

generalization capacity to recognize new examples from the same data domain.91

In this study, several statistical measurements were explored, including sensitivity (SE),

specificity (SP), positive prediction value (PPV), and overall prediction accuracy (Q)

The formulas to calculate these measurements are listed as follows:

)/(

)

where TP, FN, TN, and FP represent correctly predicted positive data, positive data

incorrectly predicted as negative, correctly predicted negative data, and negative data

incorrectly predicted as positive respectively Another measurement, Matthews

correlation coefficient (MCC), was also used to evaluate the randomness of the

prediction

) )(

)(

)(

( / ) (TP TN FP FN TP FN TP FP TN FP TN FN

where MCC ranges from -1 to 1 Negative values of MCC indicate disagreement between

prediction and measurement, while positive values of MCC indicates agreement between

prediction and measurement A zero value means the prediction is no better than random

guess

3 MHC BINDING PREDCITION

This work developed several prediction systems for 22 MHC Class I and 17 MHC Class

II alleles by SVM An original dataset without the pseudo non-binding peptides has been

tested All peptide of this dataset were collected from the database The 29520 binder

peptides and 24848 non-binder peptides were collected from IEDB have been tested with

the five-fold cross validation As a comparison, serial tests were conducted based on each

allele The pseudo non-binding peptides generated from the splitting proteins have been

included in these tests Fivefold cross validation has been applied to evaluate the

performance of these prediction systems

3.1 Data Preparation

Data collection from databases

Binding peptides and non-binding peptides of 22 MHC class I and 17 MHC class II

alleles were collected from 2 databases: IEDB (Immune Epitope Database

www.immuneepitope.org/) and SYFPEITHI (www.syfpeithi.de) A total of 70692 MHC binding peptides were collected from these two databases After removing the duplicated

binders, there were 29520 peptides left 93734 MHC non-binding peptides were collected

from these two databases After removing the duplicated non-binders, there were 24848

peptides left

It had been discovered that the number of tested peptides can severely affected the

model’s prediction performance, especially when the number is less than 150 92 Thus,

only alleles with more than 150 binding peptides had been chosen to be studied in this

project, to ensure a good performance of the prediction model

There are 452, 5015, 856, 882, 796, 1176, 1134, 65, 308, 324, 226, 547, 209, 609, 517,

488, 335, 526, 454, 252, 209, 1274, 339, 288, 254, 1993, 370, 874, 270, 238, 373, 240,

221, 498, 236, 379, 150,254, 374 binders for class I and class II allele HLA-A*0101,

HLA-A*3101, HLA-A*330, HLA-A*6801, HLA-A*6802, HLA-B*0702, HLA-B*0801,

HLA-B*1501, HLA-B*3501, HLA-B*4402, HLA-A*11, HLA-A*2, HLA-DR*1,

HLA-DR*4, HLA-DR*7, HLA-DRB1*0101, HLA-DRB1*0301, HLA-DRB1*0401,

HLA-DRB3*0301, HLA-DRB4*0101, HLA-DRB5*0101 respectively The detail

information of datasets can be found in Table 3

MHC Non-binders generation

Theoretically, an n-mer peptides can lead to 20n possible combinations Compared to

these enormous combinations, the limited number of known non-binding peptides is

much smaller than the total number of the possible combinations, which cannot

sufficiently represent the entire sequence space A similar situation happened in proteins

functional families24,92 According to other researchers’ works24,92,93, additional numbers

of proteins without the specific functions can be created by grouping these pseudo

proteins into specific domain families and populating the whole protein space by

selecting representative proteins from each group of these un-functional families Such

kinds of efforts are expected to be applicable for MHC non-binders generation

In this work, the additional non-binder peptides were generated from splitting the

representative protein from each protein family The steps are outlined as below:

1) 10082 representative proteins were selected from the 10000+ protein families

respectively

2) Each selected protein has been split into small peptides with different lengths from 8

amino acids to 25 amino acids The splitting procedure is shown as below

3) The peptides were removed from the generated peptides if they were identical to the

binder peptides from the database The purpose of this step is to ensure the binding

peptides were not included in the generated dataset 472,118 peptides were removed

from the generated peptides 78,000,000 peptides were left and can been treated as

the negative dataset

4) Because the generated non-binder dataset is too large to be used in further modeling

steps, an eligible selection procedure is necessary to be applied to select the

representative negative dataset from the entire negative dataset Peptides should be

clustered into groups based on their structural and physicochemical feature space

Then the representative peptides were randomly selected from each group to form a

training set that is sufficiently diverse and broadly distributed in the feature space

However, due to the large number of generated non-binding peptides in this work, a very

long time would be needed to cluster 78,000,000 peptides into specific groups, especially

when each peptide is described using hundreds of descriptors A classical K-means

clustering method would take several months to complete the entire clustering process

Therefore, as a more simplified clustering method, randomly selection algorithm has

been applied to select specific number of peptides from each group Representative

peptide is randomly selected from each group to form the dataset which is sufficiently

diverse and equally distributed in the feature space The representative non-binders have

been equally selected from different lengths of peptides, from 8-mer to 25-mer, and

distributed into each allele group, according to a certain ratio of binders to non-binders

3.2 Descriptor Generation

Several descriptors development methods have been designed to construct the feature

space for peptides 94,95 For instance, the peptide sequence can be straightforwardly

represented by direct sequence of amino acids

In this study, as the binders and non-binders datasets were combined by flexible lengths

of peptides, the straightforward vector representation method would create different

number of descriptors for each peptide, which is not suitable for following modeling

procedures Therefore, a feature representation method with the structural and

physicochemical properties of a peptide has been developed with a well-formulated

procedure The same number of descriptors can be developed for different lengths of

peptides by this method Given the sequence of a peptide, the physical and chemical

properties, as well as the composition of every constituent amino acid can be computed

with certain formulas and then generated to be vectors These computed amino acid

properties include hydrophobicity, normalized van der Waals volume, polarity,

polarizability, charge, surface tension, secondary structure, solvent accessibility 92 and

three global composition descriptors: composition, transition and distribution

For each of the properties, amino acids can be divided into three or six groups such that

those in a particular group are regarded to have approximately the same property For

instance, charge of amino acid can be divided into three groups: positive (KR), Neutral

(ANCQGHILMFPSTWYV), and Negative (DE) Secondary structure of amino acid can

be divided into three groups: Helix (EALMQKRH), Strand (VIYCWFT), and Coil

(GNPSD) The detailed division of amino acids can be found in Table 1

The global composition of amino acids includes three descriptors: composition (C),

transition (T), and distribution (D), C represents the number of amino acids of a specific

property divided by the number of total number of amino acids in an entire peptide T is

the percent frequency of amino acids with a particular property followed by amino acid

with different properties D characters the distribution of the properties along the

sequence within which the first, 25%, 50%, 75% and 100% of the amino acids of a

particular property are located respectively

Table 1 Division of amino acids for different physicochemical properties

6 Dimensions

Property

Divisions Group

1 Group 2 Group 3 Group 4 Group 5 Group 6 Hydrophobicity

Van der Waals

volume

0~1.6 2.43~2.78 2.95~3 3.78~4.0 4.43~4.7

7 5.89~8.08 GAS CTPD NV EQIL MHK FRYW

For instance, consider a sequence KRACQTDKDLERWTS According to the charge

division in Table 1, the charge descriptor of this peptide is encoded as

n is the number of m in the encoded sequence and N is the length of this sequence

According to the example, the number of encoded class “1” is 4, “2” is 8, “3” is 3 The

composition are 4/15=26.7%, 8/15=53.4% and 3/15=20% respectively

Its transition descriptor can be calculated as