CTLPred: A SVM & ANN Based CTL epitope Prediction method

CTLPred Algorithm

General Information about T cell Epitopes

The new paradigm in vaccine design is emerging, following essential discoveries in immunology and development of new T-cell epitope prediction tools.T cells orchestrate the immune response against the intracellular and extracellular pathogens.T cells regulate the immune response by accelerlating and deaccelerating the antibody production. These cells themselves act as a killer for the pathogens.The T cells are not able to recognise the antigen in the native form.They recognise the processed antigen which is presented on surface of Antigen Presenting Cells (APCs).There are different pathways for processing and presentation of extracellular and interacellular pathogens.

Briefly about Antigen processing & presentation

The T cell epitopes are derived from the intracellular pathogens by firstly delivering to protein degrading machinery known as proteasome.This leads to the fragmentation of the antigenic protein.Some of these peptides are transported to endoplasmic reticulum through a special transporter known as TAP protein (transporter assiociated antigen presentation).In the last step, the peptides are loaded to the MHC molecules for presntation on the surface of APC. It is observed that only 0.05% of peptides transported to ER bind to MHC class I molecules.The MHC class I molecules have a groove cloased at both ends for peptide, so they mostly binds to 8-10 amino acid long peptides.The crucial steps for T cell epitopes or more specifically CTL epitopes identification is the production of peptide fragments by protease,transport of peptides to ER and binding of the peptides to MHC molecules. The rules for these events have been described making it possible to formulate the prediction algorithms.The data about all of these crucial steps is not present in such amount that computer can make generalised rules.Very little amount of data is available for TAP and proteasomal peptides.The more data of these crucial peptides is prerequisite because the accuracy of the prediction algorithms is solely the reflection of quality and quantity of data.

The extracellular pathogens are endocytosed by the various cells. These pathogens are fragmented in the special compartment of these cells which are acidic and rich in the proteases and acidic enviorment.The peptides generated in this compartment compete for binding to MHC molecules and HLA-DM acts to facilitate this process.MHC class II molcules transport these peptides to surface. The MHC class II binds to 11-25 amino acids long peptides.

The identification of CTL epitopes is backbone for subunit vaccine designs.In the past, large number of computational tools had been developed for identification of T cell epitopes. The methods of CTL epitope prediction can be boardly divided into direct methods and indirect methods.The direct methods are based on the information of T cell epitopes whereas indirect methods based on the information of MHC binding peptides.

[Top] [Home]

Direct methods of T cell epitope Prediction

In 1980_sIn 1980s a few direct prediction methods based on structural and sequential analysis of T cell has been developed. Delisi and Berzofsky proposed that the critical requirement of T cell epitopes is its ability to form stable amphipathic structure.Based on this proposal, two softwares AMPHI and SOHHA has been developed. The accuracy of AMPHI method is around 65% (Meister et al., 1 995).
The sequential models for T cell epitope prediction has also been developed which relies on the occurrence of motifs in the primary sequence, rather than considering the secondary structure. The motifs are residues or type of residues occurring at specific positions. In 1988, Rothbard and Taylor collected nearly 57 T cell epitopes and based on the patterns in these residues, they published a list of motifs. The proposed motifs are of 3 to 4 residues consisting of glycine followed by hydrophobic residues. Recently two computational CTL epitope tools EpiMer and OptiMer have been developed based on knowledge of MHC binding motifs. The OptiMer predicts the amphipathic segments of protein with high motif density and EpiMer locates the segments of protein with high motif density.

These direct prediction methods based on structural or sequential models has low accuracy. The main cause of low accuracy may be insufficient data and lack of specificity in T Cell receptors (TCR). Thus, the chances of false positive prediction of epitopes are quite high. To counter the limitations of these direct methods, many indirect methods based have been developed.

[Top] [Home]

Indirect methods of T cell epitope Prediction

The indirect methods predicts the peptides binding with specific MHC alleles.The prediction tools can act as filter to reduce the number of peptides to be tested through wet experimentation.The tools are mostly based on structure,binding motifs,quantitative matrices,artificial neural network.Different algorithms based on these approaches are listed in the link page of this server.All the algorithms have few advantages and disadvantages.These methods prects the T cell epitopes with fair accuracy. These indirect methods only predict whether a peptide will bind to a specific MHC allele or not. There still remains the possibility that a MHC bound peptide may not stimulate T cells.To overcome these limitations there is requirment of methods which can predict the T cellepitopes with good accuracy.So,we have developed a CTL epitope prediction method by using the information from CTL epitopes and non-epitopes data. The methods is based on elegant machine learning techniques.The detailed algorithm of method is defind below.

[Top] [Home]

Extraction and preprocessing of data of CTL epitopes and non-epitopes.

Development of ANN based prediction method.

Development of SVM based prediction method.

Consensus prediction based on SVM and ANN based methods

Combined prediction based on SVM and ANN based Prediction

Extraction of data of CTL epitopes and non-epitopes:-

The data of the CTL epitopes have been obtained from MHCBN database version 1.1., a comprehensive database of MHC binding and non-binding peptides.The MHC class I alleles mostly bind to peptides of 9 amino acids, thus we have chosen the 9 amino acid long epitopes from extracted data of CTL epitopes. The overlapping non-epitopes of 9 amino acids are obtained from the extracted data of the CTL non-epitopes. During the preprocessing of the data all the duplicate and epitopes with non-natural amino acids are removed from the dataset of CTL epitopes. The final dataset of CTL epitopes is consist of unique and natural epitopes.The removal of the duplicate epitopes helps in avoiding overtarining of the network.The data of non-epitopes are also obtained from the same database.The dataset of ovelapping 9mer peptides are obtained from extracted data.All the duplicate and peptides with non-natural amino acid are removed from dataset.The number of peptides in resultatnt non-epitope dataset is very less as compared to the number of peptides in epitope dataset.To eqalize the number of CTL epitopes and non-epitopes we have added MHC calss I non-binders to non-epitope dataset.The dataset of the unique non-binders is also prepared in this study.The final dataset for prediction is consist of nearly eaqual number of CTL epitopes and non-epitopes.

Development of ANN based prediction method:-

Input data for ANNs-: The input for artificial neural network is a single sequence in the binary representation. In this representation, each amino acid at each window position is encoded by a group of 20 input units- 20 units code for each of the possible natural amino acids at that position . In each group of 20 input units, the input corresponding to the amino acid type at that window position is set to 1 and rest all other inputs to 0. for example the alanine and glycine is represented as follows
Alanine: 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Glycine: 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
In this manner a single peptide of 9 amino acids is represented by 180 input units to artificial neural networks.

Training of Artificial Neural Networks (ANNs)-:

The neural network implementation has been achieved by using Stuttgart Neural Network Simulator, SNNS 4.2 (Zell, A. and mamier, G. 1997). A linear activation function is used. At the start of each simulation, the weights are initialized with the random value. The training is carried out using error back propagation with a Sum of Squared Error function (SSE). The magnitude of the SSE on training is monitored after each cycle. The ultimate number of cycles is determined where the network converges. The determination of ultimate numbers of cycles where the trained artificial neural networks gave maximum accuracy is achieved by spending months of computation. The accuracy of the trained network is validated by using the jack-knife validation test.

The optimized parameters are as follows:
The ultimate number of cycles:400
The linear learning rate:.01
neural network function: Standard back propagation.
Number of Hidden layers:1.
Number of nodes per hidden layer=20

In cross-validation, one breaks the training data into K subsets. A classifier is trained on the the union of k-1 of subsets, and evaluated on the kth subset. The process is repeated k times,using each of subset as the testing set once. One then combines the results from test subsets to get an overall estimate of the effectiveness of the training procedure. The most extreme and most accurate version of cross validation is leave-one-out cross validation (LOOCV), i.e. doing n fold of cross validation when there are n training examples. We have used LOOCV type of cross-validation test to evaluate the performance of method.

[Top] [Home]

Development of SVM based prediction method:-

The same dataset of CTL epitopes and non-epitopes has been used for for the training of the Support Vector machine(SVM). The Support Vector machine is universal approximators that can be used to learn a variety of representations from training samples and as such, are applicable to classification tasks and regression tasks (Vapnik,1998). The unique ability to develop models with superior generalization capabilities when the number of input features is large as compared to number of training samples is making this emerging technology the tool of choice among the various supervised learning algorithms, including ANNs. Support vector machine is a class of machine learning methods that has been recently applied for classification of microarray data, protein structure prediction and other biological problems. It has proven better as compared to neural network in the prediction of the MHC binding peptides. Here we have implemented the SVM_light package of SVM to classify the data of CTL epitopes and Non-epitopes.The SVM _light is freely available at http://svmlight.joachims.org/. So to discriminate between the data of CTL epitopes and non-epitopes we have implemented SVM_light. We spent hours of computation to optimize the method.

Choosing a kernel-: The problem of choosing a kernel in SVM is analog to the problem of choosing an architecture of a neural network. We have used the different kernel to get the best results. We have also varied the kernel specific parameters. .e.g. -d,-s. The best results were obtained by using the polynomial kernel. The functional parameter (d) of polynomial kernel was optimised to 2.0.

Trade off between training error and margin(c)-: After choosing a particular kernel the value of c parameter is also tuned. The value of c parameter is optimised to 0.1.

To improve the accuracy of prediction the ANN and SVM based prediction methods were combined.We investigated variety of techniques for combining SVM and ANN based methods.The best results were obtained by keeping the SVM as the base method and then combining it with ANN based prediction results at various cutoff scores.so we have developed thses two models.

[Top] [Home]

Consensus Prediction based on SVM and ANN prediction:-

In case of consensus prediction the peptides which are predicted epitopes by both methods are considered as epitopes otherwise they are considered non-epitopes.The consensus prediction results in increase in the specificity of the prediction as well as PPV.

SNNS	SVM	Consensus
Epitope	Epitope	Epitope
Epitope	Non-epitope	Non-epitope
Non-epitope	Epitope	Non-epitope
Non-epitope	Non-epitope	Non-epitope

[Top] [Home]

Combined Prediction based on SVM and ANN prediction:-

In combined prediction the peptides which are precdicted epitopes by either of the prediction methods(SVM or ANN) or by both methods, are considered as predicted epitopes otherwise they are considered as non-epitopes. This method will result in improvment of senstivity as well as NPV of prediction.

SNNS	SVM	Consensus
Epitope	Epitope	Epitope
Epitope	Non-epitope	epitope
Non-epitope	Epitope	epitope
Non-epitope	Non-epitope	Non-epitope

This demonstrates that ANN and SVM extract different patterns from the data.So, the consensus or combined prediction by simultaneous use of the both prediction methods is more accurate.

[Top] [Home]