a neural network based MHC Class I Binding Peptide Prediction Server

Help Document

This section of the server will provide complete information to users about the method. how users can use the server in prediction of MHC binders or CTL epitopes ?.what is the performence of method.The description of various step or aspects of the prediction is as follows:-

Introduction to Server

Sources of Quantitative matrices

Instructions for Prediction

Cut off score or Threshold value

Proteasome AND Immunoproteasome Filters

Implementation of server

Results of Prediction

Parameters for comprehensive analysis of Method

Performence of nHLAPred

Analysis of newly generated 47 quantitative matices (QM).

Analysis of ANN for 30 MHC alleles.

Analysis of hybrid method (ANN+QM) 30 MHC alleles.

Testing of ComPred(hybrid approach) on PRAME protein

Introduction to server

nHLAPred is highly accurate MHC binders prediction method for the large number of class I MHC alleles. The method also allows to identify the promiscuous MHC class I binders (peptides that can bind to large number of alleles) having proteasomal cleavage site at C-terminus.This leads to identification of MHC class I restricted T cell epitopes in an antigen sequence.The server is partitioned in two parts ComPred and ANNpred .In ComPred part the prediction is based on the hybrid appoarch of Qunatitative matrices and artificial neural network. In ANNPred the prediction is solally based on artificial neural network.

ComPred

This part of server can predict the MHC binding peptides for 67 MHC alleles.The method is systematically developed as follows:-

Firstly,a quantitative matrix (QM) based method has been developed for 47 MHC class I alleles having minimum 15 binders available in MHCBN database.Further, an artificial neural network (ANN) based method has been developed for 30 out of these 47 MHC alleles having 40 or more binders. We have combined ANN and quantitative matrix based prediction for these 30 alleles to improve the accuracy of prediction. The average accuracy of combined method was 93.6% which is 5% higher than individual methods (QM or ANN) and 14% higher than existing quantitative matrix based methods. The performance of method is evalvuated by jack-knife testing. In addition, method allows prediction of binders for 20 more MHC alleles using the quantitative matrices reported in the literature.

ANNPred

The Prediction is based on the artificial neural networks for 30 MHC alleles.The user can specify cutoff score of artificial neural networks.The peptides getting score more than cutoff score are predicted as binders whereas the peptied getting score less than cutoff score are predicted as non-binders.

The both parts of the server implements proteasomal matrices to refine the MHC binders to potential CTL epitopes.The proteasomal matrices are obtained from the work of Toe et al.,2002.These matrices facilitate the filtering of MHC binders to CTL epitopes which have proteasomal cleavge site at C terminal. The inclusion of proteasomal preferences in MHC prediction results in prediction of potential CTL epitopes. The details about the determination of proteaosmal cleavage site in antigenic protein is provided in algorithm part. The server represents MHC binding regions and proteasome cleavage sites in an antigenic sequence in user-friendly text formats . These display formats are useful in detection of the promiscuous MHC binding regions in their query sequence. These regions are suitable candidates for subunit vaccine design.

[Table of Content]

Sources of quantitative matrices

MHC Alleles	source of Matrices	Type of matrices
47 MHC alleles	Generated in this study by obtaining data from MHCBN	Addition matrices
20 MHC alleles	Obtained from BIMAS server(unpublished)	multiplication matrices

[Table of Content]

Instructions for Prediction

A sample of the submission form showing appropriate places for submitting sequence and varying other measures for running prediction.

Sequence:-This server allows the submission of sequence in any of the standard format. The user can paste plain sequence in the provided text box.The server also has the facility for uploading the local sequence files. The sequence should be as the single letter amino acid codes. All the non standard charcters like [*&^%$@#!()_+~=;'",<>?.\|} are ignored from the sequence.The minimum length of the submitted sequence should be nine otherwise server will show an warning.The warning is also displayed if the user submit input from both sources or no sequence is submitted.
Format:-The server can accept both the formatted or unformatted raw sequences.The server uses ReadSeq routine to parse the input.The user should choose whether the sequence uploaded or pasted is plain or formatted before running prediction.The results of the prediction will be wrong if the format choosen is wrong.
Allele Selection:-The user can vary the number of the Allelles before running prediction.The server allows the selection of multiple alleles through Alt or Meta key.The selection of the multiple alleles is crucial for prediction of promiscious MHC binders or T cell epitopes. The error message displayed in case no MHC allele is selected.

[Table of Content]
Cutoff Score :-The user can choose the value of cutoff score from the list box. The peptides which get the score more than cutoff score are predicted as MHC binders otherwise peptides are predicted as non binders.In ComPred, the peptides predicted binder at a cutoff score are filtered by quantitative matrices at non-stringent threshold e.g.10% whereas the predicted non-binders are filtered by quantitative matrices in stringent conditions e.g. 1% threshold. This filteration results in reduction of false predictions.
Threshold :-The varation in the threshold will result a varation in the stringency of the prediction. As the server shows the thresold in value like 1%,2% so on. The threshold of 1% predict peptides in any given protein sequence which belong to the 1% best scoring natural peptides. The threshold correlate with the peptide score and therefore with HLA-ligand interaction. More importantly, threshold is an indicator of the likelihood that predicted peptide is capable of binding to a given HLA-molecule.Lower the threshold (= high stringency), lower the false positive rate and the higher the false negative rate.In contrast higher the threshold (= low stringency), higher the false positive rate and lower the false negative rate. In short, from the same protein sequence input, a threshold setting of 1% will predict a lower number of peptide sequences and for a lower number of HLA-II alleles, compared to 2% or higher thresholds; however, this will also ensure a higher likelihood of positive downstream experimental results. Normally, at least for a first round of screening, threshold values higher than 3% are not desirable, since the rate of false positives can increase the size of the predicted repertoire to an amount unacceptable for later experimental testing.

[Table of Content]
Proteasome AND Immunoproteasome Filters:-The proteasome play a critical role in deciding which of MHC binder act as CTL epitope.The testing of peptides binding to MHC is indispensible. But there still remain the possibilty that these peptides are not generated in vivo by proteasome.So , it is possible to narrow down the number of CTL epitopes by including the proteasomal preferences in MHC prediction. Here we have used the standard or/and immuno proteasome matrices to refine the MHC binders to CTL epitopes.The matrices are derived from the work of Toe et al.,2002.These matrices consider the 12 amino acid peptide and make a cleavage at center position that is six amino acids away from N-terminal.The user can vary the stringency of locating the proteasomal cleavage site by varying the threshold. The higher the value of threshold non-stringent will be the prediction.The lower is the value of threshold more stringent will be the prediction.The threshold values are calculated in similar manner as for quantitative matrices.
Q: How the threshold is derived for quantitative matrices of different alleles.
The threshold value is calculated as follows:
- Step 1 and 2 were repeated for each MHC allele in order to calculate threshold score at different percent for each allele used in nHLAPred.
  [Table of Content]

Implementation of Server:- nHLAPred has been installed on a Sun Server (420R) under UNIX (Solaris 7) environment and launched by using Apache web server.The CGI script of server is written in PERL and JavaScript. The server is partitioned in two major parts, ComPred and ANNPred. ComPred allows to predict binders for 67 alleles.The method predict MHC binders through combined method (ANN and QM) for 30 alleles and QM based method for 37 alleles (17 QM generated in this study and 20 obtained from literature). ANNPred allow prediction of binders for only 30 alleles purely based on the artificial neural network. The server can read input protein sequence in any of standard formats as it uses the ReadSeq (developed by Dr. Don Gilbert).

[Table of Content]

Results of Prediction:-The results of the prediction displayed in user-freindly text formats.Each of result display format firstly provides a compehensive account of length of input sequence,nonamers generated, threshold and number of selected alleles as shown below.

The starting part of each display format will specify the method on which the prediction is based.The first part will show following options.

ANNs+QM:- means prediction is based on hybrid approach of artificial neural networks and quantitative matrices.
QM:- means prediction is based on quantitative matrices.
ANNs:- means prediction is based on artificial neural networks.

The result will be displayed in these three formats

HTML I
HTML II
Tabular



where TP=True Binders. TN=True non-binders. FP=False binders. FN=False non-binders.

[Table of Content]

Performence of nHLAPred:-

The analysis of the prediction method provide confidence to the developers about their prediction method.Here we have analysed all the newly generated 47 matrices through single parameter (accuracy).Accuracy is a single value to measure the performence of prediction method.The accuracy represents the probability that a predicted binder or non-binder will be correctly classified from the total number of the binders and non-binders in the dataset.The comprehensive analysis of accuracy of method for different MHC alleles are shown in table below.The data for evalvuating the performence of method has been obtained from MHCBN database. The number of MHC binders for various MHC alleles vary from 15 to 1200 as shown in TableS1 of supplymentary material.The method is analysed through Jack-knife validation test.The value of accuracy vary from 55% to 100% at default threshold of 5% for all MHC alleles. The average value of accuracy for 47 MHC allelesis 85.48.9%.This is good accuracy for any prediction method. The performance of quantitative matrix based prediction method on 47 alleles.

[Table of Content]

Analysis of ANN for 30 MHC alleles:-
An artificial neural network based method have been developed for 30 out of above disscused 47 MHC alleles having minimum 40 MHC binders.A feed-forward backpropagation type of ANN with a single hidden layer ( 10 hidden units), 189 (21�9) input units and 1 output unit, has been used in this study. The input layer consists of 189 nodes to repersent the peptide of nine amino acids.The performence of this ANN based method is evalvuated at default cutoff score of 0.5 through Jack-Knief validation test.The value of accuracy for 30 MHC alleles vary from 67% to 96.5%. This is very good accuracy for prediction method. The average value of accuracy for 30 MHC alleles is 87.65.9% .The comprehensive analysis of ANN in term of senstivity,specificity,PPV,NPV and accuracy is shown below.
The performance of ANN based method on 30 MHC class I alleles at default cutoff score of 0.5.

[Table of Content]

Analysis of hybrid approach for 30 MHC alleles:-
It has been observed from analysis of the newly generated quantitative matrices and ANN that QM performe good for some alleles and poor for other alleles. The similar observation is also true in case of ANN.It has been observed in past by researchers that combination of two prediction approaches can give better prediction than any individual approach (Gulukota et al., 1997).Thus, we have developed a hybrid appoarch based on combination of ANN and QM.This hybrid appoarch results in improvemt in accuracy of prediction.The accuracy of the hybrid appoarch is also evalvuated through Jack-Knife validation test at default cutoff score 0.5 of ANN.The increase in accuracy for various MHC alleles vary from 0.1% to 14%.The hybrid method result in ~5% improvment in means accuracy as compared to both indivisual approaches.The comprehensive analysis of 30 MHC alleles by using hybrid appoarch is shown below.
The performance of hybrid method based on combination of ANN and QM .

The authors also compared the performence of newly generated quantitative matrices with matrices generated by Parker et al.,.The performence of 20 MHC alleles which are common in both methods has been compared. The mean accuracy of newly geneated matrices are nearly ~10% higher as shown below.We have also developed the hybrid method for 20 MHC alleles by using the matrices of Parker et al.,.1994. The hybrid appoarch which used the newly generated matrices performe better as compared to hybrid approach developed by using old matrices.The analysis of 20 alleles which are common in all the methods is shown below.
The performance of different methods that include the performance of individual methods and combination of ANN with old and new QM based methods for 20 common alleles.

[Table of Content]

Testing of ComPred( hybrid approach )on PRAME protein:-
The aim behind the development of nHLAPred is to effectively reduce number of wet lab experiments involved in the identification of potential T cell epitopes or suitable vaccine candidates.To test whether the aim is achieved or not authors applied the server on such antigenic protein whose MHC binders and CTL epitopes are experimentally proved.We have choosen the PRAME protein whose MHC binders and CTL epitopes are recently identified by Kessler et al., 2002. The sequence of PRAME antigenic protein was obtained from SWISSPROT database.

MHC binders identification:-The ComPred part of server has been used in identification of HLA-A*0201 binding 19 high affinity and 27 intermidiate affinity MHC binders of PRAME protein.The server has been used at different cutoff scores.The detailed performence of server in correctly identifying the MHC binders is shown in TableS5.The method correctly identified 73% high affinity and 81% intermediate affinity binders at default cutoff score 0.5.The default cutoff score means a value at which the senstivity and specificity are nearly equal.This proves that it is worth to use the server for MHC prediction.

[Table of Content]

HLA-A*0201 restricted potential CTL epitope identification:-Kessler et al., 2002 has demonstracted that MHC binders having proteasomal cleavage site at C terminal are mostly recognised by CTL cells.They have identified four regions from PRAME protein which have CTL epitopes. Thus, we have used this server in identification of experimentally proved antigenic regions. The server ComPred has been applied at default cutoff score of ANN and various thresholds of proteasomal filters. The server predict the MHC binders which have proteasomal cleavge sites at C terminal.The number of peptides predicted by above falls in regions identified as T cell epitopes by Kessler et al., 2001, is shown in following Table. ComPred was tested on 4 regions of PRAME (A: 90-116; B:133-159; C: 290-316; D: 415-441) which were identified as T cell epitopes by Kessler et al., (2001). The column 2 shows the number of predicted peptides and regions (in bracket), which agree with the experimentally identified epitopes.

The server is able to identify 75% of experimantally identified regions at default cutoff score of method and 7% threshold of proteasomal filters.. The server is even able to identify all the experimantally identified antigenic regions at 0.01 cutoff score and 10% threshold of standard proteasome filter.This demonstrates that it is worth to use this server for scanning of vaccine candidates from complex genomes.

[Table of Content]

Help Document

ComPred

ANNPred

HTML I

HTML II

Tabular

HTML I:-

HTML II:-

Tabular:-