The MHC class I binding prediction tools can be found at http://tools.immuneepitope.org/main/html/tcell_tools.html.
A RESTful interface is also available for MHC class I and class II prediction tools. This allows users to perform predictions on the IEDB server in batch mode without having to install any software on their own systems. Additionally, users are always assured that they are using the latest version of the tools.
Peptide Binding to MHC Class I Molecules
Users can select from eight different methods for predicting class I epitopes – ANN, ARB, SMM, SMMPMBEC, Comblib_Sidney2008, Consensus, NetMHCpan, and IEDB recommended, which are described further below. A check box can be selected to show only frequently occurring alleles. This allows the selection of only those alleles that occur in at least 1% of the human population or allele frequency of 1% or higher. However, un-checking the check-box will allow selection of all the alleles and corresponding peptide lengths for a particular species. Users can also upload an allele file instead of entering allele on the page one at a time.
Artificial Neural Network
Artificial neural networks (ANN) are computer algorithms modeled after the brain. They consist of many simple processing units which are wired together in a communication network. Each unit is a simplified model of a neuron which sends off a new signal if it receives a sufficiently strong input signal from the other units to which it is connected. The strength of these connections can be varied in order for the network to perform a desired pattern of node signal activity, which is learned from a set of input training data. The training data in this case are peptide sequences with quantitative affinities for a specific MHC molecule.
Many different implementations of artificial neural networks exist. The one utilized here is described for HLA-A2 binding predictions by Nielsen et al. (Protein Science, 2003) and has been applied to a number of different alleles (http://www.cbs.dtu.dk/services/NetMHC/).
Average Relative Binding (ARB)
Average relative binding (ARB) matrix binding prediction method is based on the assumption that each residue along the peptide molecule independently contributes to binding affinity. When a residue R occurs at position i in the peptide, it is assumed to contribute a constant amount of Ri to the free energy of binding of the peptide. The effect of each of the 20 possible amino acids at each possible position along the peptide sequence, therefore, can be estimated by a matrix of coefficients. The overall binding propensity of each peptide sequence, an algorithm “score”, is calculated by multiplying the Ri coefficients. Predicted IC50 values, which provide quantitative KD (IC50) predictions, are then calculated by mathematical transformations of the algorithm scores (Bui et al., Immunogenetics 2005).
Stabilized Matrix Method (SMM)
The Stabilized Matrix Method (SMM) described by Peters and Sette (BMC Bioinformatics, 2005) can be applied to calculate matrices from quantitative affinity data of peptides binding to MHC molecules. The advantage of this method is that it suppresses the noise present in the training data, caused by the inevitable experimental error as well as the limited number of data points.
Stabilized matrix Method with a Peptide:MHC Binding Energy Covariance matrix (SMMPMBEC)
SMMPMBEC is an improved version of SMM. It is different from SMM in that it addresses sparseness of peptide sequence coverage that is often found in binding data sets by using Peptide:MHC Binding Energy Covariance (PMBEC). The PMBEC matrix was derived from experimentally determined binding affinity measurements using combinatorial peptide libraries. SMMPMBEC is described in Kim et al. BMC Bioinformatics 2009.
Scoring matrices derived from combinatorial peptide libraries (Comblib_Sidney2008)
Comblib_Sidney2008 refers to a set of predictors (i.e. scoring matrices) that were derived from binding affinity measurements of combinatorial peptide libraries against a panel of MHC alleles. This work is described in Sidney et al. Immunome Res. 2008. This class of predictors is unique in that average binding energy contribution of a given residue at a position is directly measured, without worrying about limited peptide sequence coverage.
The Consensus predictor was motivated by an idea that predictions made by consulting “consensus” of individual predictions from multiple predictors may result in improved performance over that of any individual ones. For MHC-I, a work describing an early implementation can be found in Moutaftsi M et al. Nat Biotech 2006. The methods used for Consensus are ANN, SMM, and CombLib_Sidney2008. The Consensus method uses as many of these three component methods as possible, depending on their applicability for the chosen allele and length.
NetMHCpan predicts binding of peptides to a MHC class I molecule using artificial neural networks (ANN). It predicts binding for over 1,650 alleles, including HLA-A, B, C, E, G; non-human primates; mouse; pig; and user-supplied MHC sequence. Predictions can be made for peptide sequences of 8 to 11 residues in length. The method has been trained on over 110,000 peptide/MHC interactions.
IEDB recommended is the default prediction method selection. It considers all alleles and their corresponding peptide lengths for a particular species. For each allele-length combination, consensus method is used, which includes ANN, SMM, and CombLib. If none of these methods are available for the allele, NetMHCpan is used.