Application Of Kernel-function Stabilized Matrix Method On The CTL Epitope Prediction

Cytotoxic T-lymphocytes (CTLs) epitopes depend on processing pathway of endogenous peptides, which includes3major steps, i.e., generation, transport and presentation of antigenic peptides. Endogenous peptides are mainly produced by proteasomes; then these peptides are transported by the transporter associated with antigen processing (TAP) into the endoplasmic reticulum (ER). The newly assembled MHC I molecules in the ER select some peptides to form MHC T/peptide complexes. These MHC T/peptide complexes will be translocated onto the cell surface to activate CTLs to recognize and to kill the infected or the damaged cells. The3endogenous antigenic processing steps put most pressure on the peptide choice. Mathematical methods could be used for characterizing and simulating this specificity and selectivity. In this study, with the kernel functions incorporated into the stabilized matrix method (SMM) and into the pair-coefficients estimation between amino acids in different peptide positions, kernel-function stabilized matrix method (KSMM) has been developed and used for modeling the3endogenous antigenic processing steps.1. KSMM was tested by sets of peptides that had interacted with12different MHC I molecules and5-fold cross-validation method. KSMM obtained an averaged AUC value of0.914, which was superior to that of SYFPEITHI and BIMAS and was close to the average AUC value of0.924obtained by NetMHC3.4. Coordination of MHC I binding model and experimental data gave a good explanation of the binding specificities of HLA-A2.1molecule to the antigenic peptides.2. The3types of proteasomal degradation data, constitutive proteasome (CCP), immunoproteasome (ICP) in vitro cleavage data and MHC I ligands as well as5-fold cross-validation method were used for testing KSMM and suport vector machine (SVM). On the three test sets, KSMM had accuracies of74.3%,72.3%and83.1%respectively; SVM had accuracies of74%,72.4%and82.5%respectively. On the HLA-A2.1ligand benchmark, the combination of proteasomal cleavage models and the MHC I binding model raised the AUC values relative to the single MHC I binding model, from0.82up to0.91. Coordination of the CCP and ICP cleavage models and experimental data gave a good explanation of the selective specificities towards cleavage sites.3. KSMM was tested by660peptides with TAP affinity and5-fold cross-validation method. KSMM obtained a spearman’s relevant correlation coefficient (Rs) of0.88, which was comparable to state-of-the-art TAP efficiency models, i.e., TAPREG and CM (Rs:0.89and0.87respectively), and was superior to ADM and TAPPRED (Rs:0.74and0.67respectively). The integration of TAP efficiency model raised the AUC values of the single MHC I binding model on the test set of A3-restricted immunogenic peptides from0.798up to0.873. Coordination of the MHC I binding models, TAP efficiency models and experimental data gave a good explanation of the mutual promotion effect in the selectivity of TAP and MHC I molecules towards immunogenic peptides.For twelve MHC-class-I alleles binding models, three kinds of proteasomal digestion models and TAP efficiency models, online predictive services were provided (http://antigenic.oicp.net). On the corresponding benchmarks, KSMM had achieved excellent results. It demonstrated that KSMM could serve as a competitive bioinformatics tool.