Identifying MHC Class I Epitopes by Predicting the TAP Transport Efficiency of Epitope Precursors

Abstract

We are able to make reliable predictions of the efficiency with which peptides of arbitrary lengths will be transported by TAP. The pressure exerted by TAP on Ag presentation thus can be assessed by checking to what extent MHC class I (MHC-I)-presented epitopes can be discriminated from random peptides on the basis of predicted TAP transport efficiencies alone. Best discriminations were obtained when N-terminally prolonged epitope precursor peptides were included and the contribution of the N-terminal residues to the score were down-weighted in comparison with the contribution of the C terminus. We provide evidence that two factors may account for this N-terminal down-weighting: 1) the uncertainty as to which precursors are used in vivo and 2) the coevolution in the C-terminal sequence specificities of TAP and other agents in the pathway, which may vary among the various MHC-I alleles. Combining predictions of MHC-I binding affinities with predictions of TAP transport efficiency led to an improved identification of epitopes, which was not the case when predictions of MHC-I binding affinities were combined with predictions of C-terminal cleavages made by the proteasome.

Quantitative modeling of peptide binding to TAP using support vector machine

“…Previously, only the positions P1, P2, P3, and the C-terminal end of the peptide were thought to be clearly relevant for binding to TAP. 12,22,26,28,29 We have confirmed that the C-terminal end of the peptide has the largest quantitative input to TAP binding; a model trained on this residue alone reached an R p 5 0.68 AE 0.06. Nonetheless, we have shown that the N-terminal half of the peptide has a larger contribution to TAP binding than the C-terminal half of the peptide, as judged by the predictive performance of SMVs trained on peptide fragments encompassing a varying number of N-terminal and C-terminal residues of the peptides in the DS 613 dataset (Fig.…”

“…[26][27][28][29] It is worth noting that, unlike any of the related studies, we have not only evaluated the predictive performance of our models in cross-validation experiments but have also repeated the experiments 10 times and provided confidence values (standard deviations). Moreover, we have also shown that the enhanced predictive performance obtained with the model trained on the DS 613 dataset is not related to sequence similarity redundancy (Fig.…”

“…28 In contrast, in TAPREG, we compute the TAP affinity using nine residues selected from the larger peptides-those equivalent to the 9-mer peptides used for training-as we have shown that all residues in a 9-mer peptide contribute to binding. To our knowledge, this is the first machine-learning based approach that can predict the binding affinity to TAP of peptides longer than nine residues.…”

“…29 The remaining methods have been trained on the PVE 435 dataset. 28 Briefly, Peters' et al 28 method is based on a consensus matrix (CM) that was obtained from three scoring matrices, which included a poly-Alanine derived matrix and a SMM-matrix (generated using the Stabilized Matrix Method) trained on the PVE 435 dataset. The CM method achieved a reported R p of 0.782 on the PVE 435 dataset.…”

Predicting T cell recognition of MHC class I restricted neoepitopes

“…71, 72 A BLOSUM matrix contains log-odds scores for each of the possible substitutions of the 20 amino acids, where highly conserved amino acids have the highest scores, while non-conservative substitutions have negative scores. We used the BLOSUM62 matrix, which is generated based on protein sequence alignments of 62% identity or less.…”