Although MHC-peptide binding is the most selective event in epitope presentation process, the protein fragments generated by proteasomal cleavage require to be recognized by transporter associated with antigen processing (TAP) and translocated from cytosol to endoplasmic reticulum before they can be loaded into the ligand-binding groove of MHC. In this article, we report the use of a new and powerful machine learning tool called Gaussian process (GP) to model the linear and nonlinear relationships between the sequence pattern and binding affinity of peptide to TAP, and to explain the physicochemical properties and structural implications underlying the specific recognition and association of peptide with TAP. The resulting statistics are compared systematically with those obtained by sophisticated PLS, ANN and SVM. Results show that: (i) Nonlinear methods such as the ANN and GP perform much better than the linear PLS. (ii) GP is capable of handling both linearity- and nonlinearity-hybrid relationship and thus exhibits a good performance relative to other two nonlinear methods. (iii) Investigation of the GP model shows that the P1, P2, P3 and P9 of peptide are the most important positions that dominate TAP-peptide recognition, P5 contributes slightly to the peptide binding, whereas P4, P6, P7 and P8 can only exert very limited potency on the binding. (iv) Diverse properties cast remarkable effects on the interaction between TAP and peptide. In particular, hydrophobility, electronic property and hydrogen bond contribute most significantly to the binding affinity of TAP-peptide association.
Copyright © 2011 Elsevier Ltd. All rights reserved.