MOTIVATION: Various methods have been proposed to predict the binding affinities of peptides to Major Histocompatibility Complex class I (MHC-I) molecules based on experimental binding data. They can be classified into two groups: (1) AIB methods that assume independent contributions of all peptide positions to the binding to MHC-I molecule (e.g. scoring matrices) and (2) general methods which can take into account interactions between different positions (e.g. artificial neural networks). We aim to compare the prediction accuracies of these methods, and quantify the impact of interactions between peptide positions. RESULTS: We compared several previously published and widely used methods and discovered that the best AIB methods gave significantly better predictions than three previously published general methods, possibly due to the lack of a sufficient training data for the general methods. The best results, however, were achieved with our newly developed general method, which combined a matrix describing independent binding with pair coefficients describing pair-wise interactions between peptide positions. The pair coefficients consistently but only slightly improved prediction accuracy, and were much smaller than the matrix entries. This explains why neglecting them-as is done in AIB methods-can still lead to good predictions. AVAILABILITY: The new prediction model is implemented at http://zlab.bu.edu/SMM. The underlying matrix and pair coefficients are also available as supplementary materials.
MOTIVATION: Various methods have been proposed to predict the binding affinities of peptides to Major Histocompatibility Complex class I (MHC-I) molecules based on experimental binding data. They can be classified into two groups: (1) AIB methods that assume independent contributions of all peptide positions to the binding to MHC-I molecule (e.g. scoring matrices) and (2) general methods which can take into account interactions between different positions (e.g. artificial neural networks). We aim to compare the prediction accuracies of these methods, and quantify the impact of interactions between peptide positions. RESULTS: We compared several previously published and widely used methods and discovered that the best AIB methods gave significantly better predictions than three previously published general methods, possibly due to the lack of a sufficient training data for the general methods. The best results, however, were achieved with our newly developed general method, which combined a matrix describing independent binding with pair coefficients describing pair-wise interactions between peptide positions. The pair coefficients consistently but only slightly improved prediction accuracy, and were much smaller than the matrix entries. This explains why neglecting them-as is done in AIB methods-can still lead to good predictions. AVAILABILITY: The new prediction model is implemented at http://zlab.bu.edu/SMM. The underlying matrix and pair coefficients are also available as supplementary materials.
Authors: Alessandro Sette; John Sidney; Huynh-Hoa Bui; Marie-France del Guercio; Jeff Alexander; John Loffredo; David I Watkins; Bianca R Mothé Journal: Immunogenetics Date: 2005-03-04 Impact factor: 2.846