Application of support vector machine to prediction of cross-reactive allergens and recognition of viral sequences

Abstract

The generalization property of support vector machine and the choice of feature vectors are important in determining the success of a classifier. Described in this paper are two novel prediction methods utilizing support vector machines: 1. AllerHunter, a cross-reactive allergen prediction based on pairwise sequence similarity, and 2. VIPR, a viral sequence prediction method based on codon usage indices. AllerHunter performs significantly better (sensitivity=88.4%; specificity to allergen-like putative non-allergen=91.4%; specificity to divergent putative non-allergen=100%) than leading cross-reactive allergen prediction methods to date. VIPR is a unique tool that lacks comparison to date(sensitivity=90.5%; specificity=94.1%), which complements existing identification methods by suggesting a potential viral origin of sequences in samples.