Comparing SVM sequence kernels: A protein subcellular localization theme

Kernel-based machine learning algorithms are versatile tools for biological sequence data analysis. Special sequence kernels can endow Support Vector Machines with biological knowledge to perform accurate classification of diverse sequence data. The kernels relative strengths and weaknesses are difficult to evaluate on single data sets. We examine a range of recent kernels tailor-made for biological sequence data (including the Spectrum, Mismatch, Wildcard, Substitution, Local Alignment and a new Profile-based Local Alignment kernel) on a range of classification problems (protein localization in bacteria, peroxisomal protein import signals and sub-nuclear localization). The profile-based local alignment kernel ranks highest, but its computational cost is also higher than for any of the other kernels in contention. The kernels that consistently perform well and tend to produce the most distinct classifications are the Local Alignment, Substitution and Mismatch kernels, suggesting that the exploration of new problem sets should start with these three.