Statistical learning of peptide retention behavior in chromatographic separations: a new kernel-based approach for computational proteomics