Applied Bioinformatics Group


A   A   A
Sections
Home > People > Nora C Toussaint > Amino Acid String Kernels

Skip to content. | Skip to navigation

Amino Acid String Kernels

Exploiting Physico-Chemical Properties in String Kernels

Nora C. Toussaint*1, Christian Widmer2, Oliver Kohlbacher1, and Gunnar Rätsch2

1) Center for Bioinformatics, Eberhard-Karls-Universität, Sand 14, 72076 Tübingen, Germany
2) Friedrich Miescher Laboratory of the Max Planck Society, Spemannstr. 39, 72076 Tübingen, Germany
*) Corresponding author (toussaint@informatik.uni-tuebingen.de)
 
 

Background: String kernels are commonly used for the classification of biological sequences, nucleotide as well as amino acid sequences. Although string kernels are already very powerful, when it comes to amino acids they have a major short coming. They ignore an important piece of information when comparing amino acids: the physico-chemical properties such as size, hydrophobicity, or charge. This information is very valuable, especially when training data is less abundant. There have been only very few approaches so far that aim at combining these two ideas.

Results: We propose new string kernels that combine the benefits of physico-chemical descriptors for amino acids with the ones of string kernels. The benefits of the proposed kernels are assessed on two problems: MHC-peptide binding classification using position specific kernels and protein classification based on the substring spectrum of the sequences. Our experiments demonstrate that the incorporation of amino acid properties in string kernels yields improved performances compared to standard string kernels and to previously proposed non-substring kernels.

Conclusions: In summary, the proposed modifications, in particular the combination with the RBF substring kernel, consistently yield improvements without affecting the computational complexity. The proposed kernels therefore appear to be the kernels of choice for any protein sequence-based inference.

Availability: Data sets, code and additional information will be available from http://www.fml.mpg.de/raetsch/suppl/aask. The developed kernels will be part of the next Shogun toolbox release.

 

Data

 
(retrieved from http://mhcbindingpredictions.immuneepitope.org/dataset.html)
  • Protein classification data
(retrieved from http://www.kyb.tuebingen.mpg.de/bs/people/weston/semiprot/supp.html)
(retrieved from http://cbio.mskcc.org/leslielab/software/string-kernels)

Code

 
will be part of Shogun-0.9.3 (http://www.shogu-toolbox.org)