This article is part of the supplement: Selected articles from the Eleventh Asia Pacific Bioinformatics Conference (APBC 2013): Bioinformatics
Prediction of peptides binding to MHC class I and II alleles by temporal motif mining
1 Bioengineering Department, Sabancı University, 34956, Istanbul, Turkey
2 Department of Medicine, BIDMC Genomics Center, Harvard Medical School, Boston, MA 02215, USA
3 Department of Bioengineering, Istanbul Bilgi University, 34060, Istanbul, Turkey
BMC Bioinformatics 2013, 14(Suppl 2):S13 doi:10.1186/1471-2105-14-S2-S13Published: 21 January 2013
MHC (Major Histocompatibility Complex) is a key player in the immune response of most vertebrates. The computational prediction of whether a given antigenic peptide will bind to a specific MHC allele is important in the development of vaccines for emerging pathogens, the creation of possibilities for controlling immune response, and for the applications of immunotherapy. One of the problems that make this computational prediction difficult is the detection of the binding core region in peptides, coupled with the presence of bulges and loops causing variations in the total sequence length. Most machine learning methods require the sequences to be of the same length to successfully discover the binding motifs, ignoring the length variance in both motif mining and prediction steps. In order to overcome this limitation, we propose the use of time-based motif mining methods that work position-independently.
The prediction method was tested on a benchmark set of 28 different alleles for MHC class I and 27 different alleles for MHC class II. The obtained results are comparable to the state of the art methods for both MHC classes, surpassing the published results for some alleles. The average prediction AUC values are 0.897 for class I, and 0.858 for class II.
Temporal motif mining using partial periodic patterns can capture information about the sequences well enough to predict the binding of the peptides and is comparable to state of the art methods in the literature. Unlike neural networks or matrix based predictors, our proposed method does not depend on peptide length and can work with both short and long fragments. This advantage allows better use of the available training data and the prediction of peptides of uncommon lengths.