Self-taught learning for classification of mass spectrometry data: a case study of colorectal cancer

Abstract

Mass spectrometry is an important technique for chemical profiling and is a major tool in proteomics, a discipline interested in large-scale studies of proteins expressed by an organism. In this paper we propose using a sparse coding algorithm for classification of mass spectrometry serum protein profiles of colorectal cancer patients and healthy individuals following the so-called self-taught learning approach. Being applied to the dataset of 112 spectra of length 4731 bins, the sparse coding algorithm represents each of them by means of less then ten prototype spectra. The classification of spectra is done as in our previous study on the same dataset [ADM+09], using Support Vector Machines evaluated by means of the double cross-validation. However, the classifiers take as input not discrete wavelet coefficients but the sparse coding coefficients. Comparing the classification results with reference results, we show that providing the same total recognition rate, the sparse coding-based procedure leads to higher generalization performance. Moreover, we propose using the sparse coding coefficients for clustering of mass spectra and demonstrate that this approach allows one to highlight differences between the cancer spectra.