Semisupervised multiband segmentation for predictive proteomic analysis.
We present a predictive proﬁling system for functional proteomics that includes the semisupervised selection of different peak widths along the spectra regions. When dealing with mass spectrometry data, a crucial phase is represented by the peak identiﬁcation step. In general,this goal is achieved by setting a parameter identifying peak width, as a function of the data and of the spectrometer. The classiﬁcation performance can be signiﬁcantly im- proved by varying thresholds on peak width on different bands of the analyzed spectra during the peak assignment step. In this paper we discuss a new multiresolution method, based on machine learning, by combining a proteomic preprocessing pipeline with the semisupervised procedures of BioDCV system, a complete validation setup based on the ERFE-SVM classiﬁer. We present an evaluation of the new procedure on synthetic and real proteomic data.