Improving peak detection by Gaussian mixture modeling of mass spectral signal

In recent years mass spectrometry became the leading measurement technique in proteomics, giving the opportunity to construct many methods for detection of signal peaks, that are the most important elements of each spectrum. An efficient approach for detecting peaks is partitioning of mass spectrum into fragments and modeling each fragment separately using Gaussian mixture decomposition. The partitioning may be obtained using unique algorithm or any existing peak detection method. In this work two commonly used peak detection algorithms were examined, namely Cromwell and Mass Spec Wavelet. Additionally, a built-in algorithm was proposed. To show that Gaussian mixture modeling of mass spectrum can improve the peak detection performance obtained by using existing solutions, many synthetic spectra with different number of true peaks and real mass spectrometry data were analyzed. In synthetic data mixture modeling of mass spectra gave higher sensitivity and lower false discovery rate of peak detection than existing peak detection algorithms. In real data the coefficient of variation of estimated peak amplitude among biological replicates was reduced.