zlzx.net), which was designed by Yu and based on MATLAB Web Server 1.2.4 (The MathWorks Inc.). ZJU-PDAS and detailed protocols have been described in our previous report . Spectra were denoised by undecimated discrete wavelet transform, based on the version 2.4 of the Rice Wavelet
Toolbox, followed by subtraction of baseline and calibration of mass. The detected peaks were filtered by S/N more than 3 and combined peaks in relative mass by 0.3%. Peaks appeared in more than 10% of spectra were defined as peaks cluster. Then AZD4547 concentration we constructed a non-linear supportive vector machine (SVM) classifier with a radial based function kernel to discriminate the different groups. Leave-one-out cross-validation approach was applied to estimate the accuracy of the classifier. This approach leaves one sample out to
find more be test set and the remaining samples as the training set. The process continues until each sample has been held in reserve one time as a test sample. Power of each peak in discriminating different groups was evaluated by the p value of Wilcoxon Rank Sum test. The top 10 peaks with the least p value were selected and randomly input into SVM in combination. The SVM model which achieved the highest Youden’s Index was determined as the final pattern and the peaks were selected as candidate biomarkers. Receiver operating curve (ROC) and survival curve was performed with SPSS package version 11.0. Results Assay reproducibility The reproducibility of the proteomic approach was determined by repeating one sera mixture 11 times using standard procedures
described above. The average coefficient of variance (CV) for the selected peaks with normalized intensity was 17.2% and the CV for selected peak mass was 0.03%. Biomarkers for prognosis prediction and blind test Total 50 peaks were qualified for establishing prognosis pattern by comparing proteomic spectrum of 20 good-prognosis GC patients with 19 poor-prognosis GC patients in Group 1. The established prognosis pattern Baf-A1 consisted of 5 prognosis biomarkers with peaks at 4474, 4542, 6443, 4988, 6685 Da (see Additional file 1). This prognosis pattern distinguished poor-prognosis group from good-prognosis with sensitivity of 84.2% (16/19) and specificity of 85.0% (17/20), while the sensitivity and specificity of CEA only reached 52.6 (10/19) and 70.0 (14/20) correspondingly (Table 1). Moreover, the area under ROC curve for the pattern was 0.861 (95% CI, 0.735 to 0.986), significantly higher than 0.436 (95% CI, 0.246 to 0.625) for CEA (Fig 2A). Peak at 4474 Da was found to be the most selleck chemical informative biomarker with the area under ROC curve of 0.695 (95% CI, 0.527 to 0.862), and with significantly higher expression level in poor-prognosis group (Wilcoxon Rank Sum p = 0.04, Fig 3).