A data-mining approach to biomarker identification from protein profiles using discrete stationary wavelet transform

A data-mining approach to biomarker identification from protein profiles using discrete stationary wavelet transform 1 first-author 1 corresp ‡Corresponding Author ‡Corresponding Author †E-mail:Miranbmh@modares.ac.ir †E-mail:Miranbmh@modares.ac.ir Miranbmh@modares.ac.ir 2 Department of Electrical and Computer Engineering, Tarbiat Modares University, P.O. Box 14115-111, Tehran, Iran Department of Electrical and Computer Engineering, Tarbiat Modares University, P.O. Box 14115-111, Tehran, Iran Faculty of Biomedical Engineering, Amir Kabir University of Technology, P.O. Box 15875-4413, Tehran, Iran Faculty of Biomedical Engineering, Amir Kabir University of Technology, P.O. Box 15875-4413, Tehran, Iran Objective: 2 To develop a new bioinformatic tool based on a data-mining approach for extraction of the most informative proteins that could be used to find the potential biomarkers for the detection of cancer. Methods: 2 Two independent datasets from serum samples of 253 ovarian cancer and 167 breast cancer patients were used. The samples were examined by surface-enhanced laser desorption/ionization time-of-flight mass spectrometry (SELDI-TOF MS). The datasets were used to extract the informative proteins using a data-mining method in the discrete stationary wavelet transform domain. As a dimensionality reduction procedure, the hard thresholding method was applied to reduce the number of wavelet coefficients. Also, a distance measure was used to select the most discriminative coefficients. To find the potential biomarkers using the selected wavelet coefficients, we applied the inverse discrete stationary wavelet transform combined with a two-sided t -test. Results: 2 From the ovarian cancer dataset, a set of five proteins were detected as potential biomarkers that could be used to identify the cancer patients from the healthy cases with accuracy, sensitivity, and specificity of 100%. Also, from the breast cancer dataset, a set of eight proteins were found as the potential biomarkers that could separate the healthy cases from the cancer patients with accuracy of 98.26%, sensitivity of 100%, and specificity of 95.6%. Conclusion: 2 The results have shown that the new bioinformatic tool can be used in combination with the high-throughput proteomic data such as SELDI-TOF MS to find the potential biomarkers with high discriminative power. Proteomics Discrete stationary wavelet transform Data mining Feature selection Biomarker Cancer classification INTRODUCTION 1 A major problem in the treatment of cancer is the lack of a suitable technique for early diagnosis of the disease. Unfortunately, the breast and ovarian cancers are widespread within the population of women, and the early diagnosis of these cancers can greatly reduce the mortality rate (Jemal et al., 2007 2007 ). The most widely used biomarkers do not present accurate diagnosis results (Alaoui-Jamali and Xu, 2006 2006 ). Therefore, there is still a need for accurate biomarkers, including ones that can identify the ovarian and breast cancers in their early stage of development. In recent years, researchers have tried to use proteomic technologies for identifying the set of proteins or peptides that are related to the disease (Liu et al., 2002 2002 ; Ressom et al., 2005 2005 ; Bhanot et al., 2006 2006 ; Xu et al., 2006 2006 ; Shin et al., 2008 2008 ; Zhu et al., 2008 2008 ). The surface-enhanced laser desorption/ionization time-of-flight mass spectrometry (SELDI-TOF MS) is currently the most viable technique that generates the protein patterns from biological fluids such as serum, plasma, and urine (Hu et al., 2005 2005 ; Zinkin et al., 2008 2008 ). Mass spectrometry is a high-throughput tool that generates a large-scale protein profile. Due to the large number of variables and the small size of samples, the data-mining approaches are necessary to overcome the challenges such as dimensionality reduction, feature selection, and biomarker identification (Thomas et al., 2006 2006 ; Hilario and Kalousis, 2008 2008 ). In the earlier published works, SELDI-TOF MS based cancer diagnosis combined with a data-mining approach has been used to find new biomarkers with high discriminative power (Adam et al., 2002 2002 ; Petricoin et al., 2002 2002 ; Yu et al., 2005 2005 ). In our study, we have developed a data-mining approach based on discrete stationary wavelet transform (DSWT) and discriminant analysis to find highly accurate biomarkers from proteomic profiles. Our method has shown good diagnostic results in the breast and ovarian cancer datasets. MATERIALS AND METHODS 1 Data description 2 We applied our method on two publicly available proteomic datasets. These datasets, hereafter, referred to as DS1 and DS2, contained SELDI-TOF MS protein profiles of ovarian and breast cancers respectively. DS1 is freely available from proteomics databank of Food and Drug Administration of National Cancer Institute website ( http://home.ccr.cancer.gov/ncifdaproteomics/ppatterns.asp http://home.ccr.cancer.gov/ncifdaproteomics/ppatterns.asp ). DS2 is available from Bioconductor website ( http://bioconductor.org http://bioconductor.org ). DS1 consists of 253 serum spectra composed of 15 154 distinct points on the mass-to-charge ratio axis ( m / z values). DS2 consists of 167 spectra with 13 488 distinct m / z values. In these datasets, each spectrum is defined by m / z values in the range of 0~20 000 Da and corresponds to the points on the signal intensity axis representing the abundance of proteins in the serum sample. The distribution of samples for each dataset is illustrated in Table 1 1 . Datasets Normal samples Cancer samples Ovarian cancer (DS1) 91 162 Breast cancer (DS2) 77 90 Data modeling 2 From the modeling point of view, the mass spectral curve may be considered in terms of additive and independent components (Malyarenko et al., 2005 2005 ; Hilario et al., 2006 2006 ). We assumed that there were n measured spectra, each sampled in the time interval T of TOFs t j ( j =1,..., T ). The following mathematical expression can be written for the mass spectrum signal (Morris et al., 2005 2005 ): y i ( t j )= B i ( t j )+ N i S i ( t j )+ ε ij , i =1,2,..., n . In this model, the signal intensity or abundance of a molecule, y i ( t j ), refers to each distinct mass in the TOF t j . The baseline, B i ( t j ), denotes a systematic error that is mainly due to the molecules of the energy-absorbing matrix. The true signal, S i ( t j ), represents the peak profile of each molecule in the biological sample and is scaled in each spectrum by the normalization factor N i . The last term, ε ij , shows the chemical noise that is assumed to have a Gaussian distribution. Discrete stationary wavelet transform (DSWT) 2 The discrete wavelet transform (DWT) is an effective tool for dimensionality reduction and noise removal in the analysis of very high dimensionality data. In recent years, wavelets have been used for the analysis of proteomic data (Qu et al., 2003 2003 ; Vannucci et al., 2005 2005 ; Chen et al., 2007 2007 ). It has been shown that the DSWT has a good performance in finding the informative peaks from MS data (Coombes et al., 2005 2005 ). The DSWT is similar to the DWT except that the signal is never subsampled and instead filters are upsampled at each level of decomposition (Nason and Silverman, 1995 1995 ) that has a redundant effect as there is no signal down sampling but is translation-invariant. Last property of the DSWT could lead to better performance in the feature selection from original data space via the selected wavelet coefficients. Data preprocessing 2 The raw data obtained from SELDI-TOF mass spectrometer must be preprocessed before the feature selection process. The processing includes baseline removal, denoising, and normalization to reduce the systematic errors. The baseline and electrical noise components given in Eq.( 1 1 ) must be removed from the spectra. In our approach, the DSWT was used for joint baseline removal and denoising. For baseline removal, the robust baseline estimation technique was applied to the approximation coefficients (Hu et al., 2007 2007 ; Ruckstuhl et al., 2001 2001 ). The soft thresholding method was used for denoising in the wavelet domain (Donoho, 1995 1995 ). To reduce the experimental variations in the datasets, the spectral intensities of all samples were normalized according to the method described by Petricoin and Liotta ( 2004 2004 ). Dimensionality reduction 2 After applying DSWT to each spectrum, the wavelet thresholding was applied for dimensionality reduction. The wavelet shrinkage method was used to select the threshold value (Donoho and Johnstone, 1998 1998 ). After choosing the threshold θ , the coefficients whose absolute values were less than θ were set to zero and the other coefficients were retained. For each spectrum, the survived coefficients could be different in different mass spectra. The common coefficients, which survived from all the spectra, were selected for the subsequent feature extraction. The dataset can be represented by an N × M matrix D , where N is the number of samples and M is the number of m / z values. Using a voting method, a wavelet coefficient was kept if it survived in at least (1− α )× N samples, where α is a parameter in the range [0,1]. The greater α is, the more number of coefficients could be retained. Wavelet coefficients selection 2 After the thresholding step, the dimensionality of data was reduced. However, still most of the survived features were irrelevant to differentiate between cancer and normal cases. To select a subset of wavelet coefficients, it was necessary to use a distance measure that distinguished between the two groups with high discriminative power. We applied Bhattacharyya distance as the feature selection criterion (Theodoridis and Koutroumbas, 2003 2003 ). Let S denote the subset of features that has a k -dimensional Gaussian distribution. For two-class problem, the Bhattacharyya distance ( J B ) is expressed as: $$J_B (1,2;S) = \frac{1}{8}(\mu _1 - \mu _2 )^T \left( {\frac{{\Sigma _1 + \Sigma _2 }}{2}} \right)^{ - 1} (\mu _1 - \mu _2 ) + \frac{1}{2}\ln \frac{{\left| {(\Sigma _1 + \Sigma _2 )/2} \right|}}{{\sqrt {\left| {\Sigma _1 } \right|\left| {\Sigma _2 } \right|} }},$$ , where μ and Σ are the sample mean vector and covariance matrix, respectively. In practice, stepwise forward search methods were used to select the near-optimal k features defined by J B . Protein identification 2 In the MS data analysis, a set of candidate proteins will be identified to be used in the biomarker selection stage. After selecting k wavelet coefficients, we applied the inverse discrete stationary wavelet transform (IDSWT) to obtain the m / z ratio of proteins. Due to the length of mother wavelet, each coefficient could be related to some of the proteins in a mass interval. We used two-sided t -test P values to select one protein that had minimum P value in this window. Biomarker selection 2 The subset of k proteins was identified via the method described above. A recursive support vector machine (R-SVM) algorithm (Zhang et al., 2006 2006 ) was used to select the potential biomarkers that could discriminate between cancer and normal cases in the datasets successfully. To evaluate the discriminative ability of each selected feature in the training set, a 10-fold cross-validation approach was applied in the blind test set. The fold one was used as the training set in the feature selection process. RESULTS 1 To evaluate the performance of the proposed method for biomarker identification, we analyzed the datasets described in Table 1 1 . All the mass spectra were processed to remove the baseline and electrical noise according to the described procedure. In the entire preprocessing steps, the Daubechies mother wavelet was used with four vanishing moments. For discrimination purpose, training and testing sets were selected randomly for normal and cancer groups in each dataset. Due to the small number of samples in each dataset and the large number of features, we used 10-fold cross-validation to avoid any bias and error during feature selection and sample classification. Effect of wavelet thresholding 2 Using the wavelet shrinkage thresholding method and a voting procedure, the length of data was reduced to a lower dimension in the two datasets. To select the survived coefficients after thresholding, we varied the value of α from 0.1 to 0.9 with a step of 0.1, and from 0.9 to 0.99 with a step of 0.01. In each step, the accuracy of diagnosis was used as a desirable measure to choose an appropriate value for α . A plot of the accuracy versus the value of α is shown in Fig. 1 1 . We selected α =0.9 according to the first maximum point on the curve. As shown in Table 2 2 , the remaining coefficients have yet a good discrimination power compared with the complete wavelet data. We used the SVM classifier to evaluate the performance of diagnosis in the wavelet domain. 2 Dataset 2 Threshold value α =1 2 Threshold value α =0.9 n Accuracy (%) n Accuracy (%) DS1 15154 100 3888 100 DS2 13488 92.60 1923 92.18 Coefficients selection 2 Using DSWT and the proposed thresholding procedure, the dimensionality was reduced in each dataset. However, most of the survived coefficients had irrelevant discrimination performance. We applied a stepwise procedure to select the k features by maximizing a distance measure. To decide how many coefficients needed to be selected, the accuracy was used as the measure. By using 10-fold cross-validation, we selected the k variables according to the first maximum point of the accuracy in each dataset. The result is shown in Fig. 2 2 . We chose the 10 and 15 wavelet coefficients in the DS1 and DS2, respectively, corresponding to the point on the curves where the maximum value of accuracy reached. Protein identification using IDSWT 2 We applied IDSWT to each selected coefficient by discriminant analysis. Due to the order of mother wavelet (Daubechies 4 in this study), each coefficient would relate to a set of eight neighboring m / z values in the reconstructed spectra. In our data-mining approach, the two-sided t -test was used to select one protein from eight neighbor masses. Table 3 3 gives the P values of detected proteins versus m / z indices in each dataset. Fig. 3 3 shows the accuracy of identified proteins. Dataset m / z index P value DS1 1677 1.62×10 −82 1268 3.46×10 −10 5531 1.43×10 −31 1662 9.30×10 −25 2240 2.44×10 −59 1441 3.36×10 −15 2532 4.96×10 −35 2314 2.73×10 −50 2655 4.08×10 −10 1429 0.0041 DS2 1657 0.0012 373 4.47×10 −8 1569 1.03×10 −7 6812 0.0361 3258 1.15×10 −5 5380 0.0264 3281 6.76×10 −5 6531 0.0137 5028 0.0535 2036 1.81×10 −7 2037 2.09×10 −7 9108 0.0004 1863 0.0379 7984 0.0029 8521 0.0546 Selection of potential biomarkers 2 Through the 10-fold cross-validation method, we trained an SVM classifier with detected proteins listed in Table 3 3 for each dataset. The sets of five and eight proteins were finally selected as potential biomarkers by a recursive feature elimination algorithm (R-SVM) in DS1 and DS2, respectively. Table 4 4 lists the identified biomarkers for each dataset. Using the detected biomarkers, we evaluated the performance of classification in the blind test set. By two classifiers [SVM and linear discriminate analysis (LDA)], we achieved the perfect discrimination in DS1. Also, we obtained accuracy of 98.26%, sensitivity of 100%, and specificity of 95.6% for DS2. Table 5 5 shows the performance of classification. Dataset m / z index m / z value (Da) DS1 1268 139.38 2532 557.06 2240 435.85 2655 612.56 1429 177.11 DS2 1657 287.85 373 2.72 1569 256.73 3258 1151.51 3281 1168.03 2037 441.80 9108 9102.50 1863 367.37 2 Dataset 3 SVM 3 LDA Accuracy (%) Sensitivity (%) Specificity (%) Accuracy (%) Sensitivity (%) Specificity (%) DS1 100 100 100 100 100 100 DS2 98.26 100 95.6 98.26 100 95.6 In order to show the intensity differences between normal and cancer cases quantitatively, box-plots of four high-ranked selected peptides were used as shown in Fig. 4 4 . As shown, some markers have the lower height in cancer patients than normal cases and vise versa. The receiver operating characteristics (ROC) curve and the area under curve (AUC) were used to estimate the performance of each selected biomarker. By varying the decision threshold of the SVM classifier, we computed the ROC curve and AUC for biomarkers as listed in Table 4 4 . Fig. 5 5 shows the ROC curves of the selected peptides by SVM for DS2 (the AUC value is 0.9780). The AUC value is 1 for DS1 and therefore the ROC curve has not been shown. The AUC values for biomarkers listed in Table 4 4 have been plotted in Fig. 6 6 . DISCUSSION 1 Emerging advances in MS technology allow the simultaneous analysis of expression patterns for thousands of proteins in the biological specimen. In the analysis of proteomic profiles, we were faced with the high dimensionality of data and highly correlation between intensity values of mass spectrum. In addition, the appropriate processing of data could play an important role in reproducibility of results (Baggerly et al., 2004 2004 ). As it was mentioned, the DWT is an effective tool for dimensionality reduction and noise removal in the analysis of microarray and proteomic data (Vannucci et al., 2005 2005 ; Subramani et al., 2006 2006 ). The wavelets are very popular in signal processing because they are able to analyze both local and global behavior of functions. In the field of MS, the wavelet analysis could provide denoised and compressed representation of mass spectra that make the feature extraction process more efficient and accurate due to their favorable properties such as de-correlated coefficients, and a wide variety of orthogonal basis-function possibilities. In this paper, we developed a data-mining approach based on the DSWT. Due to the translation invariant property of DSWT, it shows better performance in the processing of data including the feature selection step. In our method, a voting procedure was used to reduce the dimensionality of data. The advantage of this thresholding was to keep the most significant coefficients yet achieving the dimensionality reduction. A distance measure was applied to select the survived relevant features from the thresholding stage. By IDSWT and t -test, the candidate proteins were detected from the mass spectra. The potential biomarkers were then identified by the R-SVM. To evaluate the performance of our proposed method, two independent SELDI-TOF MS datasets were analyzed to select the candidate proteins. In dataset DS1, the m / z values of five identified biomarkers were 139.38, 557.06, 435.85, 612.56, and 177.11 Da. By 10-fold cross-validation, the perfect discrimination was obtained in ovarian cancer dataset. For dataset DS2 and the m / z indices listed in Table 4 4 , the accuracy of 98.26%, sensitivity of 100%, and specificity of 95.6% were achieved in breast cancer dataset. It is worth mentioning that our approach can identified the five peptides in the range of below 700 Da for ovarian cancer and this is fairly in agreement with previously reported biomarkers for the same data (Alexe et al., 2004 2004 ; Vannucci et al., 2005 2005 ; Whelehan et al., 2006 2006 ). Also, our results have shown that the detected biomarkers were independent of the classifier used in the selection step. This is an application where the data-mining techniques can be used to identify the potential biomarkers. In conclusion, our algorithm can be used to analyze the high-throughput proteomic data for the selection of potential biomarkers with high discrimination power such as SELDI-TOF MS profiles. Our proposed method is able to identify a small subset of proteins as biomarkers in the training set that could distinguish samples in a blind test set with minimal classification error. References BL Adam , Y Qu , JW Davis , MD Ward , MA Clements , LH Cazares , OJ Semmes , PF Schellhammer , Y Yasui , Z Feng , 等 . Serum protein fingerprinting coupled with a pattern-matching algorithm distinguishes prostate cancer from benign prostate hyperplasia and healthy men . Cancer Research , 2002 . 62 3609 - 3614 . 12097261 . Serum protein fingerprinting coupled with a pattern-matching algorithm distinguishes prostate cancer from benign prostate hyperplasia and healthy men MA Alaoui-Jamali , YJ Xu . Proteomic technology for biomarker profiling in cancer: an update . Journal of Zhejiang University SCIENCE B , 2006 . 7 ( 6 ): 411 - 420 . 16625706 DOI: 10.1631/jzus.2006.B0411 http://doi.org/10.1631/jzus.2006.B0411 . Proteomic technology for biomarker profiling in cancer: an update G Alexe , S Alexe , LA Liotta , EF Petricoin , M Reiss , PL Hammer . Ovarian cancer detection by logical analysis of proteomic data . Proteomics , 2004 . 4 ( 3 ): 766 - 783 . 14997498 DOI: 10.1002/pmic.200300574 http://doi.org/10.1002/pmic.200300574 . Ovarian cancer detection by logical analysis of proteomic data KA Baggerly , JS Morris , KR Coombes . Reproducibility of SELDI-TOF protein patterns in serum: comparing datasets from different experiments . Bioinformatics , 2004 . 20 ( 5 ): 777 - 785 . 14751995 DOI: 10.1093/bioinformatics/btg484 http://doi.org/10.1093/bioinformatics/btg484 . Reproducibility of SELDI-TOF protein patterns in serum: comparing datasets from different experiments G Bhanot , G Alexe , B Venkataraghavan , AJ Levine . A robust meta-classification strategy for cancer detection from MS data . Proteomics , 2006 . 6 ( 2 ): 592 - 604 . 16342141 DOI: 10.1002/pmic.200500192 http://doi.org/10.1002/pmic.200500192 . A robust meta-classification strategy for cancer detection from MS data S Chen , D Hong , Y Shyr . Wavelet-based procedures for proteomic mass spectrometry data processing . Computational Statistics & Data Analysis , 2007 . 52 ( 1 ): 211 - 220 . DOI: 10.1016/j.csda.2007.02.022 http://doi.org/10.1016/j.csda.2007.02.022 . Wavelet-based procedures for proteomic mass spectrometry data processing KR Coombes , J Koomen , KA Baggerly , JS Morris , R Kobayashi . Improved peak detection and quantification of mass spectrometry data acquired from SELDI by denoising spectra with the undecimated discrete wavelet transform . Proteomics , 2005 . 5 ( 16 ): 4107 - 4117 . 16254928 DOI: 10.1002/pmic.200401261 http://doi.org/10.1002/pmic.200401261 . Improved peak detection and quantification of mass spectrometry data acquired from SELDI by denoising spectra with the undecimated discrete wavelet transform DL Donoho . De-noising by soft-thresholding . IEEE Transaction on Information Theory , 1995 . 41 ( 3 ): 613 - 627 . DOI: 10.1109/18.382009 http://doi.org/10.1109/18.382009 . De-noising by soft-thresholding D Donoho , L Johnstone . Minimax estimation via wavelet shrinkage . Annals of Statistics , 1998 . 26 ( 3 ): 879 - 921 . DOI: 10.1214/aos/1024691081 http://doi.org/10.1214/aos/1024691081 . Minimax estimation via wavelet shrinkage M Hilario , A Kalousis . Approaches to dimensionality reduction in proteomic biomarker studies . Briefings in Bioinformatics , 2008 . 9 ( 2 ): 102 - 118 . 18310106 DOI: 10.1093/bib/bbn005 http://doi.org/10.1093/bib/bbn005 . Approaches to dimensionality reduction in proteomic biomarker studies M Hilario , A Kalousis , C Pellegrini , M Muller . Processing and classification of protein mass spectra . Mass Spectrometry Reviews , 2006 . 25 ( 3 ): 409 - 449 . 16463283 DOI: 10.1002/mas.20072 http://doi.org/10.1002/mas.20072 . Processing and classification of protein mass spectra Y Hu , S Zhang , J Yu , J Liu , S Zheng . SELDI-TOF-MS: the proteomics and bioinformatics approaches in the diagnosis of breast cancer . The Breast , 2005 . 14 ( 4 ): 250 - 255 . 16085230 DOI: 10.1016/j.breast.2005.01.008 http://doi.org/10.1016/j.breast.2005.01.008 . SELDI-TOF-MS: the proteomics and bioinformatics approaches in the diagnosis of breast cancer Y Hu , T Jiang , A Shen , W Li , X Wang , J Hu . A background elimination method based on wavelet transform for Raman spectra . Chemometrics and Intelligent Laboratory Systems , 2007 . 85 ( 1 ): 94 - 101 . DOI: 10.1016/j.chemolab.2006.05.004 http://doi.org/10.1016/j.chemolab.2006.05.004 . A background elimination method based on wavelet transform for Raman spectra A Jemal , R Siegel , E Ward , T Murray , J Xu , MJ Thun . Cancer statistics . CA Cancer J Clin , 2007 . 57 ( 1 ): 43 - 66 . 17237035 . Cancer statistics H Liu , J Li , L Wong . A comparative study on feature selection and classification method using gene expression profiles and proteomic patterns . Genome Informatics , 2002 . 13 51 - 60 . 14571374 . A comparative study on feature selection and classification method using gene expression profiles and proteomic patterns DI Malyarenko , WE Cooke , BL Adam , G Malik , H Chen , ER Tracy , MW Trosset , M Sasinowski , OJ Semmes , DM Manos . Enhancement of sensitivity and resolution of SELDI-TOF mass spectrometric records for serum peptides using time-series analysis techniques . Clinical Chemistry , 2005 . 51 ( 1 ): 65 - 74 . 15550476 DOI: 10.1373/clinchem.2004.037283 http://doi.org/10.1373/clinchem.2004.037283 . Enhancement of sensitivity and resolution of SELDI-TOF mass spectrometric records for serum peptides using time-series analysis techniques JS Morris , KR Coombes , J Koomen , KA Baggerly , R Kobayashi . Feature extraction and quantification for mass spectrometry in biomedical applications using the mean spectrum . Bioinformatics , 2005 . 21 ( 9 ): 1764 - 1775 . 15673564 DOI: 10.1093/bioinformatics/bti254 http://doi.org/10.1093/bioinformatics/bti254 . Feature extraction and quantification for mass spectrometry in biomedical applications using the mean spectrum GP Nason , BW Silverman . The Stationary Wavelet Transforms and Statistical Applications . Lecture Notes in Statistics: Wavelets and Statistics , 1995 . Springer , 281 - 299 . . The Stationary Wavelet Transforms and Statistical Applications EFIII Petricoin , LA Liotta . SELDI-TOF-based serum proteomic pattern diagnostics for early detection of cancer . Current Opinion in Biotechnology , 2004 . 15 ( 1 ): 24 - 30 . 15102462 DOI: 10.1016/j.copbio.2004.01.005 http://doi.org/10.1016/j.copbio.2004.01.005 . SELDI-TOF-based serum proteomic pattern diagnostics for early detection of cancer EFIII Petricoin , AM Ardekani , BA Hitt , PJ Levine , VA Fusaro , SM Steinberg , GB Mills , C Simone , DA Fishman , EC Kohn , 等 . Use of proteomic patterns in serum to identify ovarian cancer . The Lancet , 2002 . 359 ( 9306 ): 572 - 577 . DOI: 10.1016/S0140-6736(02)07746-2 http://doi.org/10.1016/S0140-6736(02)07746-2 . Use of proteomic patterns in serum to identify ovarian cancer Y Qu , BL Adam , M Thornquist , JD Potter , ML Thompson , Y Yasui , JW Davis , LH Cazares , PF Schellhammer , MA Clements , 等 . Data reduction using a discrete wavelet transform in discriminant analysis of very high dimensionality data . Biometrics , 2003 . 59 ( 1 ): 143 - 151 . 12762451 DOI: 10.1111/1541-0420.00017 http://doi.org/10.1111/1541-0420.00017 . Data reduction using a discrete wavelet transform in discriminant analysis of very high dimensionality data HW Ressom , RS Varghese , M Abdel-Hamid , SAL Eissa , D Saha , L Goldman , EF Petricoin , TP Conrads , TD Veenstra , CA Loffredo , 等 . Analysis of mass spectral serum profiles for biomarker selection . Bioinformatics , 2005 . 21 ( 21 ): 4039 - 4045 . 16159919 DOI: 10.1093/bioinformatics/bti670 http://doi.org/10.1093/bioinformatics/bti670 . Analysis of mass spectral serum profiles for biomarker selection AF Ruckstuhl , MP Jacobson , RW Field , JA Dodd . Baseline subtraction using robust local regression estimation . Journal of Quantitative Spectroscopy and Radiative Transfer , 2001 . 68 ( 2 ): 179 - 193 . DOI: 10.1016/S0022-4073(00)00021-2 http://doi.org/10.1016/S0022-4073(00)00021-2 . Baseline subtraction using robust local regression estimation H Shin , B Sheu , M Joseph , MK Markey . A guilt-by-association feature selection: identifying biomarkers from proteomic profiles . Journal of Biomedical Informatics , 2008 . 41 ( 1 ): 124 - 136 . 17544868 DOI: 10.1016/j.jbi.2007.04.003 http://doi.org/10.1016/j.jbi.2007.04.003 . A guilt-by-association feature selection: identifying biomarkers from proteomic profiles P Subramani , R Sahu , S Verma . Feature selection using Haar wavelet power spectrum . BMC Bioinformatics , 2006 . 7 ( 1 ): 432 17022808 DOI: 10.1186/1471-2105-7-432 http://doi.org/10.1186/1471-2105-7-432 . Feature selection using Haar wavelet power spectrum S Theodoridis , K Koutroumbas . Pattern Recognition , 2003 . 2nd Ed Academic Press , 174 - 183 . . Pattern Recognition A Thomas , GD Tourassi , AS Elmaghraby , R Valdes , SA Jortani . Data mining in proteomic mass spectrometry . Clinical Proteomics , 2006 . 2 ( 1-2 ): 13 - 32 . DOI: 10.1385/CP:2:1:13 http://doi.org/10.1385/CP:2:1:13 . Data mining in proteomic mass spectrometry M Vannucci , N Sha , PJ Brown . NIR and mass spectra classification: baysian methods for wavelet-based feature selection . Chemometrics and Intelligent Laboratory Systems , 2005 . 77 ( 1-2 ): 139 - 148 . DOI: 10.1016/j.chemolab.2004.10.009 http://doi.org/10.1016/j.chemolab.2004.10.009 . NIR and mass spectra classification: baysian methods for wavelet-based feature selection OP Whelehan , ME Earll , E Johansson , M Toft , L Eriksson . Detection of ovarian cancer using chemometric analysis of proteomic profiles . Chemometrics and Intelligent Laboratory Systems , 2006 . 84 ( 1-2 ): 82 - 87 . DOI: 10.1016/j.chemolab.2006.03.008 http://doi.org/10.1016/j.chemolab.2006.03.008 . Detection of ovarian cancer using chemometric analysis of proteomic profiles WH Xu , YD Chen , Y Hu , JK Yu , XG Wu , TJ Jiang , S Zheng , SZ Zhang . Preoperatively molecular staging with CM10 ProteinChip and SELDI-TOF-MS for colorectal cancer patients . Journal of Zhejiang University SCIENCE B , 2006 . 7 ( 3 ): 235 - 240 . 16502512 DOI: 10.1631/jzus.2006.B0235 http://doi.org/10.1631/jzus.2006.B0235 . Preoperatively molecular staging with CM10 ProteinChip and SELDI-TOF-MS for colorectal cancer patients JS Yu , S Ongarello , R Fiedler , XW Chen , G Toffolo , C Cobelli , Z Trajanoski . Ovarian cancer identification based on dimensionality reduction for high-throughput mass spectrometry data . Bioinformatics , 2005 . 21 ( 10 ): 2200 - 2209 . 15784749 DOI: 10.1093/bioinformatics/bti370 http://doi.org/10.1093/bioinformatics/bti370 . Ovarian cancer identification based on dimensionality reduction for high-throughput mass spectrometry data X Zhang , X Lu , Q Shi , XQ Xu , HC Leung , LN Harris , JD Iglehart , A Miron , JS Liu , WH Wong . Recursive SVM feature selection and sample classification for mass-spectrometry and microarray data . BMC Bioinformatics , 2006 . 7 ( 1 ): 197 16606446 DOI: 10.1186/1471-2105-7-197 http://doi.org/10.1186/1471-2105-7-197 . Recursive SVM feature selection and sample classification for mass-spectrometry and microarray data LR Zhu , WY Zhang , L Yu , YH Zheng , J Hu , QP Liao . Proteomic patterns for endometerial cancer using SELDITOF-MS . Journal of Zhejiang University SCIENCE B , 2008 . 9 ( 4 ): 286 - 290 . 18381802 DOI: 10.1631/jzus.B0710589 http://doi.org/10.1631/jzus.B0710589 . Proteomic patterns for endometerial cancer using SELDITOF-MS NT Zinkin , F Grall , K Bhaskar , H Out , D Spentzos , B Kalmowitz , M Wells , M Guerrero , JM Asara , TA Libermann , 等 . Serum proteomics and biomarkers in hepatocellular carcinoma and chronic liver disease . Clinical Cancer Research , 2008 . 14 ( 2 ): 470 - 477 . 18223221 DOI: 10.1158/1078-0432.CCR-07-0586 http://doi.org/10.1158/1078-0432.CCR-07-0586 . Serum proteomics and biomarkers in hepatocellular carcinoma and chronic liver disease 10.1631/jzus.B0820163 2008-05-15 2008-07-30 2022-01-05 Copyright © 2008, Journal of Zhejiang University Science 2008 Journal of Zhejiang University-SCIENCE B (Biomedicine and Biotechnology) 11 09