A scale space approach for unsupervised feature selection in mass spectra classification for ovarian cancer detection
 Michele Ceccarelli^{1, 2},
 Antonio d'Acierno^{3}Email author and
 Angelo Facchiano^{3}
https://doi.org/10.1186/1471210510S12S9
© Ceccarelli et al; licensee BioMed Central Ltd. 2009
Published: 15 October 2009
Abstract
Background
Mass spectrometry spectra, widely used in proteomics studies as a screening tool for protein profiling and to detect discriminatory signals, are high dimensional data. A large number of local maxima (a.k.a. peaks) have to be analyzed as part of computational pipelines aimed at the realization of efficient predictive and screening protocols. With this kind of data dimensions and samples size the risk of overfitting and selection bias is pervasive. Therefore the development of bioinformatics methods based on unsupervised feature extraction can lead to general tools which can be applied to several fields of predictive proteomics.
Results
We propose a method for feature selection and extraction grounded on the theory of multiscale spaces for high resolution spectra derived from analysis of serum. Then we use support vector machines for classification. In particular we use a database containing 216 samples spectra divided in 115 cancer and 91 control samples. The overall accuracy averaged over a large cross validation study is 98.18. The area under the ROC curve of the best selected model is 0.9962.
Conclusion
We improved previous known results on the problem on the same data, with the advantage that the proposed method has an unsupervised feature selection phase. All the developed code, as MATLAB scripts, can be downloaded from http://medeaserver.isa.cnr.it/dacierno/spectracode.htm
Background
Proteomics studies are widely used in the biomedical research as an investigation tool to gain understanding of biological processes under specific conditions. Proteomics gives a detailed picture of the presence, integrity and/or modification of the whole mixture of proteins extracted by a source. For medical purposes, proteomics offers a diagnostic perspective for the early detection of pathologies as well as for the choice of the most effective therapy. In fact, samples as serum, plasma, and other kinds of extracts contain proteins for which the covalent structure may be modified by specific pathological states, which may induce or prevent processes as glycation, phosphorylation, methylation, or any other addition of a molecular group to the protein. As a more general case, a whole protein could be expressed or not under pathological conditions. In all these cases, the proteomics pattern analyzed by mass spectrometry techniques can evidence differences due to the pathology. Similarly, comparative proteomics can be exploited to evaluate the effects of a specific therapy.
Among the thousands of proteins and peptides present in a serum sample, which represent its proteome, few key signals may be significant markers of the pathological state, and their search within the proteome represents a still open field of research. The detection of the markers and their full characterization has a number of advantages, including the opportunity of using them for diagnostics uses and the improvement of the knowledge about the pathological effects at molecular level, required to develop new drugs and therapies.
Mass spectrometry [1] is the elective technique to characterize the proteome and its modification. The mass spectrum represents a molecular profile of the sample under analysis, obtained with increasing precision and automation techniques. Despite the large number of signals obtained in the proteome analysis, molecular modifications can be detected and markers of pathological states can be identified. MALDITOF (MatrixAssisted Laser Desorption and Ionization TimeOfFlight) is a common technology used in mass spectrometry, and SELDITOF (SurfaceEnhanced Laser Desorption and Ionization TimeOfFlight) is also used as a modified form of MALDITOF. According to these techniques, proteins are cocrystallized with UVabsorbing compounds, then a UV laser beam is used to vaporize the crystals, and ionized proteins are then accelerated in an electric field. The analysis is then completed by the TOF analyzer. Differences in the two technologies, which reside mainly in the sample preparation, make SELDITOF more reliable for biomarkers discovery and other proteomic studies in biomedicine.
Data produced by mass spectrometry are spectra, typically reported as vectors of data, describing the intensity of signals due to biomolecules with specific masstocharge (m/z) ratio values. Given the high dimensionality of spectra, given their different length and since they are often affected by errors and noise, preprocessing techniques are mandatory before any data analysis. After preprocessing (to correct noise and reduce dimensionality), several statistical and artificial intelligence based technologies could be used for mining these data.
A very important contribution of the application of mass spectrometry techniques to the classification of ovarian cancer is reported in [2] where the authors suggest a stochastic search method for selecting a subset of feature which best separates between healthy and pathological cases; the classification is based on a clustering approach using selforganizing maps and the authors show that the proposed method is able to classify all cancer cases and 95% of healthy women. The same methodology has been also successfully applied to high resolution spectrometry in [3] where the authors obtain a 100% sensitivity and specificity of classification over a random split (between training set and test set) of the data. The feature selection approach of both papers follows an optimization procedure having as objective function the discrimination ability of the adopted classifier over the subset of selected features. From the geometrical point of view it can be described as the selection of a random projection of data onto a subspace where the selected patterns are best separated. However, one should consider that in this kind of problems we have a small set of data (of order of some hundreds) in a very high dimensional space (of order of several thousands of points). Therefore, there is the risk that this good separation into the subspace could just be due to a random effect depending on the sparseness of the data [4]. In order to avoid this risk, large scale cross validation should be applied for the correct evaluation of the prediction accuracy [5].
Another important issue to be addressed in the selection of features for classification is the way of performing feature selection and cross validation together. In particular, if the feature selection step is external to the cross validation procedure, as for example in [6, 7], i.e when the feature selection is done by using all the data and the performance evaluation by cross validation is performed just for the classification phase, then the obtained results may be severely biased due to the so called selection bias effect. An interesting experiment is reported in [8] where the selection bias effect produces perfect classification even for completely fake datasets. A more proper approach should validate by cross validation both classification algorithms and feature selection, and this can be easily done by leaving the test samples out of the dataset before undergoing feature selection [9]. One of the main contribution of the present paper is the validation of the dataset of [3] with a large scale cross validation study and the adoption of an unsupervised feature extraction showing that it is possible to classify the dataset with a very high accuracy without the selection bias effect.
The dataset adopted in the present work was also used in the paper [10] where the authors developed a preprocessing based on the KolmogorovSmirnov test, restriction of coefficient of variation and wavelet analysis. The classification step is then performed, as in our case, with support vector machines. Their method achieves an average sensitivity of 97.38% (sd = 0.0125) and an average specificity of 93.30% (sd = 0.0174) in 1000 independent kfold crossvalidations; here we have a better sensitivity and specificity as reported in the Results.
A study about the classification methods for ovarian cancer detection is reported in [11] where the authors compared two feature extraction algorithms together with several classification approaches on a MALDI TOF dataset. The Tstatistic was used to rank features in terms of their relevance. Then two feature subsets were greedily selected (respectively having 15 and 25 features each). Support vector machines, random forests, linear/quadratic discriminant analysis (LDA/QDA), knearest neighbors, and bagged/boosted decision trees were subsequently used to classify the data. In addition, random forests were also used to select relevant features with previously mentioned algorithms used for classification. When the Tstatistic was used as a feature extraction technique, support vector machines, LDA and random forests classifiers obtained the top three results (with accuracy in the vicinity of 85%). On the other hand, classification improved to approximately 92% when random forests were used as both feature extractors and classifiers. While these results appear promising, the authors provide little motivation as to why 15 and 25 feature sets were selected. Here we do not fix a priori the number of features letting the algorithm select automatically the number of features as function of the percentage of energy to be preserved in the PCA and the number of peaks in the analyzed average spectrum as reported below (Methods).
The PCA dimensionality reduction approach was also used in [12] with the dataset of [2] coupled with a nearest centroid classifier for classification. When training sets were larger than 75% of the total sample size, perfect (100%) accuracy was achieved on the OCWCX2b data set. The author performed cross validation after feature selection, and as explained before, this results could be influenced by the selection bias effect. Using only 50% of data for training, the performance dropped by 0.01%. Unfortunately, the probabilistic approach used in the study can leave some samples unclassified. For the OCH4 data set, the system had a 92.45% sensitivity and 91.95% specificity when 75% of the data was used for training with only 98.60% of the data samples classified.
Results and discussion
The proposed feature extraction and classification method has been tested on a dataset available from the National Cancer Institute of the U.S. National Institute of Health consisting of 121 cancer samples and 95 control samples. Each sample is an high resolution spectrum with about 360000 points and m/z ranging from 700 to 12000. Some results on these data have been published in [3] and [10] and are useful for comparison with the method proposed here.
An interesting problem to be investigated is the feature stability of the feature extraction phase. Indeed, since for a correct cross validation procedure, the feature selection must be performed inside each fold, it is possible that different folds lead to different feature sets. Therefore the question whether there is a set of stable features which are maintained in different folds and runs arises. We run the whole feature extraction 1000 times using 10folds and, at each iteration, we compute the intersection of the selected feature positions with the previous set. The final result being the set of selected peaks which are shared among all the 10000 runs.
The table reports the m/z values selected in [3] for the classification of the same dataset with four different models
m/z features in[3]  stable feature in our model 

818.480  Yes 
1001.654  Yes 
1144.796  Yes 
1255.593  Yes 
1276.861  Yes 
2374.244  No 
4260.403  Yes 
4292.900  Yes 
4377.853  Yes 
6004.416  Yes 
6548.771  Yes 
7046.018  Yes 
7060.121  Yes 
7096.922  No 
7202.716  Yes 
8540.536  Yes 
8605.678  Yes 
8664.385  No 
8706.065  Yes 
8709.548  Yes 
9367.113  No 
9870.937  No 
System tuning
As detailed in the section Methods, there are some basic parameters which can influence the whole performance of our approach:

the maximum scale of signal smoothing, σ

the peak averaging window size, (WS),

the amount of energy, in percentage, retained in the Principal Component Analysis, E

width of the kernel functions for the SVM classifier, γ;
It is important to evaluate the influence of the parameters and how the performance of the classification depends on them. We ran a large set of experiments for this purpose having:

σ varying in the interval [0.05:1] using a step of 0.1;

WS varying in the interval [1:5];

E varying in the interval [0.70:0.98] using a step of 0.02;

γ varying in the interval [2:8].
Conclusion
The paper presented the results obtained by applying a feature extraction procedure for mass spectra classification based on a scalespace analysis of the data. The features were then used to train a statistical classifier to discriminate between normal and cancer samples. In order to compare our results with state of the art methods, we adopted a public available dataset already analyzed by other researchers. We obtained an average accuracy of 98.18 of correct classification, 98.76% sensitivity and 98.48% specificity over a large cross validation experiment designed to be free of the selection bias effect: this improves previously known results [10]. We also analyzed how stable the feature selection methods is and showed that over a large set of runs the method tends to select the same set of features.
Another advantage of the adopted method consists in the use of the multiscale properties of the spectra rather than a procedure based on the discrimination ability of the selected features. For the considered problem, with a high dimensional space and a small number of data points, optimal separating surfaces based on projection can be the results of chance rather than a subset of significant features. Finally, we used the discrimination accuracy as figure of merit just to compute the optimal parameters of the feature selection step.
Methods
Preprocessing
Before the feature selection phase, a preprocessing step is performed aimed to homogenization and correction of the spectra data. The spectral data produced by a single laser shot in a mass spectrometer consists of a vector of counts. Each count represents the number of ions hitting the detector during a small, fixed interval of time. A complete spectrum is acquired within tens of milliseconds, so a typical spectrum is a vector containing between 10,000 and 100,000 entries. Therefore, each spectrum contains many thousands of intensity measurements representing an unknown number of protein peaks, this requires extensive lowlevel processing in order to identify the locations of peaks. Inadequate or incorrect preprocessing methods, however, can result in data sets that exhibit substantial biases and make it difficult to reach meaningful biological results. In our experiments we applied the following preprocessing steps:

resampling: Gaussian kernel reconstruction of the signal in order to have a set of ddimensional vectors with equally spaced mass/charge values;

baseline correction: removes systematic artifacts, usually attributed to clusters of ionized matrix molecules hitting the detector during early portions of the experiment, or to detector overload;

normalization: corrects for differences in the total amount of protein desorbed and ionized from the sample plate;
All the details, adopted parameters and scripts of the preprocessing step can be downloaded from the accompanying web page containing all the data and code.
Feature extraction
The feature selection and description is crucial for mass spectrometry since subsequent analyses are performed only on the selected features. Several methods have been proposed which often rely on biased data sets and can reach biological conclusion difficult to be interpreted [14, 15].
Peak detection is the standard method for extracting features and several techniques to identify peaks among the background noise have been proposed (see for example [16]). Model based approaches have been adopted for the phase of feature selection of mass spectra data [17]. The model based methods typically perform a huge number of regressions to fit signal models to spectra. Here we adopt a hybrid method which is fast just as the peak selection methods, and at the same time tries to model the average spectrum at various scales. The basic principle adopted for our selection of features relies on the scale space theory of signal analysis [18, 19]. The main idea of a scalespace representation is to generate a oneparameter family of derived signals in which the finescale information is successively suppressed. This principle preserves peaks or other feature to be artificially introduced through scales and forces the analysis to be from finer scale to coarser scales.
The scalespace theory provides a wellfounded framework for representing and detecting signal structures at multiple scales, however it does not address the problem of how to select locally appropriate scales for further analysis. Whereas the problem of finding the best scales for handling a given realworld data set may be regarded as intractable unless further information is available, some approaches have been proposed, for example [19] selects the scales at which normalized measures of feature strength assume local maxima with respect to scale. For the purposes of our approach we select the best scale σ by cross validation as described in System Tuning. As can be seen in figure 7, near peaks collapses in a single local maximum.
Peak selection and reduction
In order to select a set of significant peaks to be considered as features, we use the local maxima of the average spectrum E [T_{ σ }f (x)]. These local maxima are considered as the locations of the considered peaks, see figure 8. Finally, each spectrum will be described by the mean value assumed by the original spectrum in a window centered in each of the selected local maxima.
As a last feature extraction step we perform a principal component analysis (PCA) for dimensionality reduction of the data collected as the intensity values over the selected windowed peak positions (see figure 9). It is important to point out that the feature selection must be performed inside the cross validation in order to not incur into the selection bias effect. In particular, all the parameters such as mean multiscale spectrum, peak locations and principal components are computed inside the folding, this guarantees that the all the feature selection and extraction procedure is completely blind with respect to the test data. Another important point to underline is that the feature extraction and selection described so far does not use discrimination as a measure to identify useful peaks for the classification of data. This means that the proposed feature selection method is completely unsupervised and as such it is unbiased. Even with this simplification the classification accuracy both in terms of sensitivity and specificity and AUC beats other methods such as [10] over the same dataset.
Classification
The classification problem can be naturally cast into the theory and practice of disciplines as Pattern Recognition and Machine Learning [21]. Support Vector Machine (SVM) is a technique [22] for Pattern Recognition and Data Mining classification tasks. While at present there exists no general theory that guarantees good generalization performances of SVM, but only probability bounds on its performance accuracy, there is a growing interest in this technique due to a wide literature reporting good performances in various heterogeneous fields [23].
In the case of linearly separable patterns on twoclasses vectors it is straightforward to show the basic ideas of SVM: given a set of points in ℜ^{ k }and a twoclass labels vector, SVM aims to find a linear surface that splits them in two groups according to the indicated labels, in the best possible way. Intuitively, if data are linearly separable (that is if it exists at least one hyperplane that splits them in two group), the problem becomes how to define and how to find the best possible hyperplane to do it. The SVM answer is that the best possible hyperplane is the one that maximizes the margin, that is the one that has maximal distance from both sets of points.
To generalize further, it is possible to consider surfaces that are not linear and work on a different model. By using a kernel function all data can be projected onto another space (possibly with infinite dimension) where they are linearly separable and perform the classification linearly in this new space. In practice if we look at the underlying optimization problem, it is easy to see that data appear only in the form of dot products and hence data transformed through a function ϕ appear also in the form K(x_{ i }, x_{ j }) = ϕ (x_{ i }) · ϕ (x_{ j }). Such a dot product function is called Kernel. Whatever function that satisfies dot product's constraints can be used as Kernel function, and there is an active field of research in the choice of the most suitable kernel for a given problem [24]. In this work Kernel's choice was derived form general practical considerations [25, 26]:

the Radial Basis Function (RBF) SVM has infinite capacity and hence gaussian RBF SVM of sufficiently small width can classify an arbitrarily large number of training points correctly;

the RBF kernel includes as a special case the linear kernel;

the RBF kernel behaves like the sigmoid kernel for certain parameters' values;

the RBF kernel has less hyperparameters than the polynomial kernel;

the RBF kernel has less numerical difficulties than other kernels.
Declarations
Acknowledgements
This work has been partially supported by the CNR project Bioinformatics and by Programma ItaliaUSA "Farmacogenomica Oncologica" Prog. No. 527/A/3A/5.
This article has been published as part of BMC Bioinformatics Volume 10 Supplement 12, 2009: Bioinformatics Methods for Biomedical Complex System Applications. The full contents of the supplement are available online at http://www.biomedcentral.com/14712105/10?issue=S12.
Authors’ Affiliations
References
 Aebersold R, Mann M: Mass spectrometrybased proteomics. Nature 2003, 422: 198–207. 10.1038/nature01511View ArticlePubMedGoogle Scholar
 Petricoin EF III, Ardekani AM, Hitt BA, Levine PJ, Fusaro VA, Steinberg SM, Mills GB, Simone C, Fishman DA, Kohn EC, Liotta LA: Use of proteomic patterns in serum to identify ovarian cancer. Lancet 2002, 359: 572–577. 10.1016/S01406736(02)077462View ArticlePubMedGoogle Scholar
 Conrads P, Fusaro VA, Ross S, Johann D, Rajapakse V, Hitt BA, Steinberg SM, Kohn EC, Fishman DA, Whiteley G, Barrett JC, Liotta LA, III EFP, Veenstra TD: Highresolution serum proteomic features for ovarian cancer detection. EndocrineRelated Cancer 2004, 11: 163–178. 10.1677/erc.0.0110163View ArticlePubMedGoogle Scholar
 Guyon I, Elisseeff A: An Introduction to Variable and Feature Selection. Journal of machine learning research 2003, 3: 1157–1182. 10.1162/153244303322753616Google Scholar
 Barla A, Jurman G, Riccadonna S, Merler S, Chierici M, Furlanello C: Machine learning methods for predictive proteomics. Briefings in Bioinformatics 2008, 9(2):119–28. 10.1093/bib/bbn008View ArticlePubMedGoogle Scholar
 Guyon I, Weston J, Barnhill S, Vapnik V: Gene selection for cancer classification using support vector machines. Machine Learning 2002 2002, 46: 389–422. 10.1023/A:1012487302797View ArticleGoogle Scholar
 Zhang H, Yu C, Singer B, M MX: Recursive partitioning for tumor classification with gene expression microarray data. PNAS 2001, 98: 6730–6735. 10.1073/pnas.111153698PubMed CentralView ArticlePubMedGoogle Scholar
 Zhang X, Lu X, Shi Q, Xu XQ, Leung HC, Harris L, Iglehart J, Miron A, Liu J, Wong W: Recursive SVM feature selection and sample classification for massspectrometry and microarray data. BMC Bioinformatics 2006, 7: 197. 10.1186/147121057197PubMed CentralView ArticlePubMedGoogle Scholar
 Furlanello C, Serafini M, Merler S, Jurman G: Entropybased gene ranking without selection bias for the predictive classification of microarray data. BMC Bioinformatics 2003, 4: 54–73. 10.1186/14712105454PubMed CentralView ArticlePubMedGoogle Scholar
 Yu J, Ongarello S, Fiedler R, Chen X, Toffolo G: Ovarian cancer identification based on dimensionality reduction for highthroughput mass spectrometry data. Bioinformatics 2005, 2200–2209. 10.1093/bioinformatics/bti370Google Scholar
 Wu B, Abbott T, Fishman D, McMurray W, Mor G, Stone K, Ward D, Williams K, Zhao H: Comparison of statistical methods for classification of ovarian cancer using mass spectrometry data. Bioinformatics 2003, 19(13):1636–1643. 10.1093/bioinformatics/btg210View ArticlePubMedGoogle Scholar
 Lilien R, Farid H, Donald B: Probabilistic Disease Classification of ExpressionDependent Proteomic Data from Mass Spectrometry of Humn Serum. Journal of Computational Biology 2003.Google Scholar
 Fawcett T: An introduction to ROC analysis. Pattern Recogn Lett 2006, 27(8):861–874. 10.1016/j.patrec.2005.10.010View ArticleGoogle Scholar
 Baggerly K, et al.: Reproducibility of SELDITOF protein patterns in serum: comparing datases from different experiments. Bioinformatics 2004, 20: 777–785. 10.1093/bioinformatics/btg484View ArticlePubMedGoogle Scholar
 Sorace J, Zhan M: A data review and reassessment of ovarian cancer serum proteomics profiling. BMC Bioinformatics 2003, 4: 24–32. 10.1186/14712105424PubMed CentralView ArticlePubMedGoogle Scholar
 Tibshirani R, et al.: Sample classification from protein mass spectrometry, by peack probability contrasts. Bioinformatics 2004, 20: 3034–3044. 10.1093/bioinformatics/bth357View ArticlePubMedGoogle Scholar
 Noy K, Fasulo D: Improved model based, platform independent feature extraction for mass spectrometry. Bioinformatics 2007, 23(19):2528–2535. 10.1093/bioinformatics/btm385View ArticlePubMedGoogle Scholar
 Witkin A, Terzopoulos D, Kass M: Signal matching through scale space. International Journal of Computer Vision 1987, 133–144. 10.1007/BF00123162Google Scholar
 Lindeberg T: ScaleSpace Theory in Computer Vision. Kluwer Academic Publisher; 1994.View ArticleGoogle Scholar
 Alvarez L, Lions PL, Guichard F, Morel JM: Axioms and Fundamental equations of Image Processing. Archives for Rational Mechanics and Analysis 1993, 16(9):200–257.Google Scholar
 Vapnik V: The Nature Of Statistical Learning Theory. New York: SpringerVerlag; 1995.View ArticleGoogle Scholar
 Boser B, Guyon I, Vapnik V: a training algorithm for optimal margin classifiers. Proceedings of the Fifth Annual workshop on Computational Learning Theory 1992.Google Scholar
 Schoelkopf B, Sung K, Burges C, Girosi F, Niyogi P, Poggio T, Vapnik V: Comparing Support Vector Machines with Gaussian Kernels to Radial Basis Function Classifiers. IEEE Transactions on Signal Processing 1997, 45–11: 2758–2765. 10.1109/78.650102View ArticleGoogle Scholar
 Cristianini N, Taylor JS: Kernel Methods for Pattern Analysis. Cambridge University Press; 2004.Google Scholar
 Keerthi SS, Lin CJ: Asymptotic behaviors of support vector machines with gaussian kernel. Neural Computation 2003, 15(7):1667–1689. 10.1162/089976603321891855View ArticlePubMedGoogle Scholar
 Verri A, Pontil M: Properties of support vector machines. Neural Computation 1998, 10(4):955–974. 10.1162/089976698300017575View ArticlePubMedGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.