BioMed Research International

BioMed Research International / 2017 / Article
Special Issue

Analysis and Modeling for Big Data in Cancer Research

View this Special Issue

Research Article | Open Access

Volume 2017 |Article ID 1645619 | 6 pages | https://doi.org/10.1155/2017/1645619

A Cancer Gene Selection Algorithm Based on the K-S Test and CFS

Academic Editor: Jianxin Li
Received12 Jan 2017
Accepted06 Apr 2017
Published08 May 2017

Abstract

Background. To address the challenging problem of selecting distinguished genes from cancer gene expression datasets, this paper presents a gene subset selection algorithm based on the Kolmogorov-Smirnov (K-S) test and correlation-based feature selection (CFS) principles. The algorithm selects distinguished genes first using the K-S test, and then, it uses CFS to select genes from those selected by the K-S test. Results. We adopted support vector machines (SVM) as the classification tool and used the criteria of accuracy to evaluate the performance of the classifiers on the selected gene subsets. This approach compared the proposed gene subset selection algorithm with the K-S test, CFS, minimum-redundancy maximum-relevancy (mRMR), and ReliefF algorithms. The average experimental results of the aforementioned gene selection algorithms for 5 gene expression datasets demonstrate that, based on accuracy, the performance of the new K-S and CFS-based algorithm is better than those of the K-S test, CFS, mRMR, and ReliefF algorithms. Conclusions. The experimental results show that the K-S test-CFS gene selection algorithm is a very effective and promising approach compared to the K-S test, CFS, mRMR, and ReliefF algorithms.

1. Introduction

Big data analysis technology can mine gene information related to diseases and drugs from massive gene data and provide new ideas for drug development as well as disease diagnosis and treatment. Therefore, big data has positive effects on cancer research. Genetic data analysis includes four steps: gene data acquisition, gene data pretreatment, gene selection, and classification model establishment and evaluation. Of these steps, genetic data acquisition is a biomedical process, and the other steps are data mining processes. This paper focuses on the gene selection step in genetic data analysis by exploring the challenges to gene data analysis and effective strategies and methods for gene selection.

According to its relationship with the classifier, the feature (gene) selection method is divided into the filter method, the wrapper method, and the embedded method. The filter method selects the features that contribute to the classification, which is independent of the learning process, and has a higher efficiency and a stronger generalization ability. The wrapper method selects the corresponding feature subsets according to the classification performance of the feature subsets. Depending on the learning process, the wrapper method has a higher accuracy, but it is prone to overadaptability, a poor generalization performance, and low time efficiency. The combination of the filter method and the wrapper method is a new trend in studies of feature selection.

There is a significant difference in the expression value of a discriminative gene between different genotypes. Thus, a series of filter-based gene selection methods, based on parametric statistics, was developed to detect whether there were significant differences between genotypes and to select a subset of genes with significant differences [1, 2]. However, parametric statistical methods need to assume a Gaussian distribution of the data, and the actual genetic dataset usually does not meet the Gaussian distribution hypothesis. Therefore, a nonparametric statistical method, the Wilcoxon rank sum test, is used in gene selection studies. However, the rank sum test can be used to reveal the location of two sample types (the distributions of the values of the two sample types) only when the sample size is large or the measurement level is low (the sample observations have only a small number of values). When the sample size is very small or has the same rank value as the sample with the same rank, it is not appropriate to use the rank sum test for gene selection.

The Kolmogorov-Smirnov (K-S) test is another nonparametric statistical method used to compare the distribution of two sample types. This method is very sensitive to the difference of the distribution of two sample types. It has been successfully applied in the analysis of ovarian cancer gene data, recognition, and other fields [3]. However, an independent nonparametric test method does not take into account the redundancy of the genes in the selection of genes with discriminatory power.

The correlation-based feature selection (CFS) [4, 5] method can efficiently select subsets of genes that are highly correlated with the class and that have low redundancy. However, due to the high-dimensional characteristics of gene datasets, it is very time-consuming to adopt the CFS method for gene selection directly. Therefore, a gene selection algorithm combining the K-S test and CFS is proposed in this paper. Most of the redundant and noise genes are removed by the K-S test, and the genes with a significant distinguishing ability are retained. Then, CFS is used to evaluate the genes that are highly correlated with the class and have low redundancy. A support vector machine (SVM) [6, 7] is used as the classifier to evaluate the gene subsets generated based on accuracy. Finally, the method is compared with the K-S test, minimum-redundancy maximum-relevancy (mRMR) [8], and classic ReliefF algorithms [9]. The experimental results from five gene datasets show that the K-S test-CFS gene selection method is an effective gene selection algorithm.

2. Materials and Methods

2.1. Datasets Description

In this paper, five classical cancer gene datasets are used: breast cancer [10], lung cancer [11], colon tumor [12], ovarian cancer [13], and leukemia [2]. Detailed information on the datasets is listed in Table 1. To eliminate the influence of different dimensions on the experimental results, the five datasets were -score standardized as part of the preprocessing.


DatasetSamplesGenes

Breast cancer9724481
Lung cancer18112533
Colon tumor622000
Ovarian cancer25315154
Leukemia727129

2.2. K-S Test

In this paper, the K-S test was used to determine significant differences between the genes of the tumor patients and those of normal controls. Let be a gene from the gene dataset, and the observed value is , where is the sample number of the gene dataset. According to the gene order value, the order of the observations is , and the cumulative distribution function of the gene is defined as follows:Assuming that the cumulative distribution functions of the gene to be tested in the tumor sample and the normal sample are and , where the number of observations is the number of positive and negative samples, the K-S test statistic isAccording to the K-S test theory, when (the critical value of for the level of significance α), the gene has no significant difference between the positive and negative classes when the significance level is α; if , there is a significant difference between the positive and negative samples at the confidence level.

From (2), we can see that the bigger the value, the greater the difference between the positive and negative classes of the gene, indicating a stronger ability to distinguish between the positive and negative samples.

2.3. Correlation-Based Feature Selection (CFS)

The correlation feature selection (CFS) method evaluates subsets of features according to the following hypothesis: “good feature subsets contain features that are highly correlated with the classification yet uncorrelated to each other.” The bias of the evaluation function is towards subsets containing features that are highly correlated with the class and uncorrelated with each other. Irrelevant features should be ignored because they have a low correlation with the class. Redundant features should be removed, as they will be highly correlated with one or more of the remaining features. The acceptance of a feature depends on the extent to which it predicts classes in areas of the instance space not already predicted by other features.

2.4. K-S Test-CFS Method for Gene Selection

As we previously mentioned, the K-S test is a general and successful attribute estimator and is able to effectively provide quality estimates of attributes in problems that have dependencies between attributes. However, the K-S test does not explicitly reduce the redundancy in selected genes. CFS selects genes that have the highest relevance with the target class and that are also maximally dissimilar to each other. Thus, the integration of the K-S test and CFS leads to an effective gene selection scheme.

The details of the K-S test-CFS algorithm are as follows: in the first stage, the K-S test is applied to find a candidate gene set. This approach removes many unimportant genes and reduces the computational load for CFS. In the second stage, the CFS method is applied to directly and explicitly reduce the redundancy and to select a compact yet effective gene subset from the candidate set.

2.5. Software Package

In this paper, the K-S test, T test, and Wilcoxon test algorithms are implemented using MATLAB R2012a. The CFS, mRMR, ReliefF, and SVM algorithms are implemented using Weka 3.6. Weka (http://www.cs.waikato.ac.nz/ml/weka/) is a software packaged that collects various types of learning algorithms for data mining tasks. The SVM algorithm uses a linear kernel function, and the penalty factor takes a fixed value of 1.

3. Results and Discussion

3.1. Comparison of the K-S Test with the T Test and the Wilcoxon Test

This section compares the performance of the gene selection algorithms using the K-S test, the Wilcoxon test, and the T-test. First, the significance level alpha was set, and then, each gene in the dataset was tested by the K-S test, the Wilcoxon test, and the T test to select the important genes in order to form a subset of preselected genes. In the preselected gene subset, SVM was used as the classifier to calculate the accuracy of the 10-fold cross-validation. Then, a performance comparison of the gene subsets selected by the K-S test, the Wilcoxon test, and the T test in the different alpha values was performed. Table 2 lists the number of gene subsets selected by the K-S test, the Wilcoxon test, and the T test in the five datasets with different alpha values. Table 3 shows the average classification accuracy of the 10-fold cross-validation in the gene subsets selected by the K-S test, the Wilcoxon test, and the T-test in the five datasets with different alpha values.


DatasetAlgorithmAlpha = 1Alpha = 0.05Alpha = 0.01Alpha = 0.005Alpha = 0.001

Breast cancerK-S2448135021397940349
Wilcoxon24481382915291029381
T2448132511161726273
Lung cancerK-S125332886198215881300
Wilcoxon125333225265819861528
T125333190258019961625
Colon tumorK-S200032414610544
Wilcoxon200038718814059
T200038917111353
Ovarian cancerK-S15154726834081386268
Wilcoxon15154765239271876329
T15154790038481938318
LeukemiaK-S712917161036843524
Wilcoxon712918601169962644
T712918111115931583


DatasetAlgorithmAlpha = 1Alpha = 0.05Alpha = 0.01Alpha = 0.005Alpha = 0.001

Breast cancerK-S68.683.686.386.383.2
Wilcoxon67.883.584.884.884.8
T66.780.283.5 80.580.5
Lung cancerK-S85.889.690.491.691.6
Wilcoxon85.886.988.589.589.5
T83.686.988.589.589.5
Colon tumorK-S73.481.485.985.983.2
Wilcoxon73.479.280.281.583.2
T73.479.280.281.581.5
Ovarian cancerK-S95.398.610010096.5
Wilcoxon95.397.398.610094.6
T95.397.398.610094.6
LeukemiaK-S71.675.381.482.685.6
Wilcoxon71.675.381.481.483.5
T71.675.381.481.482.2

The experimental results in Table 2 show that the number of gene subsets selected by the K-S test, the Wilcoxon test, and the T test with the same alpha value was different. As shown in Table 2, the K-S test selected a smaller subset of genes in most cases.

Table 2 also shows that the subset of genes selected by the three test algorithms was smaller when the confidence level was large and the significance level α was small. When the confidence level was 99.9%, the significance level . In the colon dataset, the size of the selected subset of genes was approximately 50, which is approximately 2.5% of the original dataset. The size of the subset of genes selected in the breast cancer dataset was approximately 1.5% of the original number of genes in the dataset. The worst case observed was with the lung cancer dataset, and at this significant level, the size of the selected gene subset for the three test algorithms was approximately 10% of the original gene number of genes in the dataset.

The above analysis shows that the K-S test is a very effective genetic importance measurement algorithm. This test selected a smaller subset of genes that had a high interclass discrimination ability.

The average classification accuracy of the subset of genes selected by the three test algorithms at the different levels of significance is shown in Table 3. For the breast cancer dataset, the significance level was 0.001, and the average classification accuracy rate of the K-S test was slightly worse than that of the Wilcoxon test; however it was better than that of the T test. When the significance level was 0.0.5, 0.01, or 0.005, the average classification accuracy rate of the K-S test was not lower than the rates of the Wilcoxon test and the T test. For the other four genetic datasets, regardless of whether the significance level was 0.05, 0.01, 0.005, or 0.001, the average classification accuracy rate of the gene subset selected by the K-S test was not lower than the rates of the Wilcoxon test and the T test. Therefore, this finding demonstrated that the K-S test could select a better gene subset.

Based on the above results, the K-S test was superior to the Wilcoxon test and the T test for gene selection.

3.2. Compare the CFS with the mRMR and ReliefF Algorithms

The CFS algorithm was compared to the mRMR and ReliefF algorithms to validate the performance of the gene selection in the preselected gene subset. First, all of the genes were prescreened by the K-S test with a significance level of 0.01, and a preselected gene subset was obtained. The CFS algorithm selected the appropriate subset of genes directly from the subset of prescreened genes. The mRMR and ReliefF algorithms selected the first 50 genes sorted by the importance of the gene. Then, a forward selection algorithm was used to select the appropriate subset of genes from those 50 genes.

In the experiment, we adopted SVM as a classifier and used the criteria of the average accuracy of a tenfold cross-validation in the dataset to evaluate the performance of the classifiers on the selected gene subsets. To obtain statistically significant experimental results, the dataset samples were randomly shuffled, the procedure was repeated 10 times, and the average of the 10 replicates was recorded and compared. Table 4 shows the average accuracy of the tenfold cross-validation of the three algorithms in the five gene datasets and the corresponding number of genes on average.


DatasetGene selection method
CFSmRMRReliefF
The number of genesAccuracyThe number of genesAccuracyThe number of genesAccuracy

Breast cancer11.787.410.485.615.959.5
Lung cancer23.291.625.788.426.787.6
Colon tumor10.790.112.686.415.384.8
Ovarian cancer33.298.531.595.637.493.2
Leukemia25.299.62.599.616.477.6

From the comparison of the average accuracy (calculated from the results of the ten replicates) of the three algorithms shown in Table 4, we can see that, for the breast cancer dataset, the CFS algorithm achieves the best performance with the least features, which is significantly better than the performance of the other algorithms. For the colon dataset, the CFS was superior to the ReliefF and mRMR algorithms. For the lung cancer, ovarian, and leukemia datasets, the performance of the CFS algorithm is similar to that of the mRMR algorithm and better than that of the ReliefF algorithm.

Based on the above results, the CFS algorithm is superior to the mRMR and ReliefF algorithms for the preselected gene subset.

3.3. Comparison of the K-S Test-CF Algorithm with the K-S Test, CFS, mRMR, and ReliefF Algorithms

We also compared the K-S test-CFS selection algorithm with other gene selection algorithms, including the K-S test, mRMR, CFS, and ReliefF. Table 5 presents the classification accuracy comparison using the SVM classifier based on the selected genes and these five feature selection methods. From Table 5, we observed the following:(i)The K-S test-CFS algorithm achieved a better performance than the other gene selection algorithms on almost all datasets. The experimental comparisons demonstrate the effectiveness of the integration of the K-S test and CFS.(ii)CFS achieved a good performance on most of the datasets. However, its performance was not always as good as that of the K-S test-CFS algorithm. It outperforms the mRMR and ReliefF algorithms.In summary, the performance of the K-S test-CFS is superior to other gene filtering algorithms. However, in the course of the experiment, we found that the runtime of the K-S test-CFS had no advantage over the other algorithms. Therefore, the focus of the next step in this work should be how to optimize the running time of the K-S test-CFS algorithm.


DatasetGene selection method
k-S test- CFSCFSk-S testmRMRReliefF
The number of genesAccuracyThe number of genesAccuracyThe number of genesAccuracyThe number of genesAccuracyThe number of genesAccuracy

Breast cancer11.787.419.680.522.578.821.882.415.959.4
Lung cancer2391.627.388.933.480.628989.833.684.7
Colon tumor10.790.16.889.719.484.55.989.71574.9
Ovarian cancer33.298.531.695.346 78.932.795.2 39.690.6
Leukemia25.279.633.378.938.772.728.675.736.477.6

4. Conclusions

In this paper, we present a K-S test-CFS selection algorithm developed by combining the K-S test and CFS. The K-S test effectively provided quality estimates of the attributes in problems that have dependencies between attributes, and the CFS method selected genes that had the highest relevance with the target class and are also maximally dissimilar to each other. The integration of the K-S test and CFS thus leads to an effective gene selection scheme. In the first stage, the K-S test is applied to find a candidate gene set. In the second stage, CFS is applied to select a compact yet effective gene subset from the candidate set. Comprehensive experiments were conducted to compare the K-S test-CFS selection algorithm to the K-S test, CFS, ReliefF, and mRMR feature selection methods using the SVM classifier on five different datasets. The experimental results show that the K-S test-CFS gene selection is an effective method compared to the K-S test, CFS, mRMR, and ReliefF algorithms.

Conflicts of Interest

The authors declare that they have no conflicts of interest.

Authors’ Contributions

Qiang Su and Yina Wang contributed equally to this work. Fuxue Chen and Wencong Lu conceived the project. Qiang Su designed the methodology, performed the experiments, and interpreted the results, and Xiaobing Jiang drafted the manuscript. Yina Wang revised the manuscript.

Acknowledgments

The present study was supported by The National Key Research and Development Program of China (Grant no. 2016YFD0501101), National Natural Science Foundation of China (81271384 and 81371623), and High Performance Computing Center Program of Shanghai University.

References

  1. S. Dudoit, J. Fridlyand, and T. P. Speed, “Comparison of discrimination methods for the classification of tumors using gene expression data,” Journal of the American Statistical Association, vol. 97, no. 457, pp. 77–87, 2002. View at: Publisher Site | Google Scholar | MathSciNet
  2. T. R. Golub, D. K. Slonim, P. Tamayo et al., “Molecular classification of cancer: class discovery and class prediction by gene expression monitoring,” Science, vol. 286, no. 5439, pp. 531–527, 1999. View at: Publisher Site | Google Scholar
  3. J. S. Yu, S. Ongarello, R. Fiedler et al., “Ovarian cancer identification based on dimensionality reduction for high-throughput mass spectrometry data,” Bioinformatics, vol. 21, no. 10, pp. 2200–2209, 2005. View at: Publisher Site | Google Scholar
  4. M. A. Hall, Correlation-Based Feature Selection for Machine Learning, The University of Waikato, Hamilton, New Zealand, 1999.
  5. B. Senliol, G. Gulgezen, L. Yu, and Z. Cataltepe, “Fast Correlation Based Filter (FCBF) with a different search strategy,” in Proceedings of the 23rd International Symposium on Computer and Information Sciences (ISCIS '08), Istanbul Technical University, Suleyman Demirel Cultural Center, Istanbul, Turkey, October 2008. View at: Publisher Site | Google Scholar
  6. A. Ben-Hur, D. Horn, and H. T. Siegelmann, “Support vector clustering,” Journal of Machine Learning Research, vol. 2, pp. 125–137, 2001. View at: Google Scholar
  7. C. Cortes and V. Vapnik, “Support-vector networks,” Machine Learning, vol. 20, no. 3, pp. 273–297, 1995. View at: Publisher Site | Google Scholar
  8. H. Peng, F. Long, and C. Ding, “Feature selection based on mutual information: criteria of max-dependency, max-relevance, and min-redundancy,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 27, no. 8, pp. 1226–1238, 2005. View at: Publisher Site | Google Scholar
  9. M. Robnik-Šikonja and I. Kononenko, “Theoretical and Empirical Analysis of ReliefF and RReliefF,” Machine Learning, vol. 53, no. 1-2, pp. 23–69, 2003. View at: Publisher Site | Google Scholar
  10. L. J. Van't Veer, H. Dai, M. J. Van de Vijver et al., “Gene expression profiling predicts clinical outcome of breast cancer,” Nature, vol. 415, no. 6871, pp. 530–536, 2002. View at: Publisher Site | Google Scholar
  11. G. J. Gordon, R. V. Jensen, L.-L. Hsiao et al., “Translation of microarray data into clinically relevant cancer diagnostic tests using gene expression ratios in lung cancer and mesothelioma,” Cancer Research, vol. 62, no. 17, pp. 4963–4967, 2002. View at: Google Scholar
  12. U. Alon, N. Barka, D. A. Notterman et al., “Broad patterns of gene expression revealed by clustering analysis of tumor and normal colon tissues probed by oligonucleotide arrays,” Proceedings of the National Academy of Sciences of the United States of America, vol. 96, no. 12, pp. 6745–6750, 1999. View at: Publisher Site | Google Scholar
  13. E. F. Petricoin, A. M. Ardekani, B. A. Hitt et al., “Use of proteomic patterns in serum to identify ovarian cancer,” Lancet, vol. 359, no. 9306, pp. 572–577, 2002. View at: Publisher Site | Google Scholar

Copyright © 2017 Qiang Su et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.


More related articles

953 Views | 303 Downloads | 6 Citations
 PDF  Download Citation  Citation
 Download other formatsMore
 Order printed copiesOrder

Related articles

We are committed to sharing findings related to COVID-19 as quickly and safely as possible. Any author submitting a COVID-19 paper should notify us at help@hindawi.com to ensure their research is fast-tracked and made available on a preprint server as soon as possible. We will be providing unlimited waivers of publication charges for accepted articles related to COVID-19. Sign up here as a reviewer to help fast-track new submissions.