Improving the prediction accuracy in classification using the combined data sets by ranks of gene expressions

Abstract

Background: The information from different data sets experimented under different conditions may be inconsistent even though they are performed with the same research objectives. More than that, even when the data sets were generated from the same platform, the data agreement may be affected by the technical variation among the laboratories. In this case, it is necessary to use the combined data set after adjusting the differences between such data sets, for detecting the more reliable information.

Results: The proposed method combines data sets posterior to the discretization of data sets based on the ranks of the gene expression ratios, and the statistical method is applied to the combined data set for predictive gene selection. The efficiency of the proposed method was evaluated using five colon cancer related data sets, which were experimented using cDNA microarrays with different RNA sources, and one experiment utilized oligonucleotide arrays. NCI-60 cell lines data sets were used, which were performed with two different platforms of cDNA microarrays and Affymetrix HU6800 oligonucleotide arrays. The combined data set by the proposed method predicted the test data sets more accurately than the separated data sets did. The biological significant genes were detected from the combined data set, which were missed on the separated data sets.

Conclusion: By transforming gene expressions using ranks, the proposed method is not influenced by systematic bias among chips and normalization method. The method may be especially more useful to find predictive genes from data sets which have different scale in gene expressions.

Figure 1

Relationship of OOB error rates to the number of significant genes. (A) Red: data A, black: data B, blue: data AB, which is a combined data set by the proposed method. (B) Variations of averaged OOB error rates among three different data sets, data A, data B, and data AB.

Figure 2

Comparison of the prediction accuracies. The number by the name of each data set represents the sample size of the data set. Seven significant genes with high prediction accuracy were used.

Figure 3

Comparison of the prediction accuracies in ME method.

Figure 4

Comparison of the prediction accuracies between the proposed method and the ME method.

Figure 5

Comparison of prediction accuracies of single and combined data sets.

Figure 6

Comparison of prediction accuracies under conditions of two test data sets and the number of informative genes. (A) Blue: test Oligo data with cDNA by proposed method; Red: test cDNA data with Oligo data by proposed method; Green: test Oligo data with cDNA by ME method; Cyan: test cDNA data with Oligo data by ME method. (B) Summary of prediction accuracies using a boxplot. cDNA_R and Oligo_R are data sets transformed by the proposed method. cDNA_ME and Oligo_ME are data sets transformed by the ME method.