Introduction

Most molecular studies on diseases have focused on just one or two data types as an attempt to show more distinct differences among target samples. However, recent efforts to produce massive data such as the TCGA (The Cancer Genome Atlas) project (Cancer Genome Atlas Research Network 2008) allowed researchers to analyze comprehensive molecular landscapes of diseases. Now various windows such as copy number variation (CNV), DNA methylation, RNA-Seq or somatic mutations are available to understand the molecular landscape of a target disease.

CNVs are structurally variant regions where copy number differences have been observed between two or more genomes (Feuk et al. 2006). CNV has received considerable interests because it is one of the important sources of genetic variation causing phenotype diversity (Henrichsen et al. 2009). It is believed that CNVs can be a predominant mechanism driving gene and genome evolution (Zhang et al. 2009). Due to its prevalence, CNVs could drive significant intraspecific genetic variation (Henrichsen et al. 2009; She et al. 2008; Shlien and Malkin 2009). For example, consistent increase in the frequency of rare CNVs was reported among breast cancer cases (Pylkas et al. 2012). It was found that CNVs are specific to cancer types and reproducible from cell to cell (Ni et al. 2013). Because human populations show extensive polymorphism (both insertions and deletions) in the number of copies of chromosomal segments (Hastings et al. 2009), CNVs have the potential for understanding underlying factors in human diseases.

DNA methylation provides stability and diversity to the cellular phenotype through chromatin marks affecting local transcriptional potential. Methylation of DNA cytosine residues at the carbon 5 position in CpG dinucleotides is a common epigenetic mark in many eukaryotes (Laird 2010). In particular, aberrant promoter hypermethylation associated with inappropriate gene silencing affects virtually every step in tumor progression (Jones and Baylin 2002). Especially, CpG island methylation plays an important role in transcriptional regulation. Between 5 and 10 % of normally unmethylated CpG promoter islands become abnormally methylated in various cancer genomes (Dawson and Kouzarides 2012). In breast cancer transcriptionally repressed genes become aberrantly methylated, and the affected genes can be used to distinguish breast tumors of epithelial and mesenchymal lineage (Sproul et al. 2011). Four DNA methylation-based subgroups of colorectal cancer were identified using cluster analyses (Hinoue et al. 2012). Significant promoter hypermethylation in at least 50 % of CpG sites in two genes, ABHD9 and HOXD3, was found in tumors from recurring patients compared with those without recurrence (Stott-Miller et al. 2014). Changes in DNA methylation patterns play a critical role in development, differentiation and diseases such as multiple sclerosis, diabetes, schizophrenia, aging, and multiple forms of cancer (Bibikova et al. 2011).

RNA-Seq is an approach for transcriptome profiling based on deep-sequencing technologies. RNA-Seq produces a genome-scale transcription map consisting of both the transcriptional structure and/or level of expression for each gene (Wang et al. 2009). Several RNA-Seq-based technologies, such as improvements in transcription start site mapping strand-specific measurements and small RNA characterization, have allowed more complete observation of RNA transcripts (Ozsolak and Milos 2011). The splicing signatures of the subtypes of breast cancer were revealed using RNA-Seq (Eswaran et al. 2013). RNA-Seq was also utilized to define the subsets of pancreatic circulating tumor cells (Ting et al. 2014). From the comprehensive landscape of the transcriptome profiles of prostate cancer in the Chinese population, it was reported that there exists wide diversity in gene fusions, long noncoding RNAs, alternative splicing and somatic mutations (Ren et al. 2012).

Although genetic mutations causing human disease can be inherited from one’s parents, most mutations that cause cancer as well as other diseases arise somatically (Poduri et al. 2013; Stratton 2011). In addition to diseases, the somatic mutation theory of aging posits that the accumulation of mutations in the genetic material of somatic cells as a function of time results in a decrease in cellular function (Kennedy et al. 2012). With the widespread use of next-generation sequencing technologies, high-throughput mutation profiling identifies frequent somatic mutations in cancers. It was observed that 14.4 % of gastric cancer patients harboring mutations (Lee et al. 2012). The integrated analysis based on 27 cancer types illustrated that the variation in mutation frequency can be partly explained by cancer types, and the mutation spectra also vary across cancer types (Watson et al. 2013). The fact that examining the patterns of somatic mutations is not enough to decipher individual mutational signatures that are operative in each sample (Alexandrov and Stratton 2014) indicates the need for multiplatform-based approach for cancer analysis.

The TCGA project (Cancer Genome Atlas Research Network 2008) offers various types of data for a number of cancers. Currently, the analyzed data of 23 types of cancers are available without limitation and nine types of cancers are available under publication limitations. For each cancer type, a user can download clinical data, images microsatellite instability, DNA sequencing, miRNA sequencing, protein expression, mRNA sequencing, total RNA sequencing array-based expression, DNA methylation and copy number data. The publication of the TCGA data enabled multiplatform-based analysis of various cancers.

There have been active researches to understand a disease from various dimensions. The Cancer Genome Atlas Research Network produced a catalogue of molecular aberrations causing ovarian cancer (Cancer Genome Atlas Research Network 2011). The landscape of somatic genomic alterations of chromophobe renal cell carcinomas was produced based on multidimensional and comprehensive characterization of the molecular basis of a target disease (Davis et al. 2014). Hoadley and colleagues identified 11 “integrated subtypes” from 12 tumor types, which were consistent with the histological classification. Among the cases, approximately 10 % were reclassified based on the multiple assay platforms with significantly increased accuracy in the prediction of clinical outcomes by the newly defined integrated subtypes (Hoadley et al. 2014). However, these researches did not consider the effectiveness of each molecular assay platform and their combinations for explaining the difference between tumor and normal tissues. In this paper, we analyzed bladder urothelial carcinoma (BLCA) and kidney renal papillary cell carcinoma (KIRP) from the viewpoint of the effectiveness of aberrations in CNV, DNA methylation, RNASeq version 2 (RNASeqV2) which is similar to RNA-Seq in terms of the employment of sequencing data but uses a different set of algorithms for determining expression levels, and somatic mutations (SNPs, insertions, and deletions of DNA bases) for in silico classification. We measured the effectiveness in terms of quantitative and qualitative performance by using information gain (Mitchell 1997), and observed the difference in classification performance by CNV, DNA methylation, RNASeqV2, somatic mutations, and their combinations.

Materials and methods

Data preparation

In this study, we used samples generated by The Cancer Genome Atlas (http://cancergenome.nih.gov/). Among the various types of data, we selected CNV (SNP array), DNA methylation, RNASeqV2, and somatic mutations. These four data types represent characteristics of the selected cancer types in terms of genomics, epigenomics, and transcriptomics. Especially for this study, we selected BLCA, and KIRP as target diseases, and downloaded 28 cases of the BLCA data (14 tissue samples for each tumor and normal case) and 42 cases of the KIRP data (21 tissue samples for each tumor and normal case). All of the 70 cases of samples were composed of observations on all of CNV, DNA methylation, RNASeqV2, and somatic mutations (Table 1). Our analyses were based on relevant genes with the above data types, which were identified by using the following pre-processing steps. Here we followed approaches available in literature.

Table 1 Data summary
Full size table

The TCGA consortium performed CNV calling and provided segment mean values for all detected CNVs. Especially the segment mean value is computed as \(\log_{2} ({\text{observed intensity}}/{\text{reference intensity}})\) and represents the extent of copy number change. We regarded CNVs with segment mean value of greater than 0.2 as amplifications and less than −0.2 as deletions, which was decided based on a previous study (Laddha et al. 2014).

To increase the likelihood of identifying differentially methylated genes, we only considered genes that met the following criteria. We first extracted a set of differentially methylated CpG sites (DMC) in a promoter region (size of 1.5 Kbp) of each gene, and then searched for differentially methylated genes based on the state of their DMCs. The degree of methylation of each probe (target CpG site) is represented using a β value. The β value is a continuous variable between 0 and 1, with β values approaching 1 (or 0) indicating complete methylation (or non-methylation) (Kim et al. 2012). This β value is used to determine a hypermethylated or hypomethylated DMC. In the case of normal tissues, the β values of ≥0.7 and ≤0.3 were used as a threshold for hyper and hypomethylation respectively. In the case of tumor tissue, high methylation values were rarely observed due to heterogeneous mix of cell types in each sample. Thus β value of 0.3 was used as a threshold for distinguishing hyper or hypomethylated states (Sproul et al. 2011). Finally, at least three quarters of multiple DMCs in the same promoter region should have the same direction of methylation and at least one DMC have at least 0.35 mean methylation difference between tumor and normal phenotypes to determine aberrant DNA methylation of a gene.

In order to find differentially expressed (DE) genes using RNASeqV2 data, we employed the R package EBSeq (Leng et al. 2013), which can compute the fold change (FC) value. We classified genes as DE genes when |FC| > 2. The DE genes were further partitioned into up- or down-regulated genes based on the FC values (FC > 2: up-regulated, FC < −2: down-regulated) (Guo et al. 2013). From the somatic mutation data, we collected genes that contain single nucleotide polymorphisms (SNPs), insertions, and/or deletions.

The above data is summarized in Table 1. There exist distinct difference between normal samples and tumor samples from both of the BLCA and KIPP data sets. In terms of the CNV data, tumor samples from the BLCA data set have more than 17.73 times and 4.58 times amplifications and deletions compared to normal samples. In contrast, the KIRP data set shows smaller differences (7.69 and 2.21 times, respectively). On average, 76.71 genes are hypermethylated in the case of tumor samples in the BLCA data set. Although more genes in normal samples from the BLCA data set is hypermethylated than genes in tumor samples of the KIRP data set, 13 times as many genes were hypermethyalted in tumor samples compared to normal samples in the case of the KIRP data set. In the RNASeqV2 data, tumor samples show fewer up-regulated genes and more down-regulated genes in both of the BLCA and KIRP data sets. The difference is less distinct in the somatic mutation data. However it is clear that tumor samples show more SNPs, insertions, and deletions per samples in both of the BLCA and KIRP data sets.

Information gain and decision tree

A gene with a specific state of CNV, DNA methylation, RNASeqV2, or somatic mutations could be regarded as an attribute capable of explaining associated tumor or normal tissue samples. A proper quantitative measure of the worth of an attribute is information gain (Mitchell 1997), which can be used as an indicator for the quantitative importance of the gene for the tumor or normal tissue samples. Information gain is defined based on the Shannon’s entropy formula if the target domain can take on c different classes,

$$Entropy(S) \equiv \mathop \sum \limits_{i = 1}^{c} - p_{i} \log_{2} p_{i}$$
(1)

where S is the set of samples (the BLCA or KIRP data sets in our case), \(p_{i}\) is the proportion of S belonging to class i (normal or tumor in our case). Entropy characterizes the purity of a collection of samples. Entropy is interpreted as the minimum number of bits of information needed to encode the classification of an arbitrary member of S. Thus, one can quantify the effectiveness of an attribute based on the expected reduction in entropy and information gain is a measure for the expected reduction in entropy. Information gain, Gain(S, A) of attribute A relative to a set of samples S is defined as,

$$Gain\left( {S,A} \right) \equiv Entropy\left( S \right) - \mathop \sum \limits_{v \in Values(A)} \frac{{\left| {S_{v} } \right|}}{S}Entorpy(S_{v} )$$
(2)

where \(Values(A)\) is the set of all possible values for attribute A, and \(S_{v}\) is the subset of S for which attribute A has value v. In our case, A denotes genes and \(Values(A)\) can have different values based on which omic data type is in consideration. By using the pre-processing steps described in the previous section, we defined the following values for each different omics data type:

  • CNV: amplification, deletion, and non-variant

  • DNA methylation: hypermethylated, hypomethylated, and non-methylated

  • RNASeqV2: up-regulated, down-regulated, and normal

  • Somatic mutations: SNPs, insertions, deletions, and no mutation

For example, if we are testing the effect of CNV, the values of A can be “amplification”, “deletion”, and “non-variant.” If we are interested in the joint effect of CNV and RNASeqV2, nine different values are possible from the three values of each CNV and RNASeqV2.

Because information gain is the measure of the expected entropy-reduction of an attribute A, information gain is utilized to build a decision tree (Mitchell 1997; Quinlan 1993). The decision tree is composed of nodes specifying a test of an attribute, and an instance is classified by sorting down the decision tree from the root to some leaf nodes. Information gain is used to select the best attribute for current decision node. In this paper, we utilized the implementation of a decision tree in WEKA (Hall et al. 2009), a public data mining software, for measuring the contribution of CNV, DNA methylation, RNASeqV2, somatic mutations, and their combinations for distinguishing tumor-normal tissue samples.

Performance measure

To evaluate the effect of omic data and their combinations on tumor and normal tissue classification, we used two widely used measures, precision and recall, in the field of data mining. In terms of true positive (TP), false negative (FN), and false positive (FP), precision and recall are defined as:

$${\text{Precision }} = \frac{TP}{TP + FP}$$
(3)
$${\text{Recall }} = \frac{TP}{TP + FN}$$
(4)

Results and discussion

Information gain and classification performance

Information theory-based approach was already introduced for transcriptomes, where gene diversity, the specialization of transcriptomes and gene specificity were defined using Shannon’s entropy (Martinez and Reyes-Valdes 2008). In this paper, we adopted similar approach to analyze the effectiveness of CNV, DNA methylation, RNASeqV2, somatic mutations and their combinations for explaining tumor-normal tissues. Specifically, for two target diseases, BLCA, and KIRP, we collected total 70 samples containing CNV, DNA methylation, RNASeqV2, and somatic mutation data with information whether each sample is normal or tumor (“Materials and methods”). We created total ten different evaluation cases based on the number and type of used omics data: CNV only (C), DNA methylation only (M), RNASeqV2 only (R), somatic mutations only (S), CNV and DNA methylation (CM), CNV and RNASeqV2 (CR), CNV and somatic mutations (CS), DNA methylation and RNASeqV2 (MR), DNA methylation and somatic mutations (MS), and RNASeqV2 and somatic mutations (RS). For each different evaluation case, we computed the information gain of each gene and re-classified the samples based on a decision tree by only using the assigned omic data, and the classification accuracy was measured by using precision and recall (“Materials and methods”). Here the combinations with more than two omics data types could not be used because we could not find genes that contain more than two omic data signatures due to our stringent rules (“Materials and methods”). We used the average of the information gain across all genes as the information gain of each evaluation case.

Table 2 summarizes the important findings in this study. In general the aggregation of omics data clearly increased information gain compared with results from single omics data. For example, in the case of the BLCA data set, the information gain of the combinations of two omics data types, except for MS, was higher than or equal to each one of single omics data. A similar pattern was also observed in the KIRP data set. In addition, the aggregation of multiple omics data is beneficial especially for omics data with lower discriminating performance. In the BLCA data set, the precision and recall of CR, CS, MR, MS, and RS are better than those values of R and S only cases. The KIRP data set also showed a similar pattern. For every measure, the aggregation of multiple omics data resulted in better performances in Table 2, which confirms the benefits of the aggregation of multiple omics data. Figure 1 reports correlation between precision/recall and the average information gain of each evaluation cases (“Materials and methods”). In general, the precision/recall and the information gain have positive correlations in both of the BLCA and KIRP data sets. For instance, the Pearson correlation coefficients between precision and information gain were 0.62 and 0.71 for BLCA and KIRP respectively, and 0.65 and 0.74 for BLCA and KIRP respectively in the comparison of recall and information gain. Overall these results confirmed the suitability of the considered platforms or their combinations for in silico classification of tumor and normal tissue samples.

Table 2 Information gain and classification accuracy
Full size table
Fig. 1
figure 1

Correlation between information gain and classification accuracy

Full size image

Effect of multi-omics data combinations

The information gain-based multi-omics data analysis produced mixed results. As described in the previous section, the aggregation of two omics data resulted in higher information gain in almost all the cases for both of the BLCA and KIRP data sets. However, the improvement in precision/recall was less clear although all average performance of multi-omics data sets outperformed the performance of single-omics datasets. In terms of precision/recall, the combinations of multiple-omics datasets produced improvement for the omics data with lower information gain. The most striking example was somatic mutations. The precision/recall performance of somatic mutations were 0.14/0.14 and 0.18/0.19 for the BLCA and KIRP datasets, respectively. These low precision/recall were improved to 0.91/0.89 and 0.93/0.93 for BLCA and KIRP respectively through the combination with DNA methylation. The low information gain due to the lack of discriminating process to identify significantly mutated genes was overcome with the combination of other omics-data. For somatic mutations, the improvement in precision/recall were also observed in the cases of CS and RS in both of BLCA and KIRP. For the remaining cases, the combination of multiple omics data tends to be beneficial for the omics data with lower information gain and not very useful for the omics data with higher information gain in general. The example cases are CM, MR, and MS for BLCA, and CM, CS, MR, MS and RS for KIRP. There were cases where the combination of multiple omics data produced better precision/recall than each single-omics data set. In BLCA, the combinations of CS and RS were beneficial for each one of single omics data. In KIRP, CR combinations produced better precision/recall than single omics data sets. Basically, the information gain represents the imbalance in the distribution of attributes across samples. Therefore, the pairing with an attribute with higher information gain is beneficial only for an attribute with lower information gain. This is because the pairing induces more skewed distribution of attributes. For the same reason, the pairing is not beneficial for an attribute with higher information gain.

An important benefit of the multi-omics data-based approach is that it can prevent a “large” platform (with a large number of features) from dominating a solution (Hoadley et al. 2014). Information gain-based approach produced similar effect. An omics-category with lower information gain suffered from the too balanced distribution of attributes among tumor-normal tissue samples. The aggregation with the imbalanced attribute such as DNA methylation induced in more skewed distributions and resulted the improvements in the in silico classification-performance.

Role of genes with higher information gain

For the qualitative analysis of the genes with higher information gain, we investigated the role of genes with higher information gain in the development of target diseases or tumorigenesis in general. For BLCA, we selected 44 genes from the ten omics data-combinations. We investigated 36 genes for KIRP from eight omics data-combinations. Those selected genes were first checked using the CaGe (Park et al. 2012) and GeneCards (http://www.genecards.org) database (Rebhan et al. 1997). A gene covered by one of two databases is regarded as target diseases or tumorigenesis-related genes, and its role was summarized. Genes not covered by two databases were also checked through literature searches. The list of relevant genes is shown in Table 3. We also investigated pathways related with the higher information gain genes by using the NCI-Nature Pathway Interaction Database (Schaefer et al. 2009). The results are listed in supplementary Table 1.

Table 3 The list of genes with higher information gain and their relevance to tumorigenesis in general
Full size table

In the case of BLCA, 32 genes were relevant to tumorigenesis in general. Among the relevant genes, SLC6A6 was associated with higher information gain in terms of CNV. It was found that SLC6A6 is important for the maintenance of side population cells and their cancer stem cell properties. It was suggested that SLC6A6 signaling is a significant player in the survival and maintenance of cancer stem cell population and its capacity for tumor initiation, starvation tolerance and multidrug resistance (Yasunaga and Matsumura 2014). ZNF154 was another informative gene in terms of DNA methylation. ZNF154 encodes a protein belonging to the zinc finger Kruppel family of transcriptional regulators, whose members are deemed to function in normal/abnormal cell growth and differentiation. The methylation of ZNF154 was validated as a tumor marker gene (Reinert et al. 2011) and it was shown that it is possible to detect a concomitant tumor recurrence with a single marker ZNF154 (Reinert et al. 2012). RBP7 was highly informative in terms of DNA methylation. RBP7 (also named CRBP-IV) is the member of the cellular retinal-binding protein family and it was shown that transcription silencing of this gene by aberrant methylation is involved in the tumorigenesis of human cancers (Kwong et al. 2005). The possibility was raised that the mutation of RBP7 impaired retinoid’s function in breast cancer cells where tamoxifen-induced ZR-75-1 cell death requires intact retinoid signaling (Zarubin et al. 2005). PI16 was associated with higher information gain in the case of RNASeqV2 and the combination of RNASeqV2 and somatic mutations. It was suggested that PI16 would be a tumor suppressor and a metastasis enhancer because the cell lines ectopically expressing PI16 display a net increase in the rate of secondary lesion (Crawford et al. 2008). MTUS1 was selected considering the combination of CNV and somatic mutations. A hypothesis was suggested that MTUS1 is involved in the regulation of tumor progression in various malignant diseases including human colon carcinoma. MTUS1 is down-regulated in undifferentiated tumor cell lines and inhibits tumor cell proliferation after recombinant over-expression (Zuern et al. 2010).

In the case of KIRP, 22 genes were regarded as relevant ones. For example, JDP2 reported higher information gain in DNA methylation. The role of JDP2 is prominent in the regulation of the differentiation and proliferation of cells. The overexpression of JDP2 inhibits the retinoic acid-dependent differentiation of embryonic carcinoma F9 cells (Jin et al. 2002). In mice, the overexpression of JDP2 induces arrest of the cell cycle. The absence of JDP2 decreases the expression of both p16Ink4a and p19Arf, which inhibits progression of the cell cycle. It is supposed that JDP2 not only inhibits the transformation of cells but also plays a role in the induction senescence. These two functions imply that JDP2 might act as an inhibitor of tumor formation (Nakade et al. 2009). TMEM207 achieved distinctively higher information gain considering RNASeqV2. TMEM207 facilitates tumor invasion possibly through binding to WWOX (WW domain-containing oxidoreductase), a molecule plays an important role in the regulation of a wide variety of cellular functions such as protein degradation, transcription, and RNA splicing. Human TMEM207 was found to be overexpressed in many aggressive gastric signet-ring cell carcinomas and TMEM207 expression is relatively restricted to the kidney physiologically (Kito et al. 2014). VCP was associated with higher information gain considering somatic mutations. VCP regulates various cellular processes such as chromatin decondensation, homotypic membrane fusion, and ubiquitin-dependent protein degradation by the proteasome. Interference of proteasome inhibitors with the ubiquitin proteasome pathway leads to the accumulation of proteins engaged in cell cycle progression, which ultimately put a halt to cancer cell division and induce apoptosis (Rastogi and Mishra 2012; Tresse et al. 2010). BCL11B was a highly informative gene in the domain of somatic mutations. The BCL11B gene is responsible for the regulation of the apoptotic process and cell proliferation. BCL11B has recently been identified as a tumor suppressor gene. In particular, BCL11B is known as a haplo-insufficient tumor suppressor, the absence of BCL11B resulted in vulnerability to DNA replication stress and damage, and down-regulation of BCL11B gene by siRNA (small interfering RNA) led to growth inhibition and apoptosis in a human T-ALL cell line (Huang et al. 2012). EPO was associated with higher information gain considering the combination of CNV and RNVSeqV2. Tumor necrosis factor-alpha (TNF-alpha) selectively kills tumor cells in vitro and in vivo. It was shown that EPO could be used to prevent TNF-alpha-induced erythroid suppression (Johnson et al. 1990) (Supplementary Material for the details of other relevant genes).

Information gain-based analysis of putative significantly mutated genes

Genomic variant causing ‘gain of function’ or ‘loss of function’ plays a key role in cancer diagnostics and targeted therapy (Krishnan and Ng 2012). Therefore, the identification of potential cancer drivers is another role of cancer genomics. The Cancer Genome Atlas Research Network analyzed urothelial bladder carcinoma. As a result, 29 significantly mutated genes and 20 genes with statistically significant focal copy number changes were identified (The Cancer Genome Atlas Research Network 2014). For 29 significantly mutated genes identified in (The Cancer Genome Atlas Research Network 2014), we collected their information gain from our experiment results and found that the average information gain was 0.254. The maximum information gain is 0.705 of RHOB and the minimum information gain is 0.0367 of FOXA1. Because information gain is quantified by the distribution of attributes in tumor and normal tissue samples, if an attribute is evenly observed in tumor and normal tissue samples then the attribute would have low information gain. As a result, the 29 identified genes reported relatively low average information gain with high standard deviation of 0.146. Among 20 genes with statistically significant focal copy number changes, the average information gain is 0.294. The maximum information gain is 0.610 of CCNE1 and the minimum information gain is 0.0728 of CCND1. The most differentially regulated 48 genes in renal cell carcinoma were identified (Beleut et al. 2012). For the identified 48 genes, we observed information gain of genes excluding genes no longer serviced by NCBI (http://www.ncbi.nlm.nih.gov/). For the remaining genes, the average information gain is 0.175. The maximum information gain is 0.759 of PTPRO and the minimum information gain is 0.0198 of SMARCA4. The complete list of significant genes and their corresponding information gain values are provided in Supplementary Tables 2 and 3.

Conclusions

In this paper, we investigated the potential of information gain for analysis of biomedical datasets generated from multiple platforms. The quantitative analysis based on the concept of information gain showed that the utilization of multiple-omics data is beneficial for in silico classification of tumor-normal instances. Furthermore, the experimental results reported that the classification power of each omics data and their combinations are very distinct. The qualitative analysis based on previous researches also verified the usefulness of the concept of information gain. We verified the relevance of genes with higher information gain to tumorigenesis through literature search.

However, this research also revealed the weakness of the information gain-based approach. Basically, the concept of information gain employed in this research utilizes the distribution of attributes across classes or categories of a target disease. As a result, our research was able to find genes whose expression pattern is biased to tumor or normal samples but it was unable to find “significantly” mutated genes as in (Beleut et al. 2012; The Cancer Genome Atlas Research Network 2014). This weakness was expressed during quantitative analysis. Although, it was confirmed that the majority of genes with higher information gain are tumorigenesis-relevant genes, the computed information gain of significantly mutated genes or genes with significant focal copy numbers reported relatively lower information gain. These findings suggest a number of useful guides for future researches. Firstly, the information gain-based approach would be useful for limiting candidates for detailed analysis. For example, DNA methylation is superior to other omics data for in silico classification of tumor-normal samples. Therefore, DNA methylation focused approach would produce tumor-intensive expressed genes. Secondly, a novel method is needed to reflect prior knowledge. Current approach is unable to recommend significantly mutated genes without incorporating prior knowledge. Finally, a novel approach is needed to utilize the unbalanced composition of biomedical data sets. For the same target cancer, the omics composition of each data instance from TCGA is very different. As a result, data instances lacking observations on the target omics data should be ignored in this study. In future works, we will focus on developing a novel method capable of utilizing data instances with different composition of data sources.