The primary study goals of cancer genomics are not simple, which include the studies to get either clinical or mechanistic insights from the cancer genome. To clarify the complicated study designs and strategies of cancer genomics, we have categorized the gene signatures obtained from cancer genome data into four classes prediction, phenotype, function, and molecular targets based on the study goals [25]. The majority of the previous studies have suggested the translational or clinical utility of the genomics data by addressing the candidate biomarkers or the prediction signatures for predicting patients clinical outcomes, such as recurrence, survival, metastasis, or response to therapies. Notwithstanding the overwhelming identification of candidate biomarkers from the cancer genome, only a handful of candidate biomarkers that have been discovered from genomic analyses can succeed in the validation of the clinical utility [18]. There are several challenges in cancer genomics that preclude clinical utility. One of them would be data reproducibility. They might be due in part to the experimental biases as well as sample cohort issues. The use of different platforms measuring gene expressions and different data processing methods could produce biased observation in each study. Increasing sample size will be one of the solutions to find proper biomarkers, overcoming the reproducibility problem. Undoubtedly, large-scale sample collection provides increased statistical power. However, previous studies, even with large sample sizes, have often failed to reproduce their findings in independent studies [26]. This might be due mostly to the use of different protocols and analysis methods. Moreover, biased sample collection may also affect the performance of prognostic biomarkers, leading to subsequent failure to validate the biomarker in another patient population [27]. For example, diagnostic biomarkers must be discovered in early-stage tumors; however, the sample collection of early tumors with enough of a sample size might be difficult in the clinical setting [18, 28]. In addition, the cost-effectiveness of the sample size enrolled in a study should be considered. Simply increasing the sample size might not be the best solution.

The sample sources and qualities are also important factors to be considered in the study design. For example, circulating DNAs or microRNAs in the plasma or urine can be used to develop "noninvasive" biomarkers in cancer patients, which might bring the technology much closer to the clinic [29, 30]. Attempts to use formalin-fixed paraffin-embedded (FFPE) tissues might also be more applicable to the clinic [31], although the quality and the quantity of the DNAs or RNA extracts from FFPE or plasma are still problematic for genomics studies requesting further elaboration.

The data complexity comes not only from the heterogeneous or biased sample composition, but also from the innate complexity of tumor biology. Previously, mounting evidence has shown the enormous heterogeneity of tumors at the molecular level. The tumor heterogeneity can be explained by two hypothetical models. One is the clonal segregation model with a multi-step process, and the other is the cancer stem cell theory. The cancer stem cell model describes the heterogeneous cellular origin of cancers from primitive progenitor cells to mature differentiated cells, which may contribute to tumor heterogeneity. In this context, genomic profiling studies could define the cancer subpopulation harboring stem-like traits in various cancer types, supporting the cancer stem cell theory [32, 33]. Similarly, we also defined the bilineal trait in a subpopulation of hepatocellular carcinoma (HCC) by comparing the gene expression profiles of HCC and cholangiocarcinoma (CC) [34]. This result showed the continuous liver cancer spectrum between HCC and CC, suggesting that stem-like or de-differentiation traits may give rise to the heterogeneous progression of HCC. We also suggested that the dysfunction of p53 machinery is associated with the acquisition of the stemness trait in HCC [35]. Recently, this association could be validated by showing p53 knockout mouse model can give rise to bilineal liver cancers [36].

In addition, various host factors may contribute to tumor heterogeneity. Interactions of the tumor cells with host eco-system, such as innate immune systems, or the reactions of surrounding microenvironment against the tumor may affect the tumor behaviors [37]. Thus, the proper detection of biomarkers might be difficult without considering the effect of host factors. Furthermore, the intratumoral heterogeneity of cancers have been notified in detail by genome-wide sequencing of multi-loci from the same tumor [38]. The comparison of primary and metastatic tumors by single cell sequencing also revealed the sequential mutation process during cancer progression [39]. Similarly, the comparison of mutations from multiple HCC tumors in the same patient could define the evolutionary lineage among tumors cells [40]. Strikingly, the development of single-cell sequencing technology could provide a more detailed and systematic view on intra-tumoral heterogeneity [41, 42]. Of interest, such attempts enabled the construction of phylogenic trees from mutational heterogeneity, which revealed evolutionary tumor growth opening a new field of "cancer evolution".

As discussed above, there are many factors contributing to tumor heterogeneity, which may impede the discovery of new biomarkers. Thus, we are now urgently in need of developing new strategies for biomarker discovery from cancer genome data. Considering the huge complexity of the cancer genome and the limitation of current technologies, it would be a reliable strategy to evaluate the functional relevance of the candidate biomarkers rather than simply showing the statistical significance of the association by enrolling larger samples or applying more stringent statistics. Although the conventional strategies for biomarker discovery do not require the functional significance of the candidates [43], current hurdles in cancer genomics request a functional validation step in the pipeline of the biomarker discovery. Our limited understanding of the complexity of cancer biology is a significant challenge for translational interpretation of the cancer genome.

More recently, systematic structuring and integration of multiple and multi-layered omics data resources, i.e. integromics, are thought of a state-of-the-art strategy. The recent establishment of The Cancer Genome Atlas (TCGA) and the International Cancer Genome Consortium (ICGC) could accelerate and facilitate the integration and sharing of cancer genome data [44]. Multi-layered integromics could define tumor heterogeneity, revealing pivotal aberrations of genetic events or signaling pathways [45]. In parallel, genomic repositories for drug activities have been established [46-48]. Linking the profiles of drug sensitivity to the cancer genome could provide a powerful platform to guide rational and personalized cancer therapeutics.

Now, as genomic data are increasing and accumulating enormously, new study designs and analysis strategies for integromics might be required, particularly with the context of tumor heterogeneity and the discovery of the functional biomarkers. As shown in Fig. 1, the first step in cancer integromics is the dissection of tumor heterogeneity. Then, the next step will be a recapitulation of the relations between clinical/biological phenotypes and molecular genotypes in cancer subpopulations, which can address novel functionalities in particular subpopulations of cancers. We suggest that the discovery of reliable candidate functional biomarkers as well as functional genetic alterations, so-called "driver events", can be achieved through performing this step-by-step evaluation of both clinical and functional utilities. This hybrid study design would open an exciting era for developing new "functional biomarkers" and preventive/therapeutic strategies with the consideration of biological backgrounds of tumor heterogeneity.