Gene transcript profiles.

Global gene expression profiling with DNA microarrays [9–11] has furthered our understanding of the regulation of biological processes and has become an indispensable tool in the study and classification of human tumors. Semiquantitative profiles of gene expression have been measured for many cancer types and subtypes [12,13]. Through unbiased comparative analysis of these profiles, a subset of genes can be found that correlate with tumor phenotype and can serve as diagnostic and prognostic markers of disease. Disease-specific regulatory programs can be studied using techniques such as chromatin immunoprecipitation (ChIP) of tumor biopsies [14]. Several large public compendiums of gene expression data generated from diverse experimental methods also exist, some examples of which are presented in Table 2.

Table 2

Gene Expression Repositories

Open in a separate window

Not all available cancer microarray data are generated from gold-standard tissues or primary cell culture. Often, immortalized cell line models of neoplastic disease are studied because they are easier to access than histopathologically characterized human tumors [15,16]. However, in vitro extended passaging of cell lines could lead to accumulation of alterations and yield less representative expression profiles and less stable disease phenotypes [17]. Additionally, the in vivo microenvironment affects tumor–host interactions and causes variations in gene expression and pathways. These changes make disease-specific expression patterns difficult to infer from cell line data alone. Nevertheless, mining and analysis of the large amount of available cell line gene expression data promises new insights into disease conditions.

Protein levels and modifications.

Mass spectrometric instruments and protein chip technology allow large-scale analysis of proteins, their quantitative expression, interactions, post-translational modifications, and localization [18–25]. Proteomic profiling of clinical samples ranging from tissues to biofluids (e.g., urine, sera, plasma, whole blood, cerebrospinal fluid, and saliva) will help assess disease development and progression, generate diagnostic and prognostic disease markers, and predict patient response to intervention. This information can be used to identify regulatory networks and activated signaling events in biological pathways, and can help characterize pathologically benign and tumor tissue samples (Figure 2). Public proteomics repositories that collect this data are now available, and relevant examples are listed in Table 3. Accurate protein identification involving processing and identification of mass spectral peaks, peptide sequencing, search algorithms, and statistical validation of correct assignment of peptides and proteins is challenging, often confounded by splice variants or other protein isoforms [26–30].

Open in a separate window

Figure 2

Protein Profiling for Cancer Diagnosis and Prognosis

Generation of protein profiles using mass spectrometry is an example of an experimental technique that produces massive amounts of data that is difficult to interpret without computational and statistical algorithms. For instance, comparison of disease versus control sample profiles can lead to identification of disease-specific protein expression signatures, which could be used as diagnostic or prognostic markers. Aggregation of such data from multiple sources and pooled analysis requires proper annotation of sample source, sample handling, and experiment information.

Table 3

Proteomics Repositories

Open in a separate window

Small RNAs.

Small RNAs add an additional layer of complexity to gene regulation [31]. Initially discovered in plants and C. elegans, at least four subfamilies exist: microRNAs (miRNAs), short interfering RNAs (siRNAs), tiny noncoding RNAs (tncRNAs), and small modulatory RNAs (smRNAs). miRNAs are small, ~21–24, nucleotide noncoding endogenous RNAs involved in translational repression and messenger RNA (mRNA) cleavage, which can potently downregulate translation of specific mRNAs by targeted 3′-UTR binding. An increasing number of miRNAs have been implicated in disease. For instance, a small cluster of miRNA genes on Chromosome 13, c13orf25, appears to be involved in B cell lymphoma [32]. Cell cycle regulation has been shown to involve a miRNA regulatory circuit where c-Myc, a known proto-oncogene upregulating E2F1 (a cell cycle regulator), also regulates the expression of six miRNAs on Chromosome 13, two of which have been shown to downregulate E2F1 [33]. In the case of acute lymphoblastic leukemia (ALL), in studies by Lu et al. (2005), miRNA transcriptional profiles have been shown to better classify tumors by type and developmental stage than mRNA profiles [34].

mRNA targets for several hundred miRNAs that are expressed in human have been computationally predicted, though few targets have been experimentally confirmed ([35,36], http://www.microrna.org). The lack of perfect base-pair complementarity between the miRNA sequence and its target and the short length of the miRNAs make accurate prediction of miRNA genes and targets difficult.

Pathways.

Pathway information is vital for understanding biological processes and how they are disrupted or reprogrammed in disease. However, collecting complex pathway information in a usable form from diverse and heterogeneous sources, including more than 220 pathway databases (http://pathguide.org), is a major challenge [37]. A number of pathway database efforts seek to ameliorate the situation by making pathway data more accessible for computational analysis (Table 4). For instance, Memorial Sloan-Kettering Cancer Center and the Institute of Bioinformatics have collaboratively developed ten large cancer-focused signaling pathways (http://cancer.cellmap.org).

Table 4

Human-Focused Pathway Repositories

Open in a separate window

Metabolite profiles.

Metabolomics involves measurement of metabolite concentrations and fluxes in cells and tissues [38]. Such measurements provide insight into the response of biological systems to genetic and environmental influences. Metabolites provide important markers for disease state and the pathways underlying drug metabolism. Metabolomic profiles can be used for classifying disease by type and stage, for prognosis, and for testing the effectiveness of therapeutics using statistical learning methods. For example, high-resolution magic angle spinning NMR (HRMAS–NMR), which can quantitatively identify a range of metabolites while leaving the tissue sample intact for further studies, has been used to profile and classify prostate tumors [39] and to detect drug efficacy in liposarcoma [40].

Biological images.

Advances in optics, digital detectors, and automation have significantly improved biological imaging technology and have led to a large increase of quantitative information extracted from digital images. Fluorescent and confocal microscopy, and whole-body imaging of model organisms [41,42] can be used to test specific hypotheses of cellular function and disease. Deep tissue–penetrating infrared light and various alterations of two-photon laser scanning microscopy (2PLSM) have been successfully used to reveal the dynamic nature and spatio–temporal aspects of hematopoietic tissue [43], organ development [44], and neurobiology [45]. Fluorescent proteins and photo bleaching techniques enable visualization of protein localization, protein–protein interactions, and protein fate in vivo [46]. In clinical settings, particularly for solid tumors, better resolution and higher contrast dyes have allowed the use of magnetic resonance imaging (MRI), computed tomography (CT) scan, positron emission tomography (PET) scan, and ultrasound for diagnosis and tracking disease progression. As imaging data is highly context-specific, data aggregation from multiple sources is possible only if sufficient metadata on samples, microscope, and data derivation algorithms are available.

Data integration.

Data integration is the combination of heterogeneous biological data encoded with different semantics. Integration of heterogeneous data is useful not only to validate and to improve confidence in experimental results but also to develop more complete models of biological systems. For instance, real time quantitative RT–PCR data are routinely used to validate cDNA array experiment results. Integration of gene expression and proteomics data, for example, could be used to identify post-transcriptional or post-translational modifications. It could also provide insights into the advantages and shortcomings of particular experimental methods.

The integration of diverse experimental data to build models of biological processes, or pathways, will boost our ability to identify clinical markers and therapeutic targets and to interpret genotype information. For instance, a marker such as prostate specific antigen (PSA) may be widely known and used clinically without much knowledge about its biology. Knowing the pathway involving the marker gene allows other pathway components, or the entire pathway, to serve as a more specialized marker.

Clinical data, securely and ethically accessed, can be integrated with molecular data from basic research to gain insight into disease state and lead to better treatments. Molecular and clinical data has been integrated for identification of clinically relevant subtypes of leukemia with 100% sensitivity and specificity [54]. Analysis of molecular profiles on biospecimens from patients before, during, and after therapy can lead to identification of drug-responsive or nonresponsive profiles that could be used to optimize choice of therapy. In the case of advanced non-small cell lung cancer (NSCLC), a significant difference in response to the kinase inhibitor gefitinib (Iressa) was observed for patients with mutations or amplifications in the EGFR gene [55,56]. Comparison of patient molecular profiles with poor and favorable outcomes can be used to predict disease outcome (prognosis). For example, in breast cancer patients, the estrogen-receptor status of the primary tumor and other clinical features have been used to construct nomograms that predict the likelihood of developing non-sentinel lymph node (non-SLN) metastases. Such information can be used to assess metastatic risk and the need for complete axillary lymph node dissection [57]. Additionally, a 32-gene expression signature distinguished p53-mutant and wild-type breast tumors of different histologies, and this strategy outperformed sequence-based assessments of p53 in predicting prognosis and therapeutic response [58]. Combination of molecular profiles with clinical profiles can also help select patients for targeted treatment.

Effective integration of heterogeneous data is difficult since important information necessary to decipher data semantics, such as context, could be missing. It is often possible for the human brain to infer this information using prior knowledge, but such tasks have remained impossible to encode into a model or rule-based computational procedure. Thus, missing information can lead to errors during integration. For example, a query may identify relevant datasets labeled with the term metastasis. However, metastatic processes could be different among tissues, so tissue information is required to avoid errors in dataset selection. Also, datasets labeled with alternate descriptions of metastasis may be missed in the search. As another example, proper mapping of genes and proteins between microarray and mass spectrometry data is an important requirement for integrating the datasets. This task could be confounded if the measured gene transcript is a different splice variant than the measured protein product. Thus, correct data integration requires semantic compatibility among datasets and context resolution.

In Box 2 we present two successful examples of how integration of heterogeneous data from different sources can help in biological discoveries. In the first example, publicly available data was used to help identify LRPPRC as a specific gene involved in Leigh syndrome, a complex hereditary disease. In the other example, the discovery of the central role of micropthalmia-associated transcription factor (MITF) amplification to malignant melanoma was accomplished using experimental and computational approaches to integrate copy number data, with publicly available gene expression and genome information. These examples provide proof of principle that there already exists a wealth of biological data in the public domain, and data integration approaches can be used to better understand biology and development of disease, including cancer.

Software systems and algorithms for data analysis.

The volume of the data generated by modern biomedical studies is too large to be processed by the human brain alone. Data storage, querying, and presentation software systems and computational algorithms are required for effective interpretation of large-scale experimental data. Automated methods are now available to find genes or pathways that are significantly differentially expressed using molecular profiles [59–63]. The GOALIE algorithm maps the temporal evolution of biological processes from time-course gene expression data [64]. An algorithm for general integration of heterogeneous data that differ in type and size has been proposed [65]. Further studies of pathways in the context of gene expression and complementary genome-scale data will lead to the discovery of new pathway components, some of which could be new vulnerable points and therapeutic targets [66].

Pathway simulations, requiring detailed cellular models, have been used in model organisms, such as Escherichia coli and budding yeast, to find pathway regulators and to design new experiments that test hypotheses about the function of the pathway [67]. Applying these methods to predict the result of a tumor-specific mutation in multicellular organisms using human pathway information is an important research direction and major challenge for computational biology.

Development of software systems that integrate diverse biomedical research data types promise to support the study of disease biology and development. For instance, the REMBRANDT (Repository for Molecular Brain Neoplasia Data) framework attempts to integrate clinical and molecular data from the Glioma Molecular Diagnostic Initiative (GMDI)—a collaborative effort of the NCI and the US National Institute of Neurological Disorders and Stroke (NINDS) (http://rembrandt.nci.nih.gov). REMBRANDT can be used to query and generate statistical reports across all component glioma (brain tumor) datasets.

Computational prediction of the biological effects of drugs based on structure–function relationships across many targets can help increase the success rate of clinical trials and may forewarn of possible adverse events associated with the small-molecule therapy. This strategy could also be used to develop drug combinations to target multiple vulnerable points to shut down tumor growth. ADME/Tox (absorption, distribution, metabolism, excretion, and toxicity) prediction based on molecular profiles can help eliminate candidate drugs that have unacceptable toxicity early in the drug discovery process and thus reduce the cost of drug development.