Perspective: Microarray Technology, Seeing More Than Spots

The publication in the mid-1990s of a new molecular tool, the DNA microarray, has led to a revolution in the way scientists approach the investigation of gene expression and regulation. This technology has made its impact upon many basic scientific disciplines including cancer biology, developmental biology, toxicology, investigation of growth-factor and hormonal signaling, and the applied areas of disease diagnostics and drug development. The ability to assay thousands of genes simultaneously in a high-throughput manner across RNA samples derived from various biological sources and treatments has increased the need for integration of higher-order statistical analyses and data management schema into molecular biology laboratories. Many investigators are struggling with interpreting large, complex datasets, communicating their findings, and developing strategies for following up numerous potential leads indicated in datasets. This perspective provides an overview of the technology and its impact, as well as strategies for study design, data analysis, and avoidance of potential pitfalls. In the mid-1990s, a multidisciplinary team at Stanford University published a landmark paper initiating a new era in the realm of gene expression studies (1). This paper described a new technology allowing the quantitative, simultaneous monitoring of the expression of thousands of genes using a new tool termed a DNA microarray. Since then, numerous papers have been published reviewing the technology and its potential applications for scientific discovery in a broad range of biological disciplines. While reviewing all of the literature and findings related to this exciting technology is not within the scope of this perspective, a broad overview of the technology will be provided with a focus on study design and data analysis. Thus, investigators who have not yet ventured into this new world of the DNA chip will have a guide to develop a satisfying microarray study. Although the DNA microarray chip is a tool most commonly used to monitor the level of expression of a gene at the RNA level (1–3), it also documents DNA copy number (4–6) and DNA protein interactions (7, 8) and sequencing applications (polymorphism detection). This perspective concentrates on the application of DNA microarrays for monitoring the expression of a gene by measurement of its ability to hybridize to a target sequence localized to a specific region on a chip. To measure this hybridization, RNA, extracted from a biological sample of interest, is reverse-transcribed into cDNA that ideally represents a quantitative copy of genes expressed at the time of sample collection. This cDNA is labeled with a tracking molecule such as a radioactive or a fluorescent nucleotide, or an affinity molecule like biotin. The labeled cDNA is then hybridized to the DNA chip that contains thousands of gene targets. Ideally, each molecule in the labeled cDNA will only bind to its appropriate complementary target sequence on the array. Quantitative imaging coupled with clone database information allows measurement of the amount of labeled cDNA that hybridized to each target sequence, resulting in the identification and relative quantification of the genes expressed in the original biological sample (3, 9, 10). Several varieties of the DNA microarray are used frequently throughout the scientific community. These include deposition or spotted cDNA, spotted oligomer, or synthesized oligomer chips. The cDNA microarray is comprised of a collection of partial gene sequences that are spotted individually into precise locations within the DNA chip (11). These sequences usually range in size from 500–2000 bp and may be chosen from different regions of the gene depending on the goal of the project. The selection of sequences from the 3′ untranslated region of the gene confers the advantage of specificity to the genes being measured. Multiple genes within a conserved functional family may have a high degree of sequence similarity, especially in domains responsible for catalytic functions. When using a microarray to monitor gene expression, ideally the investigator must detect the precise gene that is affected without cross-reactivity with other family members. By selecting a sequence that contains a majority of 3′ untranslated region, a researcher can take advantage of sequence diversity in this region that may be gene specific and not conserved among related family members. Furthermore, in recent applications, investigators have selected promoter sequences to generate custom chips to monitor binding to specific promoter regions (7, 8). While the long target sequences of cDNA microarrays provide an advantage for sensitivity of detection of gene expression, the potential problems with specificity must be considered (12). If there is significant sequence overlap between clones representative of multiple genes (>70%), such as in related enzyme families (i.e. kinases, cytochrome p450s, etc.), then the discrete transcript may not be specifically detected; rather, multiple, related sequences might be simultaneously detected (12). An additional advantage of these cDNA chips is their ease of use and attainability. Complete instructions on their manufacture within a laboratory are freely available (http://cmgm.stanford.edu/pbrown/mguide/nospecs2.html). The ability to manufacture chips within a research laboratory or an institute provides the advantages of flexibility and customization of design as necessitated by the scientific goal of the project. Another form of DNA microarrays is based on deposition or on-chip synthesis of oligonucleotides for generation of targets. Several approaches to the manufacturing of these chips (13–15) lead to the same end result—a chip that contains short oligomers ranging from 25–80 bases as the target sequences. While these shorter sequences can confer high specificity, they may have decreased sensitivity/binding compared with cDNA arrays. Users of these arrays compensate for sensitivity problems by employing multiple sequences for each gene. One disadvantage of these oligomer-based chips is their availability only from commercial manufacturers. The cost of custom development of an oligomer-based chip is beyond the budget of a majority of researchers; but more affordable, premade oligomer-based chips are available and may be sufficiently informative for investigators seeking to employ this type of array. The impact of the microarray technology has proven tremendous because it has enabled investigators to progress from studying the expression of one gene in several days to hundreds of thousands of gene expressions in a single day. For the first time, investigators can relatively quickly measure the expression of a complete genome (16–19) across a large number of environmental stimuli. This awe-inspiring technological breakthrough has the potential to impact some previously intractable scientific realms and aid in the elucidation of complex models and systems. Tracking the numbers of publications in PubMed reveals that, since 1995, over 2000 papers have included this technology. Scientists in the arena of cancer biology have been some of the most industrious at incorporating this technology into their studies (Fig. 1). While generous funding in this arena may have contributed to greater usage of microarrays, this technology is, nonetheless, ideally suited for the comparison of gene expression in various stages of tumor development (2, 20–24) and monitoring expressions of premalignant and tumorigenic cells following exposure to anticancer agents or tumor promoters (25–32). Several laboratories (25, 33) demonstrated that gene expression profiling of patient lymphoma samples improved distinction of tumor classification and provided insight on clinical outcome. These and other data indicate that one important benefit of this technology might be to inform clinicians of better, specific markers for cancer diagnosis, prognosis and treatment (23–25, 31, 34). Increased numbers of scientific studies employing microarray technology. A search of Medline by the following key words revealed a growing number of publications referencing microarray technology. Search terms were: DNA microarray (▧), DNA microarray and cancer (▪), and DNA microarray and hormone (□). Other basic scientific disciplines have also benefited from global gene expression profiling. Investigations using microarrays have provided a more comprehensive view on the complex regulation of the cell cycle (16, 35), differentiation (36–38), and aging (39–41). Other studies involving organisms associated with infectious or other noncancerous diseases have elucidated alterations in gene expression that may prove beneficial in treatment (42–44). An understanding of the mechanism of action of pharmaceutical agents, their potential adverse outcomes, and the prediction of activity of unknown agents will undoubtedly significantly impact the development of safer and more efficacious drugs (45, 46). While microrarrays allow scientists to gaze across the genome at the response of an organism to a biological stimulus, they have an equal potential to mislead if false assumptions are made or flaws in study design are overlooked. At present, there is no one standard guide or consensus for the design and conduct of an expression-profiling experiment because the ultimate goal of the project dictates the study design. Some considerations, however, are worth noting (Table 1). Important considerations in the design of a microarray experiment Important considerations in the design of a microarray experiment The data from a single microarray hybridization, while containing a plethora of valuable information, will also be filled with many potential false positives and negatives resulting from technical processing steps or part of the data analysis process. To achieve confidence in a dataset and the most sensitive limit of detection, replicate hybridizations—both biological and chip—should be analyzed (47–50). In all disciplines of the biological sciences, including gene profiling, the reproducibility of an observation has been and must continue to be a requisite for publication. A good study design will strategize to incorporate biological replicates into the chip analysis process; such inclusion is not apparent in all microarray-related publications. The use of biological replicates in the analysis schema enables detection of gene-expression changes that may be more related to environmental alterations than to the stimulus or biological parameter being monitored. In addition, examination of as many biological parameters as possible will enable the investigator to account for variations in experiments and obtain accurate results. For example, simultaneous measurements of DNA synthesis, cell cycle progress, viability, and apoptosis might be employed as standard measures to gauge the health and growth of cultured cells and ensure that these parameters are consistent within replicates. For clinical samples, notations on patient history and health, tissue handling upon necropsy or surgery, clinical chemistry, and histopathology will collectively aid in the ultimate interpretation of the gene-expression profiles by providing a phenotypic anchor or description of sample physiology that provides a biological context for the sample from which the RNA was derived (51). For example, these descriptions will provide additional information to aid in integration and interpretation of the gene expression profile as related to proliferation, apoptosis, necrosis, growth arrest, differentiation, or other physiological states. In the optimal study design, multiple chips for each biological sample should be assayed (50). This safeguard will provide the investigator the best dataset for balancing sensitivity with confidence. In the case of two-color fluorescent cDNA microarray hybridization, a dye swap should be conducted to filter out artifacts that could be attributed to unequal incorporation or quenching of the dye molecules. Historically, many investigators have used an arbitrary fold cutoff as a measure of significance of a gene expression change. This approach poses many potential problems, including the gross overestimation of the number of significant gene changes in a dataset that contains a wide distribution of alterations (e.g. two diverse cell types like normal and cancerous tissue; Fig. 2A); or the severe underestimation of gene expression changes when a dataset is derived from a tightly linked comparison (Fig. 2B), as in the case of an early time point following hormonal stimulation (51A ). Numerous, statistics-based, analytical approaches have been designed to gain the most sensitivity in detecting gene changes while simultaneously providing a measure of the potential error associated with the statistical method. Statistical measurements of significance can be employed at both the single- and multiarray levels (52–57). Properties of data distribution in various sample comparisons. A, Distribution of nonrelated biological samples. The distribution of gene expression changes between nonrelated samples is wide. Use of a 99% confidence interval indicates a statistical significant cutoff at 2.51-fold. B, Distribution of related biological samples. The distribution of gene expression changes between closely related samples is very tight. Use of a 99% confidence interval indicates a statistical significant cutoff at 1.43-fold. Image analysis was conducted using Microarray suite of IP Labs (Scanalytics, Arlington, VA). The output from a microarray hybridization is usually a large data spreadsheet filled with numbers related to the distribution of the signal and background intensities for each gene on the chip. Most investigators convert these data to a more visual format for interpretation. Many visualization tools are available, both academically and commercially. One of the first tools applied to microarray data was clustering (58), which provides a two-dimensional hierarchical grouping of data points. In one dimension, assayed samples are grouped according to their similarity; in the second dimension, genes are grouped according to the overall similarity of their expression patterns across the samples. This type of analysis is useful for visualizing groups of genes that are similarly regulated across the biological samples under study (57, 59). Higher order computational approaches are needed to support the interpretation of data for discerning gene signatures or informative gene subsets. Methods such as genetic algorithms combined with supervised clustering (23, 31, 52, 57), principle component analysis, self-organizing maps, and linear discriminant analysis are all examples of analytical methods that might be used for discovering trends of expression across samples. Contemplation of the ultimate research objective for the study prior in the context of visualization of microarray data will ensure that appropriate treatment groups are incorporated in the study design. For example, if an investigator wants to follow the signaling of a receptor-ligand interaction, inclusion of the profiling of a receptor-negative line, receptor pathway inhibitor, or a nonactive ligand to filter out potential effects not directly related to the activated pathway should allow clear determination of direct receptor mediated effects (Fig. 3). The use of a targeted treatment to help show biological specificity was demonstrated in a study (41, 60) in which the investigators wanted to examine effects of aging. They hypothesized that effects of metabolism and energy consumption contributed to the aging process in mice and used microarrays to determine age-associated gene-expression changes in these pathways. They benefited by adding a group of caloric-restricted mice in their study design and thus could filter their observed, age-associated gene-expression changes to determine precisely those genes that changed with normal aging and reversed in the extended-lifespan, calorie-restricted mice. Illustration of how experimental design and visualization work together in data interpretation. Example of a clustering diagram generated from an experiment where tissue was isolated from an animal treated with a compound, a compound inhibitor, and dual treatment of compound and inhibitor together. Red signifies genes that are induced, and green indicates genes that are repressed by treatment relative to vehicle control treated animals. The clustering visualization indicates a group of genes that are highly induced by the treatment (shown in bright red) and then are attenuated by cotreatment with the inhibitor (shown by less intense red color), indicating that affects on these genes are directly related to the compound/inhibitor pathway. Gene expression changes are usually measured relative to another sample. Therefore, another early consideration in the study design process is the strategy for pairing sample hybridizations, in the case of two-color cDNA arrays. (Traditionally in oligonucleotide arrays or arrays hybridized with radioactive probes, only one sample is hybridized per chip and sample pairing is made during the data analysis process.) Investigators have adopted several approaches in sample pairing. As in traditional experiments, many investigators use closely matched sample controls (e.g. vehicle or time matched). Other researchers choose to use universal control RNA derived from multiple samples that is designed to provide a baseline measurement on every gene being assayed (61). A third approach, employed when closely matched sample controls cannot be obtained, e.g. some clinical studies, uses a pooled-normal, or tissue-representative control (23). Each sample-pairing strategy has advantages and disadvantages, although matching as closely as possible the concurrent control to the sample under study may be most desirable (50). One advantage of the universal control is the ease of comparing all the samples assayed against it; however, comparison to datasets that were not analyzed against it may be difficult. In addition, the maintenance of a reproducible, universal RNA stock for an extended period of time is of paramount importance when employing this approach. Finally, a completely different hybridization design minimizes control measurements and implements a loop design that optimizes direct treatment comparisons (where sample A is compared with B, B is compared with C, C is compared with D, and D is compared with A). Heralding the advantages of expression profiling must be fairly balanced by stating the limitations of the technology. Obviously, this is an expression-based only of monitoring at the RNA Some signaling changes may only at the protein or level and not be detected with gene arrays. at this point microarray measurements of RNA provide the advantage of being to use one biological sample to measure thousands of sequences analyses one sample into multiple and cannot provide of the being measured. as signaling and are more potential protein changes or of signaling based on or RNA changes should One in interpreting gene expression profiling data will be between and An of a biological will gene expression changes that will be by more alterations related to The of the will be a that will be possible by experimental design incorporating a broad range of time or disease At the technical several are by of microarray in sequence lead to in gene of clone identification is For cDNA microarray may be in the of a gene of microarrays indicate a quantitative of the level of gene expression, a that for some genes may be interpretation may for some however, if precise is as for clinical applications, the level of gene expression should be by another technology. One may to employ quantitative for a quantitative measure or provide the benefit of not only a quantitative but also the number of potential detected with the chip sequence Finally, gene-expression arrays can provide a sensitive measure of gene changes within a of have that microarrays can detect gene expression in of the This signifies that, if a cell or tissue is significant gene changes may be related to only a of the and investigators may to with an in The for microarrays is as they will undoubtedly continue to be used in for of compound and drug development (45, 46). As the cost of experiments more investigators will include this technology in their of tools Microarray analysis should not be considered as the to an but as a discovery mechanism to help determine which to variations of the microarray like the search for promoter (7, 8) or of for expressed genes may within the will be by the of standard data more global and development of interactions of with scientists and can that the format of microarrays will to more and and the to assay changes using this technology will only be The to and for their and in the of this to and for of the and to for