DNA millichips as a low-cost platform for gene expression analysis.

Our goal was to create a DNA chip that is as easy, convenient, and inexpensive as an agarose gel. For a first-generation solution, we describe a low-cost, easy-to-use de novo synthesis oligonucleotide microarray technology that draws on the inherent flexibility of the maskless array synthesizer for in situ synthesis of thousands of photolithographically produced oligonucleotides covalently attached to a microscope slide. The method involves physically subdividing the slide into 1 × 1 mm millichips that are hybridized to fluorescent RNA or DNA of biological origin, in a microfuge tube at an ordinary laboratory benchtop, rather than in dedicated hybridization chambers. Fluorescence intensity is then measured with a standard microscope rather than sophisticated DNA chip scanners. For proof of principle, we measured changes in the transcriptome of Arabidopsis (Arabidopsis thaliana) plants induced by growth in the presence of three major environmental abiotic stresses (temperature, light, and water status), in all possible combinations. Validation by comparison with quantitative reverse transcription PCR showed a high correlation coefficient and analysis of variance indicated a high technical reproducibility. These experiments demonstrate that low-cost DNA millichips can be made and reliably used at the benchtop in a normal laboratory setting, without assistance of core facilities containing costly specialized instrumentation.

Over the past decade, genomic research has seen an explosion in the quantity of data being produced from high-density microarrays and next-generation DNA sequencing. These technologies have been critical to an accelerating pace of knowledge acquisition and innovation. Although the use of increasingly larger data sets is crucial for elucidating the complexity of biological pathways, the generation, manipulation, and evaluation of these data require specialized equipment (e.g. hybridization chambers, scanners) along with advanced computer programming and statistical analysis skills typically found only in dedicated core facilities or groups dedicated to this type of research. Furthermore, the costs associated with these techniques can place them outside the reach of a typical bench scientist on a day-to-day basis.
The goal was to devise a method that would make DNA chips as accessible, inexpensive, and convenient for individual laboratories to use as agarose gels. For this purpose the maskless array synthesizer (MAS) was used to create DNA chips with a high density of long, single-stranded DNA sequences. The MAS is an automated instrument that utilizes the Texas Instruments digital micromirror device (DMD; Sampsell, 1994) to perform photolithography and is uniquely capable of de novo synthesis of long (e.g. 60-mer) oligonucleotides on a glass surface with a density that exceeds 1 million or more probes patterned within a 2-cm squared glass surface and affords the capability of programming any desired probe sequence of reasonable length (Singh-Gasson et al., 1999;Cerrina et al., 2002;Nuwaysir et al., 2002).
Because the cost of reagents for an entire/whole MAS-derived DNA chip on a microscope slide is around $450 it was reasoned that if the surface of a full-scale microarray slide was subdivided into many smaller but reproducible pieces, one could obtain a chip with a few thousand custom probe sequences at a price rivaling that of an agarose gel (i.e. in the order of tens of dollars rather than hundreds of dollars). Furthermore, if standard laboratory equipment could be used for processing the millichips (e.g. thermal cycler versus custom hybridization chambers, fluorescent microscope versus scanner, etc.) this would eliminate the need to purchase specialized equipment. Together these changes increase the accessibility of microarray technology for routine, daily usage in a standard molecular biology laboratory setting.

The Millichip Design
A critical goal of this study was to include enough probes on each millichip to make the resulting experiments scientifically valuable yet provide enough physical separation of the chip segments to allow adequate millichip yield from a standard glass microarray slide. As a first step, a 1 3 1 mm millichip was found to meet these criteria (Fig. 1). A millichip of this size was accommodated by standard PCR tubes (e.g. 0.2 mL), was easily held with a reverse-acting tweezers, and could contain a 67 3 67 matrix of probe spots (i.e. 4,489 probe spots). From a yield perspective, it was determined that the best method for separating the millichips from the slide was to (1) score the slides prior to probe synthesis on the MAS; (2) align the scored slides to the MAS DMD and perform the probe synthesis; and (3) postsynthesis, manually separate the individual millichips from the slide by depressing a finetipped metal stylus against the slide surface opposite to the scoring (i.e. the probe surface of the slide). Initial yields of usable millichips from this manual process ranged from 50% to 70%. Although not explored as part of this study, process improvements (e.g. optimizing the prescoring depth and cut shape) along with automation could undoubtedly improve this yield.
A method was sought to eliminate the need for specialized equipment, such as hybridization chambers and scanners, to hybridize and measure the fluorescence intensity of the DNA millichips. It was determined that the 1 3 1 mm millichip could be conveniently placed in the bottom of a 0.2-mL PCR tube for hybridization of fluorescently tagged DNA or RNA targets to the complementary sequence synthesized on the glass surface. Various reaction volumes and conditions were explored with emphasis placed on minimizing volume to keep reagent expense low. As little as 10 mL of hybridization solution adequately covered the entire millichip and provided sufficient volume for a successful hybridization. To wash nonspecific label off the glass surface, a reverse-acting tweezers was used to hold the millichips securely and to allow their transfer between wash solutions contained in 250-mL standard glass laboratory beakers. This procedure minimized the risk of marks and other damage to the surface that would negatively impact optical measurements of fluorescent intensity.
To provide a routine and inexpensive method for measuring fluorescent intensity and data extraction, the use of a standard fluorescent microscope rather than a dedicated (i.e. expensive) scanner, as is currently used for DNA chips, was explored. By operating at 1003 magnification with the fluorescent microscope, it is possible to capture the fluorescent intensity of each millichip with four overlapping 16-bit TIFF images. Software was written to merge data from the four images into a single dataset and fluorescent measurements could be quantified and dropped into a spreadsheet file for evaluation by the user. This program is freely available for download at http:// www.biotech.wisc.edu/sussmanlab/Downloads. This stand-alone software operates with Microsoft Windows (i.e. requires no other software to run) and performs the following functions: accommodates images from a standard fluorescent microscope (e.g. 16-bit TIFF images); adjusts the angle, size, and position of an image; automatically adjusts the spacing between spots to compensate for stretched images; gives the user the option of excluding damaged or contaminated areas of the millichip from further evaluation (Fig. 2); and manipulates the raw pixel data to calculate minimum, maximum, average, median, and SD as well as background correct probe data. This study employed a method of background subtraction that utilizes blank spots adjacent to probe spots, but additional approaches are currently being incorporated. The experiments described here looked only for large changes in expression levels, allowing the use of simpler techniques. In future work, mismatch probes and other standard techniques will be incorporated into the array design to allow for more advanced background correction and normalization methods. This system of measuring and recording fluorescent intensity with a microscope found in routine use in most biology laboratories enables an untrained user to evaluate the Figure 1. Millichip process. A, Standard glass microarray slides are prescored to a depth of 0.7 mm prior to synthesis on the MAS. B, The glass slide is broken down into milllichips along the prescored grid. C, Hybridization is carried out in 0.2-mL PCR tubes with as little as 10 mL solution. D, Millichips are held with reverse-acting tweezers for the washing steps after hybridization. E, Each millichip is imaged using a standard fluorescence microscope. The image shown is the array of a single millichip. transcriptome of an organism or tissue type without the need for specialized scanners.

Examining a Stress Response in Arabidopsis (Arabidopsis thaliana)
An experiment was designed to test the ability of millichips to provide biologically useful information and demonstrate the capability to readily produce and use a large number of individual chips in a single study. As the initial test of the millichip platform, we chose to examine several environmental perturbations under which plants have evolved that are related to the drought response: (1) temperature (heat, cold, or normal), (2) water status (dehydration, salt, or normal), and (3) light intensity (normal or high). In addition, we separately examined the effect of the dormancy hormone (+)-abscisic acid (ABA) to provide a comparison treatment for evaluating the millichip performance with previously published research. The genes represented on the array were chosen based on previous stress response studies in which arrays were used to monitor expression changes for a single treatment (Kreps et al., 2002;Rossel et al., 2002;Seki et al., 2002aSeki et al., , 2002bKimura et al., 2003;Taji et al., 2004;Huang et al., 2008;Kant et al., 2008;Matsui et al., 2008;Wohlbach et al., 2008;Abdeen et al., 2010). By choosing genes that showed large changes in expression under each single treatment, we were able to ensure that a range of expression levels would be present on each millichip because genes that show large changes in expression under one treatment may show little or no change under another. One advantage of smaller, more cost-efficient arrays is the ability to look at a smaller set of genes under a larger set of conditions. In this study it was possible to investigate the response to not only single stress treatments, but also every double and triple combination of these stresses. Table I describes the individual conditions as well as the double and triple combinations of conditions that were tested in this study. There were a total of 18 treatments plus two controls (i.e. normal/untreated and ethanol treated for the ABA comparison) used for this initial study.
With the ability to produce hundreds of millichips, it is possible to look at combinations of treatments in a way that was previously inaccessible to many labs. The objective of this study was to demonstrate the use of the millichip as a quantitative tool to screen for large expression changes and not for the qualitative study of expression levels. To this end, simpler methods of normalization and analysis were used in the screening experiments with the ability to verify interesting and significant results by quantitative PCR, additional chip experiments, or by other methods. Thus, expression fold changes of less than two were categorized as no change.

Verification of the Millichip Platform
To determine statistical reliability, we performed an average of three to four technical replicates and six biological replicates for each treatment and control group, resulting in a total of 457 individual millichip experiments for this study.
For the initial investigation, a simplified millichip utilizing around 1,000 out of 4,489 available spots was designed to not only provide an easy method of background correction exploiting blank spots, but also for the ease of image analysis while the software extraction and normalization functions were in development.
All millichips were identical in design with 1,156 total probe spots consisting of 911 unique gene probes derived from genes chosen from stress response studies as described previously; 17 actin genes represented by 32 probes; probes for reference genes ubiquitin, GAPDH, and 5SRRNA; four probes matching the Affymetrix GeneChip Poly-A RNA control kit sequences used to monitor target labeling; and five probe sequences that were the reverse complements of labeled spike-in oligos used to verify concentrationdependent signal response on each millichip. The actin, reference gene, and Affymetrix GeneChip label control probes were all replicated five times on each millichip whereas the spike-in concentration oligo probes were replicated nine times. To provide a more Figure 2. A, An entire millichip consisting of four fluorescent microscope images that have been stitched together using a series of standard alignment marks on the outside corners and edges of the millichip. The oversized marker at the top lefthand corner provides an orientation for the rest of the markers. B, One of four millichip images showing the overlap of the data extraction software with the image and spots marked in red that have been identified by the user for exclusion due to contamination and scratches.
valuable comparison, the majority of probes were designed based on commercially available arrays from Roche Nimblegen (Thibaud-Nissen et al., 2006), whereas an additional 67 probes were designed using the OligoArray software (Rouillard et al., 2003). These 911 gene probes did not have any replicates on individual millichips (i.e. replication of the 911 unique gene probes was a function of the number of technical replicates performed for each biological replicate).
The custom software previously described was used to extract the median pixel value for each probe and blank (i.e. empty/nonprobe) position in the millichip microarray matrix. An average value was then calculated for all the blank spots immediately adjacent to a probe position. The background-adjusted probe value was calculated by subtracting the average blank value from the probe value. This method of background adjustment to the probe values not only corrected for the higher background noise associated with using a standard fluorescent microscope versus a scanner but also corrected for fluorescent lamp uniformity variance inherent with the microscope's arc lamp source. Background-adjusted probe values were converted to their log base-2 values, which were then normalized using the ratio of the probe distribution medians. Supplemental Table S1 provides a statistical summary for each probe under each treatment condition including treated and untreated normalized, absolute probe values, log2 ratio, fold change, and P value. The entire data set is available from the National Center for Biotechnology Information Gene Expression Omnibus database (accession no. GSE37118). Figure 3A illustrates the concentration-dependent signal response for four representative millichip experiments, which as expected depicts increasing signal with increasing spike-in oligo concentration. This concentration-dependent signal response was checked for each millichip. The median coefficient of variation for the probes replicated on each individual millichip (i.e. actin, GAPDH, ubiquitin, and 5SRRNA) was 12% as shown in Figure 3B.
Histograms for the technical replicate signal intensity coefficient of variation (i.e. signal SD divided by signal average) show a 27% average coefficient of variation that while not quite as good as that obtained using traditional DNA chips is reasonable when one considers the low cost and ease of use (Fig. 4A). The average biological replicate signal intensity coefficient of variation of 26% observed in this experiment is also well within the range (approximately 20%) generally seen with whole-plant studies (Fig. 4B). Quantitative real-time PCR was used to validate fold changes derived from the millichip experiments (Fig. 5). A strong correlation between the two sets of data were observed, comparing favorably to previously published evaluations of normal high-density microarrays (Nuwaysir et al., 2002). Overall these observations demonstrate that in its initial formulation, the millichip platform performs well for the identification of genes that show a 2-fold or greater change in expression in a much larger number of treatments than is currently affordable or possible with resources found in a traditional molecular biology laboratory.
When possible, results were compared with previously published array data for other stress response studies in Arabidopsis to look for agreement in upregulated and down-regulated genes. For the purpose of this study, fold changes less than 2-fold were considered to be no change. When comparing the data for cold, drought, and ABA treatment, agreement with Table I. Single and combination treatments Matrix of test conditions and combinations of conditions evaluated. Combinations such as heat + cold or salt + dehydration were not tested due to inherent incompatibilities. ABA was tested separately as a standalone comparison with previously published data. The first column gives the abbreviation used for each single or combination treatment.

Treatment
High Light Heat Cold Salt Dehydration ABA High light x Heat x Cold x Salt x Dehydration x ABA x previously published data were approximately 70%. Discrepancies in these types of comparisons arise from differences in experimental conditions such as treatment times and methods as well as the problems inherent in making comparisons between different array platforms.

Data Clustering
One objective of the biological experiment performed in this study was to investigate whether combining different environmental perturbations resulted in gene expression changes that were not predictable based on a study of those treatments performed individually. We believe this highlights an important use of the millichip platform. The ability to do hundreds or even thousands of unique chip experiments makes it possible to study more perturbations to a biological system. To examine this question, hierarchal clustering was performed (TIBCO Spotfire) with the normalized dataset (i.e. fold changes in gene expression are described by the ratio transcript signal intensity from treated plant tissue relative to untreated control plant tissue) by complete linkage with Euclidean similarity measure ( Fig. 6; Kaushal and Naeve, 2004). This analysis of the mRNA changes observed in the various single, double, and triple treatments showed several relationships that were expected based on prior knowledge of the underlying biology. There were many genes on the chip that were positive controls based on known effects in the literature. For example, we observed that an ABA-responsive proteinrelated gene (AT5G52300.1) is up-regulated 80-fold in the ABA treatment, with relatively no change in the control treatment. Also noted was the heat shock protein-related gene (AT2G29500.1) that was upregulated in the heat treatment, as well as all combination treatments that included heat. Additional examples of genes or groups of genes behaving as noted in prior literature are described in more detail in Supplemental Table S2.
In general, there were two major clades observed: clade I representing high sharing the cold variable. A dominant effect of temperature on gene expression, compared with the other treatments examined in this limited study can be explained by the realization that unlike warm-blooded animals that maintain a constant body temperature under many different external temperatures, the metabolism and growth of plant cells is strictly dependent on and greatly affected by the temperature of the environment. As known from many physiological studies, and the simple observation of annual tree rings from trees grown in environments with large seasonal swings in temperature, at low temperature plant cells stop dividing and the metabolism is greatly reduced. At the same time, at higher temperature, there is a nearly universal heat shock response in which gene expression is focused on producing proteins that offer protection against thermal denaturation. For example, three of the most highly up-regulated genes across the heat treatments code for heat shock proteins (AT2G29500, AT1G74310, and AT3G12580). The fold changes seen across the heat treatments for these genes range from 15-to 232-fold, relative to the control, in agreement with previous work (Kilian et al., 2007). It should be kept in mind that only 5% of the genes in the Arabidopsis genome are represented on this millichip, and they were chosen mainly to reflect those genes known or suspected to be involved in the plant drought response. The relationships observed from the cladogram may artificially reflect this bias rather than portraying underlying mechanisms in the organism's response. In addition, although this hierarchal clustering appears to be the best fit given the parameters chosen for the analysis, there may be other equally plausible cluster results possible. As shown in Figure  5, there is no relationship between the number of genes that showed significant changes, and the position of that treatment within the cladogram (Table II). The single treatment with the fewest changes was the high-light treatment (24), followed by the cold-salt treatment (76), and the triple treatment of high-lightcold-salt (78). Salt treatment produced a much larger number (250) of genes that change significantly. Although salt is often considered a drought-inducing agent, the greater response to salt compared with drought (91) is not unexpected because in addition to acting as an osmolyte that pulls water out of cells, sodium chloride is permeant and the entry of large amounts of sodium and chloride ions in the cell is likely to cause a much greater change in the transcriptome than simple dehydration induced by air drying. Cold alone produced 97 changes whereas heat alone gave 250. As mentioned previously, from the number and pattern of genes induced by either low or high temperature, it is clear that the Arabidopsis plants are qualitatively more responsive to the temperature than any other single treatment.

Combination Treatments
Of particular interest is the question of whether there were genes that exhibited changes in the double and triple treatments that were not expected based on changes known to occur in single treatment conditions. Very few examples of combined treatment comparisons exist and, to the best of our knowledge, there are currently no examples of triple treatment comparisons, possibly due to the prohibitively high cost for the necessary number of arrays (Rizhsky et al., 2004). The millichip provides researchers with the ability to do these and similar experiments requiring hundreds or thousands of DNA chips cheaply and easily. There are several examples of unexpected results in expression in which the combined treatment shows expression in one direction and the corresponding single treatments show expression in the opposite direction. When the double treatment of high-light and cold (LC) is compared with the single high-light (L) and cold (C) treatments there are 160 genes up-regulated in both double and single treatments and 242 genes downregulated in both double and single treatments (Supplemental Table S3). However, there are also 81 genes up-regulated in the high-light-cold treatment that are down-regulated in both the single high-light   Table S4. The same comparison was made with triple treatments as related to double and single treatments and is summarized in Supplemental Table S5 with more detailed information in Supplemental  Table S6.
In conclusion, it is clear that when environmental perturbations are grouped together in all possible double and triple combinations, a significant fraction of the genome sampled on this chip do not display changes in a direction predictable based on the single or even double treatments alone. Given (1) the size of the Arabidopsis genome (there are over 500 transcription factors known, out of the approximately 30,000 genes), (2) the complexity of the environment in which Arabidopsis evolved (light, temperature, nutrient starvation, gas concentration, to mention just a few), and (3) the large number of genes devoted to network cross-talk and other regulatory aspects of sensory perception and response (the largest single gene family in Arabidopsis is the protein kinase gene family, containing 1,000 members) it may not be surprising that the transcriptional apparatus does not respond in a simple fashion when the plant is perturbed.

DISCUSSION
Although DNA chips are still routinely used for DNA and RNA analyses, next-generation DNA sequencing technologies are also now becoming more commonly utilized for the same purposes. As DNA sequencing instrumentation offer higher throughput and multiplexing capabilities in next-generation experiments, they may provide opportunities for lower cost. However, ease of use and requirements for specialized instrumentation remain obstacles for efforts to make sequencing a viable alternative to the millichip technology described herein. In addition, it should be noted that DNA arrays are often utilized together with next-generation sequencing, for capture experiments that extend the utility and reduce the price of nextgeneration sequencing studies (Okou et al., 2007). Millichips should also be amenable for use in capture experiments, and may offer new low-cost and convenient opportunities for high-throughput studies that require the isolation of specific DNA or RNA sequences. Methods for millichip capture and other applications are currently under development.
We have shown that the millichip platform is a useful tool for screening genes that show a 2-fold or greater change in expression with a far greater number of biological variables possible. When it is desirable to identify genes with statistically significant changes less than 2-fold in magnitude, more traditional microarray approaches should be used. This technology shows great promise as a primary screen to examine gene expression in many separate treatments (e.g. 100 or more) than is currently possible. For most applications, the millichip flexibility and reliability combined with low cost and ease of use outweigh any limitations to detect minute changes.
Excluding labor, total material costs in this study were between $33 and $37 per millichip including all material costs associated with reverse transcription, amplification, and fluorescent labeling of the starting total RNA ($21 per millichip); synthesis of the millichips on the MAS ($11-$15 per millichip depending on the slide break efficiency as discussed earlier); and hybridization of the labeled sample to the millichips ($1 per millichip). Combining process improvements (e.g. improved slide break efficiency) with a lower-cost reverse transcription and labeling process (e.g. Promega ChipShot indirect labeling and clean-up system), we believe that a total material cost of $20 or less per millichip experiment is readily achievable. In particular, material costs for synthesis of millichips on the MAS should be closer to $7 to $8 per millichip with process improvements currently under consideration.
The MAS tool used for these experiments utilized a DMD with about 768,000 mirrors to de novo synthesize oligonucleotide probes on a 2-cm squared surface. Newer versions of this tool include higher density DMDs with smaller features to create over 1 million probes per slide. The usable surface area for each millichip allowed a 67 3 67 matrix (i.e. 4,489 spots) of 60-mer single-stranded DNA probe locations. For this initial feasibility study we erred on the side of caution and isolated each probe location from adjacent probes by blanks (i.e. empty spots), resulting in slightly more than 1,000 oligo locations per millichip. If maximal probe density is used with the higher density DMD, a millichip containing 7,225 (i.e. 85 3 85 matrix) different 60-mer ssDNA sequences is possible. In addition to these upgrades, the 1 3 1 mm millichip size can also be changed to some extent, resulting in yet another increase in the total number of possible probes per chip. Thus with currently available technology, a much higher density of probes than was performed in these experiments can be achieved. With further modifications in optics even higher densities should be attainable in the near future. Because the Arabidopsis and other advanced multicellular eukaryotic genomes including humans have 20,000 to 40,000 genes, each millichip is capable at this time of containing a significant fraction of the targeted transcriptome and whole-genome millichips for advanced eukaryotes are within reach. CONCLUSION We have described a method to produce low-cost DNA chips that can be used for hybridization experiments using standard equipment available in most molecular biology laboratories. A pilot study with mRNA isolated from Arabidopsis plants grown under a variety of treatments has demonstrated that these low-cost and easy-to-use chips produce data that is statistically robust and reliable. They also demonstrate that current studies of how cells respond to perturbations applied one at a time, necessitated by the high cost and inconvenience of DNA chips and next-generation sequencers, may not be a true reflection of how the gene networks are responding when many such perturbations are applied simultaneously. Because evolution has acted on organisms present in the environment in complex ecosystems over long periods of time, with many various chemical and environmental changes occurring all at once to varying degrees, it is clear that a complete understanding of how genomes work may not occur until transcriptome studies with DNA chips become as routine, cheap, and convenient as agarose gels.

Plant Growth Conditions
Seeds of Arabidopsis (Arabidopsis thaliana), ecotype Columbia, were surface sterilized by agitation in 70% ethanol for 15 min, followed by 30% bleach and 0.1% Triton X-100 for 30 min. Seeds were then rinsed four times with sterile water and stratified in a fresh change of sterile water at 4°C in the dark for 3 d. After stratification, seeds were transferred in a minimal volume of sterile water to 45-mL polypropylene flat-bottom tubes with flip-top lids containing 15 mL of one-half-strength Murashige and Skoog (MS) salts, 0.05% (w/v) MES, and 0.5% (w/v) Suc adjusted to pH 5.7 with 1 M KOH (Murashige and Skoog, 1962). Seedlings were germinated at room temperature under constant illumination (40 mmol m 22 s 21 , traceable dual-range light meter, Control Company) with gentle shaking.

Water-Deficit Treatments
After 10 d of hydroponic growth, seedlings were exposed to several single and combinatorial treatments intended to mimic environmental water-deficit conditions: (+)-ABA, heat, cold, salt, dehydration, and high light. Treatments were carried out in a growth chamber (Enconair) under normal light conditions (80 mmol m 22 s 21 ) or at 5-fold-higher light conditions (400 mmol m 22 s 21 ). For drought treatment, approximately 100 mg fresh weight of seedling tissue was removed from a polypropylene tube, blotted dry on tissue paper, and exposed to a water deficit treatment by placement on a piece of filter paper dampened with MS media in a polypropylene tube for 5 h. In parallel, nondrought control blotted seedlings were added to 10 mL of fresh MS media. For the ABA treatment, seedlings were added to 10 mL fresh MS media containing 100 mM (+)-ABA (Sigma) in 10 mL 95% ethanol. For the ABA control treatment, seedlings were added to 10 mL of fresh MS media containing 10 mL solvent Plant Physiol. Vol. 159, 2012 555 alone (95% ethanol). Heat-treated seedlings were incubated in a growth chamber maintained at 37°C, whereas cold-treated seedlings were incubated on ice in a growth chamber maintained at 10°C. For the salt treatment, seedlings were added to 10 mL of fresh MS media containing 200 mM NaCl.

RNA Isolation
Treated seedling tissue was blotted dry on tissue paper, weighed, flash frozen in liquid nitrogen, and stored at 280°C until RNA could be isolated. Total RNA was isolated from ground tissue using the RNeasy plant mini kit (Qiagen), including the on-column DNase digestion step.

Slide Preparation Prior to Microarray Synthesis
A series of cuts extending partially through the thickness of a standard glass microarray slide (ArrayIt) were machined at 1.5-mm intervals across the length and width of the slide surface (Mindrum Precision, Inc.). These cuts defined the approximate 1 3 1 mm square millichips on the slide surface and served as break-lines during cleavage of the millichips from the slide after microarray synthesis. Cuts were machined approximately 75% through the slide thickness and created a grid pattern encompassing the millichip pattern projected by the MAS during microarray synthesis. A stock solution was prepared with 0.1% (v/v) glacial acetic acid in 95% (v/v) ethanol:water. At room temperature slides were immersed with rotation for 4 h in a solution consisting of 2% (v/v) N-(3triethoxysilylpropyl)-4-hydroxybutryamide (Gelest) in stock solution. Slides were then rinsed with rotation twice in fresh stock solution for 20 min. After a final diethyl ether rinse, slides were placed in a 120°C oven for 1 h followed by overnight curing under vacuum at 120°C (12-16 h). Functionalized slides were sealed and stored dry until used for synthesis (Beier and Hoheisel, 1999).

Microarray Synthesis
Light-directed synthesis was performed using the MAS technology described previously (Singh-Gasson et al., 1999). Prescored slides were mounted in the standard slide holder used on the synthesizer and the cuts on the slide surface were aligned to the millichip images projected on to slide through the use of a CCD camera (Watec) coupled to an infinity-corrected long working distance 203 M Plan APO objective (Mititoyo) via a standard InfiniTube (Infinity Photo-Optical) mounted behind the transparent quartz reaction cell. Manual stages holding the reaction cell and slide assembly were used to position the cuts on the slide relative to the projected millichip image by observing their relative positions with the CCD camera assembly. Oligonucleotide probe synthesis was performed using DNA synthesis protocols previously described (Singh-Gasson et al., 1999;Cerrina et al., 2002;Nuwaysir et al., 2002) where the removal of the photo-labile protecting group 2-nitrophenylpropyloxycarbonyl was performed by exposure to broadband UVlight wavelengths of g, h, and i lines produced by a 350 Hg arc lamp (Newport) and nucleotide base attachment achieved using standard phosphoramidite chemistry (Fodor et al., 1991). After completion of microarray synthesis, the slide was broken into individual millichips by pressing a small-diameter precision drill bit (0.127-mm diameter) against the glass surface opposite to the slide cuts. A stereomicroscope was used to manually align the bit and slide cut during breaking.

Millichip Hybridization
Total RNA was amplified and labeled with Cy3 dye using the Amino Allyl MessageAmp II aRNA amplification kit from Ambion followed by fragmentation with Ambion fragmentation reagents. The Affymetrix GeneChip Poly-A RNA control kit was used to monitor the target amplification and labeling process. MES-based prehybridization and hybridization buffers were made with the following final, standard component concentrations: 100 mM MES, 1 M NaCl, 20 mM EDTA, 0.01% (v/v) Tween20, 0.1 mg mL 21 herring sperm DNA, and 0.5 mg mL 21 bovine serum albumin. (All reagents were purchased from Sigma Aldrich with the exception of herring sperm DNA and bovine serum albumin, which were purchased from Promega). Each millichip was placed in a dedicated 0.2 mL, flat-topped PCR tube and incubated for 15 min in 25 mL of prehybridization buffer using an MJ Research thermocycler (PTC-200) set at 45°C with lid temperature set at 50°C. At the conclusion of the 15 min, the prehybridization buffer was removed with a pipette and immediately replaced with 10 to 25 mL of hybridization solution per millichip. Total Cy3labeled RNA per millichip was, on average, 2.7 mg. PCR tubes were loaded a second time into the thermocycler with their top covers left open and a low-density polyethylene plug cut to size with a razor blade was inserted into each PCR tube with the bottom of the plug approximately 3 mm above the surface of the hybridization buffer covering the millichip. The PCR tube top covers were then closed, the thermocycler lid was closed, and the millichips were incubated overnight (16-20.5 h) with the thermocycler set at 45°C with a constant lid temperature of 50°C. After hybridization individual millichips were washed with the Roche NimbleGen microarray wash buffer kit following the protocol specified with the kit. Each millichip was held with a reverseacting tweezer in 250-mL beakers containing 120 mL of each wash buffer. Agitation was provided by magnetic stir bars in the beakers. After washing the millichips were dried with argon.

Fluorescent Microscope Imaging
Four fluorescent images of each millichip were captured with an upright Nikon E800 microscope equipped with a Nikon Plan APO 103 objective and Roper Scientific CCD camera. MetaMorph version 6.2r4 was used to capture the four images per millichip and save them as 16-bit TIF files.

Reverse Transcription Quantitative PCR
Residual DNA was eliminated from 1 mg of total RNA in 10 mL of diethylpyrocarbonate-treated water using 1 unit of amplification grade DNase I (DNase I, Amp Grade, Invitrogen) followed by reverse transcription with the SuperScript III first-strand synthesis system for reverse transcription-PCR (Invitrogen). Real-time PCR reactions were performed with iQ SYBR green supermix, clear 0.2-mL low-profile eight-tube strips, and optically clear flat eight-cap strips (Bio-Rad) in a Bio-Rad C1000 thermal cycler equipped with a Bio-Rad CFX96 real-time PCR detection system and CFX Manager software. Primers were designed with QuantPrime (Arvidsson et al., 2008) and purchased from a commercial vendor (IDT). All reactions were performed in accordance with the vendor-supplied protocols. Threshold cycle (Ct) values were extracted with the default threshold setting in the CFX Manager software and fold changes were calculated using the 2 2DDCt quantitative PCR data transformation (Schmittgen and Livak, 2008). The GAPDH gene was used as the reference gene.
Microarray data from this article can be found in the National Center for Biotechnology Information Gene Expression Omnibus data libraries under accession number GSE37118.

Supplemental Data
The following materials are available in the online version of this article.
Supplemental Table S1. Stress response data set.
Supplemental Table S2. Data for genes with known function.
Supplemental Table S3. Comparison of double and single treatments.
Supplemental Table S4. Comparison of double and single treatments for specific genes.
Supplemental Table S5. Comparison of triple, double, and single treatments.
Supplemental Table S6. Comparison of triple, double, and single treatments for specific genes.