Microarray design

We created an oligonucleotide microarray for cotton using fiber genes in Genbank, a reported EST assembly of >150,000 ESTs [10], and a recent assembly of >210,000 ESTs [11] as templates for probe design. From these sequences, we designed three sets of oligonucleotide probes (1,154, 12,006, and 9,629, respectively) and included all three sets (22,787 total oligonucleotides) on a single, publicly available microarray [16]. The first set of 1,154 oligonucleotide probes was designed from Cotton ESTs with homology to Arabidopsis genes with roles as regulators of chromatin, transcription, cell wall biosynthesis, and cell cycle [17].

The third set ofoligonucleotides probes was designed from CGI8 [11] that contained 55,673 unique sequences using Picky v2.0. Where possible, identical Picky parameters were used to design the 3^rd oligonucleotide set as the 2^nd set. Probes that targeted the same genes as in the first two probe sets were excluded from further analysis. In total, 9,629 additional oligonucleotides probeswere generated (66 bp average length, 3.6 s.d.; 76.82 T_m, s.d. 1.93) and added to the previous 2 probe sets.

Two essential considerations of microarray quality include the number of targeted genes and the broad utility of the microarray for specific tissues or treatments. Regarding the first consideration, the 22,778 genes described here include perhaps 46–60% of the total genic diversity, given that the total number of genes in the cotton genome may be approximately 40,000–50,000 [19, 20], Indeed, 44% and 40% of the oligonucleotides were designed from singletons from the first [10] and second [11] assemblies, respectively. However, some probes were not completely target-specific, perhaps due to imperfect EST assemblies and due to separate sets of probes designed from different assemblies. Approximately 1,800 of potential non-specific homologies were identified in the recent Gene Index of 55,673 unigenes [11] using vmatch [21] with a ~95% sequence percent identity threshold. Nevertheless, most of the oligonucleotides had a single target, as designed [see Additional file 2].

Regarding the second consideration, these microarray probes have a broad utility for specific tissues or treatments. For example, a detailed analysis of the second probe set revealed that ~7,300 probes represented genes expressed in specific tissues or under specific conditions [see Additional file 3]. The number of specifically expressed genes was determined by summing the number of contigs mostly composed of ESTs from a single library (90%) and the corresponding singletons. In total, 56% of the 2^nd oligonucleotide set represent genes from a specific library; however, a large number of those genes (24%) are from the two G. raimondii libraries that were prepared from heterogeneous tissues and which were more deeply sampled that most of the other cotton cDNA libraries. More than 1,000 oligonucleotides appear to target genes found only in a 7–10 days post-anthesis fiber library [9], and 733 appear to target transcripts uniquely identified following cyclohexamide treatment of ovules [22]. These two considerations suggest that the oligonucleotides selected for the cotton oligonucleotide microarray have a broad diagnostic utility while potentially targeting tissue-specific transcripts expressed under a variety of conditions. The sequences and annotations of all the probes are publicly available via a web-based query [16] or by request.

Microarray hybridizations

Many potential sources of error can have a large impact on microarray experiments, such as inconsistency among multiple RNA extractions, reverse transcription, RNA amplification, and labeling, as well as different levels of background noise for each microarray. Many of these sources of error can be resolved by appropriate experimental design [23, 24]. and careful laboratory technique; however, the quality of the microarray must often be assumed and often is not under the control of the investigator. We investigated the oligonucleotide performance within the first version of the cotton microarray [Gene Expression Omnibus (GEO) database: GPL4305] containing the 1^st and 2^nd replicated spots printed for each probe in these two sets. We found a low level of within-microarray variation, reproducible 'self vs. self' hybridizations with bud tissue, and reproducible expression differences between bud and leaf treatments. The results from the first version of the cotton microarray suggest that the current version with 9,629 additional probes will provide an robust, reproducible platform for transcription profiling in cotton.

Variation within microarrays between the two replicate features was estimated for each oligonucleotide probe, as a consequence of including two microarray features for each oligonucleotide probe in two separate sections of the microarray. The average log-adjusted difference between two replicated features was 0.03 (s.d. 0.69), 0.02 (s.d. 0.80), and 0.02 (s.d. 0.83) for replications 1, 2, and 3, respectively (Figure 1A). The slightly positive value of the average spot differences suggests that the first pin-touch on the microarray deposited slightly more oligonucleotide probe in a slightly larger spot on the slide than the second pin-touch.

Finally, if we test for expression differences between the leaf RNA pool and each of the bud RNA pools (bud and bud'), many genes were found to be consistently differentially expressed between bud vs. leaf and bud' vs. leaf, as seen by the high correlation among the most significant p-values (Figure 1B). If our goal was to describe differentially expressed genes between leaf and bud, we would have a large list of putatively, differentially expressed genes (Table 2); however, in this particular case, we can only claim that these genes were differentially expressed between the two amplified RNA samples considered in our experiment. We intentionally did not include biological replicates because our interest was in quantifying technical variation. Biological replication would be necessary to conclude that the expression differences are inherent differences between leaf and bud rather than simply differences between the particular leaf RNA pool and the particular bud and bud' RNA pools prepared for this experiment.

Because a global assembly of ESTs was used to design this first version of the microarray, genes expressed in many tissues of the plants including cotton fibers, are represented. A total of 1,864 genes were found to be differentially expressed between our leaf RNA pool and our bud RNA pools (false discovery rate = 0.001; Figure 1C). (Again note that these are differences between our RNA pools rather than inherent expression differences between general leaf and bud tissues.) Of these differences, slightly less than 10% belonged to set of probes that had no BLASTX hit (< 1 × 10^-20) and were designed solely based on an ESTScan prediction, thereby providing an element of functional gene discovery to the microarray. Of the 1,864 hypothetical genes that were differentially expressed, 602 were identified by probes derived from cotton fiber cDNA libraries [9]; Ben Burr, unpublished data; Candace Haigler, unpublished data). This number did not include probes designed for known cotton genes, nor other genes that were removed from the analysis (low expressed fiber genes that may not be detected in leaf or bud aRNA) that may also represent genes specifically expressed in fiber tissue.

Microarray probe selection

A total of 22,787 probes were printed on the cotton oligonucleotide microarray, composed of 12,006 oligonucleotides selected from recently reported contigs and singletons derived from a global assembly of cotton ESTs [10], and 9,629 designed against the latest TIGR cotton transcript assembly (CGI8). We also included 1,154 oligonucleotides designed in the Z. J. Chen laboratory based on previously available sequence data [27].

Similar probe design strategies were used for all threes sets of probes. Lee et al. [17] provided a description for the design first set of oligonucleotides. The second and third set of oligonucleotides were designed from the EST exemplar sequences [10] and the TIGR Gene Index [11] with Picky (v 1.0) [18] and Picky (2.0), respectively. Only EST sequences with predicted a protein from ESTScan [28] or a high protein homology (70% percent identity) to an Arabidopsis gene were considered candidates for oligonucleotide selection.

Three different criteria were invoked in the iterative selection of the microarray probes from the EST assemblies: (1) candidate probes identified by Picky by their unique sequences, complexity, and T_m; (2) characterized cotton mRNAs in GenBank and genes of special interest to the cotton community; and (3) complementation to previously synthesized probe set(s). Thus, for each oligonucleotide set, candidate long-oligonucleotide probes (60 – 70 mers) were separately generated based on criteria 1 and 2, then a final list of probes was selected for oligonucleotide synthesis by cross-checking the new list of probes with previously synthesized oligonucleotides.

For design of both the second and third probe sets, we solicited input from the cotton community to identify genes of interest for microarray probe design. Most requests were for known cotton genes with sequences in Genbank, in addition to candidate genes identified in our EST assembly by high homology to genes characterized in other organisms. Probes for these genes were designed with the sequences from the EST assemblies as 'background' to identify the most unique probes possible.

Both transcription factors and genes with little or no annotation were represented on the cotton microarray. Based on widespread interest in transcription factor expression levels, we selected probes targeting genes that had either a transcription related GO ontology [29], or a transcript factor domain as predicted by PFAM [30] (Table 1). Genes with little or no annotation represent a gene discovery component to future microarray experiments. When genes targeted by the 2^nd probe second were compared to the Arabidopsis TAIR protein dataset, 1,200 of them did not have a significant BLASTX hit (< 1e-20) but they did have a coding frame as predicted by ESTScan [28].

Oligonucleotide synthesis and microarray printing

Each set of oligonucleotides was synthesized and aliquoted into 3 replicate plates by IDT Technologies (Coralville, IA, USA). An aliquot of 384-well plates from all three sets of oligonucleotides was hydrated in water then diluted to the printing concentration with 3× SSC. Positive and negative controls were included on the printed microarrays. To assess microarray quality, two spots of each oligo from the same pin-dip were printed in separate slide sections on Corning epoxy slides at the Washington University Microarray Core facility using a locally constructed linear servo arrayer (after the DeRisi model [31]) creating the first version of the cotton oligonucleotide microarray [GEO: GPL4305]. After printing, slides were allowed to dry in 50–70% humidity for 12–16 hrs (~25°C) and cross-linked at 150 mJoules. Two slides from the print batch were checked using SpotCheck (Genetix). Printed cotton microarrays, and images of each print batch are publicly available [16]. Experiments using the preliminary platform [GEO: GPL4305] or this new platform [GEO: GPL4808] can be found at GEO [32].

RNA extraction

One leaf and one bud (10 – 14 days before anthesis) tissue sample of G. hirsutum cv. Acala Maxxa were collected from three separate replications of 4 – 8 plants grown in Horticulture Greenhouse at Iowa State University under supplemental lighting (16 hr. days). RNA was extracted from each of the six samples using a modified hot-borate method [33], quantified, and checked for integrity using a Bioanalyzer (Agilent, Inc., Palo Alto, CA, USA). Equimolar amounts of RNA (A₂₆₀) from three separate extractions were pooled into a single leaf and single bud sample, respectively.

RNA amplificationand labeling

An indirect labeling procedure of amplified aminoallyl.aRNA (TargetAmp™, Epicentre Biotechnologies, Madison, WI, USA) was used for one leaf RNA sample and one bud RNA sample. 0.5 ug of total RNA was used as starting material for 1 round of aRNA amplification, resulting in 26 ug and 51 ug of aRNA from leaves and buds, respectively.

Cy3 and Cy5 dyes (Amersham Biosciences, Pittsburgh, PA, USA) were coupled to two aliquots of 13 and 16 ug of both aRNA samples, respectively. The Cy3- and Cy5-labeled aRNA probes were purified using the Qiagen RNA easy Mini kit (Qiagen, Germantown MD, USA). and sufficient incorporation Cy3 (550 nm) and Cye5 (650 nm) dyes was verified.

Microarray hybridization and image analysis

For microarray hybridization, 300 ng of Cy3 and Cy5 labeled aRNA was used per each slide using the Pronto!™ Plus system protocol (Promega Corporation, Madison WI, USA) with minor changes as described below. Slides from each rep (3) were immersed in 200 ml of Pronto Universal Pre-Soak solution containing 2 ml of liquid Sodium Borohydride for 20 min at 42°C. Slides were transferred to fresh containers with Wash Solution 2 at room temperature for 2 min and then immersed in 200 ml of hybridization buffer (5 × SSC; 0.1 × SDS; BSA 0.1 mg/ml). Slides then were incubated with a fresh Wash Solution 2 at room temperature for 2 min, and were washed 2 additional times with Wash Solution 3 at room temperature for 2 min each. Following immersion in nuclease-free water, slides were dried by centrifugation at 1,600 g for 3 min. All hybridizations and post-hybridization washes were performed exactly as described in the Pronto!™ Plus system protocol.

Microarray images were captured using an arrayWoRx^® Biochip Reader (Applied Precision, Issaquah, WA, USA) using an exposure of 0.5 sec for each channel (Cy5 and Cy3) at ~10 um resolution. GenPix^® Pro (v 5.1, Molecular Devices, Sunnyvale, CA, USA) was used to extract the background-adjusted intensity of each spot. Features that were 'absent', 'not-found', or that had a negative intensity after background adjustment were excluded from the analysis. Data files from this experiment can be found in GEO data set [GSE5875].

Experimental design and statistical analysis

Three replications of a three treatment loop design (bud → leaf, leaf → bud', and bud' → bud) were hybridized on nine microarrays, where bud and bud' simply represent different aliquots of the same aRNA. The signal intensity data were natural log transformed and median normalized, and the 9,654 genes with complete data were examined for expression differences among the three sample types (leaf, bud, and bud'). We considered a standard mixed linear model for the data from any single gene given by

y _ijk = μ + δ _i + τ _j + s _k + e _ijk ,

where y _ijk denotes the normalized log-scale signal intensity (averaged over duplicate spots) for dye i, sample type j, and slide k; μ denotes a an intercept parameter; δ _i denotes the effect of dye i; τ _j denotes the effect of sample type j; s _k denotes the random effect of slide k; and e _ijk denotes a random error term that is intended to capture all other sources of variability. (Note that although we considered a separate model for each gene, we have suppressed a gene subscript on each term to simplify notation.) Here i = 1, 2 (Cy3 and Cy5); j = 1, 2, 3 (bud, bud', and leaf); and k = 1, ..., 9 (microarray slides 1 – 9). On the basis of this model, t-tests for differential expression between each pair of sample types (leaf vs. bud, leaf vs. bud', and bud vs. bud') were conducted. The 9,654 p-values from each of these comparisons were converted to q-values using the method of Story and Tibshirani [34]. These q-values were used to identify the number of differentially expressed genes for a given comparison when controlling the false discovery rate at various levels.