The amplification procedure we used is based on protocols described by Eberwine [21,22,24], Wang et al. [25], and Baugh et al. [23], and details are given in Materials and methods. Briefly, double-stranded cDNA was synthesized either from total RNA isolated from imaginal discs or from poly(A)⁺-purified embryonic RNA. The cDNA was used as a template for in vitro transcription, routinely generating an RNA product that represented an amplification of approximately 1,000 fold. This RNA was subjected to a second round of RT and IVT. The yield from the second round represented an approximately 100-fold amplification, resulting in an overall amplification of 1-5 × 10⁵. Probes for hybridization were synthesized by RT and labeled with either Cy5 (red) or Cy3 (green). Arrays were hybridized simultaneously with Cy5- and Cy3-labeled probes, and the intensities and ratios of bound fluor were measured for each spot. Because of inherent variability in the hybridization efficiency of individual microarrays, only the ratios of the red and green fluorescence were interpreted to indicate relative transcript abundance in the starting pool of poly(A)⁺ RNA.

A known drawback of linear amplification is that sequences at the 5' end of transcripts can be preferentially lost. This reduction in sequence complexity can probably be attributed to incomplete RT, to priming from subterminal locations, and to incomplete IVT. To overcome this problem, we supplemented the random hexanucleotides that were used to prime the second RT reaction with a template-switch (TS) primer [25]. This primer includes a guanosine triplet that is designed to pair with a 3' cytidine overhang created by the reverse transcriptase (Clontech [25]). Control amplifications with the TS primer alone or random hexanucleotides alone yielded similar quantities of aRNA. When these aRNA preparations were used for microarray hybridizations, many spots had significantly different signal intensities (data not shown). The 25 cDNAs with the most intense signals for the probe amplified with the TS primer alone had an average length of 1.6 kilobases (kb). In contrast, after amplification with hexanucleotides alone, the 25 cDNAs with the most intense signals had an average length of 1.1 kb. We interpret the bias towards long transcripts in TS-primer amplified probes as indicating better preservation of 5'-complexity.

The most important feature of RNA pools that amplification must preserve is proportionality, the maintenance of the relative concentration of different RNAs. We established the quality of our amplification method in several ways. First, we isolated poly(A)⁺ RNA from embryos and subjected 100 pg, 1 ng and 10 ng aliquots to two rounds of linear amplification. The size distribution of the aRNA ranged from about 100 nucleotides to high-molecular-weight products, with the majority between 240 and 2,000 nucleotides (Figure 2a). aRNA from independent amplifications was then used to generate probes and to compare hybridization efficiencies of pairs of probes. Such separate amplifications of aliquots of 1 ng of the same embryonic poly(A)⁺ RNA produced signals with similar intensities. A scatterplot of the Cy3 and Cy5 signal intensities showed a tight correlation coefficient (CC) of 0.97, with a low standard deviations (SD) of the red/green (R/G) ratios for all spots (SD 0.25, log₂-transformed) (Figure 2b). Comparisons of probes made from 100 pg and 1 ng and from 1 ng and 10 ng poly(A)⁺ RNA yielded essentially the same level of reproducibility (CC 0.95, SD 0.29; data not shown). We conclude that the amplifications were reproducible and that good proportionality of RNA was maintained during amplification even if the concentration of poly(A)⁺ RNA differed by tenfold.

Comparisons between amplified and unamplified RNA indicated that essentially no transcripts were lost during amplificaton. Of the spots that hybridized to probes generated from amplified and unamplified embryonic poly(A)⁺ RNA 99% (5,514 out of 5,574) did so with both probes. Differences in intensities were, nevertheless, significant (CC 0.53 and SD 1.13, data not shown). We do not know whether these differences in intensities were a result of changes in relative abundance or in transcript length, both of which will affect hybridization signals. Nevertheless, the good correspondence of the hybridizing spots shows that the aRNA is an essentially complete representation of the population of transcripts present in the initial pool. We conclude that comparisons of two different probes on microarrays are valid if the RNA pools have been amplified in the same way.

We carried out 43 separate hybridizations, comparing 13 different pairwise combinations of tissues (Table 1). This analysis identified many genes that are preferentially expressed in specific tissues; it also provided an independent and quantitative measure of the distinct nature of larval tissue and imaginal disc cells.

Hybridizations that were carried out were of two types: with probes generated from different imaginal discs; and with probes from imaginal discs and larval organs. Comparisons between imaginal disc probes revealed a high degree of similarity, but comparisons between imaginal disc and larval organ probes identified many differentially expressed genes. We now give a general description of these results, starting with those tissues that were most alike. It should be noted that these results were obtained with microarrays that represent a pool of cDNAs obtained from the Berkeley Drosophila Genome Project (BDGP) containing approximately 44% (6,000/13,600) of the transcription units predicted to be in the Drosophila genome. As this pool may not be a truly random subset, extrapolations to the whole genome may be of limited value. Indeed, a preliminary comparison of our cDNA arrays with microarrays that contain short cDNA sequences for all annotated genes suggests that the total number of genes that produce a signal with these probes is similar. It is possible that the set of 6,000 cDNAs distributed by the BDGP is biased toward genes expressed in the tissues we analyzed.

In order to compare and contrast the expression profiles of larval organs and imaginal discs, we made probes from the poly(A)⁺ RNA isolated from several larval organs. We directly compared probes from the wing imaginal disc with: the salivary gland, an anterior fraction of the midgut, parts of the fat body and the brain hemisphere, including the optic lobe primordium. For each experiment, probes prepared from a single larva were compared. With the exception of the comparisons between the first leg disc and fat body which was carried out once, each of the other comparisons was repeated at least three times.

Visual inspection of the array scans of imaginal disc to larval tissue comparisons revealed a large number of red and green spots, indicative of a high degree of divergence. A representative block from one wing disc-to-fat body scan is shown in Figure 7b. These results contrast with disc to disc comparisons which produced mostly yellow and only few red or green spots on the arrays. For purposes of illustration, the same block of spots from a leg1 disc-to-leg3 disc scan is shown in Figure 7a. When the intensity of Cy3 and Cy5 signals for all cDNAs was plotted on a scatterplot, only a small number of data points are positioned away from the bisector on which signals of equal intensity in the two channels fall (Figure 7c). This is indicative of strong similarity in expression profiles and is reflected in a high CC (0.93). In contrast, the wing disc-to-fat body comparison had many data points located on each side of the bisector (CC 0.37) (Figure 7d). This disparity between leg disc-to-leg disc and wing disc-to-fat body comparison was also observed for all other disc-to-disc and disc-to-larval organ comparisons we carried out. Figure 8a illustrates the numbers of cDNAs that were expressed differentially in the various pairwise comparisons as a bar diagram. All disc-to-disc comparisons showed small numbers of differentially expressed cDNAs, whereas the disc-to-nondisc comparisons revealed many. The only exceptions were the comparisons of the wing imaginal disc to the brain/optic lobe (see below). To score for the differences in expression for all 6,000 genes, the standard deviations of the R/G ratios (log₂-transformed) for each gene were calculated for all 42 experiments and represented in a scatterplot (Figure 8b). Standard deviations from the datasets derived from disc-to-disc comparisons (25 experiments) were all < 1, indicating the high level of similarity among the different imaginal discs. In contrast, the disc-to-nondisc comparisons (17 experiments) were, with a single exception (see below) > 1, reflecting a high degree of divergence (Figure 8b). These observations are consistent with the distinct developmental programs of imaginal and larval cells.

The only exceptions to emerge in this analysis were the cells of the optic lobe/brain hemisphere. Comparisons of these neuronal tissues with wing discs revealed only a small number of differentially expressed genes (Figure 8a) and had low standard deviations that were in the range of disc-to-disc comparisons (Figure 8b). The optic lobe primordium and brain of late third instar larvae have only limited functions in larval photoreception and contain only a small number of mature neurons. The larval brain is populated with clusters of immature neurons, neuroblasts and ganglion mother cells whose descendants will generate the neuronal and glial populations of the adult optic neuropil and the brain hemisphere. The brain hemisphere also differs from other larval tissues by not degenerating during metamorphosis. Instead, most neurons of the larva join with newly formed adult-specific neurons to build the CNS of the adult ([36] and references therein). We interpret the high degree of similarity in the expression profiles of imaginal discs and the brain/optic lobe preparations as a manifestation of the apparent commonalities of their developmental programs.

The number of genes that were found to be induced in the larval organs is significantly greater than the number of genes in the sets defined by the disc-to-disc comparisons. Most of the genes in these sets have not been characterized previously and therefore have no known function, but the known genes in these clusters illustrate the predictive power of the method; we mention three examples.

Many of the genes that are preferentially expressed in the optic lobe/brain hemisphere have predicted or described expression and/or function in neurons or glial cells (data not shown). One of these is lola, which encodes a transcription factor involved in axon guidance and is expressed in the embryonic optic lobe placodes as well as other tissues [37]. Another is beta amyloid protein precursor-like (Appl), with known expression in the larval optic lobe [38].

The expression profile of the cyclin-dependent kinase 1 (Cdk-1/cdc2), is also noteworthy. Cdk-1 is believed to regulate the G2 checkpoint of the cell cycle and is required to maintain diploidy. Mutations in Cdk-1 drive cells that would normally remain diploid into endoreplication [39]. We found Cdk-1 (CG5363) to be repressed in larval tissues when compared to imaginal discs (Figure 9). As most of the cells in the larval fat body, salivary gland and midgut are polyploid (reviewed in [40]), the array hybridizations are in good agreement with the known function of Cdk-1.

Two methods for comparative hybridizations to DNA microarrays have been used in previous studies. Most commonly, experimental samples have been compared to a common reference sample that contains an appropriate diversity of cell types and states. An alternative method, which we have used in this study, is to directly compare probes generated from two experimental samples. This second method should be the more sensitive. The amplification technique we used was such that product generated from a single imaginal disc was sufficient to make several probes and to carry out multiple array hybridizations. By comparing samples that had been isolated from a single animal, this method minimized differences due to biological variability. By repeating experiments several times and by reversing the label (dye-flip) for experimental samples in repeat hybridizations, we also attempted to minimize experimental noise.

In addition to these design aspects of the experimental protocols, we applied a number of filters to the data analysis. We selected only those genes with hybridization signals over background in 80% of all experiments and with a 1.74-fold difference between the two channels in at least 45%. Applying this method to 11 arrays that compared wing and eye-antennal discs, we found two subclusters representing genes preferentially expressed in either disc. These lists of genes are largely validated by the previously described expression pattern in either tissue and by in situ hybridization analysis. We understand that the use of any such threshold settings is likely to filter out biologically relevant genes. For example, when less stringent settings were used (for example, a 1.6-fold intensity difference increased the gene sets by 20%), additional genes such as twin of eyeless were included in the eye-antennal cluster and Antp and CG6680 (Figure 6e) in the wing cluster. Relaxing the requirement to produce a ratio in at least 80% of all experiments to 70% identifies yet another group of genes, and includes LD11162 (Figure 6j). Conversely, setting a more stringent threshold to 1.9-fold intensity difference decreased the gene sets by 25% and removed genes such as Arrestin2 and CG11798 (Figure 6d,6f).

Setting a threshold level for the clusters is arbitrary, and as some candidate genes may be expressed by only a small subset of cells in the experimental sample, each analysis must be evaluated separately. Nevertheless, we interpret the good correlation between the array analysis and expression patterns we have described as indicating that the method is both efficient and reliable. Subsequent to the analyses described above, we carried out single-disc comparisons with microarrays that represent a nearly complete set of the approximately 14,000 annotated Drosophila genes. The short cDNAs spotted on these arrays were produced with specific primer pairs obtained from Incyte Genomics, Inc. Use of these arrays for disc-to-disc comparisons also identified only a relatively small number of differentially expressed genes (data not shown). The list of induced genes was apparently more complete, however, and included, for example, vestigal in the wing-to-leg disc comparisons (A.K., G. Schubiger and T.B.K., unpublished observations).

Although the absolute number of genes included in the lists of tissue-specific genes is not meaningful, the relatively small number of genes that differ between imaginal tissues and the significantly greater number of genes that differ between imaginal and larval tissues certainly is.

Spotted cDNA microarrays were produced essentially as described at [41]. In brief, the Drosophila Gene Collection (DGC 1.0, kindly provided by the Berkeley Drosophila Genome Project) of 5,849 nonredundant cDNAs was amplified using universal primers essentially as described at [42]. The amplification products were purified with 96-well format Qiagen PCR purification columns. All PCR products were analyzed on agarose gels, and reactions with no detectable product, multiple bands or bands of unexpected size were repeated using Herculase (Amersham) or Expand polymerase (Roche). One hundred and seventy-five cDNAs from our lab collection were amplified and added to the set, resulting in a total of 6,024 cDNAs. Reactions were arrayed into 384-well plates and printed on poly-L-lysine-coated glass slides using a linear servo arrayer and ArrayMaker Version 2 control software. To minimize background caused by oxidation of the polylysine coating, slides were pretreated by a protocol suggested by Paul Ebert. Before post-processing, slides were incubated for 5 min in 3× SSC; 0.2% SDS at 65°C, washed first in water and then in 95% ethanol, followed by centrifugation to dry the slides. Slides were then treated, and the hybridization reaction was carried out as described at [42]. We indirectly labeled the hybridization probes by incorporation of amino-allyl modified nucleotides in a first-strand cDNA RT reaction. Monofunctional Cy5 or Cy3 dye (Amersham) was subsequently coupled to the reactive residues. Multiple hybridizations were carried out for most experiments, and dye labeling was reversed to avoid systematic bias.

Third instar wandering larvae were dissected and washed several times in Ringer solution (130 mM NaCl, 5 mM KCl, 2 mM Na₂HPO₄, 1.5 mM CaCl₂, 0.37 mM KH₂PO₄). Fat-body fragments were mainly anterior plates. Midgut was the anterior portion just posterior to the gastric caeca. Brain/optic lobe consisted of one brain hemisphere including the optic lobe primordium.

Total RNA from larval tissues was extracted using the Mini RNA Isolation Kit (Zymo Research, Orange, CA) and was eluted with RNase-free water. Total RNA from embryos was extracted with Trizol reagent (Invitrogen). Embryonic poly(A)⁺ RNA was purified using the Oligotex mRNA Mini Kit (Qiagen). Poly(A)⁺ RNA was not purified before amplification of disc RNA because of the small amounts (see Figure 1).

In situ hybridizations were carried out as described in O'Neill and Bier [43]. The plasmids used to generate in situ probes were derived by subcloning an aliquot of the PCR product that was printed onto the microarray with Topo TA cloning (Invitrogen). All cloned products were sequenced and separate hybridizations were carried out for sense and antisense probes.

Methods for amplification were adapted from Wang et al. [25] and Baugh et al. [23]. Total RNA isolated from dissected larval tissues or poly(A)⁺ RNA purified from embryos was incubated with 100 ng oligo(dT)₂₄ -T7 primer (GGCCAGTGAATTGTAATACGACTCACTATAGGGAGGCG-GTTTTTTTTTTTTTTTTTTTTTTTT) and 125 ng TS primer (AAGCAGTGGTAACAACGCAGAGTACGCGGG) in 5 μl at 65°C for 10 min and cooled on ice. Six microliters first-strand premix were added and incubated at 42°C for 2 h (6 μl mix was made up with 2 μl 5× first-strand buffer, 1 μl 0.1 M DTT, 1 μl dNTPs (10 mM each), 0.3 μl T4gp32 (USB, 13.82 mg/ml), 0.5 μl RNasin (Promega), 1 μl Superscript II (Invitrogen)). The reaction was incubated at 65°C for 10 min and cooled on ice. Cold second-strand premix (64.5 μl prepared on ice: 45 μl RNase-free water, 15 μl 5× second-strand buffer (Invitrogen), 1.5 μl dNTPs (10 mM each), 0.5 μl E. coli ligase (10 U/μl), 2 μl E. coli polymerase (10 U/μl), 0.5 μl E. coli RNaseH (2 U/μl)). Enzymes for the second-strand premix (Invitrogen) were added and incubated at 16°C for 2 h, followed by the addition of 2 units T4 DNA polymerase (Promega) and incubation at 16°C for 15 min before heat inactivation at 70°C for 10 min. Clean-up was performed with DNA clean and concentrator-5 (Zymo Research), eluting twice with 8 μl RNase-free water. The total volume was adjusted to 8 μl in a Speed Vac. The first IVT was performed with Megascript T7 (Ambion) in a 20 μl volume for 5-6 h, followed by DNA digestion. aRNA was purified using Mini RNA Isolation Kit or RNA Clean-up kit (Zymo Research), eluting in 2 × 8 μl RNase-free water. Random hexanucleotides (250 ng) and TS primer (1 μg) were added and the volume adjusted to 5 μl. The mix was incubated at 65°C for 10 min before 5 μl first-strand premix was added followed by incubation at 42°C for 2 h. Before the second-strand synthesis, 200-500 ng oligo(dT)₂₄ -T7 primer was added and denatured at 65°C for 10 min. The second-round second-strand synthesis, clean-up and IVT were carried out as for the first round, except that no ligase was added to the second-strand premix. One-quarter to one-third of the total aRNA was applied to the clean-up columns after the second IVT to avoid overloading. All reactions were carried out in a thermocycler with heated lid or air incubator to avoid evaporation.

Hybridized microarrays were scanned with a GenePix 4000 A Microarray Scanner (Axon Instruments, Union City, CA). Data were analyzed and displayed with Cluster and Treeview [31], AMAD [41], Genepix PRO (Axon Instruments), and Microsoft Excel. Normalization for cluster analysis and calculation of standard deviations were done with AMAD. Only genes that qualified with a combined median intensity > 300 above background in both channels in at least 80% of the repeated experiments were included in the analysis. A threshold of > 1.74 (= 0.8 of the log₂-transformed ratios) was chosen for all comparisons. A requirement to show induction above threshold in 100% of the arrays was applied for experiments that were repeated three times or less. For experiments that were repeated more than three times this requirement was relaxed to 45%. Standard deviations were calculated on the log₂-transformed normalized R/G ratios using the 'nonbiased' method. To calculate the correlation coefficients, the intensities of both channels were normalized to equalize the sums of the intensities.