To determine the embryo sac transcriptome, we used coa and wild-type pistil samples (late 11 to late 12 floral stages 39) in three biologic replicates, and followed the Affymetrix standard procedures from cRNA synthesis to hybridization on the chip. Finally, raw microarray data from the coa and wild-type samples in triplicate were retrieved after scanning the Arabidopsis ATH1 'whole genome' chips, which represent 24,000 annotated genes, and they were subjected to statistical analyses. The normalized data were examined for their quality using cluster analysis 40. There was strong positive correlation between samples within the three replicates of wild-type and coa (Pearson coefficients: r = 0.967 for for wild-type and r = 0.973 for coa). Therefore, the data were considered to be of good quality for further analyses. It was necessary to ensure that the arrays of both the wild type and coa did not differ in RNA quality and hybridization efficiency. The hybridization signal intensities of internal control gene probes were not significantly altered across the analysed arrays, hence assuring the reliability of the results (data not shown). The quality of data for the spl mutant and wild-type microarray was described previously 34. Subsequently, differentially expressed genes were identified using three independent microarray data analysis software packages.

To identify genes that are expressed in the female gametophyte, we subtracted the transcriptomes of coa or spl from the corresponding wild type. Genes that were identified as being upregulated in wild-type gynoecia are candidates for female gametophytic expression, and genes highly expressed in coa and spl are probable candidates for gain-of-expression in the sporophyte of these mutants. However, this comparison was not straightforward because we were not in a position to compare the mere four cell types of the mature embryo sac with the same number of sporophytic cells. Whether using whole pistils or isolated ovules, a large excess of sporophytic cells surrounds the embryo sac. The contaminating cells originate from the ovule tissues such as endothelium, integuments and funiculus, or those surrounding the ovules such as stigma, style, transmitting tract, placenta, carpel wall and replum. Therefore, we anticipated that the transcript subtraction for embryo sac expression would suffer from high experimental noise. We examined the log transformed data points from the coa and spl datasets (with their corresponding wild-type data) in volcano plots. This procedure allows us to visualize the trade-offs between the fold change and the statistical significance. As we anticipated, the data points from the sporophytic gain outnumbered the embryo sac transcriptome data points on a high-stringency scale (data not shown). This problem of dilution in our data for embryo sac gene discovery was more pronounced in the coa dataset than that of spl, because we did not dissect out the ovules from the carpel. Therefore, we made the following decisions in analyzing the gametophytic data: to use advanced statistical packages that use different principles in their treatment of the data; and to set a lowest meaningful fold change in data comparison, in contrast to the usual twofold change as recommended in the literature.

In the recent past, many new pre-processing methods for Affymetrix GeneChip data have been developed, and there are conflicting reports about the performance of each algorithm 414243. Because there is no consensus about the most accurate analysis methods, contrasting methods can be combined for gene discovery 44. We used the following three methods in data analyses: the microarray suite software (MAS; Affymetrix) and Genspring; the DNA Chip analyzer (dCHIP) package 45; and GC robust multi-array average analysis (gcRMA) 46. MAS uses a nonparametric statistical method in data analyses, whereas dCHIP uses an intensity modeling approach 47. dCHIP removes outlier probe intensities, and reduces the between-replicate variation 48. A more recent method, gcRMA uses a model-based background correction and a robust linear model to calculate signal intensities. Depending on the particular question to be addressed, one may wish to identify genes that are expressed in the embryo sac with the highest probability possible and to use a very stringent statistical treatment (for example, dCHIP), or one may wish to obtain the widest possible range of genes that are potentially expressed in the embryo sac and employ a less stringent method (for example, MAS). We did not wish to discriminate between the three methods in our analysis, and we provide data for all of them.

Although conventionally twofold change criteria have been followed in a number of microarray studies, it has been disputed whether fold change should be used at all to study differential gene expression (for review, see 49). Based on studies correlating both microarray and quantitative RT-PCR data, it was suggested that genes exhibiting 1.4-fold change could be used reliably 5051. Tung and coworkers used a minimum fold change as low as 1.2 in order to identify differentially expressed genes in Arabidopsis pistils within specific cell types, and the results were spatially validated 52. In order to make a decision on our fold change criterion in the data analysis, we examined the dataset for validation of embryo sac expressed genes that had previously been reported. We found that genes such as CyclinA2;4 (coa dataset) and ORC2 (spl dataset) were identified at a fold change of 1.28 (Additional data file 1). In addition, out of the 43 predicted genes at 1.28-fold change from coa and spl datasets, 33% were present in triplicate datasets from laser captured central cells (Wuest S, Vijverberg K, Grossniklaus U, unpublished data), independently confirming their expression in at least one cell of the embryo sac. Therefore, the baseline cut-off for subtraction was set at 1.28-fold in the wild type, and a total of 1,260 genes were identified as putative candidates for expression in the female gametophyte (Additional data files 2 and 3).

However, it must be noted that lowering the fold change potentially increases the incidence of false-positive findings. By setting the baseline to 1.28, we could predict that false discovery rates (FDRs) would range between 0.05% and 3.00%, based on dCHIP and gcRMA analyses (data not shown). Convincingly, we we able to observe 24 essential genes and 17 embryo sac expressed genes at a fold change range between 1.28 and 1.6 (Additional data files 1 and 4, and references therein). Moreover, our data on homology of candidate genes to expressed sequence tags (ESTs) from monocot embryo sacs will facilitate careful manual omission of false-positive findings. The usefulness of this approach is also demonstrated by the observation that 84% of the essential genes and genes validated for embryo sac expression (n = 51) present in our datasets exhibited homology to the monocot embryo sac ESTs. Therefore, our practical strategy of using a low fold change cut-off probably helped in identifying low-abundance signals, which would otherwise be ignored or handled in an ad hoc manner.

In contrast to the embryo sac datasets, we applied a more stringent twofold higher expression as a baseline for comparison of the mutant sporophyte with the wild type. This is because we had large amounts of sporophytic cells available for comparison. In all, 527 genes were identified as candidate genes for gain of sporophytic expression in coa and spl mutant ovules (Additional data file 5). Because the transcriptome identified by three independent statistical methods and the resultant overlaps were rather different in size for both the gametophytic and sporophytic datasets, we report all the data across the three methods (Additional data file 6). This approach is validated by the fact that candidate genes found using only one statistical method can indeed be embryo sac expressed (see Additional data file 7). Furthermore, only 8% of the validated genes (n = 51) were consistently identified by all three methods, demonstrating the need for independent statistical treatments (Additional data file 7). In short, our data analyses demonstrate the usefulness of employing different statistical treatments for microarray data.

Another practical consideration following our data analyses was the very limited overlap between coa and spl datasets. Although both mutants are genetically and phenotypically similar, the overlap is only 35 genes between the embryo sac datasets and 13 genes between the sporophytic datasets (Additional data files 2, 3, and 5). In light of the validation in expression for 12 genes from the coa dataset, which were not identified from the spl dataset, we suggest that the limited overlap is not merely due to experimental errors. It is likely that the embryo sac transcriptome is substantial (several thousands of genes 2), and two independent experiments identified different subsets of the same transcriptome. This is apparent from our validation of expression for several genes, which were exclusively found in only one microarray dataset (Additional data file 1). In terms of the sporophytic gene expression, we have shown that three sporophytic genes initially identified only in the spl microarray dataset were indeed over-expressed in coa tissues (discussed below). In short, despite the limited overlap between datasets, both the embryo sac and sporophytic datasets will be very useful in elucidating embryo sac development and its control of sporophytic gene expression.

The genes identified as embryo sac expressed or over-expressed sporophytic candidates were grouped into eight functional categories based on a classification system reported previously 53 (Figure 1). The gene annotations were improved based on the Gene Ontology annotations available from 'The Arabidopsis Information Resource' (TAIR). The largest group in both gene datasets consisted of genes with unknown function (35% of embryo sac expressed genes and 37% of over-expressed sporophytic candidate genes), and the next largest was the class of metabolic genes (24% and 27%; Figure 1). Overall, both the gametophytic and sporophytic datasets comprised similar percentages of genes within each functional category (Figure 1). In both datasets, we found genes that are predicted to be involved in transport facilitation and cell wall biogenesis (15% of embryo sac expressed genes and 13% of over-expressed sporophytic candidate genes), transcriptional regulation (10% and 9%), signaling (7% and 6%), translation and protein fate (5% each), RNA synthesis and modification (3% and 1%), and cell cycle and chromosome dynamics (1% each).

The efficacy of the comparative profiling approach used here was first confirmed by the presence of 18 genes that were previously identified as being expressed in the embryo sac (Additional data file 1). They included embryo sac expressed genes such as PROLIFERA, PAB2 and PAB5 (which encode poly-A binding proteins) and MEDEA, and genes with cell-specific expression such as central cell expressed FIS2 and FWA, synergid cell expressed MYB98, and antipodal cell expressed AT1G36340 (Additional data file 1 and references therein). Therefore, our comparative profiling approach potentially identified novel genes that could be expressed either throughout the embryo sac or in an expression pattern that is restricted to specific cell types.

In order to validate the spatial expression of candidate genes in the wild-type embryo sac, the six following genes were chosen for mRNA in situ hybridization on paraffin-embedded pistils: AT5G40260 (encoding nodulin; 1.99-fold) and AT4G30590 (encoding plastocyanin; 1.88-fold); AT5G60270 (encoding a receptor-like kinase; 1.56-fold) and AT3G61740 (encoding TRITHORAX-LIKE 3 [ATX3]; 1.47-fold); and AT5G50915 (encoding a TCP transcription factor; 1.36-fold) and AT1G78940 (encoding a protein kinase; 1.35-fold). Broad expression in all cells of the mature embryo sac was observed for genes AT5G40260, AT4G30590, AT5G60270, and AT4G01970 (Figure 2). The trithorax group gene ATX3 and AT5G50915 were predominantly expressed in the egg and the central cell, and the expression of the receptor-like kinase gene AT5G60270 was found to be restricted to the egg cell alone (Figure 2). In addition to the in situ hybridization experiments, we examined the expression of transgenes where specific promoters drive the expression of the bacterial uidA gene encoding β-galacturonidase (GUS) or in enhancer detector lines. We show that CYCLIN A2;4 (1.28-fold) and AT4G01970 (encoding a galactosyl-transferase; about 1.51-fold) were broadly expressed in the embryo sac, and that PUP3 (encoding a purine permease; 1.3-fold) was specifically expressed in the synergids (Figure 2). CYCLIN A2;4 appears to be expressed also in the endothelial layer surrounding the embryo sac (Figure 2e). Diffusion of GUS activity did not permit us to distinguish unambiguously embryo sac expression from endothelial expression. In short, both broader and cell type specific expression patterns in the embryo sac were observed for the nine candidate genes. Hence, we could validate the minimal fold change cut-off of 1.28 and the statistical methods employed in this study.

Our strategic approach to exploring the embryo sac transcriptome was twofold: we aimed first to identify embryo sac expressed genes; second to describe the gametophyte enriched (male and female) transcriptome; and finally to define the embryo sac enriched (female only) transcriptome. Although the first category does not consider whether an embryo sac expressed gene is also expressed in the sporophyte, the second class of genes are grouped for their enriched expression in the male (pollen) and female gametophyte, but not in the sporophyte. The embryo sac enriched transcriptome is a subset of the gametophyte enriched transcriptome, wherein male gametophyte expressed genes are omitted. Of the embryo sac expressed genes, 32% were also present in the mature pollen transcriptome, and the vast majority (77%) were expressed in immature siliques as expected (Additional data files 2 and 3). Because large-scale female gametophytic cell expressed transcriptome data of Arabidopsis based on microarray or EST analyses are not yet available, we compared our data with the publicly available cell specific ESTs from maize and wheat by basic local alignment search tool (BLAST) analysis. Large-scale monocot ESTs are available only for the embryo sac and egg cells but not for the central cells (only 30 central cell derived ESTs from 54). Therefore, we included the ESTs from immature endosperm cells at 6 days after pollination in the data comparison (Additional data file 8 and the references therein). Of our candidate genes, 38% were similar to the monocot embryo sac ESTs, 33% to the egg ESTs, and 53% to the central cell and endosperm ESTs (Additional data files 2 and 3).

Genes that were enriched in both the male and female gametophytes, or only in the embryo sac, were identified by subtracting these transcriptomes from a vast array of plant sporophytic transcriptomes of leaves, roots, whole seedlings, floral organs, pollen, and so on (Additional data file 9). The transcriptomes of the immature siliques were omitted in this subtraction scheme because often the gametophyte enriched genes are also present in the developing embryo and endosperm. We found 129 gametophyte enriched and 108 embryo sac enriched genes, accounting for 10% and 8.6%, respectively, of the embryo sac expressed genes (Table 1). Among the embryo sac enriched genes, 52% are uncategorized, 17% are enzymes or genes that are involved in metabolism, 15% are involved in cell structure and transport, 8% are transcriptional regulators, 4% are involved in translational initiation and modification, 3% are predicted to be involved in RNA synthesis and modification, and 2% in signaling (Figure 1 and Table 1). Of the embryo sac enriched transcripts, 31% were present in the immature siliques, suggesting their expression in the embryo and endosperm (Table 1). Furthemore, 26% of the embryo sac enriched genes were similar to monocot ESTs from the embryo sac or egg, and 41% were similar to central cell and endosperm ESTs (Table 1).