We isolated RNA from ten tissues and different developmental stages, ranging from young seedlings to senescing leaves, and roots to fruits of the A. thaliana Col-0 referenced strain. In addition, we made use of inflorescence apices from the clavata3 (clv3) mutant 37 to enrich for shoot and floral meristems (Additional data file 1). We used both GeneChip^® Tiling 1.0R and ATH1 gene expression arrays to obtain triplicate expression estimates from all samples. Because our priority was to detect transcribed regions, we decided to use double-stranded DNA (dsDNA) as hybridization targets for the tiling arrays. Consequently, we did not obtain information about the strand from which a signal originates. However, several recent reports have raised the question of how reliable the detection of antisense transcripts on tiling arrays is 3334. Another advantage is that DNA targets exhibit higher specificity than RNA targets 38.

To profile the expression of annotated genes on tiling arrays, we extracted probe information for all genes that can be analyzed in a robust manner (see Materials and methods [below] for details). Consequently, we ignored small transcription units such as tRNA genes, which are represented by an insufficient number of probes. Having each gene represented by a set of probes allowed us to apply a standard algorithm, robust multichip analysis (RMA), to both microarray platforms, thereby minimizing differences resulting from different analytical procedures 39. A total of 20,583 genes were represented on both platforms; an additional 136 and 9,645 genes were exclusively represented on ATH1 and the tiling array, respectively. Resulting RMA log2 expression values for tiling and ATH1 arrays spanned 11 to 12 log2 units in both cases.

To compare the expression values derived from ATH1 array and tiling array, we generated scatter plots and calculated pair-wise Pearson correlation coefficients (PCCs) for all samples (Figure 1a,b and Table 1). Expression values for all genes in a given sample were well correlated across platforms, with PCCs ranging from 0.854 to 0.882 (P < 10^-15), indicating that both produce comparable results. Transcripts with expression estimates close to background correlate the least between platforms, as a result of higher variance of tiling array estimates (Figure 1a,b).

We were particularly interested in the power of the tiling array to detect differential gene expression. To this end, we compared two samples, roots and inflorescences, which are known to have very different expression profiles 5. Applying the RankProduct method (RankProd) 4041, we detected 2,484 and 2,294 differentially expressed genes (P < 0.05) on ATH1 and tiling arrays, respectively, with 1,780 genes in common. A PCC of 0.92 (P < 10^-15) indicated a good agreement for detecting expression differences of individual genes across platforms (Figure 1c). In addition, we generated a 'correspondence at the top' (CAT) plot using P values to rank the genes (Figure 1d) 42. In the top 200 and 1,500 lists, 150 and 1,308 genes, respectively, were found in common, further supporting high concordance between the two types of arrays.

Comparing the platforms across all samples, we found that more than 70% of all genes showed a correlation of 0.8 or greater (Figure 2a). Genes with low correlation between platforms tend to be those that are represented by a comparably small number of tiling probes (Figure 2b). Qualitatively, the same is true for genes that, because of the improved annotation, are represented by only a limited number of probes on the ATH1 array (Additional data file 4) or by strongly overlapping probes on ATH1 (Figure 2b). These results indicate that gene expression estimates based on ten or more tiling array probes are highly robust. More than 27,000 annotated genes fulfill this requirement for the Affymetrix Arabidopsis 1.0R tiling array, making it a powerful tool for gene expression studies.

The tiling array allows the analysis of 9,645 genes, corresponding to 31.9% of all annotated genes, that are not represented on the ATH1 array. The average expression levels of these genes across all 11 samples are clearly lower than of those that are also present on the ATH1 array. Although only 15% of genes represented on both the tiling and ATH1 array platform have average expression level of less than six log₂ units, this applies to more than 50% of the genes found only on the tiling array (Figure 3a). This is consistent with priority during the ATH1 design being given to genes with prior expression evidence 9. Nevertheless, many genes absent from ATH1 are expressed more highly in at least one sample (Figure 3b).

Of the 9,645 genes, 1,065 genes had z scores exceeding 2.5 across the 11 samples, making them good candidates for having tissue-specific or stage-specific expression patterns (Additional data file 9, Table 1, and Figure 3c). The number of easily detectable transcripts was higher in roots or senescing leaves than in young leaves or seedlings, which is in agreement with previous observations 5.

To identify transcripts that are not present in the current genome annotation, we adopted a computational method, margin-based segmentation of tiling array data (mSTAD), for the segmentation of tiling array data into exonic, intronic, and intergenic regions 35. Extending a segmentation method developed for yeast tiling arrays 43, we modeled spliced transcripts with ten discrete expression levels and incorporated a more flexible error model. Moreover, mSTAD is a supervised machine-learning algorithm with internal parameters that are estimated on hybridization data together with information on the location of annotated genes. After training, it can make predictions based on hybridization data alone.

When comparing a genome-wide sample of all mSTAD exon predictions with annotated genes, we found that the predictions were generally accurate for the more highly expressed half of genes (Figure 4a; see Materials and methods [below] for details). For each sample, we further analyzed a set of high-confidence exon predictions (Figure 4b and Additional data file 5). These contained a minimum number of four probes, had predicted discrete expression level between 6 and 10, and had at most 25% repetitive probes. From these high-confidence exon predictions, which make up 37% to 50% of the total length of all predictions depending on the tissue analyzed, more than 97% overlap at least 25 bp with annotated exons (Figure 4c). Between 26% and 36% of the remainder overlap with cDNAs and expressed sequence tags (ESTs) but not with annotated transcripts.

In summary there are between 1,107 and 1,947 predicted high-confidence exons per sample, for a total length of 242 to 406 kilobases (kb), that are neither included in the current annotation nor covered by sequenced cDNA clones. A complete list of all high-confidence exons with chromosome start and end position can be downloaded from the At-TAX homepage 36. Among the unannotated high-confidence predictions, 14% to 31% are specifically detected in a single sample, with inflorescences and senescing leaves showing the highest proportion (Figure 4d). Whether these predictions indeed correspond to expressed transcripts was tested for some of these by RT-PCR. From high-confidence predictions that do not overlap with known cDNAs or ESTs, a subset of 47 segments was selected so that different lengths as well as different predicted expression strengths were covered. We could confirm by RT-PCR that more than three-quarters (37) of these 47 predicted segments as transcribed (Figure 4e and Additional data file 6).

Previous analyses with whole-genome tiling arrays have focused on the polyadenylated portion of the Arabidopsis transcriptome 14303132. However, studies conducted in several other organisms have suggested that there is a large fraction of nonpolyadenylated RNAs (for example 1920). In order to revisit this question in Arabidopsis, we isolated total RNA from two different tissues, whole seedlings and inflorescences, and depleted it for rRNA using a mix of locked nucleic acid (LNA) oligonucleotides. This RNA preparation was used for reverse transcription with either an oligo-dT primer (which targets only polyA [+] RNA) or random primers (which target both polyA [+]and polyA [-] RNAs). After conversion to dsDNA, samples were hybridized to tiling arrays. For both tissues analyzed, there was a good correlation between polyA(+) samples and polyA(±)samples (PCC = 0.84; P < 10^-15; Figure 5a). Nevertheless, we found many transcripts that were more easily detected in polyA(+) samples than in polyA(±) samples. This probably reflects the fact that mean signal intensities are for unknown reasons generally lower toward the 3' end after random priming (Additional data file 7). Hence, expression values of short transcripts in particular may be underestimated with random-primed hybridization targets.

Only a small proportion of annotated genes produced a much higher polyA(±) signal compared with the polyA(+) fraction (Table 2). Large differences were detected for two structural RNAs: a U12 small nuclear RNA and an H/ACA-box small nucleolar RNA (Table 2). The majority of snRNAs undergo 3' end processing that is very distinct from polyadenylation 4445, indicating that our method appears suitable for detecting nonpolyadenylated transcripts. Most other transcripts that were much more abundant in polyA(±) than in polyA(+) samples emanate from transposons and pseudogenes (Table 2). These results suggest that in Arabidopsis the overwhelming majority of known protein coding transcripts possess a polyA tail.

We also applied the above described mSTAD algorithm to the two polyA(±) samples, to detect transcription from unannotated regions. When we subtracted high-confidence segments found in at least one polyA(+) sample from the segments found in both polyA(±) samples, segments totaling less than 100 kb were identified as potential polyA(-) transcripts (Figure 5b). These regions represent less than 0.1% of the entire genome, which appears to be very low compared with results reported for C. elegans tiling array studies using the transfrag method 19. To rule out the possibility that this discrepancy is a computational artifact, we applied the transfrag method to our tiling array data also 46. This method led to similar estimates of polyA(±) specific transcribed fragments (transfrags), with a combined length of about 250 kb, or 0.2% of the genome (Figure 5b). These results imply that nonpolyadenylated transcripts are much less abundant in Arabidopsis than in C. elegans and humans 2047.

To make our results easily accessible to the research community, we created an online resource that consists of two parts: a web-tool that reports expression values for user-specified genes, and a customized generic genome browser 48.

The At-TAX gene expression visualization tool can be fed with TAIR (The Arabidopsis Information Resource) locus IDs 49. Expression estimates for input gene(s) are displayed in all analyzed samples and on both ATH1 and tiling arrays, where available (Figure 6a). This not only provides a convenient means of analyzing genes not represented on the ATH1 array, but also allows simple cross-platform comparison. The generic genome browser displays transcriptional active regions as predicted by mSTAD across the genome, as well as all raw expression values for each probe in all analyzed samples 50 (Figure 6b).

Tiling arrays have several advantages compared with focused gene expression arrays such as the ATH1 platform, because tiling arrays allow detection of all transcripts irrespective of their annotation status as well as different splice forms. However, because probes have not been optimized in a similar manner, especially for uniform isothermal hybridization behavior, it has been unclear how broadly suitable they are for routine expression analysis. To address this issue, we used both array types to analyze 11 different samples covering different tissues and developmental stages. The resulting gene expression estimates on both array platforms are highly correlated, including measures of expression changes between tissues. We conclude that whole genome tiling arrays are indeed an appropriate tool for standard gene expression analyses. However, expression estimates derived from the two different platforms can differ for various reasons, indicating that expression data must be interpreted carefully. Discrepancies are often due to the selection of probes on the ATH1 arrays, which are biased towards the 3' end of transcripts and sometimes overlap, thus violating assumptions of independence. Conversely, expression analysis with tiling arrays can be inaccurate for small genes represented by very few probes, especially if these have unfavorable hybridization properties. Uncertainty in gene annotations is another source of error, because expression may erroneously be measured from intronic probes.

Compared with the ATH1 array, a disproportionately high number of genes that are represented only on the tiling array produced very low hybridization signals. This is not unexpected because the genes selected for the ATH1 array were supported by cDNAs and ESTs, whereas the tiling array includes hypothetical genes that lack any experimental evidence of expression. In addition, the number of annotated pseudogenes in A. thaliana has been increasing dramatically. The first annotation released in 2001 (TIGR1) contained 1,274 pseudogenes, whereas the recent TAIR7 annotation includes 3,889 pseudogenes 11.

Identification of unannotated transcribed regions is a major motivation for tiling array experiments. That our segmentation algorithm generated highly reliable predictions is evident from the observation that there was very good overlap with annotated genes as well as high success rates for RT-PCR validation experiments. Despite extensive cDNA cloning and previous use of tiling arrays (for example, 14), we could detect more than 1,000 additional transcripts. We found that exonic regions in the different tissues comprise on average about one-third of the genome. Despite the finding of unannotated transcripts, the ratio of annotated exons to polyA(+) transcripts detectable on tiling arrays appears to be much higher in Arabidposis than in some other organisms 51. Interestingly, tiling array analysis of Arabidopsis mutants impaired in DNA methylation or RNA quality control has revealed more than 200 noncoding transcripts that are normally transcriptionally silenced, indicating that the Arabidopsis genome has at least the potential to generate a large number of transcripts from intergenic regions 3132.

Tiling array studies of human and C. elegans indicated that about half of all transcripts are not polyadenylated 20. In contrast, our data suggest that nonpolyadenylated RNAs make a more limited contribution to the Arabidopsis transcriptome. It is already known that specific classes of plants transcripts are generated in a different manner than in animals. For example, some human microRNA precursors are transcribed by RNA polymerase III and hence are not polyadenylated, whereas Arabidopsis microRNA precursors feature characteristics of RNA polymerase II generated transcripts 5253. Another reason might be differences in 3' end processing. For example, histone mRNAs in land plants are polyadenylated, which is in contrast to histone mRNAs in animals, which are subject to a unique form of 3' end processing resulting in a hairpin that protects the 3' end from RNA degrading enzymes 5455565758.

We found that many nonpolyadenylated RNAs in Arabidopsis are derived from pseudogenes and transposons. Several examples of actively transcribed pseudogenes have been reported 59, and many pseudogenes become transcriptionally activated in methylation-deficient mutants 31. Known mechanisms for the transcriptional silencing of pseudogenes involve small RNAs that are generated through the RNA-dependent-RNA-polymerase (RDR)2/DICER-LIKE 3 biogenesis pathway 6061. Interestingly, improperly terminated, nonpolyadenylated RNAs derived from transgenes can be subject to a silencing pathway that involves another RNA-dependent-RNA-polymerase, namely RDR6, which can use both polyadenylated and nonpolyadenylated transcripts as a substrate 6263. Therefore, our observation that RNAs corresponding to a subset of pseudogenes are much more abundant in the polyA(±) fraction is compatible with a scenario in which these pseudogenes are transcribed into polyA(-) RNAs that subsequently serve as template for RDR-dependent amplification. However, transcripts from some pseudogenes are also detectable in polyA(+) samples. These pseudogenes might either be transcribed into both polyA(+) RNAs and polyA(-) RNA or, alternatively, polyA(-) RNAs derived from polyA(+) RNAs accumulate during RNA amplification and processing steps.

We have demonstrated that the use of the GeneChip^® Arabidopsis Tiling 1.0R Array for routine expression analyses does not have any apparent disadvantages compared with the ATH1 array. Rather, it has many advantages, including the ability to provide information on genes that are not represented on ATH1, as well as the ability to analyze additional aspects of gene expression, such as alternative transcript initiation and 3' end formation or splicing, all of which might be under physiological or developmental control 6465. Tiling arrays might be the platform of choice to further resolve transcriptional activity over developmental stages and cell types, especially when combined with techniques for the isolation of specific cells by laser microdissection or cell sorting (for review 66).

Wild-type Col-0 and clv3-7 plants 37 were grown on soil or on solid MS medium under continuous light at 23°C. Additional data file 1 describes each sample. Tissue samples were frozen in liquid nitrogen and total RNA was isolated using the RNeasy Plant Mini Kit (Qiagen, Hilden, Germany). RNA integrity was determined on a Bioanalyzer with the RNA 6000 Series II Nano kit (Agilent, Santa Clara, CA, USA).

For synthesis of probes (targets) for ATH1 and tiling arrays, 1 μg of total RNA was used as template for generation of cRNA using the MessageAmp II-Biotin Enhanced Kit (Ambion, Austin, TX, USA). We followed the manufacturer's instructions with one exception; for tiling arrays, biotinylated NTPs were replaced by unmodified NTPs (stock solution 25 mmol/l each). Sixteen micrograms of biotinylated cRNA (for ATH1 arrays) was fragmented using 5× Fragmentation Buffer. Seven micrograms of unmodified cRNA (for tiling arrays) was converted into dsDNA (GeneChip^® WT Double-Stranded cDNA Synthesis Kit; Affymetrix Inc.) and dsDNA was purified using the MinElute Reaction Cleanup Kit (Qiagen). A total of 7.5 μg dsDNA was fragmented and labeled using the GeneChip^® WT Double-Stranded DNA Terminal Labeling Kit (Affymetrix Inc.). Targets were hybridized to ATH1 and Arabidopsis Tiling 1.0R arrays for 14 hours at 42°C, washed (Fluidics Station 450, wash protocol EukGE-WS2_V4 for ATH1 arrays or wash protocol FS450_0001 for tiling arrays) and scanned using a GeneChip^® Scanner 3000 7 G.

For comparison of polyA(+) and polyA(±), rRNA was depleted from 10 μg total RNA using RiboMinus™ Yeast Transcriptome Isolation Kit (Invitrogen) and an Arabidopsis specific RiboMinus™ LNA oligonucleotide mix kindly provided by Invitrogen, Carlsbad, CA, USA. rRNA depleted RNA was precipitated and resuspended in 12 μl water, from which 11 μl were used for reverse transcription using MessageAmp II-Biotin Enhanced Kit (Ambion) with an oligo-dT-T7 primer (MessageAmp II-Biotin Enhanced Kit) or a random-T7 primer (included in the GeneChip^® WT Amplified Double-Stranded cDNA Synthesis Kit; Affymetrix Inc.). All subsequent steps were performed exactly as described above.

In order to analyze annotated genes, we mapped tiling probes to Arabidopsis gene models as per TAIR7 annotation 68. Probe sets for individual genes were defined as follows. From all probes mapped to exons (either coding or untranslated region) in their entire length, we retained only those for expression analysis that correspond to constitutive exons in all annotated splice forms of the same gene. We further excluded probes that mapped to more than one (overlapping) gene model, and in order to reduce cross-hybridization artifacts we also removed repetitive probes whose 25 mer sequence occurred multiple times in the genome. For expression measurements from tiling arrays we only considered the set of 30,228 annotated genes that are represented by at least three probes.

For the ATH1 array, probe sets were defined according to the A. thaliana CDF version 10 provided by the Microarray Lab at the Molecular and Behavioral Neurobiology Institute of the University of Michigan 69.

In order to minimize artificial expression level differences between platforms only resulting from differences in the computational analyses procedures, the RMA method was applied to hybridization data from both platforms 39. RMA proceeds in three steps. First, background correction and quantile normalization were applied before gene expression levels were calculated with the median polish method. Data from ATH1 arrays were analyzed using the RMA implementation in the Bioconductor package affy 707172. For the analysis of tiling arrays, we constructed a pipeline that combined the same background and quantile normalization methods from Bioconductor (BufferedMatrixMethods package by BM Bolstad), with the median polish routine extracted from Bioconductor sources (preprocessCore package by BM Bolstad) and adopted for the analysis of custom probe sets.

We applied the Rank Product method (Bioconductor package RankProduct) 4073 to identify significantly expressed genes at a cut-off of P < 0.05. The P values were also used to assess platform concordance by CAT analysis, in which gene lists ordered by P value were compared between platforms. The proportion of most significant genes in common between platforms was plotted as a function of list sizes increasing in steps of ten 42. As a measure of tissue-specific expression, z scores were calculated as described by Schmid and coworkers 5.

We preprocessed the hybridization signal to reduce a bias due to divergent probe sequences using a transcript normalization method 3574 and subsequently applied a modified version of the mSTAD algorithm 35.

For each sample, we trained mSTAD separately on mean intensities across replicates and used the trained instance only for prediction of array data from the same sample. To obtain unbiased whole-genome predictions we employed cross-validation. After splitting the genome between pairs of neighboring genes, one instance of mSTAD was trained on 500 of these genic regions and hyper-parameters were tuned on another 500 genic regions. We trained and tuned a second instance of mSTAD on two further disjoint sets of 500 genes each. For region-wise whole-genome predictions, we chose the mSTAD instance that had not seen the particular region during training and hyperparameter tuning (or a random instance if neither of them had). From the predicted labeling of tiling probes we extracted exon segments by assigning the genomic coordinates corresponding to the start of the first and the end of the last probe of a run of consecutive exon labels. The resulting segmentations are available as gff files and visualized in the At-TAX Generic Genome Browser.

Prediction accuracy was determined on genomic regions that had not been used for training or parameter tuning of the mSTAD instance evaluated. Sensitivity and specificity were assessed in comparison to annotated genes on a per-probe level as well as for the overlap between annotated and predicted exons. Figure 4a shows mean performance across 1,000 genic test regions (with at least five probes annotated as exonic and at least ten probes in total) chosen randomly for each of the mSTAD instances used to make whole-genome predictions for root data. Accuracy on probe level was also calculated for whole-genome (test) predictions for all other samples (see Additional data file 2).

To determine overlap with annotated regions, we used the TAIR7 annotation 11 and direct alignments with EST and cDNA sequences (downloaded from TAIR on 15 August 2007) 75. Sample-specific segments were obtained as residual after computing the overlap between predicted exon segments in the tissue of interest to those from all other tissues (Figure 4d). Similarly, we obtained predictions specifically made for polyA(±) conditions as exon segments that were predicted for both polyA(±) samples (ones that overlapped between samples), but did not overlap to predictions for any polyA(+) sample (Figure 5c).

One microgram of RNA from seedlings and young leaves was treated with DNaseI (MBI Fermentas, St. Leon-Rot, Germany) and converted into cDNA using the RevertAid™ First Strand cDNA Synthesis Kit (MBI Fermentas). One microliter of the resulting cDNA solution was used as a template in a PCR reaction with primers lying within the predicted transcribed region. The sizes of PCR products ranged from about 150 to 300 bp. A complete list of all used primers is available in Additional data file 3.

As an independent method to compare transcriptional activity between polyA(+) and polyA(±) samples, we computed transfrags as described previously 76 and implemented in the Affymetrix Tiling Analysis Software version 1.1 build 2. In order to select optimal parameters, we evaluated transfrags generated for root tissues for 900 different combinations of parameters in comparison with annotated genes (bandwidth in steps of 25 between 50 and 150, signal threshold between 5 and 13, minimum run in steps of 20 between 20 and 100, and maximum gap in steps of 20 between 40 and 100). As optimal setting for all transfrag computations we chose the one with maximal sensitivity at a precision similar to mSTAD predictions (bandwidth 100, signal threshold 6, minimum run 100, maximum gap 40; see Additional data file 8). Among nonrepetitive transfrags (at most 25% repetitive probes) comprising at least four probes and without overlap to annotated transcripts, the ones specific to polyA(+) or polyA(±) samples were computed the same way as for high-confidence mSTAD predictions (Figure 5d).