To search for cis-encoded Arabidopsis natural antisense transcripts, we aligned all Arabidopsis full-length cDNA sequences collected in the UniGene and RIKEN datasets with the Arabidopsis genome sequences. Pairs of transcripts that satisfied the following criteria were selected as cis-encoded natural sense-antisense transcript pairs (referred to as NAT pairs hereafter): first, cDNAs of both transcripts can be uniquely mapped to the Arabidopsis genome with at least 96% sequence identity; second, the two transcripts are derived from opposite strands of the genome; third, both transcripts are encoded by overlapping genomic loci, and the overlap length is longer than 50 nucleotides; fourth, the sense and antisense transcripts have distinct splicing patterns. Applying all of the above criteria, we identified 332 sense-antisense pairs from Arabidopsis full-length cDNAs. These NAT pairs are referred to as cDNA-NATs.

The 332 pairs of cDNA-NATs can be grouped into two categories. The first category contained 145 NAT pairs in which both the sense and antisense transcripts had nearly perfect annotated gene matches. The second category contained 187 NAT pairs in which at least one transcript had no corresponding annotated gene. This observation led us to hypothesize that additional NAT pairs, whose corresponding cDNAs were not included in the UniGene and RIKEN Arabidopsis full-length cDNA datasets, could be identified using the Arabidopsis genome annotation.

To identify potential NAT pairs without full-length cDNA evidence, we compared the genomic loci of all Arabidopsis annotated genes to search for gene pairs that overlap in an antiparallel manner. Using the criteria described in Materials and methods, 952 putative NAT pairs were identified from the Arabidopsis genome and were named genomic-NATs. Among the 952 genomic-NATs, 145 pairs had corresponding full-length cDNA for both the sense and antisense genes, and therefore were also included in the cDNA-NAT set. The remaining 807 new NAT pairs were predicted using the Arabidopsis genome annotation only and are referred as the unique genomic-NAT set in the following analysis (Figure 1a).

For most NAT pairs in the second category of the cDNA-NAT set, only one transcript in each pair matched an annotated gene. This indicates that transcripts of some full-length cDNAs could form cis-NAT pairs with other transcripts, although their corresponding genes are not included in the current Arabidopsis genome annotation. In a search of such NAT pairs, we compared the genomic loci of the UniGene and RIKEN Arabidopsis full-length cDNAs with those of annotated genes and identified 1,291 full-length cDNAs whose transcripts could form cis-NAT pairs with potential transcripts of annotated genes (see Materials and methods for criteria). The 1,291 genomic-cDNA-NAT pairs included the 332 cDNA-NAT pairs and 758 unique genomic-NAT pairs. Therefore, 201 unique NAT pairs were predicted by the cDNA-genome comparison approach and are referred to as unique genomic-cDNA-NAT pairs hereafter (Figure 1b).

In total, we have found 1,340 potential NAT pairs from three categories: 332 pairs with cDNA evidence for both sense and antisense transcripts; 807 pairs based on the Arabidopsis genome annotation (including 758 pairs with full-length cDNA evidence for one strand) and another 201 genomic-cDNA pairs by combining genome annotation with full-length cDNA sequence information.

We classified the 1,340 unique NAT pairs according to the exon-intron structures of each transcript and their overlapping patterns (Table 1). The overlapping patterns of NAT pairs were determined by comparing the exon positions of both transcripts using sim4 29 alignment results. Consistent with previous reports of NAT pairs in other organisms 2324252627, the majority of Arabidopsis NAT pairs (72.1%) overlapped at their 3' end. For almost all NAT pairs (99%), the overlapping region included exon sequences, with a few exceptions in which one transcript was transcribed entirely from the intronic sequences of the other. Figure 2 shows the distribution of overlap lengths of NATs. No obvious chromosomal bias was observed for the genomic distribution of NATs (Table 2) 30.

The sim4 cDNA alignment results showed that some Arabidopsis full-length cDNAs are non-spliced transcripts. To assess the quality of full-length cDNAs, we systematically compared the splicing pattern and coding potential of all full-length cDNAs used in this study to all predicted Arabidopsis genes. Our result showed that the proportion of non-spliced transcripts in UniGene and RIKEN full-length cDNAs was lower than the proportion of non-spliced transcripts in annotated genes, indicating non-spliced cDNAs are likely to be derived from bona fide transcripts rather than genomic DNA contamination (Table 3).

To investigate the expression of our predicted NAT pairs, we used the public Arabidopsis MPSS data at the University of Delaware 31. MPSS is a bead-based sequencing technology that identifies a sequence of 17-20 nucleotides from each transcript. This sequencing technique is capable of identifying new, rarely expressed transcripts. MPSS can also quantitatively measure the expression level of a transcript because the transcripts per million (TMP) value for a transcript in the sequencing results reflect its in vivo abundance 3233.

The public Arabidopsis MPSS database contains 87,705 'trusted' signature sequences from 14 cDNA libraries. By aligning these MPSS sequences to the Arabidopsis genome and the 1,340 NAT pairs, we identified 455 NAT pairs with unique MPSS matches on both the sense and antisense strands, including 103 cDNA-NAT pairs, 293 genomic-NAT pairs and 59 genomic-cDNA-NAT pairs. Because MPSS signatures are short 17-nucleotide sequences identified from each transcript, sequences with multiple genomic loci were excluded from our analysis to avoid ambiguity with respect to the origin of a MPSS signature and to ensure fidelity of assigning a MPSS signature to its corresponding transcript (see Materials and methods for details). Among the 455 NAT pairs with unambiguous MPSS data for both transcripts, expression of both transcripts of 78 pairs was only found in distinct libraries, indicating these NAT pairs might have an exclusive transcription relationship. For the other 377 NAT pairs, expression of the sense and antisense transcripts was mainly observed in different libraries or one transcript was dominantly expressed when both transcripts could be detected in the same library (Tables 4 and 5). For a pair of NATs found in the same library, if the TPM value of one transcript is at least three times as high as that of the other transcript, we consider that transcript as dominantly expressed. The number of coexpressed and dominantly expressed transcripts in each library was shown in Figure 3. On average, coexpression was only observed in two of the 14 tested sample libraries for each of the 377 NAT pairs, whereas dominant expression of one transcript was observed in 9 libraries. No expression was detected in the remaining libraries.

We also found additional 222 genomic-NAT pairs and 51 genomic-cDNA-NAT pairs with full-length cDNA evidence for one transcript and MPSS data for the other transcript. Together with the 332 cDNA-NAT pairs, we have obtained either full-length cDNA or MPSS expression evidence for both transcripts of 957 NAT pairs, corresponding to 71.4% of the total 1340 pairs ((455 - 103) + 332 + 222 + 51 = 957).

We compared short interfering RNA (siRNA) sequences collected in the Arabidopsis small RNA database to investigate the possibility that cis-NAT pairs may generate siRNAs. Similar to the MPSS alignment process, only siRNAs with unique loci on the Arabidopsis genome were used in the comparison to ensure unambiguous assignment. We found 11 pairs of NATs had siRNA sequences mapped uniquely to their overlapping region (Table 6). SiRNAs of all but one NAT pairs originated from their overlap region, the only exception being pair At#S18901030 and At#S18898439, whose overlap length was only 52 nucleotides long.

To examine whether NAT pairs might be conserved during evolution, we compared the protein sequences of the 1,340 putative Arabidopsis NAT pairs with the protein sequences of the 687 predicted rice NAT pairs 27. Orthologs of two Arabidopsis NAT pairs were also encoded by antiparallel genes originated from the same locus in rice (Table 7). In addition, homologs of one transcript of 392 Arabidopsis NAT pairs were also found in the rice NAT set.

The 1,340 Arabidopsis NAT pairs were identified from three sources. First, by aligning full-length cDNA sequences to the Arabidopsis genome, we identified 332 cDNA-NAT pairs. However, comparison of these 332 cDNA-NAT pairs with Arabidopsis annotated genes showed that more than half of these NAT pairs had one partner that was not included in the current Arabidopsis genome annotation. Because traditional genome annotation mainly aims at the identification of protein coding genes within a genome, there is the possibility that non-coding antisense transcripts may be overlooked by currently trained gene finders. A recent report using a genome tiling array to examine the transcription activity of the entire Arabidopsis genome also supports this notion 28.

To search for potential NAT pairs not included in the current full-length Arabidopsis cDNA library, we compared the genomic coordinates of all annotated genes with each other and with those of full-length cDNAs. This approach uncovered another 807 overlapping genomic-NAT pairs based on the annotation of their corresponding genes, and 201 genomic-cDNA-NAT pairs, each including a transcript derived from an annotated gene on one strand and a transcript represented in the full-length cDNA database on the other strand. The full-length cDNAs included in genomic-cDNA-NAT pairs either had no annotated gene match or their corresponding transcripts cannot form cis-NAT pairs with transcripts of other genes based on their annotation. These results indicate that although the Arabidopsis genome is currently one of the best annotated eukaryotic genomes, a lot of information is still missing. The identification in eukaryotes of several classes of regulatory RNA genes, such as those encoding natural antisense transcripts, which are the focus here, will not only further our understanding of genome structure and gene regulation, but will also open a new window for improved genome annotation.

Most antisense prediction work reported to date has focused on identifying NATs from expressed cDNAs and ESTs 2324252627. In this work, we avoided using ESTs because of the ambiguous orientation of some sequences. We also included sequence information of annotated Arabidopsis genes in our NAT prediction in order to provide a more complete picture of antisense transcripts in Arabidopsis. The reliability of our approach is supported by the following lines of evidence: first, the expression of both sense and antisense transcripts of 293 pairs of genomic-NATs (36.3% of a total of 807) was observed in the public MPSS data, and another 222 genomic-NAT pairs (27.5% of a total of 807) have full-length cDNA evidence for one transcript and associated MPSS data for the other transcript; second, the two NAT pairs which were conserved in rice were also identified in our Arabidopsis genomic-NAT dataset; third, it is known that imprinted genes are likely subject to antisense regulation; three of the six reported Arabidopsis imprinted genes 3536373839, FIE, FIS2 and MSI1, are included in our genomic-NAT sets. However, it remains possible that some genomic-NAT pairs are false positives if the lengths of their untranslated regions (UTRs) were annotated inaccurately.

In rice, both transcripts of 86% of the NAT pairs have coding sequence (CDS) regions whereas 28% of the predicted Arabidopsis NAT pairs include at least one transcript without coding potential. Non-protein-coding transcripts are more prevalent in cDNA-and genomic-cDNA-NAT pairs in that 170 cDNA NAT pairs and 156 genomic-cDNA-NAT pairs include one non-protein-coding transcript. We used Genescan to evaluate the coding potential of each transcript by screening their corresponding genomic DNA sequence for valid gene structures. Using annotated genes as controls, we estimated the false-negative rate of our definition of coding potential to be 2.3%. Unlike CDS-containing antisense transcripts that may be translated into proteins under certain conditions, transcripts without any protein-coding potential could possess solely regulatory functions.

In our work described here, and in all other genome-wide antisense transcript identification papers published so far 2324252627, the investigation was focused on cis-antisense RNAs, which are transcribed from the same genomic loci as their sense RNAs, but on the opposite genome strand. To ensure the cis-antisense relationship of NATs reported here, only cDNAs with unique genomic loci were included in this study. We note that certain number of trans-antisense transcripts also exist in cells. Examples include miRNAs and siRNAs which are widely studied in most model organisms 6. Genome-wide identification of trans-antisense transcripts in Arabidopsis is being attempted.

The non-gel-based properties of MPSS technology render it an ideal resource for evaluating the expression profile of NAT pairs for the following reasons: first, because the MPSS technology captures almost all polyadenylated transcripts within cells, this technology is theoretically capable of identifying new, rarely expressed transcripts without prior knowledge of their corresponding genes; second, the digital result of MPSS reflects the expression pattern of a sequenced RNA molecule, and therefore provides a quantitative relationship between the sense and antisense transcript of a NAT pair in different tissues. This information was not available in any of the previous NAT prediction work 3233.

Using the full-length cDNA and public Arabidopsis MPSS data, we were able to obtain expression evidence for both transcripts of 957 NAT pairs. The digital nature of MPSS data enabled us to evaluate the expression relationship of the sense and antisense transcripts directly. Our results showed that the sense and antisense transcripts of a NAT pair tend to be expressed in different tissues or under different conditions. In addition, in cases where the sense and antisense transcripts of a NAT pair were expressed in the same library, one type of transcript was usually more abundant than the other. On average, transcripts of NAT pairs were found to be coexpressed in only two libraries, whereas dominant expression (the expression level of one transcript was at least three times higher than that of the other transcript) or absolute expression (only one transcript of a NAT pair was expressed) was observed in nine libraries. The tissue-specific expression of sense and antisense transcripts observed in this study is consistent with the Arabidopsis genome transcription study using a whole genome-tiling array, in which about 7,600 genes were found to have tissue-specific sense and antisense expression 28. Although a detailed list of these 7,600 genes is not yet available, it is possible that for some genes not included in our list, the antisense transcription activity was contributed by trans-antisense transcripts. This could explain why we predicted fewer NAT pairs than the previous work, as our work only focuses on cis-antisense transcripts.

To ensure the MPSS sequences were indeed generated by their matching transcripts, all MPSS data were first aligned with the Arabidopsis genome and all annotated mRNAs to remove signatures with multiple genomic loci. Therefore, unless an MPSS signature sequence was derived from the joint-exon region of some transcripts that are not included in the current genome annotation, it should originate from its corresponding transcript.

One possible function of NATs is to trigger the degradation of their sense transcripts via the RNAi pathway. However, in our study, we found only 11 NAT pairs with known siRNA matches. There are two possible explanations for this observation. First, the current public Arabidopsis siRNA database, which only contains 1,822 unique siRNA sequences, is small and does not cover all siRNAs associated with sequences of the NAT pairs reported here. Second, all NATs identified in this work are cis-antisense transcripts. siRNAs are used to downregulate expression levels of their target mRNA to achieve a low protein concentration. Cis-antisense transcripts can accomplish the same goal by interfering with the transcription of their sense transcripts, and this might be a more energy-efficient mechanism to achieve local gene regulation. This hypothesis predicts that we would expect to find more siRNAs associated with trans-antisense transcripts.

For most NAT pairs with associated MPSS data for both transcripts, the expression of sense and antisense transcripts tends to occur in different tissues. In these cases, we could speculate that transcription of genes encoding these NAT transcript pairs may be regulated by similar factors but that the production of antisense transcripts might interfere with the transcription of their sense transcripts, resulting in reciprocal expression patterns. Another possibility is that the two genes of a NAT pair are subject to different transcriptional regulation and consequently they are never expressed in the same tissue at the same time. Functional analysis of all NAT pairs using gene ontology reveals no over-representation of any functional category compared to the Arabidopsis genome, indicating that cis-antisense regulation might be a global mechanism for all gene families. Further experiments are needed to investigate the validity of these hypotheses.

Antiparallel transcription and antisense transcripts are known to be involved in genomic imprinting of Xist gene in mouse and human 21. There is supporting evidence that the MEA and PHE genes of Arabidopsis are imprinted 35, and in addition, FIS2, FIE, MSI1 and FWA may also be imprinted, although the evidence for these four other genes is not unequivocal 36373839. Nonetheless, we found antisense transcription units for FIS2, FIE and FWA, suggesting that transcription of these three genes might be regulated by antisense transcripts, or their antisense transcripts might be involved in silencing their expression. Genomic imprinting usually involves a chromosomal locus and, in certain cases, may even extend overa chromosomal region. Given the close proximity of the sense-antisense gene transcripts if one member of the pair is imprinted, it is likely that the other would be subject to the same regulation. Unfortunately, because of the absence of data on imprinted genes in rice, we were unable to examine whether imprinted genes were also subject to antisense regulation in rice.

We found that two Arabidopsis NAT pairs are conserved in rice. These conserved NAT pairs could be used to study the antisense regulatory mechanism and the origin of NATs in plants. Given over 150 million years of evolutionary distance between Arabidopsis and rice, the gene order on the two genomes has diverged quite significantly. Therefore, the conservation of these two NAT pairs might have some functional relevance. A closer comparison of the Arabidopsis and rice NAT pairs and the identification of additional conserved NAT pairs could help address this issue.

Taken together, our results provide the first genome-wide identification and prediction of NATs in Arabidopsis. These results will facilitate functional studies of NATs in this model plant, as well as in other plant species, and help to unravel complex gene regulatory networks in eukaryotes.

The Arabidopsis UniGene (Build 45) dataset (file named At.seq.all) was downloaded from the National Center for Biotechnology Information (NCBI) UniGene Resources 4041. A total of 20,683 full-length cDNA sequences were extracted from the UniGene dataset by selecting sequences marked as 'Full-length/full-length cDNA'. The RIKEN Arabidopsis full-length cDNA dataset, which contains 13,181 sequences, was downloaded from the RIKEN BioResource Center (BRC) 4243. The 20,683 UniGene and 13,181 RIKEN full-length cDNAs were aligned to the Arabidopsis genome sequences from The Institute for Genomic Research (TIGR) (release version 5) 44 by BLAT. The splicing pattern of the transcript derived from each cDNA was further confirmed using the sim4 sequence alignment program 294445. cDNAs with at least 96% sequence identity to the Arabidopsis genome were used in the following analysis. For pairs of cDNAs encoded by opposite strands of the Arabidopsis genome and sharing overlapping genomic loci, if both their corresponding sense and antisense transcripts had no other genomic locations and exhibited different splicing patterns, they were selected as encoding sense-antisense transcript pairs and are referred to as cDNA-NAT pairs in the text.

We used the A. thaliana genome annotations from TIGR (release version 5) in this study 4445. Putative NAT pairs were identified on the basis of annotated genomic loci of Arabidopsis genes. If a pair of overlapping genes were located on opposite strands of the Arabidopsis genome and at least one gene had no annotated UTR at the overlap end, their encoded transcripts were selected as a putative NAT pair regardless of the overlap length of the encoded transcripts. Otherwise, if a pair of antiparallel overlapping genes both have annotated UTR regions at the overlap end, the overlap length of their encoded transcripts must be longer than 50 nucleotides to qualify as NAT pairs. NAT pairs from the above two categories are both referred to as genomic-NAT pairs in the text.

Genomic-cDNA-NAT pairs were identified by comparing the genomic loci of full-length cDNAs with those of annotated genes. UniGene and RIKEN full-length cDNAs with unique genomic locations and at least 96% sequence identity to the Arabidopsis genome were used in this step. Using the same criteria for genomic NATs, if an annotated gene had a overlap cDNA match on the opposite strand and the transcript of the annotated gene and that derived from the antisense cDNA had different splicing patterns, the gene and its matching cDNA were selected as a genomic-cDNA-NAT pair.

Splicing patterns of transcripts encoded by full-length cDNAs were obtained by aligning the cDNA sequences to the Arabidopsis genome using the sim4 program 29. Splicing patterns of transcripts derived from Arabidopsis annotated genes were extracted from the TIGR Arabidopsis genome annotation (release version 5) 44. To evaluate the coding potential of full-length cDNAs, their corresponding genomic sequences (determined by BLAT and sim4 result) were extracted and screened by GeneScan 46.

We used the public Arabidopsis MPSS data at the University of Delaware 31 to evaluate the expression of NAT pairs. MPSS sequences from 14 different libraries of Arabidopsis Columbia-0 (Col-0) ecotype were downloaded from 31. Each MPSS library contained signature sequences identified from the same tissue. The quality of these MPSS sequences was evaluated according to the information provided by the database. Only MPSS sequences with 'reliable' (present in more than one sequencing run) and 'significant' (TPM ≥ 4) expression pattern were considered as 'trusted' signatures and used in this analysis.

The public MPSS database contained 87,705 trusted signatures that satisfied the above expression criteria. These signatures were aligned to the sequences of the 1,340 putative NAT pairs to identify MPSS sequences derived from them. Signatures with multiple perfect matches to the Arabidopsis genome or to cDNAs had ambiguous origins and were not considered further. For a NAT pair, if both the sense and antisense transcripts had associated MPSS data and their expression values were both significant in one or more libraries, transcripts in this NAT pair were considered as coexpressed in the same tissue. On the other hand, if both transcripts had MPSS data but had no significant coexpression in any of the examined libraries, then the transcripts were considered as expressed, but in different libraries.

Full-length cDNA sequences of the 687 rice NAT pairs were downloaded from the website described in 27. To facilitate protein sequence comparison, the rice and Arabidopsis cDNAs were mapped to their corresponding genomes by BLAT 45. Both the A. thaliana and O. sativa genomes were downloaded from TIGR 44. The corresponding genomic sequences of each cDNA were extracted according to their genomic coordinates from the BLAT results. Protein sequences were obtained by evaluating the genomic sequences of those cDNAs using GENSCAN 46. The protein sequences of rice NATs were aligned with those of Arabidopsis NATs using blastp 47. High similarity pairs with E-value less than 10^-30 and alignment coverage greater than 50% of query sequence were considered as homologous sequences.