A complete annotated genome sequence of Arabidopsis thaliana was released by the Arabidopsis Genome Initiative (AGI) in the year 2000, the first completed plant genome[1]. Since then, our understanding of the Arabidopsis genome structure and transcriptome has been improved through the release of 4 sequential updates to the annotation, culminating in The Institute for Genomic Research's release 5 (TIGR5), which forms the basis of the work presented here. Following the TIGR5 annotation release, responsibility for maintaining and updating the Arabidopsis annotation was turned over to The Arabidopsis Information Resource (TAIR), which has since released version 6 of the Arabidopsis annotation (TAIR6). Over the course of the TIGR annotation releases, the number of annotated protein-coding genes of Arabidopsis has increased from 25,498 (a number that included transposons and pseudogenes) to a final total of 26,207 protein coding genes plus 3,786 regions annotated as transposon-related or other pseudogenes in the final TIGR release. At the same time, the size of the Arabidopsis pseudomolecules has increased from 115 MB in the initial 2000 release, to 119 MB in TIGR5 due to the inclusion of additional finished and unfinished BACs.

While the sequential TIGR re-annotations of the genome have been relatively stable in terms of overall gene density and gene structure statistics, the major benefits of the re-annotation efforts have come from the incorporation of expressed sequence tags (ESTs) and full length complementary DNA (FL-cDNA) clone sequences into the Arabidopsis annotation, improving the accuracy of individual gene structures [2–4]. However, transcripts from the most lowly expressed genes, or genes specifically expressed in important but relatively minor cell types such as meristems or the Arabidopsis gametophyte stage may very likely be under-represented in the over half million ESTs available through GenBank. To provide experimental support for genes lacking EST or other cDNA evidence, we have previously carried out high-throughput Rapid Amplification of cDNA Ends (RACE) experiments and generated partial or complete sequence for over 1000 genes, leading to the improvement of many gene structures [5, 6].

Genome annotation is never complete or final. Since its release in January of 2004, various lines of evidence have come to light which suggest that the TIGR5 annotation still paints an incomplete picture of the Arabidopsis gene space and transcriptome. Continued submission of ESTs and other sequence information to GenBank reveals the existence of transcripts that do not map to currently annotated genes [7, 8]. These may represent novel protein coding genes, genes which code small unknown peptides, or may also represent non-coding RNA. Additionally, evidence of transcription in un-annotated intergenic regions of the genome has been seen through Massively Parallel Signature Sequencing (MPSS) efforts which reported several thousand transcript signatures from un-annotated intergenic regions [9]. Analysis of whole-genome tiling arrays to examine the Arabidopsis transcriptome have also provided strong indications for the presence of over five thousand novel transcriptional units [10, 11]. A survey of the Arabidopsis genome for a family of divergent cysteine rich anti-microbial defensin-like peptides yielded over 300 genes, 80% of which were absent from TIGR's Arabidopsis annotation [12].

The wealth of new sequence data for other plant species that has become available since the landmark release of the Arabidopsis genome has now allowed for the refining and improvement of gene detection based upon comparative genomic analysis. Comparative genomics techniques have been proven extremely valuable for identifying conserved genes and regulatory elements in a variety of closely related species and has been already been applied effectively to the human genome [13–15], as well as the malaria parasite genome [16] and the C. elegans genome [17], among others. The Arabidopsis genome annotation is also beginning to benefit from comparative genomic analysis. A comparative study of Arabidopsis thaliana and Brassica oleracea yielded a large number of Conserved Arabidopsis Genome regions (CAGS), 72% of which aligned with predicted genes [18]. The remaining intergenic CAGS suggest the existence of several thousand currently un-annotated genes. RACE experiments have demonstrated transcriptional activity at 58 of 192 targeted CAGS, demonstrating that the CAGS may correspond to conserved un-annotated genes. A separate, similar investigation comparing the same Brassica olreacea sequence set with the Arabidopsis genome resulted in the identification of 25 genes that were missed by the TIGR annotation [19].

Two relatively new gene prediction tools that make use of comparative genomics have been deployed for analysis of the Arabidopsis Genome: Twinscan [20] and EuGene [21]. Twinscan is a gene-structure prediction program that extends the probabilistic models employed by the ab initio gene finding program GENSCAN [22]. Twinscan exploits cross-species homology between closely related genomes to produce improved gene models. EuGene is another gene prediction program developed to make use of comparative genomics for improved gene models. EuGene makes use of multiple homologous sequences (including ESTs, protein sequences and genomic homologous sequences) from closely related organisms, tblastx analysis, splice site analysis and probabilistic models to provide gene predictions. Both the Twinscan and EuGene programs have been applied to the Arabidopsis genome, resulting in predictions for 30,635 and 28,530 protein coding genes, respectively. Both Twinscan and EuGene predict more genes than currently exist in the TIGR5 annotation, and there is significant overlap between the predictions produced by each program and with the TIGR5 annotation. It is only within the more challenging un-annotated regions of genome where less evidence exists to support gene predictions where the overlap between Twinscan and EuGene predictions begins to decrease significantly. In addition to this work, Twinscan has been applied broadly and with success to refine the annotation of the C. elegans [23], chicken [24], and rat [25] genomes. The target for EuGene at present is plant genomes and in addition to Arabidopsis, has been deployed for poplar [26] and barrel medic [27].

Like the hypothetical genes of Arabidopsis that we have studied previously [5, 6], most of the novel genes predicted by Twinscan and EuGene lack experimental support. To assess the validity of these Twinscan and EuGene predictions we have applied a high-throughput RACE pipeline and have verified the presence, structure, and expression of several hundred currently un-annotated genes, of both deducible and potentially novel function.

For the most part, the Arabidopsis genome annotation has relied heavily on the presence of ESTs and FL-cDNA sequences, along with ab initio gene predictors such as Genscan [22] and Genemark.hmm [31] to identify the genome's set of protein coding genes. Both Twinscan and EuGene have, in direct ways, used sequence information from related species to expand gene prediction in the Arabidopsis genome. We have used our high-throughput RACE pipeline to assess the reliability of these predictions and have verified the presence of several hundred currently un-annotated genes that were predicted by the Twinscan and/or EuGene programs.

Our decision to use RACE to verify the expression and structure of these un-annotated genes was necessitated by their low level of expression and uncertain gene structures. These genes were not captured by previous large scale EST sequencing efforts and were also not represented in any significant proportion in a normalized Arabidopsis cDNA library sequenced in house (unpublished data). By specifically targeting these most lowly expressed genes with 5' and 3' RACE, we were able to capture transcripts for over a third of the targeted un-annotated genes. However, since our targets excluded many small EuGene predictions of less than 180 bp in length, we can not be certain that the success rates we observed with RACE can be extrapolated to the total set of intergenic predictions. It is unclear whether the remainder of the un-captured targets were not expressed, differed significantly from their predictions, were not present in our cDNA populations at high enough levels to ensure reliable amplification or were not captured due to failure of PCR. In this study, we did not attempt to re-target un-annotated genes with our RACE pipeline, although previous attempts to re-target hypothetical genes by RACE resulted in successful amplification of ~ 40% of the re-targeted genes when using new cDNA populations, suggesting that the relative abundance of signal and the heterogeneity of the cDNA population may likely be a success-limiting factor (unpublished data).

In addition to verifying expression of novel genes by RACE, we have also demonstrated tissue specific activity of intergenic promoters using promoter-reporter fusions, as well as by examining enhancer trap tagged mutants obtained from Cold Spring Harbor Laboratory's Trapper collection. With our promoter-reporter fusions, we observed tissue specific reporter gene expression with one of six promoters tested. With lines obtained from CSHL, we observed reporter gene expression from lines tagging 2 of 3 genes. The lack of expression observed with five of our promoter-reporter lines as well as a gene trap line obtained from CSHL's collection is likely due to either a very low level of expression directed by those promoters, or a very specific pattern, timing, or condition for expression that was not tested by our assays. The pools of cDNA which served as our RACE template originated from several biotic and abiotic treatments that we did not examine in relation to our promoter-reporter constructs.

Overall, we verified the expression and structure of 378 un-annotated genes, 27% of which do not display similarity to any annotated proteins. Furthermore, nearly 50% of the genes described herein are most similar to hypothetical proteins or other proteins of unknown function. Examining the putative functional roles of the remaining un-annotated genes, we can begin to speculate the reasons that many were overlooked by previous annotation efforts. Single copy, lowly expressed genes such as the putative alpha 1,-6, mannosyltransferase corresponding to Twinscan prediction At.chr1.1.117, which does not have similarity with other Arabidopsis genes was likely overlooked due to the lack of support from conserved EST sequences. On the other hand, At.chr1.6.385, which encodes a CLAVATA3/ESR related protein and is a member of a divergent multi-gene family, was likely overlooked due to a combination of lack of EST support [30], and a relatively small coding size. Similarly, we have identified members of a large and divergent gene family encoding Cysteine Rich Peptides. The small size and divergent sequences of this family have contributed to their under-representation in the genome annotation [12]. Additionally, approximately 50% of the top protein hits for the newly verified genes are to hypothetical proteins, or other poorly characterized proteins having unknown function and low expression levels. These examples underscore the limitations of gene prediction programs that still rely to a large extent on training sets derived from relatively abundantly expressed proteins. The incorporation of genomic sequence data from related organisms allows for the easier identification of such genes.

We observed alternative splicing in over 30% of the genes validated with these efforts. This percentage is comparable to that found previously in our work targeting Arabidopsis hypothetical proteins, and higher than the 21.8% of Arabidopsis genes recently reported to be alternatively spliced by Wang and Brendel based upon large scale EST analysis [32]. Our specific RACE-based approach to verifying these genes, using a pool of cDNA from several tissues and treatments resulted in deep sequence coverage and allowed us to capture multiple splicing isoforms. The depth of data that we obtained by sequencing up to 24 clones per gene also allowed us to observe splice isoforms with more regularity than past sequencing efforts. It is uncertain whether the alternatively spliced isoforms identified play a specific biological role, or are simply mis-spliced transcripts.

Both Twinscan and EuGene performed well at identifying un-annotated genes within the Arabidopsis genome. Our success rate for capturing full length sequence information for these genes is in line with our past success in targeting annotated hypothetical proteins. In terms of accuracy, Twinscan and EuGene predicted gene structures with comparable success, though EuGene predicts many smaller, single and multi-exon genes than Twinscan. Based upon the number of intergenic loci predicted by these programs, and our success rates when targeting these loci with 5' and 3' RACE, it is likely that the Arabidopsis genome contains many new genes and other transcribed regions that have yet to be identified.