Genetically modified organisms have been engineered through the stable integration of a recombinant genetic cassette into the genome of a recipient organism. The purpose of generating a genetically modified organism (GMOs) is, like breeding in general, to provide the new variety with novel features, and for introduced traits to be inheritable, the nuclear or organellar genome has to be altered. Protein coding mRNAs represent a causal starting point for most metabolic processes and structural components of a cell, and a cell's pattern of RNA transcription reflects the coding potential of its genome. For a genetic modification to have an effect, it is thus also vital that it changes the coding capacity of the recipient cell.

The strategy most commonly used when generating genetically modified plants that are commercially relevant is to introduce a genetic construct that either confers some kind of advantage when it comes to farming/storage or increases the nutritional quality of the end product. Among the most widely used genetic features are genes that encode herbicide tolerance, insect resistance or improve content of key nutrients http://www.agbios.com/. In addition to these trait genes, various selection markers are also usually introduced in order to simplify the process of GMO generation. These genes include herbicide resistance genes such as the bialaphos resistance gene (bar) from Streptomyces hygroscopicus [1], antibiotic resistance genes such as the neomycin phosphotransferase II gene (nptII) from Escherichia coli found in the Flavr Savr tomato [2] or positive selection markers such as the phosphomannose isomerase gene (pmi) from E. coli (used in for instance Golden Rice, see [3]). Careful examination of the pool of transcripts found in a plant should therefore reveal whether or not a plant has been genetically modified.

Recently, a new strategy for identification of foreign nucleic acids (DNA or RNA) called computational subtraction has been described for pathogen discovery in human diseases of unknown etiology [4]. In short, the approach takes advantage of the fact that for a growing number of species the complete genomic sequence has now been generated, and sequencing costs have been dropping dramatically in recent years. Using sequence similarity search algorithms it is thus possible to analyze DNA or RNA sequence data from a sample, compare the sequences against a set of reference sequences, and filter away all the endogenous ('expected') reads, leaving a small collection of sequences that do not appear to stem from the organism in question. This principle appears to work well even when subtracting short sequence tags [5], and should be an efficient way to identify for instance unexpected transcripts.

We have attempted to use high massively parallel pyrosequencing and the concept of computational subtraction to look for allochthonous transcripts in a transgenic line of Arabidopsis thaliana. We also explore the concept of computational subtraction in silico using expressed sequence tag (EST) data from transgenic rice and papaya.

Most of the methods currently used for characterization of (unknown) genetic modifications rely on PCR [7]. This approach assumes some knowledge about the target sequence, as it relies on primer design. High density array-based methods that make fewer assumptions about the nucleic acids to be detected have been suggested and developed [8, 9], but even here some basic assumptions have to be made. By using high throughput sequencing of either a cDNA or a genomic/organellar DNA library, it should be possible to detect any novel transcript or genetic construct. The exception would be if one works with cDNA and the target organisms' only novel feature on the expression level is the increased or reduced expression of an otherwise endogenous gene [10].

Computational subtraction might also be performed using genomic DNA instead of mRNA. The number of sequences that need to be derived for computational subtraction to be effective when working with transcripts will depend upon the frequency and length of the transgenic mRNA versus the pool of endogenous mRNA and small transgenic transcripts and/or a low level of expression will require deeper sequencing. The same principle applies to computational subtraction using genomic DNA, but here the size of the inserted construct relative to the target genome will be the most important factor [11]. Using A. thaliana transformed with pBI121 as an example, the insert size is 6,192 bases (GenBank accession number AF485783) and the genome size of A. thaliana is 125,000,000 basepairs [12] (excluding mitochondrial and chloroplast genome). If we had sequenced 58,933 genomic tags, we could have expected only to find <3 sequence tags that had sequence overlap with the insert.

One way to increase the likelihood of picking up GM-specific nucleic acids would be to do an in vitro physical subtraction of the DNA/RNA before library preparation. This would reduce the amount of nucleic acids that the sample would have in common with a (wildtype) reference and increase the relative amount of the GM-associated DNA or transcripts. There are kits available for performing suppressive subtractive hybridization based on published techniques [2] and subtractions can also be performed commercially (offered by for instance by Eurofins MWG/Operon, see Products & Services at http://www.eurofinsdna.com/).

Regardless of what the starting material for library preparation is, the target organisms' transcriptome and/or genome must be well characterized. Sequence filtering might be done using data from a close relative (see for instance the use of mouse data in [4]), but this alone will not be sufficient when working with a high number of sequence reads. At time of writing, ten large plant genome sequencing projects had been completed: Arabidopsis thaliana (thale cress), Glycine max (soybean), Phoenix dactylifera (date palm tree), Medicago truncatula (barrel medic), Oryza sativa (rice), Populus trichocarpa (black cottonwood), Sorghum bicolor (sorghum), Vitis vinifera (wine grape), Carica papaya (papaya) and Zea mays (corn). Many more species are slated to be sequenced in the near future http://www.ncbi.nlm.nih.gov/genomes/PLANTS/PlantList.html, so we believe that for the major crop species this will not be a limiting factor for long.

A possible example of the potential benefits of such an approach was observed in our collection of downloaded EST libraries where a library from a unpublished project entitled 'Subtractive cloning of differentially expressed mRNA from transgenic rice plants' was found (library name: Oryza sativa cv. Pusa Basmati-1). This library comprised only 9 sequences, but even with this small number a reads, a transgenic EST could be detected. The 242 basepair sequence found (accession number AJ309294) was a 100% match with the gene trapping Ds/T-DNA vector pDsG8 designed for insertion mutagenesis in rice [13].

The data generated in this study can also be used to search for other novel transcripts than those that represent transgenic candidates. Careful examination of the 5.568 transcripts that were found that did not match the reference A. thaliana transcriptome but matched the genome sequence well (Table 1; 5,727-159 = 5,568), revealed several potentially novel, spliced endogenous genes (data not shown). We do not believe that these transcripts are directly linked to the genetic modification, but this merely demonstrates how versatile data generated using high throughput sequencing of cDNA libraries can be.

As the amount of available sequences data increases and DNA sequencing costs drop, we believe that a sequencing-based approach using computational subtraction will be feasible for the detection, characterization and risk assessment of genetic modifications. In this pilot study we have shown that transgenic cDNA can be detected using genetically modified plants as a model system.