All genomes are composed of nucleotides, which are represented abstractly as letters (Adenine (A), Guanine (G), Cytosine (C), and Thymine (T)). Strings of such letters can be conceptualized as words, which provide the blueprints for organisms. Each word is found a specific number of times in a particular genome. Note that the expected frequency of a word is inversely related to the word's length. Some nucleotides appear more frequently than others (e.g. A/T in Arabidopsis), giving each genome a distinct (G+C)% content and biasing expected word frequencies. Higher order frequencies (dinucleotide and trinucleotide) also show distinct biases beyond those expected for single nucleotide frequencies .
Distinct selective pressures shape words positioned in different genomic regions. For example, a word in an open reading frame (ORF) has a direct influence on the primary amino acid sequence of a protein and hence is under strong selective pressure. In contrast, words in introns are likely to be under more relaxed selective constraints, unless they are important for gene functions, for example by providing docking sites for splicing factors  or for enzymes involved in the post-transcriptional processing of a transcript [3, 4]. The gene sections corresponding to the 5' and 3' untranslated regions (5'UTRs and 3'UTRs, respectively) are also likely to be under less selective constraints than the ORFs, yet signatures of strong selection in UTRs have been described (reviewed in ). The constant formation of DNA microsatellites through slippage by the replication machinery, and the action of viruses and transposons, also complicate the word landscape, especially in regions with lower selective constraints (such as introns, UTRs and intergenic regions) [6, 7].
This manuscript describes the results of a genome-wide analysis to discover putative regulatory words. Within this context, we define the cis-regulatory apparatus as all the DNA segments that are located proximal to a gene, and that also contribute to the gene's expression. It is the function of transcription factors, miRNAs, or other molecules that interact with DNA, to interpret the words (sequence code) hardwired in the cis-regulatory apparatus and to 'execute' them, thereby generating signals to the basal transcription machinery that result in changes to the rate of RNA production by the corresponding DNA-dependent RNA polymerases. When located upstream of the transcription start site (TSS), the cis-regulatory apparatus is often referred to as the promoter of a gene.
Promoters are typically divided into three regions: core, proximal and distal. The core promoter, a region at location [+1;-100] relative to the TSS, performs a central role in the formation of pre-initiation transcriptional complexes. Immediately upstream of the core promoter is the proximal promoter, which is located at position [-101;-1000] relative to the TSS and serves as a docking site for transcription factors. The distal promoter is located at [-1001;-3000] relative to the TSS and contains the regulatory elements that are commonly known as enhancers and silencers. The participation of a particular DNA segment in the regulation of gene expression can only be demonstrated experimentally. Thus, understanding the rules at play in deciphering the transcriptional regulatory code remains one of the most significant challenges in biology today.
Although most regulatory elements are present in the UTRs and upstream regions, due to their proximity to the TSS, studies have shown the presence of regulatory elements in introns, and, to a much lesser extent, in coding regions [2, 8–16]. Building on this knowledge, a segment-based analysis was performed that is focused on non-coding regions within the open reading frames (i.e., introns) and flanking non-coding regions (i.e., UTRs and upstream regions). The coding regions were omitted from this analysis because they are under other selection pressures corresponding to the amino acid sequences of the proteins they produce, and thus they are subjected to biases other than regulation.
Arabidopsis thaliana provides an ideal reference organism to investigate the word landscape of a plant genome, and to relate said landscape to important biological features. The Arabidopsis genome consists of 125 Mbp arranged into five chromosomes [17, 18]. The genome is well annotated and regions corresponding to introns, 3'UTRs, 5' UTRs, and intergenic genomic spaces are all available from The Arabidopsis Information Resource (TAIR, http://www.arabidopsis.org) .
Many studies have characterized Arabidopsis DNA sequence motifs that participate in the regulation of particular genes (e.g., [20–23]), and public databases such as AthaMap  and AGRIS  provide comprehensive collections of cis-regulatory elements likely to participate in the regulation of gene expression. However, a systematic analysis of all the words present in the Arabidopsis genome is still lacking.
To analyze the different segments of the Arabidopsis genome, an enumerative word discovery approach was used to detect statistically overrepresented words. Similar approaches have been successfully applied over the last decade in the area of motif discovery [26–37]. In a 2005 study, Tompa et al.  showed that enumerative methods outperformed heuristic methods in many cases. They are particularly applicable in this research, because they allow the study of the entire 'word landscape' of a genomic data set.
Our approach scans the sequences and produces a set of words and word frequencies. This information is employed by a Markov model to compute expected word frequencies. Words with unexpectedly high frequencies are putative functional elements, and thus they are further characterized by comparing word frequencies and positions to gene induction or suppression using the method of Geisler et al. . Additionally, clusters of similar words are formed and used to create motifs for putative transcription factor binding sites. Sequences that contain the same functional elements are grouped together into putative 'nodes' of regulatory networks. Words that co-occur often are identified as putative transcription factor binding modules.