Trypanosomatid protozoan parasites have diverged early on in eukaryotic evolution [3]. Their evolutionary closeness to bacterial ancestors is delineated by intrinsic cellular characteristics such as tandem arranged genes [4], polycistronic transcription [5,6], mitochondrial RNA editing [7], lack of transcriptional control [8], infrequent introns [9], and trans-splicing [10]. The latter consists of the 5' cleavage of polycistronic RNA precursors into individual mRNA transcripts by addition of an exogenous 39 to 41 base long capped RNA fragment, namely the splice leader (SL) or mini-exon, provided by a highly abundant SL-RNA [11], yet similar processes have also been discovered in nematodes and even in mammals [12,13]. This process is somewhat similar to cis-splicing in other organisms, as RNA is cleaved at an AG dinucleotide downstream of a polypyrimidine stretch.

In addition to co-transcriptional trans-splicing, polyadenylation of the upstream transcript is also required in order to generate monocistronic mRNAs in these organisms. Trypanosomatid protozoa are believed to lack a conserved polyadenylation (poly(A)) signal, in contrast to other eukaryotes who generally require a cytoplasmic polyadenylation motif for successful polyadenylation [14]. Several studies support that polyadenylation is mechanistically coupled to trans-splicing and that it depends upon the presence of polypyrimidine tracts [15-19], thus leading to the belief that the spliceosome complex interacts with the polyadenylation machinery in trypanosomatids. It has also been reported that distant pyrimidine tracts may be responsible for polyadenylated positions further away from the downstream 5' splice site in trypanosomes [17,20]. These analyses also convey the non-specific nature of poly(A) site selection in trypanosomatids, as polyadenylation seems to occur in a given region rather than at a specific position.

The apparent heterogeneity of kinetoplastid mRNA polyadenylation and its dependence on successful trans-splicing make 3'-untranslated region (3'UTR) length predictions troublesome. Currently, there exists a 3'UTR prediction method for Trypanosoma brucei derived from the statistical analysis of mRNA transcript extremity lengths from expressed sequence tag (EST) data [20]. The prediction is essentially obtained by selecting the position located at an empirical distance (100 bases) upstream of the polypyrimidine tract closest to the open reading frame (ORF). The authors claim a 38% prediction rate within a 73-nucleotide window. These metrics are somewhat inappropriate for predictions in the Leishmania genus since the species flaunt larger intergenic (IR) sequences, higher average UTR lengths, and less stringent splice acceptors [4,21].

In addition to the statistical analysis of transcript length distributions for 3'UTRs, 5'UTR prediction has been submitted to supplementary investigation [22-24]. Prediction algorithms that essentially focus on selecting the first AG dinucleotide after the longest polypyrimidine stretch can reportedly identify 62% of valid splice junctions in trypanosomes and 51% in Leishmania [20,23]. For Leishmania, it has been shown that by fragmenting the non-coding sequence upstream of a start codon at every occurrence of AG, the AG following the longest fragment corresponds to a valid splice junction in 60% of the cases. When combining this method with a linear discriminant analysis of dinucleotide composition, the later method can obtain a prediction accuracy as high as 92% on selected high-scoring sequences [23].

Considering that regulation of gene expression in kinetoplastids occurs mostly at the post-transcriptional level, it has become apparent that UTRs bear essential regulatory tags [8,25-29]. From the standpoint of computational motif discovery, it is imperative to discriminate between functional and non-functional sequences in order to successfully identify novel conserved regulatory regions. This premise is most important when dealing with non-coding RNA as it is exposed to less stringent evolutionary pressure than open reading frames [30]. It can be expected that limiting sequence and structure motif searches to legitimate mRNA UTRs will generate more informative results while reducing the inherent computational cost of search algorithms.

When compared to previously published trans-splicing prediction rates [23], the models we propose here appear to be just as effective at predicting known trans-splicing sites when tested on the same search space (Table ​(Table1).1). Their accuracy remains significant even when increasing the query sequence size (1.75× increase in search space at the cost of 0.9× accuracy). The augmented search space is in order to ensure that the full inter-AG fragments upstream of putative splice sites are considered. Overlapping into the downstream coding sequence is vindicated by erroneous genome annotations; it is not uncommon that the furthest in-frame ATG is selected as a start codon. Also, our scoring function rates all inter-AG fragments, unlike the previously proposed study that selects high-scoring fragments based upon their dinucleotide composition [23]. As shown in Table ​Table3,3, a scoring threshold can be implemented to ensure that few false-positives are unsuitably identified as splice-junctions at the cost of slightly lower specificity. However, a threshold will necessarily neglect certain sequences, which may be objectionable when dealing with few or essential queries. Since our method is more dependent on correct annotations, it is conceivable that coupling it to linear discriminant analysis would generate even better predictions at the cost of higher complexity.

Predicting poly(A) sites with PSSM's have previously been shown to successfully predict poly(A) sites in humans [38]. Capturing the global nucleotide composition surrounding known poly(A) sites and utilizing it as a comparative predictor has also proven to be a successful prediction procedure in Leishmania. Albeit the public EST data appears to be of questionable quality, stringent screening has permitted to reveal specific polyadenylated sequence traits. Given the nature of the sequence data, smaller mRNA transcripts may be favoured and this should be considered when analyzing results. Nonetheless, PSSM scanning is more than 10 times more effective at identifying poly(A) sites than the distance-only approach when precision is fundamental (Figures ​(Figures2A2A and ​and4).4). This result can be interpreted as evidence that distance is not as powerful for targeting poly(A) sites in Leishmania than in trypanosomes.

For Leishmania, precision may not be essential when predicting 3'UTR extremities given that several mappings display heterogeneous poly(A) positions [15]. This observation motivates the use of an error margin, which is interpreted as lowering the resolution of sensitivity testing in this work. Allowing correct predictions to be within a certain range of the mapped position emulates the identification of a polyadenylation region. We also tested a window scanning approach, where the cumulative bit-scores for a given range were averaged over the size of the window instead of considering each position independently. Such an approach yielded weaker overall predictions than the position-specific approach (data not shown), perhaps because the extent of polyadenylation regions varies among different transcripts.

The best 3'UTR predictions emanate from the grouping of distance limitation and scanning with dual PSSMs. Combining both metrics proved to be more effective than either one individually (Figures ​(Figures2,2, ​,4,4, and ​and5),5), a result that hints at the importance of each factor when predicting poly-A sites in Leishmania. For restraining PSSM scanning, we tested various distances instead of using a specific confidence interval since spacer sequences display somewhat of a bias towards longer fragments. Although the data is partially derived from estimations, such a shift in the distribution supports the notion that polyadenylation does not occur randomly in Leishmania. Poly(A) sites further away from the splice junction may be an effect of distant polypyrimidine tracts, a situation that has already been observed in trypanosomes [20]. One must also consider that the longer non-coding regions in Leishmania may contain non-annotated genes or provide alternative stage-specific polyadenylation sites, which could explain the longer spacer sequences. These are considerations that motivated the exclusion of intergenic sequences longer than 5000 nucleotides for sensitivity testing.

To our knowledge, no other method can predict poly(A) sites as effectively in Leishmania spp. as the one described in this work. Even enforcing a highly-selective threshold only faintly affects this method's specificity (Table ​(Table3).3). The rather unusual and non-specific nature of kinetoplastid polyadenylation is a line of reasoning to substantiate low computational prediction rates. Although over-represented A-rich hexamer motifs are found (Additional File 2), these are not however present in all the genomic poly(A) sites, which suggests that they may not play a central role in driving polyadenylation in Leishmania. In addition, the genomic alignment of polyadenylated EST mappings cannot be used to mark out a precise consensus sequence, as it is impossible to distinguish the exact cleavage site among multiple consecutive adenosines on the unprocessed transcript. The heterogeneity of poly(A) sites in Leishmania mRNA transcripts is extra incentive for using PSSMs that embody a global trend in nucleotide composition. Furthermore, neglecting secondary structure and stage-specificity are additional factors that make it difficult to conceive obtaining higher prediction accuracies at this point.

Notwithstanding the possibility that no consensus motif drives polyadenylation in kinetoplastids, there is evidence for a biological model based on sequence context. The low sensitivity obtained from a poly(A) prediction algorithm based on spacing metrics alone is an evidence for a more dynamic biological model. Also, the correlation between certain regions of the genomic alignment and their respective prediction rates is most interesting, as best illustrated by the sensitivity surface plots (Figure ​(Figure2).2). The data is presented in order to asses the innate characteristics that have an impact on poly(A) targeting.

Two main common sequence features appear to directly influence the prediction sensitivities. Firstly, the adenosine-rich region within close range to the mapped poly(A) site. Secondly, the pyrimidine-rich region 300 to 600 positions downstream. The latter, which represents the polypyrimidine tracts known to be crucial for trans-splicing, generates the best predictions when loosening the accuracy and bounding the search space. In turn, the A-rich region is responsible for the best predictions when precision is fundamental. Interestingly, the affluence of polypyrimidines (most notably thymines) in the -50 to -25 region (Figure ​(Figure1)1) may play a role in 3'UTR cleavage since its exclusion from scanning matrices reduces the sensitivity at close range (Figure ​(Figure2).2). The matrix encoding the sequence information of zero upstream bases and 25 downstream (0A25) is somewhat futile at predicting poly(A) sites, a rather surprising observation seeing as the adenosine concentration is comparable. Upon closer inspection, it is apparent that adenosine-rich regions are not a fundamental marker because many sequences do not contain profuse adenosine residues at their poly(A) site.

PSSMs can be regarded as a simplistic representation of the interaction between an enzymatic complex and a strand of nucleic acids. The highest scoring position corresponds to a region that is most similar to the consensus of all poly(A) sites, which relates to a high affinity region for the polyadenylation complex. In this perspective and based on our results, it is enticing to contemplate a generic biological model where adenosine richness (possibly contrasted by a pyrimidine-rich upstream region) helps to direct the polyadenylation of specific positions downstream of polypyrimidine tracts in unprocessed mRNA transcripts. Deletion studies directed at these features followed by mapping the modified transcript's poly(A) site could shed additional light into the biological process. Moreover, in-vitro UV cross-linking could help identifying novel ribonucleoproteins (RNPs) that might interact with the trans-splicing/polyadenylation complexes.