Translation initiation in eukaryotic mRNA molecules typically follows the basic mechanism postulated by the scanning hypothesis [1], according to which the 40S ribosomal subunit binds to the 5'-cap of an mRNA, scans in the 5' → 3' direction until the first AUG is encountered, stalls to recruit the 60S subunit, and forms the 80S ribosomal particle, which then proceeds unencumbered with translation to render a protein product (reviewed in [2]). Roughly 10% of eukaryotic transcripts are subject to so-called leaky scanning [3], in which the ribosome continues scanning beyond the first AUG codon until it encounters one in a more favorable context [4]. Alternative methods to initiate translation from certain RNAs of viral origin exist, including, one, the formation of kissing stem-loops to facilitate translation initiation from a 5'-proximal methionine codon [5] and, two, usage of internal ribosomal entry sites [6]. Efficient translation initiation from non-methionine codons is also possible in eukaryotes [7, 8]. In the present work, we are concerned only with modeling 5'-cap-dependent translation initiation occurring at AUG codons in eukaryotic protein coding genes of non-viral origin.

A variety of approaches to in silico translation initiation site (TIS) detection in nucleotide sequences have been previously considered, including perceptrons [9], single, multilayer artificial neural networks (ANNs) [10], multiple, multilayer ANNs [11], linear discriminant analysis [12], mixture Gaussian models [13], unsupervised clustering algorithms [14], support vector machines [15–17], expectation maximization [18], and hidden Markov models [19]. Unfortunately, none of these methods are conveniently available in the form of open source, distributed software. In part, our motivation for this work is to provide a software framework for the implementation and testing of a variety of different algorithmic approaches to TIS identification. Software systems such as ESTScan [20, 21] and Diogenes [22], originally developed for detecting significant open reading frames in (potentially errant) cDNA sequences, have also been used to identify TISs, although empirical results suggest that these methods are inappropriate for the task [23]. One strategy for integrating TIS detection methods into computational gene finding pipelines, as opposed to predicting TISs in mRNA sequences per se, is to refine results produced from a separate gene finding tool. For example, the TICO tool [14, 24] was developed to refine prokaryotic gene structure annotations generated by the GLIMMER program [25, 26]. The mechanism of translation initiation in prokaryotes differs considerably from that of eukaryotes [27]. Here, we describe the MetWAMer system, developed primarily for post-processing spliced alignment-based eukaryotic gene annotation results provided in the gthXML format [28]. A variant of the MetWAMer code is abstracted from any specific gene prediction system and allows TIS prediction in eukaryotic reading frames as generated by any procedure, thus facilitating integration into other gene prediction software and workflows.

In the following we first describe MetWAMer and its incorporated TIS-finding algorithms and then discuss applications to annotating transcripts from the model plant Arabidopsis thaliana. MetWAMer currently implements five distinct methods for TIS detection. Among these, the best performer is the perceptron-based flank-contrasting weighted log-likelihood ratio routine (PFCWLLKR), which combines local TIS feature scores and scores probing the contrast in coding potential of sequences flanking a site. MetWAMer allows the user to develop and apply stratified parameter sets for an arbitrary number of data clusters. We demonstrate the potential for stratified parameter deployment to yield considerable increases in TIS prediction accuracy relative to a homogeneous parameter strategy. Also discussed are strategies for parameter selection in practice, depending on prior assessment of the likelihood that the transcript under consideration is or is not 5'-complete. Source code implementing this package is released under the ISC license, and is available for download from [29]. It is also registered as Additional File 1 in this report.

In the following subsection, we briefly describe the components of the MetWAMer software. Then we discuss the distinct algorithms implemented for TIS-identification and report our training and testing approaches for Arabidopsis data.

The MetWAMer system

The MetWAMer code, written in the C programming language, implements the executable files MetWAMer.CDS and MetWAMer.gthXML. MetWAMer.CDS is the generic application for TIS prediction in eukaryotic open reading frames, as derived via any computational procedure. MetWAMer.gthXML is a special-purpose variant of the software, specifically tailored to refine gene structure predictions generated by the GenomeThreader [30] and GeneSeqer [31] programs for spliced alignment-based gene structure annotation. GenomeThreader and GeneSeqer, like most other spliced-alignment tools, do not make explicit predictions concerning translation initiation sites, but rather are restricted to the identification of reading frames in genomic sequences for which transcript evidence or homologous sequences suggest a protein coding function. MetWAMer.gthXML extends the 5'- and 3'-most termini of these annotated reading frames such that a maximal (non-stop) open reading frame (ORF) is realized. (No distinction between MetWAMer.gthXML and the more generic MetWAMer.CDS variant exists subsequent to reading frame maximization; we therefore refer to the system as "MetWAMer" for the remainder of this article.) MetWAMer scans for methionine-encoding sites in this maximal reading frame, considering their potential as translation initiation sites under a variety of scoring schemes, described below, in an attempt to identify a TIS for the gene structure under consideration. At most one prediction per maximal ORF is made, if and only if the optimal solution rendered exceeds some method-specific quality threshold.

Common to all detection methods implemented in MetWAMer is utilization of a start-methionine signal-specific weight array matrix (WAM) that records position-specific base transition frequencies proximal to methionine codons in protein coding sequences. Here, WAMs characterize position-specific dinucleotide abundances; see the Stratified training and testing section below for a more detailed description. The train_MetWAM utility from the MetWAMer package can be used to develop such a WAM, given appropriate training data. The first in-frame methionine codon encountered, subsequent to a specified offset in the training sequence, is considered to be the true TIS for that sequence. Training of the methionine weight array matrix proceeds by tabulating dinucleotide frequencies from five positions upstream of the adenine through three positions downstream of the guanine residue of the pertinent methionine codon. Next, the system advances 105 bases in the training instance, to resume scanning for in-frame methionine codons, each of which will be classified as a false TIS; dinucleotide frequencies proximal to these false TISs are tabulated in the same manner as true TISs (see Figure 1). Following tabulation of dinucleotide frequencies at true and false TISs in training data, these are converted to relative frequencies, yielding the WAM, which enables calculation of the likelihood that a site in question is a true or false TIS.

Methionine log-likelihood ratios

The log-likelihood ratio (LLKR) approach to TIS prediction functions by scanning the ORF for in-frame ATG codons. (We use ATG to denote a methionine codon, as opposed to AUG, because MetWAMer scans for potential TISs in conceptually spliced genomic sequences.) A constraint is imposed on the protein length implied by any potential start-methionine such that if the ATG served as a true translation initiation site, the resulting protein must exceed 50 amino acid residues. Using the trained methionine-WAM, the method scores each such feasible site by calculating the likelihood that it is a true initiation site and taking the ratio of this value relative to the likelihood that it is not a true start site. The system identifies the methionine codon yielding the optimal value among such likelihood ratios, and provided the log of this ratio is non-negative, the LLKR routine returns it as the predicted start-methionine. The non-negativity constraint implements a classification threshold, imposed because we require the likelihood of the potential start site to favor its actually being a true TIS. If the system fails to identify any in-frame ATG codons, or the best-scoring site's score is negative-valued, then LLKR returns no prediction for the maximal ORF being surveyed.

Weighted methionine log-likelihood ratios

The weighted log-likelihood ratio approach (WLLKR) is identical to LLKR, but each in-frame ATG's log-likelihood ratio score is scaled as a function of the induced protein product's coverage of the maximal ORF. Precisely, coverage x is defined as the ratio of the length of the implied amino acid chain starting from the TIS under consideration over the length of the maximal ORF. For a true TIS, we expect the coverage value to be close to unity, as it would be unusual for a long, uninterrupted reading frame to be evolutionarily maintained in a genome, yet not be encoding an expressed, functional protein product. Empirically, we settled on weights calculated as w(x) = x³ (other convex functions give commensurate results). The WLLKR routine optimizes over weighted log-likelihood ratios for all in-frame ATG codons, returning a predicted start-methionine if and only if the optimal such value is non-negative.

Multiplicative-based flank-contrasting with weighted methionine log-likelihood ratios

MetWAMer also implements an approach to start-methionine prediction that considers two descriptive features of potential TISs: weighted methionine log-likelihood ratio scores as used by the WLLKR routine (signal sensing) and the ratio of coding potential in a swath of sequence downstream from the site to that of a swath upstream of it, evaluated under a coding hypothesis (content sensing). Intuitively, we expect that the coding potential of the sequence downstream from a true site – which is, by definition, coding – would exceed that upstream of it – which is, by definition, non-coding – and that the ratio of the former to the latter should be greater in true sites as opposed to false. Coding probabilities of sequence swaths (96 nucleotides in length) are computed using a fifth-order χ²-interpolated Markov chain model [25, 26] as implemented in the IMMpractical library [32]. The idea of integrating both content- and signal-based features into TIS prediction has been explored before [11, 12, 33], although the methodologies used here are distinct from previous studies.

For the multiplicative-based flank-contrasting with weighted methionine log-likelihood ratios (MFCWLLKR) method, the signal- and content-based scores, expressed in log space, are added. The system optimizes over these scores at viable, in-frame start-methionine sites, and if the best-scoring site's score is non-negative, it is returned by the routine as its TIS prediction.

Perceptron-based flank-contrasting with weighted methionine log-likelihood ratios

The perceptron-based flank-contrasting with weighted methionine log-likelihood ratios (PFCWLLKR) routine considers the same descriptive features as MFCWLLKR, but uses a perceptron as a multivariate utility function, as opposed to the multiplication operator. Perceptrons implement linear discriminants, and as such require linearly (or near-linearly) separable data sets to provide good classification performance (see, e.g., §4.1.7 of [34]). Intuitively, we expect that the two dimensions corresponding to the signal- and content-based features exhibit linear (or near-linear) separability: both weighted log-likelihood ratios of methionine sites and log-likelihood ratios of the coding potentials of downstream to upstream content swaths should be greater-valued in true start methionines as opposed to false, non-start ones. Linear and sigmoid units are used to implement perceptrons in the MetWAMer system; each of these neural elements can learn a continuous-valued function that can be thresholded to enable discrete, binary classification; excellent discussions of these methods can be found in §4.4.3 of [35] and §20.5 of [36]. Thus, linear and sigmoid units can be used to optimize over viable candidate start-methionine codons.

PFCWLLKR returns the best such potential TIS if and only if it is classified as being a true site by the perceptron. Although Stormo et al. used a perceptron to classify translation initiation sites in bacteria in a pioneering study [9], they considered an entirely distinct feature set.

Bayesian TIS prediction

Lastly, we also considered a Bayesian approach (BAYES) to predicting TIS sites. Each viable start-methionine in the maximal reading frame is considered under two separate models, one that the ATG is a true translation start codon and the other that it is not. The maximum a posteriori (MAP) hypothesis among this set of possibilities is computed, and if the site it denotes is represented as being a true TIS, BAYES returns this result as its TIS prediction. Otherwise, the method refrains from making any predictions. Calculation of the MAP hypothesis is formulated as follows. A prior distribution is derived for each maximal reading frame being surveyed: each in-frame ATG, under the model of its being a true initiation site, is given a prior probability proportional to the relative length of the peptide it induces compared with that of the maximal reading frame. Similarly, under the model of not being a TIS, each such site is assigned a prior probability proportional to the complement of its prior probability of being a true one. These values are normalized so as to collectively represent a valid probability mass function over all putative start-methionine sites, under both models. The likelihood of data is modeled using log-likelihood scores computed with the methionine-WAM.

Test design

A five-fold cross-validation strategy was used to assess the methods on the task of translation initiation site detection competency. Because the TIS-containing instances consist only of known coding sequences from gene structures, and in practice MetWAMer scans for potential TISs across a maximal ORF, we extended the coding sequences at their 5'-termini to achieve a maximal, non-stop reading frame, thereby presenting the system with the challenge of disambiguating spurious (in-frame) methionine codons in the extended reading frame from true start codons – we make no considerations as to whether these extended sequences are derived from 5'-UTRs, introns in 5'-UTRs, or intergenic sequences. Methionine-WAMs and Markov chains were trained on each cross-validation replicate and then used to train a sigmoidal perceptron, using a learning rate of 1 × 10^-5. Sigmoid units outperformed linear units in all experiments we conducted (data not show), so we do not consider the latter further. As a baseline for comparison of the implemented models, we consider also the 1 ^st -ATG method, which predicts the first in-frame ATG it encounters in the maximized reading frame as a TIS. Tests using non-TIS-containing instances were conducted similarly, though reading frames were not maximally extended at their 5'-terminus. The testing procedure is shown pictorially in Figure 2.

To assess the performance of MetWAMer relative to prior art in translation initiation prediction, we compared our system with the NetStart [10], TIS Miner [43], TISHunter [17] and ATGpr [12] programs.

Our results on the TIS-containing data set suggest that, compared with the methods implemented in MetWAMer, a policy of labeling the first ATG as TIS in a maximal ORF wil achieve quite good (though imperfect) results. However, in practice we cannot always assert whether a maximal ORF has sufficient 5'-coverage so as to include the gene's true TIS, or whether a spurious in-frame ATG occurs upstream from it. In such cases, the 1 ^st -ATG strategy fails, as it does in cases of leaky scanning, thus sustaining the importance of further development of statistical TIS prediction methodologies that capture the sequence features recognized by the ribosome in translation initiation. In this work, we present a number of distinct models for TIS prediction, the most successful of which mixes content- and signal-based features of putative TISs using a perceptron (PFCWLLKR). Furthermore, we demonstrate that, in the model plant Arabidopsis, TIS prediction can be enhanced by integration of class-specific parameter sets, regardless of the prediction method utilized.

We attribute the well-balanced performance of PFCWLLKR to the biological plausibility of the features provided to it as inputs. As a signal-based feature, weighted log-likelihood ratios considerably improve the specificity of TIS prediction (e.g., contrast WLLKR and LLKR in Tables 1 and 2, likely because our weighting function, w(x) = x³ for induced protein length to maximal ORF coverage x, appears to empirically approximate the epistemology of eukaryotic translation initiation fairly well: according to the (leaky) ribosomal scanning hypothesis [3], one would expect that more upstream AUG sites – especially those occurring in a favorable signaling context – in a maximal reading frame would be more likely to function as bona fide translation initiation sites. Also, it is unusual for a long, uninterrupted reading frame to be maintained, yet not expressed as part of a functional protein product. Our weighting scheme has been explicitly designed to reflect these biologically-informed biases.

During the post-scanning phase of translation initiation, the small ribosomal subunit stalls at a TIS to recruit the large subunit, thereby forming the 80S ribosomal particle. The scanning process, as conducted by the small ribosomal subunit in concert with various eukaryotic initiation factors, does not appear to take more global nucleotide compositional features of the mRNA molecule into account, notwithstanding the possibility of secondary structures causing steric interference with scanning itself. That we might utilize contrast in coding potential of sequences flanking a TIS for modeling purposes is a consequence of the fact that sequences upstream of a TIS are non-coding, and those downstream, coding, though this plays no known role in the recognition of TISs in vivo. The use of Markov chains in a classification setting was shown to distinguish exons from introns with good accuracy in plant systems [32], and our expectation that these content-sensing tools could be gainfully transferred to the TIS prediction domain was born out by the performance results shown. Similar inclusion of coding potential contrast has also been employed to increase splice site prediction accuracy [31, 45].

Our data set was developed from gene models flagged as curated in the current Arabidopsis annotations, though it should not be overlooked that potential errors in these structures might have distorted our results. Manual inspection of several genes whose TISs were predicted incorrectly by the PFCWLLKR routine indicate possible problems with existing annotations. For example, in gene model At4g34080.1 http://www.plantgdb.org/AtGDB-cgi/getRegion.pl?dbid=2&chr=4&l_pos=16326388&r_pos=16328548, our system predicted the TIS as that from the TAIR version 6 gene annotation, rather than that of version 7, which occurs downstream. Similarly, we predict the version 6 TIS of gene model At5g35580.1 http://www.plantgdb.org/AtGDB-cgi/getRegion.pl?dbid=2&chr=5&l_pos=13778674&r_pos=13781581 as correct, rather than the revised TIS from the version 7 model. Partial protein sequencing using Edman degradation could potentially resolve such ambiguities in the annotations (e.g., [7]), as might consideration of homologous proteins with matching N-termini whose translation initiations sites had previously been determined; such efforts are beyond the scope of this work, however.

Although we were unable to achieve the performance levels of a priori-known cluster-specific parameter deployment with our parameter set indexing schemes, stratified parameter deployment can nevertheless be used effectively in practice, pending certain characteristics of the test data: if these are expected to be moderately enriched for 5'-complete sequences, then static WAM-based indexing should recover a larger fraction of true TISs than would homogeneous deployment. However, if complete 5'-coverage is expected to be quite sparse, homogeneous parameter deployment should be utilized instead. This affords a complete prescription of how to most effectively identify TISs in transcript data: 1 ^st -ATG would be the best method for use in data sets with a high degree of 5'-completeness, static WAM-based PFCWLLKR in moderately enriched data sets, and homogeneous deployment of PFCWLLKR in data sets likely to contain few 5'-complete sequences.

We have replicated our experiments using a data set based on the most recent GenBank annotations for the nematode Caenorhabditis elegans (dated 16 February 2006), the results of which are similar to those presented here for Arabidopsis (available as supplementary material at [29]), suggesting that our method is not specific to plant taxa, and can be used for eukaryotic TIS prediction in general. Also available as supplementary material are homogeneous parameter deployment-based results for a small set of TIS-containing human genes culled from the Consensus CDS project [46]; these results imply that the system can be utilized for vertebrate taxa, as well.

As a demonstration of MetWAMer's applicability for post-processing gene structures predicted by separate tools, we refined maize gene annotations generated by the GeneSeqer spliced alignment program [47]. 11,742 full length maize cDNA sequences were obtained from the Maize Full Length cDNA project [48] and aligned via GeneSeqer to a set of 17,163 BAC sequences downloaded from PlantGDB [49]. These results were post-processed with MetWAMer's PFCWLLKR routine under homogeneous parameter deployment, using parameters trained with Arabidopsis data. We considered only predicted protein sequences such that at least one full length cDNA supporting its annotation exhibited an overall GeneSeqer alignment score of at least 0.9 and the predicted TIS occurred in or upstream from the first exon identified by spliced alignment. The resulting set of 6,926 proteins was aligned against a collection of 36,338 annotated sorghum proteins downloaded from the Phytozome project [50] using BLASTP. BLASTP output was inspected using the MuSeqBox program [51] in order to select only those inferred maize proteins of at least 150 amino acids in length whose best hit in the sorghum data, also at least 150 amino acids long, shared high-scoring segment pairs (HSPs) of at least 20% identity apiece such that the sum of these non-overlapping HSPs was not less than 90% of the length of either sequence. Furthermore, at most five amino acids at both the N-and C-termini, for both sequences, were allowed to be disjoint from an HSP. These 2,315 proteins were then made non-redundant using BLASTClust with default settings. In summary, the resultant set of 1,665 maize proteins on 1,463 distinct BACs identifed by GeneSeqer in concert with MetWAMer represents a reliable collection of high-quality, non-redundant full length maize proteins that could not have been identified by GeneSeqer alone, thereby demonstrating the practical utility of this approach to modern genome annotation projects. Our results are available as supplementary data at [29].

We compared annotation results of our pipeline with those achieved by a current state-of-the-art ab initio gene prediction tool, AUGUSTUS [52]. The BAC sequences containing our annotated maize genes were fed to the program and processed using its maize-specific parameters. We note that a fair comparison between the two approaches is basically impossible, since the search space probed by pure ab initio gene finders is quite distinct from that explored by spliced alignment annotation systems such as GeneSeqer+MetWAMer, so we disregard false positive predictions generated by AUGUSTUS. In summary, of the 1,665 maize proteins we identified, AUGUSTUS correctly predicted 1,232 (≈74%) TISs and 581 (≈35%) complete gene structures. These results underscore the necessity that a complete and robust gene annotation pipeline should integrate evidence from multiple data sources, gene prediction software and even manual gene curation results, as is achieved by various higher-order systems including AUGUSTUS+ [53], the Ensembl pipeline [54], EuGéne [55], and JigSaw [56, 57]. Our efforts to integrate a variety of retrained, state-of-the-art gene finding tools using such systems in the context of various plant genomes will be presented in a forthcoming report.

MetWAMer performance results, particularly for PFCWLLKR, suggest that the method can be used with good success for the task of annotating TISs in eukaryotes. However, our data are not precisely comparable with those provided by a number of previous studies, e.g., [10–13, 15, 19, 33], just as results between those papers are essentially incomparable, as well. This is due to differing experimental designs (some studies focus on the number of ATG codons correctly classified as true or false TISs, and others on the number of genes for which the TIS was correctly identified) and different data sets (some studies used human genes, some cyanobacterial, etc., and these corpora were often of very different sizes).

Comparing these published methods with our own, using our data and experimental design, was often not practical: the availability of software implementing methods developed for eukaryotic TIS prediction per se is very limited at present. Among the papers addressing intrinsic TIS detection methods, only the ATGpr [12], StartScan [33], DIANA-TIS [11], TISHunter [17], NetStart [10], and TIS Miner [43] systems are described as "available" software. We were only able to utilize the NetStart, TIS Miner, TISHunter and ATGpr systems to compare against our software system, though we note that it is impossible to re-train any of these programs. StartScan is available via a web interface (currently trained only for human), but an important distinction from our tool is that StartScan is for TIS recognition in genomic sequences, a much different task than that addressed by MetWAMer. Although not mentioned in its reference paper, [11], we were able to locate a web interface to the DIANA-TIS system at the author's web page http://diana.pcbi.upenn.edu. However, documentation for the interface is unavailable, and most prohibitive is that it only allows a pictorial representation of its predictions, which is unrealistic for processing data sets of the scale used in this study. GeneHackerTL is mentioned in [19], but it is not described as being publicly available, nor were we able to locate it in any web-accessible forum.

The paucity of freely available, functioning programs for TIS prediction comprises an important gap in the software infrastructure for computational biology. Our MetWAMer package represents a well-documented, extensible, and open source software system that can be modified for differing applications and extended with existing and novel TIS prediction methods to support further research in this area; this is, to our knowledge, the first such contribution made to the eukaryotic TIS prediction community at-large. There are certain limitations to the existing scope of MetWAMer, however, which may present opportunities for future work. We have explicitly ignored the possibility of non-AUG start codons, although these are known to occur in various eukaryotic organisms [7, 8]. Also, the system does not explicitly integrate extrinsic information, such as homologous proteins, which is reportedly successful [58]; however, due to evolutionary forces operating on homologous genes, it is possible that translation initiation sites differ, and the use of such information for prediction could be misleading. We have explicitly ignored the possibility of translation initiation proceeding by a re-initiation mechanism, whereby a short ORF upstream of the more significant ORF is translated, and the ribosome resumes translation at a downstream AUG [59]. For MetWAMer, however, this is not a potentially obfuscating phenomena: because the system scans for TISs in a maximal reading frame, there is no possibility to predict a start codon upstream of the significant ORF that is succeeded by a stop codon a short distance thereafter. Another open problem is the prediction of alternative TISs in various gene structures [60].

The ability to train TIS models in a species-specific manner is an important strength of MetWAMer, because differences in translation initiation processes among distinct taxa are known to occur [61]. To the extent that cross-specific TISs are representative of some target species, these could in principle be used as a proxy if species-specific data are not available; the performance of our system in such a scenario will be reported in a forthcoming study in which we refine gene structure annotations of a variety of cereal crop genomes. Results presented here also indicate that improvements in TIS prediction accuracy are possible when taking the class of potential start-methionines into account. Our software readily accommodates these needs, and can be integrated into other gene annotation programs and/or pipelines with straightforward modifications.