Overlapping genes are common in RNA viruses where they serve as a mechanism to optimize the coding potential of compact genomes. However, annotation of overlapping genes can be difficult using conventional gene-finding software [7]. Recently we have been using a number of complementary approaches to systematically identify new overlapping genes in virus genomes [7-11]. When we applied these methods to the toroviruses, we found strong evidence for a new coding sequence - overlapping the 5'-terminal region of ORF1a (Figure (Figure1).1). Here we describe the bioinformatic analyses.

Relatively little sequence data is available for the relevant 5'-terminal region of the torovirus genome. In fact there are only two non-identical sequences in GenBank (tblastn [12] of translated NC_007447 ORF1a; 2 Aug 2009) for the region of interest: [GenBank:NC_007447] - Breda virus or Bovine torovirus (derived from [GenBank:AY427798]) [5], and [GenBank:DQ310701] - Berne virus or Equine torovirus [4]. However these two viruses are reasonably divergent (mean nucleotide identity within ORF1a ~68%), thus providing robust statistics for comparative methods of gene prediction. The NC_007447 and DQ310701 ORF1a amino acid sequences were aligned with CLUSTALW [13] and back-translated to produce a nucleotide sequence alignment, which was analyzed with a number of techniques.

The first piece of evidence for an overlapping coding sequence is the presence of an unusually long open reading frame (229 codons; hereafter ORFX) at the 5' end of ORF1a but in the +2 reading frame relative to ORF1a (Figure (Figure1B,1B, panels 2-4). In fact ORF1a in Breda virus has 589 stop codons in the +2 frame (out of a total of 4444 codons), while Berne virus has 569 stop codons (out of 4568). In other words, approximately one in every eight codons in the +2 reading frame is a stop codon (see, for example, the last three alignment blocks in Figure Figure2).2). Thus the probability of obtaining an uninterupted 229-codon +2 frame ORF simply by chance is vanishingly small (if +2 frame stop codons within ORF1a are assumed to be randomly distributed, then the probability is of order p < 10^-10). Moreover, there are 141 point nucleotide differences between Breda virus and Berne virus within ORFX, and yet the open reading frame is preserved in both viruses. The absence of stop codons may be linked to local nucleotide biases - indeed the mean nucleotide frequencies within ORFX (Breda virus) are A 28%, C 24%, G 20% and U 27% compared with A 27%, C 14%, G 23% and U 36% in the rest of ORF1a, so that the ORFX region is relatively C-rich and U-poor. However the simplest explanation for these nucleotide biases is simply the presence of an overlapping gene (i.e. ORFX) and the constraints imposed by having to code in multiple reading frames.

Next, the ORF1a alignment was analysed for conservation at synonymous sites, as described in [11] (but inspired by ref. [14]). The procedure takes into account whether synonymous site codons are 1-, 2-, 3-, 4- or 6-fold degenerate and the differing probabilities of transitions and transversions. There was a striking, and highly statistically significant (p < 10^-17 for the total conservation within ORFX), peak in ORF1a-frame synonymous site conservation at the 5' end of the alignment, corresponding precisely to the conserved open reading frame, ORFX (Figure (Figure1B,1B, panels 5-6). Peaks in synonymous sites conservation are generally indicative of functionally important overlapping elements, though such elements may be either coding or non-coding. In fact, high synonymous site conservation at the 5' end of long polyprotein-encoding sequences is a feature common to a number of RNA viruses and can not, in itself, be taken as evidence of an overlapping coding sequence. However the extent (229 codons) and degree (Figure (Figure1B,1B, panel 6) of the conservation here is unusual and, furthermore, the high conservation is not matched in the related coronaviruses. Thus an overlapping gene, viz. ORFX, provides the most obvious explanation for the high conservation seen here. (An alternative explanation is recombination, as in ref. [15]. However recombination does not provide an explanation for the other evidence presented in this report.)

Finally, we analysed the alignment with MLOGD - a gene-finding program which was designed specifically for identifying overlapping coding sequences, and which includes explicit models for sequence evolution in multiply-coding regions [7,8] (Figure (Figure1B,1B, panels 7-9). In contrast to the synonymous site conservation index above, MLOGD, when applied in the sliding window mode, does not depend on the degree of conservation per se (the sequence divergence parameter is fitted independently for each window). With just two input sequences, the MLOGD signal proved to be somewhat noisy (e.g. there are a number of positively scoring windows that clearly do not correspond to potential overlapping genes in, for example, the +2 frame; Figure Figure1B,1B, panel 9). However the signal for ORFX was clear - with consecutive positively scoring windows throughout the ORFX region in the +2 frame - indicating, again, that ORFX is indeed a coding sequence. Moreover, the MLOGD score in the +2/ORFX frame within the ORFX region was significantly greater than the score in the +0/ORF1a frame, indicating that the ORFX product is subject to stronger functional constraints than the product of the overlapping region of ORF1a (which indeed has a negative MLOGD score towards the 5'-terminal half of the ORFX region). Consistently, further inspection showed that, in the region where ORFX and ORF1a overlap, ORFX has higher amino acid conservation than ORF1a (182/229 identities for ORFX, 153/229 identities for ORF1a).

In Breda virus (NC_007447), the annotated ORF1a AUG initiation codon is at nucleotide coordinates 859..861 and the first ORFX-frame AUG codon is at coordinates 1110..1112. However leaky scanning to this AUG codon is unlikely, due to intervening AUG codons in the ORF1a frame (1 in NC_007447, 3 in DQ310701; Figure Figure2).2). Instead we propose that ORFX initiation takes place at a CUG codon located upstream of the ORF1a AUG codon, at coordinates 774..776 (Figure (Figure2).2). CUG is, apparently, the most commonly used non-AUG initiation codon in mammalian systems (reviewed in [16]), and this particular CUG codon is conserved, and has a strong Kozak context ('A' at -3, 'G' at +4; [17]), in both Breda and Berne viruses. The downstream sequence is predicted to fold into a hairpin structure that is identical between Breda and Berne viruses - despite a number of base variations - and that is separated from the CUG codon by 13 nt (Figure (Figure2).2). Such structures - particularly at this spacing - have been shown to greatly enhance initiation at non-AUG codons [18]. Moreover, inspection of the sequence alignment upstream of the ORF1a initiation site shows that the majority (14/18) of base variations occur in the 3rd nucleotide positions of ORFX-frame codons, indicative of an ORFX-frame coding sequence (Figure (Figure2).2). This pattern of base variation continues right up to the proposed CUG initiation codon. Initiation at a site further upstream is precluded by ORFX-frame termination codons and, consistently, the sequence further upstream does not maintain the reading frame and base variations no longer favour the 3rd position (Figure (Figure22).

Initiation at the upstream CUG codon would give ORFX the nucleotide coordinates 774..1547 in NC_007447 and 776..1549 in DQ310701, resulting in a 258 amino acid product with a molecular mass of 30 kDa which, for want of a better designation, we tentatively name '30K'. The full predicted amino acid sequences are shown in Figure Figure3.3. Note that the product has only one methionine residue, making detection with [³⁵S]Met difficult. Application of blastp [12] to the amino acid sequences revealed no similar sequences in GenBank (3 Aug 2009) - as expected for a gene created de novo via out-of-frame 'overprinting' of a preexisting gene [19,20]. Similarly, application of InterProScan [21] also returned no hits (protein motifs, domains etc).

It is expected that a large proportion of ribosomes should scan past the CUG codon and initiate at the ORF1a AUG codon - thus allowing synthesis of the replicase polyprotein - though the additional possibility that the CUG-initiation efficiency may be temporally regulated as part of the virus lifecycle can not currently be discounted [16,22].