In order to study the functional integration of TE sequences into host genes, we identified TE insertions present in mature transcripts of the D. melanogaster euchromatic Release 3 genome sequence. We call such transcripts "chimeric" as each of them has one component from a host gene and one from a TE insertion. In addition to using the standard methods in the field for directly finding genes and TEs that share overlapping coordinates or querying annotated transcripts directly for TE sequences 81011, we also sought evidence for chimeric transcripts using a novel three-step process based on expressed sequence tags (ESTs) (see Materials and Methods). This indirect method of identifying gene-TE chimeras was necessary to avoid annotation biases resulting from the fact that "coding exons were not annotated in sequences with homology to transposable elements" 14 in the D. melanogaster genome.

In total, we found 63 protein-coding genes that produce chimeric transcripts supported by EST evidence (Table 1; for more information [see Additional file 1]). These chimeric transcripts involve 63 different TE insertions, but the relationship is not simply one-to-one: in two cases, TE insertions (FBti0019107 and FBti0020178, Table 1) occur in overlapping 3'UTRs of convergently transcribed neighboring genes producing two separate chimeric transcripts each (see Figure 1A); and in one case, three TE insertions are found in a chimeric transcript for a single gene (CG32021) on the 4^th chromosome. In addition, we found one noncoding transcript, the αγ-element 15, which is generated by two TE insertions within a larger nest of TEs situated between the Hsp70 Ba and Bbb genes. Our screen appears to have high sensitivity as evidenced by the fact that we identified four of the five exonic TE insertions previously reported in 12 (we found no supporting EST evidence for the fifth gene CG7900); the single exonic jockey insertion in the gene CG6191 reported in 16; and the chimeric transcript generated by a Doc insertion into the gene CHKov1 (CG10618) reported in 1718. We did not identify the Bari-1 insertion in cyp12a4 recently reported in 19, which is supported by EST evidence, since the region of overlap (18 bp) does not pass our length threshold.

We note that six of the 65 chimeric TE insertions identified by BLAST-based methods do not have corresponding TEs in the Release 3.2 annotation. However, unannotated TEs of the correct family can be found in the genome sequence for these chimeric TE insertions (Table 1). This result indicates that an unknown proportion of real TE insertions has not been annotated in the Release 3 genome sequence (see below). To be able to analyze aspects of chimeric TEs in the context of the genome annotation, we excluded these six TE insertions from the "annotated set" of 59 TE insertions, although we do consider them to be bona fide members of the "total set" of 65 potential gene-TE chimeras in the D. melanogaster genome.

Most of the 63 genes generating the total set of chimeric transcripts are of unknown function, but we did identify chimeric transcripts in 23 characterized protein-coding genes including brown (bw), a gene that appears to be a hot-spot for natural TE insertions 20 and is known to carry a viable mutation (bw¹) in the sequenced strain 14. Our in silico screen also identified a chimeric TE insertion generated by the serine protease encoding gene Tequila that has recently been shown to impair the transcription of this gene, but with no apparent phenotypic consequences 21. A general analysis of the molecular function and cellular localization of the total set of genes with chimeric transcripts, however, did not indicate a significant enrichment of any particular Gene Ontology (GO) category (data not shown).

Relative to other non-chimeric TEs inserted in transcribed regions (i.e. intronic TE insertions), the annotated set of TEs present in chimeric transcripts is significantly enriched for LTR insertions (Figure 2A). This observation largely accounts for the fact that the annotated set is also enriched in long TEs (Figure 2B), since LTR insertions tend to be longer than other classes of TE insertion in the genome 14. Furthermore, chimeric TEs have a greater tendency to be present in high-recombination areas of the genome than non-chimeric, intronic TE insertions (Figure 2C). However, the overabundance of chimeric TEs in regions of high recombination is not caused simply by the fact that chimeric transcripts are preferentially formed by LTR insertions, since high-recombination TE insertions are over-represented among the chimeric non-LTR (i.e. LINE-like, TIR and FB) elements even more strongly than among the chimeric LTRs (data not shown).

TE sequences are found in UTRs in most of the chimeric transcripts they generate: 38 of the 63 TE insertions are found in 3'UTRs, 23 in the 5'UTRs and 4 in coding exons. We note that these numbers total more than 63 because two TE insertions (chimeras 47 and 61 [see Additional File 1]) fall into multiple categories. The higher incidence of TEs in UTRs and specifically in 3'UTRs parallels findings in the human and mouse genomes 1011. The increased prevalence of TE insertions in 3'UTRs may be attributed to the increased average length of 3'UTRs (442 bp) relative to 5'UTRs (265 bp) in Drosophila 22 (as has been suggested previously to explain such patterns in the human genome 10), or to the lower density of functional signals in 3' regions relative to 5' regions of genes. This pattern does not appear to result from biases in the EST libraries, since over 10 times more 5' ESTs were analyzed than 3' ESTs 23.

Surprisingly, the genes involved in chimeric transcripts are not always those nearest to the sites of the corresponding TE insertions. Four chimeric transcripts skip one or more genes between the gene and TE components of the transcript (chimeras 12, 18, 23 and 50; Table 1, Figure 1B and 1C), thereby creating nested or intercalated gene arrangements. The process of gene- or exon-skipping in chimeric transcript formation suggests a novel mutational mechanism to explain the surprisingly large proportion of nested genes in the D. melanogaster genome (many of which bear no hallmark of retroposition) 2224, as well as the evolution of complex intercalated gene structures that cannot arise via simple mechanisms of gene duplication.

Of the 1,566 valid TEs in the Release 3.2 annotation of the D. melanogaster genome sequence, we estimate that 59 are chimeric TE insertions with some component co-transcribed in an exon, 414 are transcribed but entirely contained within spliced intronic sequences, and 1,093 are entirely contained within intergenic sequences not currently annotated as transcribed. A similar rank order pattern of TE abundance in different functional compartments has been observed in the Arabidopsis thaliana genome 25. These numbers of TE insertions deviate significantly from their expected proportions based on the genome annotation of the 116.8 Mb Release 3 sequence (p < 1 × 10^-15) (Table 2). This deviation from expectations is the result of two factors: there are fewer TEs in transcribed regions than in intergenic regions (p < 1 × 10^-15) 14, and there is a further reduction in exonic regions relative to intronic regions (p < 1 × 10^-15). The reduction in transcribed regions, however, is not solely caused by under-representation in exonic sequences, since the number of intronic TE insertions is reduced relative to the number in intergenic regions (p < 1 × 10^-15). Together, these results indicate that there is a paucity of chimeric TE insertions in the genome, and that the causes of this paucity go above and beyond the effects of simply being transcribed.

To estimate the extent to which the number of exonic TE insertions is reduced while controlling for the effect of transcription per se on the distribution of TEs, we use the number of intronic TEs and the length of the intronic compartment of the genome to estimate the proportion of unobserved chimeric TE insertions. The total length of intronic regions in the D. melanogaster genome is approximately 37.7 Mb and the total length of exonic regions is 28.2 Mb 2226. If the selective pressures on exonic TEs were similar in magnitude to those on intronic TEs we would expect to find approximately 414*(28.2/37.7) = 310 TE insertions in the predicted exonic (coding plus untranslated) regions of the genome. The fact that we detect only 59 chimeric TEs out of an expected 310 (or 19%) indicates that a chimeric TE insertion is much more likely to be highly deleterious to the organism than a non-chimeric TE insertion that is spliced out of a mature transcript. These results are consistent with previous findings in the human genome, that the proportion of TE-derived sequence increases with increasing distance upstream from the start of transcription 10.

These calculations are based on a comparison of the annotated set of chimeric TE insertions relative to the total set of annotated TE insertions. As noted above, however, our results reveal that an unknown proportion of TEs in the Release 3 sequence were not annotated in 14. If we assume that the frequency of unannotated TEs in intronic regions is proportional to that of the unannotated TE insertions in our sample (~10%), the expected number of TE insertions in exonic regions would increase to 310*1.10 = 341. Thus, using the total set under this proportionality assumption, the percentage of chimeric TE insertions detected relative to expectation is little changed (65 out of 341, 19%). To the extent that the number of unannotated TE insertions in introns is proportionally higher than in our sample, the percentage of observed chimeric TE insertions decreases even further, strengthening the claim for a paucity of chimeric TE insertions relative to expectation.

We estimated that ~80% of the TEs that have been inserted into mature genic transcripts are immediately purged from the genome by strong purifying selection, and therefore are not observed in the sequenced strain. What about the remaining ~20% of chimeric TE insertions that we do detect? We can envisage three scenarios to explain the existence of these chimeric TE insertions: 1) they are under strong purifying selection, like the TE insertions we do not observe; 2) they are adaptive, contributing useful sequences to the host genome; or 3) they are neither particularly deleterious nor particularly advantageous in comparison to the observed non-chimeric TE insertions in the genome.

In order to evaluate these possibilities, we surveyed the frequencies of chimeric TE insertions in wild D. melanogaster populations. The presence of each TE was tested in six pools of 8–12 North American strains (Table 1, rightmost column) using a PCR procedure custom-designed for each chimeric TE (see Methods for details). Pool frequencies were used to estimate confidence bounds on population frequencies using a maximum likelihood procedure (Table 3, [see Additional file 2]; see Methods for details).

We were able to generate population data for 48 of the 59 annotated chimeric TE insertions. Twenty-seven chimeric TE insertions were found only in the sequenced strain, seven were found in all six-strain pools and 14 had intermediate pool frequencies. These proportions of absent (56%) and polymorphic (44%) chimeric TEs are very similar to a combined, non-random sample of 92 non-chimeric TE insertions with previously reported population frequency data that map to annotated Release 3 TEs: absent (58%) and polymorphic (42%) 1217272829. The negative effects of intronic TE insertions on transcription do not strongly affect this non-chimeric sample, since similar proportions of absent and polymorphic TE insertions are observed in intronic (60% absent, 40% polymorphic; n = 30) and intergenic (56% absent, 44% polymorphic; n = 62) regions.

To determine whether the chimeric TE insertions are, on aggregate, subject to unusual selective constraints, we compared each of their pool frequencies to those of similar, non-chimeric TE insertions (Table 4). By "similar," we mean that these TE insertions came from the same family as their chimeric counterparts, that they had similar lengths, and were inserted in areas with similar recombination rates (see Methods for details). Since the selective constraint on a TE insertion is expected to increase with its length and the recombination rate of its genomic neighborhood 1730, we tried to bracket each chimeric TE with a pair of similar non-chimeric family members: one with slightly higher, and one with slightly lower, length and recombination rate (columns 4 and 6 of Table 4, respectively). Our null hypothesis was that the chimeric TE insertions are neither particularly deleterious nor particularly advantageous in comparison with their non-chimeric counterparts. If this null hypothesis is true, we expect the pool frequencies of non-chimeric TE insertions in column 5 of Table 4 to be no higher, and the pool frequencies in column 7 to be no lower, than those of the chimeric TE insertions in column 3.

For the set of 48 TE insertions for which we have population data, we cannot reject the null hypothesis of no difference in pool frequencies between chimeric and non-chimeric TE insertions. Neither the Wilcoxon one-sided test nor the Kruskal-Wallis test reject the null hypothesis in favor of the alternative that pool frequencies of chimeric TEs are significantly higher than those of their counterparts with greater lengths and recombination rates (p = 0.38 and p = 0.75, respectively; tests performed on the n = 34 TEs in Table 4 that have the appropriate counterparts). This indicates that, in general, the fact that a TE insertion is chimeric does not increase the likelihood that it is at higher population frequency and is therefore potentially adaptive. Similarly, we find no evidence that chimeric TEs in general have pool frequencies lower than those with shorter lengths and lower recombination rates (p = 0.15 for the one-sided Wilcoxon rank sum test, p = 0.30 for the Kruskal-Wallis rank sum test; n = 46). Thus, the fact that an observed TE insertion is chimeric does not increase the likelihood that it is deleterious.

While we do not provide evidence for unusual selection pressures acting on chimeric TE insertions overall, we do find a few exceptions to this general rule when TE insertions are analyzed on an individual basis. As shown in Figure 3, by comparing pool frequencies of chimeric TEs to those of the two types of non-chimeric counterparts, we detect evidence for two exceptional chimeric TE insertions. One, a Doc insertion (FBti0019430), which creates a truncated version of the putative choline transferase gene CHKov1 (CG10618), has a significantly elevated population frequency (chimera 44, Figure 3A) and has been reported previously to be a putatively adaptive TE insertion 1718. The second, a pogo (FBti0019206) insertion into the fructose-bisphosphate encoding gene fbp, has a significantly decreased population frequency (chimera 21, Figure 3B) and is likely to be more deleterious than similar non-chimeric pogo insertions.

Chimeric gene-TE transcripts were identified by three independent methods (with the following number codes used in Additional file 1): 1) a genomic coordinate intersection analysis; 2) a TE-to-gene BLAST analysis; and 3) a TE-to-EST-to-gene BLAST analysis. Coordinate overlaps were evaluated using the UCSC D. melanogaster table browser 48 and finding the intersection between the "FlyBase genes" and "FlyBase noncoding genes" tables, with a subsequent filter for those TE-gene overlaps >25 bp supported by EST evidence. For the TE-to-gene BLAST analysis, we sought chimeric TEs directly by querying each canonical TE sequence in version 7.1 of the BDGP TE data set 49 that had a representative in the Release 3.1 euchromatic genome annotation against the Release 3.1 annotated transcripts. For this analysis we used the combined output of hits from WU-BLASTN B = 10000 V = 10000 X = 3 M = 3-lcfilter-filter dust of >50 bp and >85% identity 50 together with NCBI-BLAST2 51 hits of E < 1 × 10^-10.

For the TE-to-EST-to-gene BLAST analysis, we developed a three-step process using WU-BLASTN with the following parameters: B = 10000 V = 10000 X = 3 M = 3-lcfilter-filter dust. First, each TE in the BDGP TE data set was used to query the BDGP EST database (ca. Dec 2002) containing 281,297 ESTs and complete cDNAs 2352. Second, ESTs with TE homology of >25 bp and >85% identity were aligned to the canonical TE sequence, and the non-TE component of the sequence was used to match the EST back to the corresponding host gene by querying transcripts in the Release 3.1 genome annotation 22. Finally, the annotated host gene (±5000 bp) was used to query the TE database to ensure that a TE of the appropriate family is present in the genomic region, thereby filtering artifacts generated by EST library construction.

Transcripts from heterochromatic regions of the Release 3 genome were excluded from this analysis, as were genes labeled as "pseudogene" or unnamed genes with "existence uncertain" status in FlyBase. We also note that as in 14, we excluded from this analysis the enigmatic INE-1 element 53 that can be found in many transcripts 54, since this repetitive sequence is structurally distinct from all other TEs in the Drosophila genome.

A population of 64 individual strains from North America was combined into a total of 6 pools of 8 or 12 strains. The final concentration of each pool was 2.5 ng DNA of each individual strain per PCR reaction. The composition of each pool was as follows: Wi pool: Wi1, Wi3, Wi15, Wi18, Wi41, Wi45, Wi68, Wi77, Wi83, Wi98, Wi137, Wi148 – these strains were collected at the Wolfskill Orchard, Davis, CA and have been subjected to over 30 generations of brother-sister matings (gift and personal communication by Sergey Nuzhdin); We1 pool: We4, We7, We10, We11, We25, We44, We47, We50, We57, We60, WE67, We80; We2 pool: We13, We17, We21, We28, We33, We37, We63, WE70, We75, We83, We88, We91 – the strains in the two We pools were collected in Raleigh, NC and have been subjected to 10 – 15 generations of brother-sister matings (gift and personal communication by Greg Gibson); NA pool: Broward13, Broward5, Lake5, Okee14, Okee5, Orange1, Orange2, Paho4, Paho6, Paho9, Sebring12, Sebring17 – these isofemale strains were collected at various locations throughout North America (gift and personal communication by Jeff Birdsley); NB pool: NB1, NB6, NB7, NB8, NB12, NB13, NB14, NB16 – these isofemale strains were collected in New Buffalo, Michigan (gift and personal communication by Bettina Harr); CSW pool: 3B, 6D, 11D, 20C, 23D, 25C, 29B, 36D – these isofemale strains were collected at Countryside Winery, Blountville, Tennessee (gift and personal communication by Lev Yampolsky).

We note that some of the isofemale strains above may be heterozygous for a given TE insertion. This would lead to a slight increase in the effective number of strains in any given pool. However, such an increase is unlikely to have an effect on the qualitative nature of our results, as the addition of several strains to a pool generally has no significant effect on the confidence limits of the population frequency of a TE. For instance, in the section on population frequency estimation (below, also [see Additional file 2]), we show the extent to which the population frequency estimate remains the same when we treat 8-strain and 12-strain pools as if they were equivalent to each other.

The presence/absence of TEs in all strain pools was determined using the polymerase chain reaction (PCR). All PCR primers were designed using Primer 3 55 and were checked with Virtual PCR 56. All primers have a melting temperature of 63°C (+/-0.2°C) and were synthesized by Operon Biotechnologies, Inc. in 96 well plates. The primers are intended to assay for the presence of the TE insertion and consist of a "Left" primer that lies within the TE sequence and a "Right" primer that lies in the flanking region to the right of the TE insertion. Primer sequences used in this study can be found in Additional file 3. The presence of the TE insertion should produce a band of approximately 500 bp and the absence of the TE insertion should result in the absence of any band. On each plate there are 3 internal controls that should always produce a single band of predetermined size, designed to control for quality of PCR.

We also verified that the DNA concentrations were sufficient to detect the presence of TE in a single strain out of the 12 or 8 strains tested in the pool. Each plate of primers was assayed with a control pool comprising one of three North American pools (Wi, We1, or We2) with the addition of y; cn, bw, sp (sequenced strain) to control for primer design problems. The addition of y; cn, bw, sp should give a result indicating the presence of the TE insertion being assayed in all cases where primers were designed correctly. To be conservative, the concentration of the DNA from the y; cn, bw, sp strain was somewhat lower than that from the assayed strains. The PCR reaction mix was made using Redtaq Readymix from Sigma Aldrich (#R2523) and primers at a final concentration of 1 μmol/μl. The PCR conditions were: 94° for 5 s, 27 cycles of: 94° for 30 s, 62° for 30 s and 72° for 1 min. We note that for 83 TEs, the positive control PCR did not fail in any such cases, showing the presence of the TE; PCR with the same pool DNA lacking any TE showed its absence.

Given that a TE insertion is present in some of the North American strain pools and absent from others (i.e. given its pool frequency), we wished to calculate the likeliest frequency of this insertion in the entire North American population, as well as suitable confidence bounds around such a frequency estimate.

Let x₁ (a number between 0 and 2) and x₂ (a number between 0 and 4) be the respective numbers of 8-strain and 12-strain pools in which a particular element is present. Let y be the theoretical frequency of this element in the North American D. melanogaster population. The likelihood L, of any particular value of y given the observed values of x₁ and x₂ is proportional to the probability of obtaining such x₁ and x₂ if y has that value. That is,

L(y|x₁, x₂) ∝ Pr(x₁ | y) × Pr(x₂ | y)     (1)

Where Pr(x₁|y) is the probability that x₁ out of two 8-strain pools contain the element and Pr(x₂|y) is the probability that x₂ out of four 12-strain pools contain the element, given that its overall frequency in the population is y.

The first term on the right hand side of equation (1) is equal to:

Where (1-y)⁸ is the probability that an element is not found in a given 8-strain pool, 1-(1-y)⁸ is the probability that it is, and the first term on the right hand side is the appropriate binomial coefficient. Similarly, the second term of equation (1) is equal to:

Substituting (2) and (3) into (1) and simplifying, we find that

Where k is an arbitrary multiplicative constant that absorbs the binomial coefficients in (2) and (3), since they are independent of the parameter y. In accordance with common practice, we make use of the log-likelihood function ln(L), which entails an arbitrary additive constant ln(k).

Additional file 2 provides three examples of the resulting log-likelihood functions. These functions correspond to the three possible combinations of x's that yield a total of four pools with detected element presence (i.e. for (x₁, x₂) equal to (0, 4), (1, 3) and (2, 2)). This file demonstrates that, given that the element is present in four out of six pools, estimation of population frequencies is relatively insensitive to the number of pools that contain eight or 12 strains. Therefore, to simplify the analysis, we combined all combinations of x₁ and x₂ under a common category such that x₁ + x₂ = 4.

For each log-likelihood function, the maximum likelihood estimate of the population frequency is the value of y at which the function reaches its maximum (middle column of Table 3). The confidence limits are determined by a likelihood ratio test of the values of y where the function drops below its maximus minus two (rightmost column of Table 3). The test statistic is the likelihood ratio of the 0-parameter model where y is fixed at the value of its maximum likelihood estimate to the 1-parameter model where y is allowed to vary. This statistic is distributed as a χ² distribution with one degree of freedom. When the difference in log-likelihoods increases above 2, the likelihood ratio increases above e² = 7.39, where e is the base of the natural logarithm. This value is the 99.3% quantile of the χ² distribution (corresponding to p = 0.007, 1 d.f.). These confidence limits were used to set the error bars in Figures 3A and 3B. Note that in situations with more than one possible combination of x₁ and x₂ the two rightmost columns of Table 3 list values that are averaged over all possible combinations (see explanation for x₁ + x₂ = 4 above).

We estimated the recombination rate at each TE insertion site method using a method previously developed for the D. melanogaster genome 54. This method combines the known physical and genetic distances between D. melanogaster genes to estimate the recombination rate profile of each chromosome as a second-degree polynomial function. An explanation of the method, and a tool that demonstrates its use, can be found on the world-wide web 57.

In Figure 2C, we classify chromosomal sites where the polynomial functions in 54 drop below zero as areas with "zero" recombination. We find that for the TE insertions in non-zero recombination areas, the median recombination rate is 2.75 cM / Mbp. Accordingly, we classify chromosomal sites with recombination rates above 0 and below 2.75 as areas with "low" recombination rates. The remaining chromosomal regions are labeled as areas of "high" recombination.