Around 40 percent of the mouse genome consists of RTE sequence, slightly lower than observed for the human genome, although this presumably is a result of the higher substitution rate in mouse, limiting the identification of old RTE sequence 1213. RTEs are divided according to the presence or absence of long terminal repeats (LTR). Mammalian LTR elements consist mainly of endogenous retroviruses (ERV) that at some point during evolution have been inserted in the germline and fixed. Although the amount of sequence occupied by LTR elements is comparable between human and mouse, the level of de novo mutations caused by LTR element activity is extensively higher in mouse than in human 814. The most abundant ERV class in the mouse genome (~5.5%) is the Class III ERVs, which in the RepeatMasker 15 annotation - upon which this study is based - is broadly divided in two groups, the ERVL and MaLR elements. The latter is a non-autonomous transposon, meaning that the elements do not encode the enzymatic machinery required for its own transposition. The Class II ERVs (~4% of the genome), annotated as ERVK in RepeatMasker, is believed to be younger than Class III ERVs 16 and consists of a broad range of clades, including the IAP elements (Intracisternal A-type Particles). Class I ERVs (ERV1 in RepeatMasker) cover less than 1% of the mouse genome. Through ectopic recombination between the flanking LTR sequences, solitary LTR sequences may be formed. In RepeatMasker, terminal LTR sequences and the internal sequences (residing between the terminal LTRs in a complete LTR retrotransposon) are annotated independently. Although the terminal and internal sequences may in many cases be determined to form a single LTR retrotransposon, for simplicity, the two annotations (termed 'LTRter' and 'LTRint', respectively) are analysed independently in this study.

Non-LTR retrotransposons are divided into LINEs (Long INterspersed Elements) that are autonomous, and SINEs (Short INterspersed Elements) that are non-autonomous. LINEs occupy roughly 20% of the mouse genome. The majority of mouse LINE elements belong to the L1 superfamily, which contains sub-families that are still active 171819. Despite the comparable levels of genome occupied by LINE sequences in human and mouse 1213, there are more than 15 times as many full-length L1 elements with intact open reading frames in the mouse genome 20. Almost 1.5 million SINE elements are present in the mouse genome, making up approximately 8% of the total genome size. Unlike the human genome where a single SINE, the Alus, is dominating 21, the mouse genome harbours two successful superfamilies of SINEs, Alu and B2 that are present in equal numbers 12. The evolutionary histories of the mouse SINEs are truly different; Alus are derived from a 7SL RNA, whereas B2s evolved from a tRNA sequence 2122.

Piwi-associated RNAs (piRNAs) are small (24-30 nucleotides long) RNAs that bind Piwi proteins of the Argonaute family 2324. The mouse genome encodes 3 Piwi proteins, MILI, MIWI and MIWI2 that all binds piRNAs in the male germline 2526. Initially, piRNAs from adult mouse testis were found to contain less RTE sequence than would be expected from the genomic content of RTEs, suggesting that piRNAs were not specifically targeting RTEs 2728. However, a later study on piRNAs from an earlier (pre-pachytene) stage showed a high content of RTE sequence in piRNAs 29. Further evidence for the involvement of mouse piRNAs in controlling RTE activity came with the finding that knockout of Mili and Miwi2 resulted in reduced piRNA levels and increased RTE transcription 2930. Knockout mice further showed decreased DNA methylation levels at RTE loci 3132. As the temporal expression of Piwi proteins in developing mouse testis coincides with the resetting of genomic methylation 33, it is hypothesised that piRNAs act as guides for the methylation machinery 293132.

By analysing the piRNAs bound to MIWI2 and MILI, Aravin and colleagues 31 suggested the following scenario: In prenatal development (16.5 days postcoitum, dpc), transcripts from full-length active RTEs are the main substrates for piRNAs that primarily associate with MILI (and to a lesser extent to MIWI2). Available transcripts containing antisense RTE sequence bind this complex and antisense RTE piRNAs are formed which in turn associate primarily with MIWI2 (and MILI, respectively). Both complexes may bind complementary RTE transcripts, entering the so-called ping-pong amplification cycle of piRNAs, in which Piwi-bound piRNAs pair with complementary transcripts that are subsequently cut into new piRNAs having a 10 nucleotide overlap with the template piRNAs 3134. In prenatal development, piRNAs are primarily targeting L1 and IAP RTEs, for which activity has been reported at this stage 3536. In postnatal development (10 days postpartum, dpp) MIWI2 is no longer detectable, whereas MILI is present throughout germline development 313738. The overall level of piRNA targeting of RTE sequences drops at 10 dpp, but interestingly, a relative increase in the piRNAs targeting B1 SINEs (members of the Alu superfamily) was observed 31.

This raises two fundamental questions. Firstly, do Piwi proteins discriminate between transcripts and how is RTE sequences then distinguished from other transcripts? The finding of piRNAs targeting supported a scenario with limited discrimination 31. Secondly, what lies behind the apparent shift in RTEs being targeted by piRNAs during development in male mouse germline? By analysing to extent to which genomic RTE loci are targeted by piRNAs in developing mouse testes, the present study aims at assessing the transcriptional dynamics of RTE during development, and consider the relationship between RTE activity and piRNA generation further.

The data for such analysis should meet a range of criteria. Although numerous mouse RNA libraries are available, only a small subset is derived from wild-type developing mouse testes 39. Further, as the prevalent transcription of RTEs will results in a large population of fragmented transcripts of sizes similar to piRNAs, analysis should be restricted to libraries of RNAs associated with Piwi proteins. This limits the available data to libraries from the above-mentioned study by Aravin and colleagues 31.

Three libraries of small RNAs bound to MIWI2 and MILI proteins in mouse male germline 31 were reanalysed; one library with MIWI2-bound RNAs from 16.5 dpc (henceforth referred to as 'MIWI2 early') and two libraries with MILI-bound RNAs from 16.5 dpc and 10 dpp ('MILI early' and 'MILI late', respectively) (Table 1).

To analyse in detail the theoretical piRNA coverage of each individual RTE loci, reads from each RNA library were mapped onto the mouse genome. Only perfectly mapping reads were considered. Due to the inherent redundancy of RTE sequences, reads mapping uniquely to RTE loci are scarce and therefore all mapping reads were considered. The number of reads mapping to each RTE loci was weighted by the uniqueness of the reads, so that the count from each read was simply divided by the number of times that particular read mapped to the entire genome. Further, read coverage was weighted by library size and length of the locus, in an attempt to allow direct comparisons between libraries and RTE types. As seen from Figure 1, differences in this theoretical piRNA coverage differ markedly between RTE families and between libraries. When only considering reads that are exclusively mapping within a particular superfamily of RTEs (e.g. L1 LINEs) the overall patterns of coverage still remain (middle column in Figure 1). For LTR elements (both internal sequence and terminal repeats), most major trends are still observable when only considering reads mapping entirely within a single family of RTE (e.g. internal sequence of IAP-d elements) (right column in Figure 1). Unless otherwise noted, in the following all reads are used for analysis, and only RTE families with a least a thousand genomic members are considered (see Additional File 1, Table S1 for a list of these 318 RTE families). The previously reported decrease in piRNA coverage of IAP LTRs and L1 LINEs and the corresponding increase in SINE coverage during development 31 are clearly evident from Figure 1 (left column). Also consistent with earlier findings 31, the median MILI early piRNA coverage of individual RTE families is highly correlated with the median coverage from MIWI2 early piRNAs, but not with the coverage from MILI late piRNAs (Additional File 1, Figure S1).

The RepBase database of repeated sequences 40 contain consensus sequences for individual RTE families, and RepeatMasker 15 annotation is based on sequence similarity to these consensus sequences. As the vast majority of RTE loci are under no negative selection (but see, for example 414243) the level of divergence between genomic loci and the RepBase consensus sequence can be taken as a proxy for the age of the RTE family. When plotting median piRNA coverage of RTE families against their median divergence, a clear trend of highly covered RTEs being relatively young is observed irrespective of RTE type (Figure 2). For all types of RTEs, the average piRNA coverage of younger elements is significantly higher than coverage of older elements (Additional File 1, Table S2). Also, the high level of variation in piRNA coverage between individual RTE loci is evident from the percentiles shown in Figure 2.

To test the piRNA coverage in the genomic context of protein-coding genes, all known genes with at least 20 kb (kilo base pairs) to the nearest neighbouring gene (in both directions) and with available Affymetrix expression data from testis tissue were retrieved 44. These 3307 genes were grouped into highly expressed genes (25% highest expression signals, 827 genes), lowly expressed genes (25% lowest, 829 genes), and medium expressed genes (rest, 1651 genes). For each gene set, the piRNA coverage of RTE residing within 10 kb upstream of their annotated transcription start sites to within 10 kb of their stop sites was recorded (Figure 3). For all types of RTE elements, piRNA coverage in the context of highly expressed genes was significantly higher than both medium and lowly expressed genes. This was found for all piRNA libraries (Additional File 1, Table S3).

For some RTE types, the relative number of young and old loci differ between the vicinities of highly expressed genes and lowly expressed genes, (Additional File 1, Figure S2), suggesting that the higher levels of piRNA coverage of RTE near highly expressed genes could simply be explained by the age of the RTE sequences. However, when repeating the analysis without the youngest RTE sequences, essentially similar results and significance levels are found (Additional File 1, Table S3).

Interestingly, when assessing the piRNA coverage of RTE sequence near transcription start sites (TSS), peaks are observed immediately upstream of TSS on the reverse strand, and for piRNAs not targeting RTEs, also immediately downstream of TSS on the forward strand (Additional File 1, Figure S3). Such a pattern resembles that of the recently discovered short transcripts generated around TSS (the TSS-associated RNAs) 4546 suggesting that these piRNAs may in fact be TSS-associated RNAs. It is uncertain if this represents experimental contamination of non-piRNAs or if these TSS-associated RNAs provide the transcripts that a processed into piRNAs, although the presence of RNA reads smaller than the usual 24-30 nucleotides - especially among early MILI piRNAs - hints that a contribution from the former scenario cannot be ruled out (Additional File 1, Figure S4). Assuming all RNAs mapping within 1000 base pairs of an annotated TSS are TSS-associated RNAs and removing these from the analysis does not change any of the presented conclusions (data not shown).

The age of RTEs and their genomic distance to protein-coding genes is not independent 1347. If RTEs residing near genes are in general relatively young, one would expect these RTEs to display high levels of piRNA coverage as a result of this. To test if proximity to genes affected piRNA coverage independent of RTE age, members of each RTE family were divided into three equally sized groups based on their divergence from their consensus sequence (called 'young', 'median' and 'old' loci, respectively). Within each age-group, members were further divided into sub-groups according to genomic location; i) RTE loci residing inside the boundaries of known genes (termed 'genic'), ii) RTE loci in intergenic regions in proximity to known genes ('proximal'), and iii) RTE loci in intergenic regions distal to known genes ('distal'). The groups 'proximal' and 'distal' defined as loci closer or further away from genes, respectively, than the median distance of all non-'genic' loci from the RTE family. An overview of the grouping procedure is presented in Figure 4. The assumption that younger RTE members tend to reside closer to genes are confirmed by the observation that for 95% of all RTE families, the fraction of loci being proximal to genes is higher for young loci than for old loci (Wilcoxon Signed Rank, p < 2.2 × 10^-16; values not shown).

RTE loci proximal to genes have - irrespective of age - significantly higher piRNA coverage than similar RTE loci distal to genes (Figure 5). With internal LTR sequences belonging to the 'old' group as the only exception, loci proximal to genes have significantly higher coverage than genic loci in prenatal development. Interestingly, in postnatal development no RTE group displays a significantly higher coverage for proximal loci than for genic loci, and furthermore, for all RTE groups, genic loci have significantly higher coverage than loci distal to genes (Figure 5). Thus, coverage by MILI late piRNAs is enriched in genic regions, an observation that is further supported by the fact that the total coverage of MILI late piRNAs mapping to genic regions is increased for all types of RTEs (Figure 6).

Aravin and colleagues 31 showed that in early development, the substrate for piRNA generation is provided by actively transcribed RTE elements. Later in development, active transcription of RTEs should then be repressed, and mRNA sequences from active full-length RTE loci are no longer widespread. This suggests that co-transcription of RTE sequences along with mRNAs from protein-coding genes (predominantly in intronic regions) could now take on a relatively larger role in providing RTE sequence transcripts. A prerequisite for the generation of piRNAs is the presence of transcripts with complementary sequences. Although active RTEs need to be transcribed from their forward strand, the RTE sequences scattered around the highly transcribed genome could produce transcripts in both orientations. But if as suggested, transcribed RTE sequences in postnatal mouse testes are mainly provided from co-transcription with genes, the transcriptional orientation of a given genic RTE loci is to a large extent determined by the orientation of the host gene. One can therefore test if the strand of piRNAs mapping uniquely to genic RTE loci corresponds to the orientation of the RTE relative to the host gene. Of course, amplification from the ping-pong cycle may generate multiple piRNAs, which may map on both strands of a genic RTE locus (although the efficiency of the ping-pong cycle may decrease in postnatal development as MIWI2 is no longer expressed 31), potentially blurring the picture. As seen from Figure 7, a clear pattern of high sense coverage of genic RTEs in the forward orientation and high antisense coverage of genic RTEs in the reverse orientation is seen for postnatal piRNAs, but not for prenatal piRNAs.

Small RNA libraries accession numbers GSM319953 (MILI late), GSM319956 (MILI early) and GSM319957 (MIWI2 early) were retrieved through DeepBase (http://deepbase.sysu.edu.cn/) 39 and mapped to the mouse genome (mm9 assembly) using bowtie 65. Prior to mapping, reads were filtered and sequences with ambiguous base calls and low-complexity sequences were removed. The latter was done measuring the linguistic complexity 66 of the sequences in 16 bp windows, and excluding reads with an average complexity of less than 0.75. Preliminary tests showed that this would remove highly repetitive reads with very large numbers of genomic mappings (not shown). For each library, this procedure filtered out between 0.14 and 0.17% of all raw reads. RepeatMasker and known gene annotations were downloaded from the UCSC Genome Browser 154067. A set of non-overlapping TSS was selected by grouping all known genes according to their assignment to ENSEMBL genes 68. For each ENSEMBL gene, the most abundant genomic start point was selected. If more than one point was found to have the highest abundance, the one furthest upstream of these was chosen. Gene expression levels were assessed from the 'testis' signal intensities in the Mouse GNF1M Gene Atlas from BioGPS (http://biogps.gnf.org/) 44.

For each RTE loci the number of reads mapped within the locus were recorded. Reads were assigned a score of 1/(number of genomic mappings of read), so that only uniquely mapping reads scored 1. The read score were then divided by the size of the RTE loci (in kilo-base pairs). Finally, scores were divided by the total number (in millions) of mapped reads from the library in question.

To test for difference between RTE loci from different genomic regions (data presented in Figure 5), all RTE families were first split in 3 groups based on age where after members in each group were divided according to their genomic context (genic, proximal, distal). Hence, for each RTE family, nine sets of loci were formed, and the average piRNA coverage for each set was recorded. To test for difference between two groups (for example, between old LINE loci being genic or distal), pairs of average values were collected for the 90 LINE families (Additional File 1, Table S1) and tested using a Wilcoxon Signed-Rank Test. Bonferroni corrections (n = 108 in Figure 5) were calculated as: p_corrected = 1-(1-p)ⁿ. All statistical analyses were carried out using R 69.