Because of the replication process characteristic of LTR retrotransposons, the LTRs of a given retroelement are sequentially identical at the time the element inserts into the host genome [18]. Thereafter, as an element accumulates mutations, its LTRs become increasing different from each other as substitutions specific for each of the two LTRs increase in number. The level of nucleotide identity seen between LTRs of a particular element, usually referred to as intra-element percent LTR nucleotide identity (%LNI), can be used in determining the relative ages of LTR retrotransposon families [7]. In rice, comparison of the two LTRs of the same element often showed the presence of a 10 to 30 bp regional duplication present in one LTR but not the other. In calculating %LNI, we have considered such duplications as single mutation events.

As the neutral nucleotide substitution rate has yet to be computed for rice, we cannot presently equate %LNI with a divergence time in years. However, the generally low level of sequence divergence between flanking LTRs of rice LTR retrotransposons (1.7%) indicates that most of the euchromatic full-length LTR retrotransposons in rice are relatively young, although significantly older elements were also identified. The seeming preponderance of young full-length LTR retrotransposons in the euchromatin of rice is similar to previous reports on yeast [19,20], Caenorhabditis elegans [7], A. thaliana [21] and Drosophila melanogaster [12]. This contrasts with findings in Z. mays [12] and humans [22].

To date, 23 families of copia-like elements have been reported for rice (S. Wang, submission to EMBL, N. Jwa, submission to GenBank, and [8,9,19,23,24]). Several have been described under more than one name. For example, the amino-acid sequence given for Tos4 in Hirochika et al. [23] is the same as that given for Tos1 in GenBank (accession number S22455) so they are really the same. Rire5 described by Kumekawa et al. [25] is the same family as Tos14 previously described by Hirochika et al. [23]. The equivalence between Tos14 and Rire5 became evident when we found the LTR sequence reported by Kumekawa et al. in elements that also contained the RT sequence given by Hirochika for Tos14. In our survey of GenBank and MRDB, we have identified an additional 16 copia-like families that have not been described by previous workers. In addition, exemplars for each of the previously identified families were found (except in the case of certain families that exist at such low copy numbers that no full-length element exists in GenBank or MRDB).

One of the most interesting new finds in our survey was Osr8, one of the oldest families of LTR retrotransposons in the rice genome. On the basis of a survey of the available portion of the GBRD and MRD, we estimate the copy number of Osr8 to be around 1,100 (more than any other copia-like family). Osr8 elements exist far more frequently as fragments (ratio of 10:1) and they display relatively low levels of %LNI in their full-length copies (mean %LNI for the five full-length Osr8 elements present in the GBRD is 97.2%). The RT of Osr8 is 60% similar to an unnamed polyprotein in Z. mays (AAD20307). A closely related family, Osr10 has two full-length copies in the GBRD but scans of the MRD suggest this element, also previously unrecognized, has the third highest copy number (~400) among copia-like elements. Outside rice, the RT of Osr10 shows highest similarity (~65%) to that of the maize retrotransposon Opie-2 (T04112). The broader clade that includes Osr7, Osr8, Osr9, and Osr10 is closely related to Endovir1-1 (AAG52949) of Arabidopsis (Figure 1, Table 2). These elements are also related (~60% similar) to maize's PREM-2 as well as to tomato's ToRTL1. Both Osr7 and Osr9 are present in very low copy number (one full-length and a few fragments in the GBRD).

Although it is present at only about a quarter of the copy number of Osr8, the unrelated Osr14/Tos1/Tos4 is also composed primarily of highly fragmented elements. Those that are full length have low %LNI (family mean 97.6%). Thus, Osr14/Tos1/Tos4 and Osr8 seem to be of similar age and to have followed a similar evolutionary pattern, albeit with less intense amplification in the case of Osr14/Tos1/Tos4. Osr14/Tos1/Tos4, Osr15/Tos12, and Osr53/Tos18 form a well defined clade and are more closely related to Ta1-2 (S23315) of Arabidopsis than to any other rice retroelement family outside their clade (Figure 1, Table 2). Osr15/Tos12 and Osr53 are only just sufficiently different to constitute distinct families.

Osr1/Tos14/Rire5, Osr13/Tos5, Osr51/Tos15, and Osr52/Tos16 have been described as distinct families but, inasmuch as their RTs are all 85% similar to each other, these groups are only marginally distinct. Searches of GenBank show that elements in this group are much more closely related to (75-80% at the amino-acid level) to maize retrotransposon Fourf (AAK73108) than to any rice LTR retrotransposon outside their clade. If the elements belonging to this group were considered to be a single family, it would be almost as large (~900 elements) as Osr8. In the GBRD the majority of these elements are fragmentary, but the estimated copy number of full-length elements in the rice genome for this quartet still exceeds 100.

Osr18, Osr19, Osr20, Osr22, Osr23, Osr24, Osr45/Tos7, and Osr46/Tos8 form a clade of low copy number families composed primarily of fragmentary copies. Our results suggest that each of these families has a copy number in the range of 50-100 elements. Members of this clade are closely related to maize's Hopscotch element (T04112) (Figure 1, Table 2).

Osr2 and Osr12 are low-copy families and are represented in the GBRD by two and three copies respectively, all of which are full length (although one copy of Osr12 contains a large internal deletion), suggesting that these elements may have recently invaded the rice genome. The high level of LTR nucleotide identity (≥ 99%) seen in these elements is consistent with this recent invasion hypothesis. Members of Osr12 and Osr2 are potentially active because they have large, intact polyprotein ORFs, usually in excess of 1,000 amino acids. All three Osr12 elements detected in the GBRD are on chromosome 10. Similarly, both Osr2 elements are inserted within 50 kb of each other on chromosome 4. Nonetheless, these two families are not closely related (their RT sequences are only ~50% similar at the amino-acid level). Osr12 RTs differ from those of all other rice copia-like elements by 50%. And yet RT sequences of elements in Osr12 are 60% similar to certain elements in the maize genome (Zmr1 (S27768) and mzecopia (M94481.1)).

One full-length, and one fragmented copy of Osr6 are present in the GBRD. Osr5 is slightly more common than Osr6, to which it is most closely related, but it is currently represented in the GBRD by only a single full-length copy and a few fragments. Osr5 is 60% similar to the tobacco retrotransposon Tnt1-94 at the amino-acid level (RT comparison). Osr4 is another low-copy family. It has several fragmented representatives in the GBRD, and is probably somewhat older than Osr12 and Osr2, but it has only three full-length copies in the GBRD, Osr4 elements have an exceptionally large polyprotein ORF (~1,600 amino acids). The RT of Osr4 shows 50% similarity to that of retroelements in the Arabidopsis genome (for example, BAB01972, NP_175303).

Although the RT of Osr3 was detected during our survey, elements in this family are fragments with ill defined LTRs. TBLASTN reveals the RT of Osr3 to be the single representative of its type in the GBRD. Both Osr3 and the equally aberrant Osr21/Tos17 differ from those of other copia-like elements found in our study by about 55%. Osr11/Rire1 is a low-copy family closely related (75% similarity) to a retroelement in the Arabidopsis genome (Atr-2, T01860). Two other closely related families are Osr16/Tos6 and Osr17, both of which are similar to Sto-4 (T17429) of maize (Figure 1, Table 2). Nine additional low-copy families identified by earlier workers are Osr47/Tos9, Osr48/Tos10, Osr49/Tos11, Osr50/Tos13, Osr54/Tos19, Osr55/Tos20, Osr57/Rtr3, Osr58/Rrt5, and Osr59/Rrt8. Source references for each of these nine families are given in Table 3.

Osr27/Rire9 [26] is the third largest family in the rice genome, with an estimated copy number of 900 elements, mostly full length. Li et al. [26] estimated the copy number of this family at 1,600. The typical Osr27/Rire9 element is quite large (~12.8 kb total length). Having intact polyprotein ORFs and high mean %LNI (99%), these elements probably are, or recently have been, actively transposing. Yet the presence of a few members of this family that are more mutated (short ORFs, low LTR-LTR nucleotide identity) suggests that this may also be an ancient family. Two other families, Osr40 and Osr41, are also members of the same clade as Osr27/Rire9, Osr25 and Osr26/Rire2 (Osr25 and Osr26/Rire2 are discussed below), but both have RTs that are about 30% different from those of Osr26/Rire2 and Osr27/Rire9. Neither Osr40 nor Osr41 has been previously identified, but with approximate copy numbers of 600 and 300, respectively, these are both large families. The RTs of members of this clade show about 60% similarity to that of Retrosor1 (Sorghum bicolor, AAD19359).

With approximately 1,500 elements, Osr30 constitutes 14% of all LTR retrotransposons in the rice genome. Although Osr30 is the largest family of LTR retroelements in the genome, it has not been previously named. These elements are slightly larger (~13.1 kb) than those of Osr27/Rire9. A higher proportion of fragmented copies and lower level of LTR-LTR nucleotide identity suggest that Osr30 is older than Osr27/Rire9. Osr29, which is closely allied to Osr30, is also a large family with more than 500 member elements. Taken together, the elements of the Osr29 and Osr30 clade are unusual, because they are as closely related to other major rice clades as they are to any elements outside rice. Osr28 is a low-copy family that is most closely related to Osr29 and Osr30 (Figure 2).

Two other large gypsy-like families are Osr33/Rire8 [25] and Osr34. These two families each have copy numbers of approximately 500. Two low-copy families belonging to the same clade are Osr32 and Osrs6/Rire3 [27] (Figure 2). Members of these families have large LTRs, typically in the range 3,000-3,500 bp. RTs of families in this clade show high sequence similarity to an LTR retrotransposon in pineapple (~70% to Acr-1; CAA73042) and to one in Sorghum bicolor (~77% to Retrosor3, AAD221153) (Figure 2).

Osr31/Rire7 is an aberrant low-copy family that is much more closely related (77% similarity) to an Arabidopsis element, Atr-4 (see Table 2), than to any other LTR retroelement families in the rice genome (Figure 2). In the clade of five low-copy families, composed of Osr35, Osr36, Osr38, Osr39, and Osr42, an RT was found in the GBRD for only two families, Osr35 and Osr36. The other elements were identified in scans of the MRD and their full sequences have since been submitted to GenBank (for accession numbers, see Table 1). This clade is closely related to Arabidopsis element Atr-5 (Figure 2, Table 2).

Members of family Osr25 are all internally deleted and thus non-autonomous (mean length 4.3 kb). Although Osr25 elements have typical LTRs, PBS, and PPT, the inter-LTR region contains only non-coding, repetitive DNA. The LTRs of Osr25 display 65-70% sequence similarity to the autonomous elements of the gypsy-like family Osr26/Rire2. Elements with LTRs having such a high degree of similarity are usually considered members of the same family. Nevertheless, because members of Osr26/Rire2 have the usual coding structure typical of other gypsy-like elements (while Osr25 elements entirely lack typical retroviral genes) and members of these two families fall into two sharply distinct, non-overlapping clades, we report these two types of elements as separate families. Estimates based on scans of the MRD and the GBRD suggest that the rice genome contains about 500 copies each of Osr25 and Osr26/Rire2. Osr25 and Osr26/Rire2 display 98.9 and 97.9% LNI respectively.

Osr37/Rire4 is also aberrant compared to other rice LTR retrotransposon families. The typical element in this family is 4.4 kb long, about the same length as Osr25 elements. Members of Osr37/Rire4 usually carry a large ORF (up to 600 amino acids) just upstream of the 3' LTR. This ORF shows no significant similarity to any known RT sequence. Up to the present in the GBRD, where these ORFs are generally identified simply as hypothetical proteins, the large ORF of Osr37/Rire4 seems not to have been recognized as a retroviral gene. This ORF may serve an integrase function as BLAST searches show that it has low homology to a putative integrase in A. thaliana (28%; AC005171). There are about 600 copies of Osr37/Rire4 in the entire rice genome.

In addition to the foregoing copia- and gypsy-like families, our scans identified two families, Osr43 and Osr44, of small elements (overall length < 2,000 bp). With LTRs only 148 bp long and an overall length of 1,207 bp, Osr44 elements are especially small. Members of Osr43 and Osr44 are unique because, although they possess all of the canonical LTR-retrotransposon structural features (LTRs, PBS, PPT, and TSRs), they are internally deleted and either completely lack or encode only very small ORFs with no similarity to any known protein. Both families contain on the order of 100 copies genome-wide.

LTR_STRUC identifies new LTR retrotransposons on the basis of the presence of characteristic retroelement features (E.M.M. and J.F.M., unpublished work). It scans nucleotide sequence data for putative LTR pairs, aligns the putative pairs, and scores them on the basis of the presence/absence of expected motifs such as TSRs, canonical dinucleotides, PBS, PPT, and so on. When a given pair receives a score above a (user-specified) cut-off, an output record is generated that specifies salient information about the putative element, such as the length of the transposon and its LTRs, its position within the contig, an alignment of its LTRs, the nucleotide sequence of the transposon, its LTRs and target-site repeats, as well as a file listing all ORFs. In our study, once putative elements were identified, sequence analysis was carried out on the individual output files to identify those that described actual LTR retrotransposons. Additional elements were identified by BLAST searches using elements located by LTR_STRUC as queries.

Initial scans with LTR_STRUC were conducted on a dataset consisting of the 29.8 Mb of O. sativa BAC-derived sequence data available in GenBank at the time of the initial scan (December 2000). This dataset (TDS) was obtained from the TIGR website [30]. Subsequently, LTR_STRUC was used to scan the non-redundant MRD, a product of the Monsanto Rice Genome Sequencing Project. The MRD is based on an initial dataset of 3,391 BACs distributed across the genome of O. sativa cv. Nipponbare - the same cultivar used by the International Rice Genome Sequencing Project. Removal of contaminants and redundancies from this initial dataset produced the MRD (consisting of 52,202 contigs, totaling 259 Mb of the 430-Mb rice genome). More recently, in an effort to determine the relative copy numbers of the various families and identify additional elements not picked up in our initial survey with LTR_STRUC, we have used representative sequences from each retrotransposon family identified in this study as queries to conduct BLAST searches against both the MRD and the GBRD. Thus, the results reported here constitute a reasonably unbiased survey of LTR-retrotransposon diversity in rice. Both the MRD and GBRD are heavily weighted toward euchromatic sequences. The amount of data scanned was significantly less than the total amount of nucleotide sequence contained in the MRD and GBRD. Much of the MRD (~36%) is composed of contigs that are less than 10 kb long and are therefore of limited utility for the LTR_STRUC program, which finds only full-length elements (rice gypsy-like elements are typically longer than 10 kb and are not entirely contained in such short contigs). In the case of the GBRD, the amount of rice nucleotide sequence available for search was less than one-third of the 174 Mb released to the public (because of 15% redundancy, the GBRD sequences amounted to a total of only about 150 Mb, of which only some 50 Mb were actually available for BLAST search because most of these sequences were in the process of being 'finished'). RT sequences were identified according to previously described criteria [31,32].

The RT domains of the Osr elements were aligned with previously reported RT sequences (Table 2). The ClustalW analysis [33] extension to MacVector 7.0 was used to generate two amino-acid alignments, one for gypsy-like, and one for copia-like elements. Draw N-J Tree and Bootstrap N-J commands of ClustalW were then used to generate non-bootstrapped and bootstrapped trees, respectively.