We analyzed a dataset of 626 multigene families, each family with at least three paralogs. In a total of 5274 genes, we detected 377 gene conversion events involving 513 genes in 189 families. Approximately 66% of these conversions involved sequences shorter than 100 nucleotides (Figure 1). The number of conversions identified by the detection algorithm declines rapidly as sequence length approaches zero. (See Methods for a description of the detection algorithm.) The average length of all conversions is 130 nucleotides and ranges from 4 to 1237 nucleotides. In general, the length distribution of converted regions in the rice genome is similar to those found in other species 67. Short conversions with fewer than about 10 nucleotides are usually considered to be artifacts (Stanley Sawyer, personal communication). Our analysis included six conversion events involving match lengths with less than 10 nucleotides; although these six events have low reliability, their presence does not influence the interpretation of our results.

Gene conversion with pseudogenes may accelerate gene family evolution; in this model, pseudogenes are postulated as a source of genetic information. The introduction of genetic material from pseudogenes may lead to higher divergence in orthologs between rice and related species. To test this hypothesis, we established rice gene families as having either low similarity (LS) or high similarity (HS) to Arabidopsis thaliana. HS gene families are defined as having statistically significant sequence similarity to Arabidopsis genes. Conversely, LS gene families have low similarity to Arabidopsis. It follows that LS gene families are more likely to be rice specific.

We categorized the 377 conversions detected in 50 LS and 327 HS gene families as described in Methods. Among the 377 conversion events, we identified those conversions involving pseudogenes in LS and HS families. The fraction of pseudogenes was approximately 56% in the LS families and 21% in the HS families (Table 1).

To rule out effects from assembly artifacts on our study, we performed a similar analysis on the japonica rice genome published by The International Rice Genome Sequencing Project (IRGSP) 17. The gene conversion characteristics were indistinguishable between the two assemblies (data not shown).

Our analysis of the rice genome detected 513 genes likely to be involved in 377 conversion events. To determine the distribution of gene conversion across the genome, we mapped the converted genes to chromosomes (Table 2). We were also interested in chromosomes with more conversions than the average number of conversions per unit length. To assess this, we estimated that the average number of conversions per million nucleotides for the entire genome was ~1.361. Based on the average frequency of 1.361 conversions per megabase, the number of conversions was found to be relatively uniform as a function of length for each chromosome. In the 377 events, 148 conversion pairings occurred within the same chromosome (~39%) as intra-chromosomal conversions; 229 conversion events occurred between chromosomes (~61%) as inter-chromosomal conversions.

From a total of 3844 gene pairs in the 626 multigene families, 2903 pairs were located on different chromosomes, with 941 pairs co-located on the same chromosome. It follows that the proportion of inter-chromosomal pairs involved in conversions was ~8% (229/2903), and intra-chromosomal pairs involved in conversions was ~16% (148/941). Although the number of inter-chromosomal conversion events was higher than intra-chromosomal conversion, the inter-chromosomal fraction was lower with respect to the potential for conversion based on total gene pairs. Thus, candidate pairs on the same chromosome apparently result in a higher likelihood for gene conversion. This does not necessarily represent an intrinsic bias for intra-chromosomal gene conversion; alternatively, it may represent an opportunistic positioning for gene conversion to occur.

The 229 inter-chromosomal conversion events are distributed non-uniformly among the twelve chromosomes of the rice genome. The conversion events are significantly more frequent between chromosomes 1 and 5, 2 and 6, and 3 and 5 than the average (Z-test, P < 0.05). The conversion distributions are shown in Figure 2 for each chromosome.

To determine the extent of segmental duplication on the observed conversion events, we compared conversion data with duplicated segments in the Beijing indica assembly 16. For all 229 inter-chromosome conversion events, only 21 out of 229 (~9%) inter-chromosome conversion events were involved in the segmental duplications.

To compare the sequence similarity of gene pairs involved in conversions to gene pairs not involved in conversions, we aligned entire gene sequences as described in Methods. The fraction of gene pairs involved in conversion events had significantly higher sequence similarity compared to all gene pairs (Figure 3). To test if the converted regions themselves influenced the higher similarity in the conversion group, we also calculated similarities of genes in conversion pairs, omitting the converted regions from the analysis. No significant difference was found between before and after omitting the converted regions (data not shown); the higher sequence similarity of gene pairs involved in conversion was not only a feature of the converted regions. In fact, most gene pairs not involved in conversions share statistically significant similarities in about 30–60% of their sequences with an average sequence similarity of 45%. In contrast, gene pairs involved in conversions exhibit a greater sequence similarity in the range of 60–80%. These results are consistent with findings that gene conversion is favored between similar genes 18.

For the 148 intra-chromosomal conversions, the conversion frequency decreased with the physical distance between gene pairs along the chromosome (Figure 4). The gene pairs separated by less than 5 kb demonstrated the highest conversion frequency; these could be considered closely linked genes. For gene pairs separated by more than ~35 kb, conversion was infrequent.

The directions between gene pairs were also examined. Among the 3844 gene pairs, 2172 had the same direction and 1672 had reverse directions. In the 377 conversion events, 277 (60%) occurred between gene pairs with the same direction (See details in Additional File 1). The proportions of conversions with same-direction gene pairs (227/2172 = ~10%) and reverse-direction gene pairs (150/1672 = ~9%) were similar. The larger number of gene conversion events in the same direction coincides with the larger number of gene pairs in the same direction; however, from our data we cannot determine whether the conversion bias is an intrinsic preference.

To determine if gene pairs involved in conversions are subject to evolutionary selection pressure, synonymous substitution rates (Ks) and non-synonymous substitution rates (Ka) were used. The Ka/Ks ratio can reflect the selection pressure between gene pairs caused by evolutionary processes. We calculated and compared the Ka/Ks ratios for two groups: (1) the closest homologs in each multigene family where at least one homolog was involved in gene conversion and (2) all close homologs in each multigene family (see Methods). The Ka/Ks profiles for the two groups were indistinguishable (data not shown).

We assessed the function of genes involved in conversions using the protein nr database at NCBI 19. Although the function of many genes is presently unknown, we identified approximately14 gene families involved in conversions to be related to disease or bacteria resistance. These include genes coding for phospholipase D, cytochrome P450, receptor-like kinase and receptor kinase-like proteins. Some conversions were also found in related Arabidopsis gene families 1320. (See details in Additional File 1). The highest conversion frequency was found in the phospholipase D (AK070203) gene family. Phospholipase D has been identified as an enzyme generating secondary messengers in plants, triggering defense against bacterial attacks 21.

The initial 28,469 full-length rice cDNA sequences were downloaded from RIKEN and FAIS centers 30. We aligned these cDNAs to the Oryza sativa indica genome sequence using BLAT 31. We removed the following sequences: redundant genes, namely those that are the smaller cDNA of two alignments with overlap of at least 100 bp; unlikely protein coding genes with open reading frames less than 100 amino acids; and those sequences with more than 10% transposon-like content identified by using RepeatMasker with RepBase 32. Eventually, 13,089 reliable protein-coding genes were obtained 33 and used as consensus sequences.

Paralogs in the rice genome (Indica 9311) were identified based on similarity with the 13,089 reference protein sequences using BLAST 34. The pipeline used to identify paralogs is similar to the FGF approach 35. The reference protein sequence was defined as the consensus sequence for each family. The members of gene families were defined by alignments with at least 80% similarity over 70% of the corresponding consensus sequence. If the subject sequence exhibited significant similarity to more than one consensus protein, we considered only the highest-scoring match. The family members were aligned together with the consensus protein sequence using GeneWise 36.

Families with more than two genes were defined as multigene families. In the 626 multigene families, we identified 3824 gene pairs; a gene pair is defined as one gene and its closest homolog in the same family. We defined high-similarity (HS) genes as those with a homolog in Arabidopsis and low-similarity (LS) as those without. LS genes and HS genes were identified based on tblastn searches using an E-value threshold of 10^-7as previously reported 16, involving at least 50% of a given Arabidopsis protein or 100 amino acids. The cDNAs were aligned to protein sequences also using GeneWise. Those cDNAs containing multiple stop codons or frame-shift mutations were considered to be pseudogenes.

The aligned sequences within the multigene families were used to detect conversion events using the GENECONV program version 1.81 37. This program detects pairs of sequences that share unusually long stretches of similarity in regions of overall lower similarity 38. The methods used by GENECONV make it difficult to detect conversion events as candidate region lengths approach zero, i.e., for very short sequences. For example, if a conversion event contains only 3 bp of information, we cannot confidently assert the criterion of unusually long stretches of similarity, a signature of gene conversion. Both global and pairwise P-values were calculated based on 1000 permutations of original data and a BLAST-like searching algorithm.

We only used P-values from global fragments, which were multiple comparisons corrected for all possible sequence pairs. The g2 option provided by GENECONV was used to allow some accumulation of mutations in conversion regions following a candidate conversion event. With the g2 option, all multigene families that had global P-values less than 0.05 for inner fragments were considered as statistical significance requirement for conversion events; 215 of the detected pairs had P-values less than 0.01, and 162 had with P-value higher than 0.01 but less than 0.05. The inner fragment indicates a possible gene conversion event between ancestors of the two sequences in the alignment where the outer fragments have diverged.

The lengths of the converted regions were also determined using the multiple-sequence alignment used by GENECONV. We removed the gaps involved in the conversion regions of both paralogs to compensate for multiple-sequence alignment effects thereby improving the length estimate for the conversion regions. The lengths were also checked manually against pairwise alignments using randomly selected pairs without a significant difference with our multiple-sequence alignment analysis.

We compared all paralogs in each gene family. If the paralogs of a family shared more than 95% similarity, we only considered one of them as a conversion partner with other high-scoring alignments designated as members of the gene family. The goal of this step was to eliminate conversion copies resulting from subfamilies with recent duplications.

The similarities between gene pairs were estimated based on their cDNA sequences. Sequence alignments and similarity scores were calculated between all paralogs within each multigene family. The genes were mapped onto the genome sequence using TBLASTN; direction and distance between gene pairs were determined from the physical map. The expected number of genes involved in conversion events on each chromosome was estimated based on the length of chromosomes. In computing sequence similarity, we compared the sequences with and without the identified converted region to exclude the contribution of converted sequence to the overall score.

Ka and Ks values of gene pairs were calculated using the LPB93 method 38. Ka/Ks ratios were compared between gene pairs for gene-conversion partners consisting of 3824 pairs in 626 multigene families. To increase the calculation sensitivity, gene pairs with Ks > 1 were removed because the Ka/Ks calculations by LPB93 may not be reliable for highly diverged gene pairs. Functional information for gene families involved in gene conversion were obtained from NCBI 39.