Lateral transfers of large DNA fragments spread functional genes among grasses

We sequenced and assembled a 0.75-Gb chromosome-level reference genome for a single Australian individual of A. semialata (SI Appendix, Supplementary Methods and Fig. S1). In total, 97.5% of the genome assembly was contained in nine scaffolds, which corresponds to the expected number of chromosomes for this individual (2n = 18) (41). Synteny is well conserved between the genome of A. semialata and those of other Panicoideae grasses, such as the Cenchrinae Setaria italica (42) and, to a lesser extent, the Andropogoneae Sorghum bicolor (43) (SI Appendix, Fig. S2); this is as expected, given high overall synteny among grasses (44, 45).

The genome contains 22,043 high-confidence annotated protein-coding genes, with BLAST matches in both SwissProt and at least one of the two other Panicoideae genomes (S. italica or S. bicolor). The gene density was plotted across the genome in 1-Mb windows. Each chromosome has a region of reduced gene density, and these regions are assumed to correspond to centromeres (SI Appendix, Fig. S3A). All putative centromeres are located roughly in the middle of the chromosome, apart from chromosomes 7 and 9 which appear acrocentric (SI Appendix, Fig. S3A).

We adopted a tiered approach to determine the proportion of the 22,043 high-confidence annotated genes within the genome that were laterally acquired from other plant species (Fig. 1 and SI Appendix, Supplementary Methods). By specifically focusing on plant-to-plant transfers, we limit the risks of contamination and false positives associated with LGT studies involving microorganisms (4, 6). We first performed BLAST searches against angiosperm genomes to determine whether any A. semialata gene is more similar to a nongrass angiosperm than to other grasses. No such evidence for gene transfer involving a nongrass angiosperm donor was found, and we therefore focused our searches on grass-to-grass LGTs. Our strategy used existing and novel genomic resources for 147 grass species in a pipeline combining similarity analyses with phylogenetic validation (Fig. 1). The analytical approach is analogous to previous scans for LGT (e.g., refs. 29 and 32), but extra validation steps were made possible by the availability of additional genomic information (Fig. 1). We first used a read mapping strategy for a selection of species with high-coverage data (n = 20; mean = 42.63 Gb; SD = 5.59 Gb) to identify all high-confidence genes in the A. semialata reference genome with a higher percentage identity to one of 17 potential donors [Themeda triandra sequenced here and 16 species from National Center for Biotechnology Information (NCBI) Sequence Read Archive (SRA); SI Appendix, Supplementary Methods and Dataset S1] than to the three conspecifics or congeners. This initial genome scan identified 1,148 LGT candidates (5.21% of high-confidence protein-coding genes), although these likely contain many false positives, for instance caused by gene losses in the close relatives. We therefore subjected all 1,148 LGT candidates to phylogenetic investigation, using two successive steps, which first identified and subsequently validated well-supported LGT candidates (Fig. 1).

First, sequences from the A. semialata genome were compared with those of 16 completely sequenced grass genomes and 20 grass transcriptomes (SI Appendix, Supplementary Methods and Dataset S1). A coalescence multigene “species” phylogenetic tree was inferred for these 36 taxa and A. semialata using 200 universal single-copy orthologs identified using BUSCO (46). The 37-taxa species tree was generally well resolved, and overall congruent with previous analyses (Fig. 2) (47, 48). While there is some gene tree discordance, particularly for the nodes at the base of the Paniceae, there are well-supported groups spread throughout the phylogenetic tree that are recovered by a majority of the nuclear markers (Fig. 2 and SI Appendix, Table S1). Genes of A. semialata positioned within any of these well-supported clades should therefore be considered strong LGT candidates. For each of these 1,148 LGT candidates, we then inferred a maximum likelihood phylogenetic tree using the same 37 taxa. If there was support (>50% bootstrap replicates) for the LGT candidate being nested within (not just sister to) one of the groups retrieved by the coalescence species tree, it was retained for further analysis. We further retained sequences of A. semialata sister to groups represented by a single sequence, such as Melinidinae, and those outside of the core Panicoideae not assigned to a well-supported clade (i.e., Stipagrostis hirtigluma, Danthonia californica, or Chasmanthium latifolium; Fig. 2) if their combined sister group was as expected based on the species tree (SI Appendix, Supplementary Methods). The phylogenetic analyses and topological filtering retained 55 LGT candidates (Fig. 1 and SI Appendix, Table S1).

Second, we validated these 55 candidates by inferring gene trees with dense phylogenetic sampling using sequence information extracted from 238 genome and transcriptome datasets for 147 grass species (including A. semialata; Fig. 3, SI Appendix, Supplementary Methods, and Dataset S2). To allow comparison with the candidate LGT gene tree, a coalescence species tree was inferred for these 147 species with the same 200 BUSCO markers as used above (SI Appendix, Supplementary Methods and Fig. S4). The topology of the species tree recovered the main taxonomic groups that have been identified in the reduced phylogenetic tree (Fig. 2) and previous published grass phylogenetic trees based on plastids or a few nuclear markers (47, 48). The 55 candidate LGT gene trees were manually inspected to verify that (i) the positioning of A. semialata sequences within a group of distant relatives was well supported (>70% bootstrap values), and (ii) the number of species represented by different paralogs (identified by comparing the gene and species trees) was sufficient to draw conclusions regarding phylogenetic relationships. In total, five genes were discarded because fewer than three species were represented outside of Alloteropsis and the putative donor clades, or in the donor clade. A further 14 genes were discarded because they inferred relationships that strongly differed from the species tree and were deemed phylogenetically unreliable. In two cases, the A. semialata gene was not nested in a distant clade with the denser sampling, and in six phylogenetic trees, the nesting was not supported by bootstrap values above 70%. The remaining 28 candidates were subjected to further validation to rule out alternative scenarios.

We first demonstrated phylogenetic support for the LGT using approximately unbiased (AU) tests, which confirmed that the topology inferring a LGT was significantly better than a topology indicating no LGT in all but two cases (Bonferroni-corrected P value <0.05; SI Appendix, Table S2). Second, phylogenetic biases due to adaptive evolution were ruled out by showing that phylogenetic trees based on third codon positions supported the same groupings for all remaining 26 candidates (Dataset S3). Third, unrecognized paralogies were excluded by demonstrating that genes from S. italica and S. bicolor are generally syntenic with the copy of A. semialata positioned as expected based on the species tree (i.e., native copy), but not with the A. semialata copy in a phylogenetic position suggestive of LGT (Fig. 4 and SI Appendix, Figs. S2 and S5). Fourth, the existence of the 26 remaining candidates was confirmed by independent sequencing runs from multiple samples, ruling out contamination. The 26 candidates passing all of the tests are considered unambiguous LGTs.

For each unambiguous LGT, the putative donor was identified based on the gene trees. Some of the genes could be assigned to precise lineages, such as the Cenchrus genus within Cenchrinae and T. triandra in Andropogoneae (Figs. 3 and 5 and Dataset S2). Because these two groups were suggested as potential donors of previously reported LGTs in Alloteropsis (32), they were purposely included in the genomic datasets. In the case of Themeda, the inclusion of multiple accessions further identifies the genomic origin of the donor (Australian T. triandra as opposed to African populations; Fig. 3). When the donor species is not included in one of the available genomic datasets, it cannot be identified to the species level, but can be assigned to one of the higher-level groups supported by the coalescence species tree (SI Appendix, Fig. S4), with an accuracy that depends on the resolution of the gene tree and the sampling density in the clade containing the donor. It is therefore possible that events associated with the same higher-level group correspond to different species, which would increase the number of donors. Taking into account only the nonoverlapping groups, the phylogenetic positioning of the unambiguous candidates indicates at least nine different donors from the Andropogoneae, Cenchrinae, Melinidinae, and Chloridoideae groups (Fig. 2 and SI Appendix, Fig. S4). These donors diverged from Alloteropsis between 20 and 40 My ago and are separated by multiple speciation events that generated thousands of descending species (32, 47, 48).

The locations of the 26 unambiguous LGTs in the reference genome were determined, and the amount of DNA involved in the transfers was assessed by mapping reads of close relatives and putative donors onto the reference genome (Fig. 6, SI Appendix, Supplementary Methods, and Dataset S4). These 26 primary LGTs are located on 23 different fragments of foreign DNA distributed throughout the reference genome, including multiple chromosomes (Fig. 4). We identified protein-coding genes surrounding these primary LGTs and inferred phylogentic trees for them. In total, 26 of the phylogenetic trees built for the surrounding genes supported an LGT scenario involving the same donor as the adjacent primary candidate (SI Appendix, Table S3 and Dataset S5). For a further seven genes surrounding the primary candidates, homologs were present in an insufficient number of species or the gene was truncated and too short to infer well-supported phylogenetic trees. We therefore used a combination of mapping of reads from relatives of the putative donor (Dataset S4), BLAST searches, and synteny analyses to support their LGT origin (SI Appendix, Table S3), bringing the total of secondary candidates to 33 (SI Appendix, Supplementary Methods). Seven of these 33 candidates had been detected by our pipeline, but discarded in the second filter because of a lack of resolution of the trees or low statistical support. The identity of the donor was refined, taking into account all genes in each fragment.

Besides multiple protein-coding genes, some of the identified LGT fragments contain long stretches of noncoding DNA with identity to the putative donors above 90% (approximate cutoff in identity for read mapping; Fig. 6 and Dataset S4). In the case of recent LGTs from donors closely related to those included in our dataset, the fragments of laterally acquired DNA could be delimited with confidence and were up to 169,972 bp long with >99% identical coding regions (e.g., for the two genes in LGT fragment A), and highly similar intergenic regions (e.g., 97.2% identical over 45.7 kb in fragment A; Fig. 6). Identifying laterally acquired noncoding DNA is difficult for fragments acquired a long time ago or for which close relatives of the donor have not been sampled, but some of them could be much larger (e.g., fragment N in Fig. 6), while others are limited to one protein-coding gene and at most a small amount of flanking DNA (fragment E in Fig. 6).

The number of protein-coding genes contained within the laterally acquired fragments ranged from 1 (13 fragments) to 10 (in fragment C) (Fig. 4). When multiple genes were identified in a laterally acquired fragment, the genomic segment shows evidence of synteny with the genomes of the two other Panicoideae investigated (Fig. 4 and SI Appendix, Fig. S2 and Table S3). We stress, however, that the native and laterally acquired genes of A. semialata are always found in different parts of the genome, with almost all native copies being syntenic with the position of the ortholog in the S. italica and S. bicolor genomes (Fig. 4 and SI Appendix, Figs. S2 and S5). By contrast, the LGT are not syntenic with orthologs from S. italica and S. bicolor, with one exception (fragment O; Fig. 4 and SI Appendix, Figs. S2 and S5).

We assembled a partial genome of T. triandra and identified T. triandra and A. semialata TEs in their respective genome assemblies (SI Appendix, Supplementary Methods). In total, 186,270 TEs were annotated in the A. semialata reference genome, accounting for 51% of the assembly (0.39 Gb; SI Appendix, Table S4). As expected, they appear prevalent near centromeres (SI Appendix, Fig. S3). The TE sequences were then clustered and phylogenetic trees inferred. Combined with coverage analyses, these led to the identification of 92 TEs acquired from T. triandra in the A. semialata genome (SI Appendix, Supplementary Methods). Several of these TEs are located within the two fragments acquired from T. triandra containing protein-coding genes and potentially have daughter copies located in other genomic locations (SI Appendix, Fig. S3).

Using resequencing (∼10× coverage) and high-coverage (∼40× coverage) data for an additional 24 Alloteropsis populations representing the genetic diversity in this species (SI Appendix, Fig. S4), we were able to establish the distribution of the LGTs across the Alloteropsis phylogenetic tree, thereby estimating the relative timing of the transfers (SI Appendix, Supplementary Methods). Fragment H was unique to the Australian reference genome, and four other fragments (A–D) are restricted to the other Australian accession (AUS2) as well as the accession from the Philippines (PHI1), indicating recent acquisitions. The other LGTs were also observed in a number of individuals from A. semialata and/or the sister species Alloteropsis angusta (Fig. 4). Some are present in most A. semialata accessions (T, I, F, and G), suggesting an early acquisition followed by retention of the genes. Others are distributed among more distantly related accessions, suggesting early acquisitions followed by losses or introgression among A. semialata populations after the transfer (Fig. 4).

RNA-Seq data for 16 populations of A. semialata and one of the sister species A. angusta were mapped to the coding sequences from the A. semialata reference genome, and the expression levels of LGTs and their native orthologs were estimated in leaf and root samples (SI Appendix, Supplementary Methods). Over 59% (35 out of 59) of the primary and secondary LGTs were expressed [>1 reads per kilobase of transcript per million mapped reads (rpkm)] in at least one population (Fig. 4 and Dataset S1). The mean expression level of 12 laterally acquired genes was higher than their respective native homolog in at least one A. semialata population (Fig. 4 and Dataset S1), and in one example the expression of the native ortholog seems to have been replaced by the LGT (ASEM_AUS1_12633; Dataset S1). The functions of the laterally acquired genes include those known to be involved in C₄ photosynthesis as well as loci associated with disease resistance and abiotic stress tolerance (SI Appendix, Table S5).

Our phylogenetic trees suggest LGTs involving recipients other than A. semialata. Indeed, the visual inspection of the gene trees for primary LGTs including the 147 species identified 10 genes with bootstrap support for positions suggestive of LGT in grasses other than the reference genome (SI Appendix, Table S6 and Dataset S2). For three of them, the lack of sequencing replicates for these samples generated in other projects meant that contaminations could not be ruled out (SI Appendix, Table S6). However, seven genes from individuals other than the reference genome were statistically supported in a position suggestive of LGT by AU tests (Bonferroni-corrected P value <0.05; SI Appendix, Table S7), and this conclusion was supported by replicate datasets (SI Appendix, Table S6). This included two genes encoding a C₄ enzyme that had been previously identified in some non-Australian populations of A. semialata (Dataset S2) (32, 36). In addition we identified LGTs in five other Panicoideae genomes (SI Appendix, Table S6).

For two genes containing non-A. semialata LGTs we had sufficient genome data for the recipient species to assemble full-length gene sequences. These data were supplemented with similar full-length sequences from published genomes and the NCBI nucleotide database to infer gene trees using intron and exon sequences. This confirmed that the gene encoding phosphoenolpyruvate carbokykinase (PCK; ASEM_AUS1_17510 on fragment L), an enzyme involved in some subtypes of C₄ photosynthesis (49), was laterally acquired by A. semialata, Echinochloa, and Cymbopogon (Fig. 5). The three LGTs of PCK, supported by multiple datasets, did not form a monophyletic group, suggesting they result from independent transfer events (Fig. 5). Transcriptome datasets for these species indicate that both Cymbopogon and Echinochloa express the LGTs in their leaves, where they likely play a role in their C₄ photosynthetic pathway (50, 51). Despite the smaller sample size, the phylogenetic tree inferred from sequences homologous to ASEM_AUS1_20550 similarly confirmed the LGT scenario for this gene observed in Alloteropsis cimicina (SI Appendix, Fig. S6). For ASEM_AUS1_20550 the donor for the LGT detected in the reference genome belonged to the Andropogoneae, whereas the gene detected in A. cimicina was acquired from a Cenchrinae (SI Appendix, Fig. S6). The native copy of A. cimicina was also detected, and transcriptome data show that the LGT is expressed at a higher level than the native copy.

Using a combination of stringent phylogenetic and genomic analyses (Fig. 1), we identify 59 genes in the genome of the grass A. semialata that were laterally acquired from other grasses. Our pipeline was designed to rule out the three main alternative explanations to LGT, namely: (i) unrecognized parology by comparing synteny with other grass genomes, (ii) contamination by having independent sequencing supporting the existence of the genes, and (iii) convergent evolution by using different data partitions. In addition, long reads spanning the laterally acquired and native DNA prove that the fragments are integrated in the genome of A. semialata (SI Appendix, Fig. S7). The LGTs are moreover supported a posteriori by our genomic analyses, which show that half of the primary candidates are flanked by coding, and in some cases noncoding, regions with a high similarity to the same putative donors (Fig. 6 and Dataset S4). The 10 fragments that contain multiple protein-coding genes represent the most unequivocal cases of LGTs and demonstrate that large stretches of DNA containing numerous genes can be transferred among distantly related grass species.

The number of LGTs reported here is likely an underestimate. For example, LGTs that lack phylogenetic informativeness because the genes are too short, or are not present in a sufficient number of grass species, would be excluded by our analysis. More importantly, we focused on gene transfers among distant relatives to differentiate LGTs from other processes that can create discordance between gene and species trees, such as incomplete lineage sorting and hybridization (Fig. 2). Similarly, we focused on relatively recent LGTs that have accumulated during the diversification of the Alloteropsis genus, because ancient LGT would alter deep branching patterns that can be detected only when the donor and recipient are extremely distant [e.g., ferns and mosses (33)]. As a direct consequence of the methodology, all transfers involving donors that are part of poorly resolved clades within the Paniceae tribe (Fig. 2) or transfers that happened before the diversification of Alloteropsis would remain undetected. In addition, our power to detect LGTs depends directly on the availability of genomic data for close relatives of the donor, particularly when it comes to accurately determining the size of the acquired DNA fragment (Fig. 6). Thus, the 59 LGTs reported here could be only a subset of those existing in the genome of A. semialata.

In total, the detected LGTs belong to 23 genomic DNA fragments (Fig. 4). These fragments were laterally acquired from at least nine different grass donors, although the number might be higher if grasses from the same lineage independently provided LGTs (Fig. 4). Using genomic datasets from multiple Alloteropsis populations allowed us to establish the distribution of each of these fragments within the species. Inferring the presence of a gene from resequencing data can be problematic if the gene is truncated, or located in regions of the genome with reduced sequencing depth. Based on our estimates, fragment H is unique to the reference genome and likely represents the most recent acquisition. Several fragments (A–D) were likely acquired around the time A. semialata colonized Australia, as they are restricted to Australian and Filipino accessions, with the latter probably a result of recent admixture from Australia (Fig. 4). Other LGT fragments are shared by a majority of A. semialata accessions and were likely acquired near the origin of this species ∼2 My ago (e.g., fragments T and F; Fig. 4) (36, 41). In some cases, the patchy distribution of LGT fragments could be due to secondary losses after ancient acquisitions, as supported by observed genetic variation among the different A. semialata populations that has likely accumulated since the LGT was acquired (e.g., fragment E; Dataset S2). In other cases low levels of genetic variation among accessions of A. semialata and A. angusta suggest that the patchy distribution results from more recent acquisition followed by introgression into different populations (e.g., fragments K and N; Figs. 4 and 6 and Datasets S2 and S5) (52). Overall, the evidence of fragments being acquired at different points suggests that the diversification of Alloteropsis has been punctuated by repeated bouts of LGT (Fig. 4A).

The 23 laterally acquired fragments are widely distributed across the genome of A. semialata (Fig. 4C). While divergence of noncoding DNA hampers a precise delimitation of the acquired fragments in more ancient LGTs, or those for which genome data of a close relative of the donor are missing (e.g., fragments E and N in Fig. 6), recent LGTs with a sampled donor can be shown to be at least 170 kb long and are composed of genic as well as noncoding regions (e.g., fragments A and C in Fig. 6). In the case of two fragments acquired from T. triandra (A and B), TE phylogenetic trees and inferences of their recent activity show that some TEs, which were acquired as part of the large DNA fragments, have subsequently transposed to new regions (SI Appendix, Fig. S3B). The laterally acquired fragments also have TEs specific to Alloteropsis, which were likely inserted after the acquisition, highlighting rearrangements between the native and foreign parts of the genome. Other laterally acquired TEs were detected around the genome outside of the large block of DNA and might represent TEs that escaped from large DNA blocks or elements that were acquired on their own.

Genomic rearrangements that happened after the transfer are also visible in the gene content of the laterally acquired fragments. Erosion is evidenced by gene loss in some accessions (e.g., part of fragment C in one of the Australian samples; Figs. 4 and 6), as well as pseudogenizing mutations in others (SI Appendix, Fig. S8). However, 59% of the laterally acquired genes are expressed in at least one population under the conditions evaluated (Fig. 4). In all cases, the donor and the recipient evolved independently for tens of millions of years before the transfers, leaving ample time for adaptive evolution and functional diversification. The genetic exchanges therefore potentially brought in genes for novel attributes. Eight of the fragments contain at least one gene that is either novel, has replaced the function of the native copy that became a pseudogene, or is expressed at a higher level than the native copy (Fig. 4 and Dataset S1). The functional LGTs have therefore added novelty to the genetic apparatus of Alloteropsis, including a variety of disease resistance and abiotic stress response loci (SI Appendix, Table S5), in addition to the previously reported photosynthetic genes (32). The advantage of each of the laterally acquired genes is yet to be determined by targeted functional studies, but selection for the novel functions is likely responsible for the retention of LGTs through time (Fig. 4). Genes that underwent adaptive shifts in some subgroups of grasses might be especially likely to be retained after transfer to other groups, potentially contributing to the three independent LGTs for the C₄ enzyme PCK (Fig. 5).

The mechanisms underlying the reported grass-to-grass LGTs remain elusive and might vary between events. The nonsyntenic localization of LGTs and their coexistence with native copies argues against classical introgression involving sexual reproduction and chromosomal recombination. The movement of genes could have occurred among genomes following the transient cohabitation of chromosomes from different species within the same nucleus, for example in allopolyploids that can subsequently backcross with diploids (53). In grasses, the growth of pollen tubes on the stigma of distant relatives can be used experimentally to trigger embryo development with only occasional transfer of paternal DNA (54, 55), providing a more likely mechanism for chromosomal exchanges. Occasional interspecific cell-to-cell contact could also occur through root-to-root interactions between grasses growing in multispecies clumps as observed in savannas or among pollen tubes of different species growing on the same stigma. Cell-to-cell contacts are known to allow movement of DNA across distinct nuclei in grafts (56) or host/parasite interactions (22–30). Since A. semialata can propagate vegetatively via rhizomes, both transfers into the seeds and into parts of the root system would allow the long-term integration of LGTs into the germline. Independently of the exact mechanism, our genomic investigations show that genes are recurrently passed among distant species of grasses, and provide a novel source of genetic variation for selection to act upon.

While we screened for nonangiosperm donors, all of the detected LGTs came from grasses. This bias in the donor identity might represent some genomic or physical incompatibilities (e.g., lack of wind dispersal of pollen or different root architecture) that limited exchanges with nongrass angiosperms. All grass donors were from the subfamily Panicoideae, with one exception. Within Panicoideae, two groups (Andropogoneae and Cenchrinae) have contributed the vast majority of fragments, while other groups with similar genomic resources (e.g., Panicinae and Paspaleae; Fig. 2 and SI Appendix, Table S1) were not involved in any of the detected LGTs. This bias can largely be explained by geographic patterns, as both Andropogoneae and Cenchrinae species frequently co-occur with A. semialata in large populations throughout Africa, Asia, and Australia, whereas in contrast, Paspaleae species are mainly found in South America where A. semialata does not occur. Whether other groups (e.g., Panicinae) had opportunities to exchange genes with Alloteropsis remains to be formally assessed and may ultimately contribute to identifying the properties that promote LGT.