Main

Modelling ancestral genomes at internal nodes of a species phylogeny is a powerful tool to trace the genetic events that shaped genome evolution. This is often done via ancestral gene repertoire reconstructions, which provide gene lists as proxies for ancestral genomes1,2. However, these methods do not account for gene contiguity across genomes and thus cannot capture patterns of genomic rearrangements. Ancestral gene order inference methods have emerged to fill this gap, helping detect rearrangements associated with speciation or identify functionally associated genes residing in conserved genomic neighbourhoods3,4,5,6,7,8. A major milestone was achieved by Muffato et al.3 with their ‘Algorithm for Gene Order Reconstruction in Ancestors’ (AGORA) method and the reconstruction of gene orders for 624 ancestral genomes across five independently processed clades (200 vertebrates, 117 non-vertebrate Metazoa, 99 plants, 478 fungi and 136 protists). However, state-of-the-art methods such as AGORA rely on computationally expensive reconciled gene trees and pairwise gene order comparisons and typically struggle to process large phylogenies with more than hundreds of genomes9. This scalability limitation affects both the accuracy and evolutionary scope of analyses, as including more extant genomes permits a higher resolution in reconstructing ancestral gene order and tracing their evolutionary histories further back in time. This limitation is highlighted by large-scale sequencing efforts—such as the Earth BioGenome Project10, which aims to deliver annotated genomes for ~9,000 eukaryotic taxonomic families within the next decade—as they are rapidly outpacing the development of methods capable of harnessing the wealth of data they generate.

Here, we introduce edgeHOG, a method for reconstructing ancestral gene orders across large phylogenies while maintaining and at times even exceeding the levels of resolution and accuracy set by AGORA. Unlike approaches relying on computationally intensive reconciled gene trees, edgeHOG uses hierarchical orthologous groups (HOGs) (Fig. 1)—which are faster to infer, computable on arbitrary datasets and widely available through databases such as OMA11, Hieranoid12 or EggNOG13—to anchor comparisons of gene adjacencies between genomes and propagate gene order predictions along the species tree. A proof of concept showed that HOGs are reliable for ancestral gene order inference: when using OMA-derived HOGs, AGORA reconstructed a Boreoeutherian ancestral genome similar to that inferred with Ensembl Compara’s reconciled gene trees3, available only for a restricted number of eukaryotic clades due to computational limitations14. Applying edgeHOG to the 2,845 extant genomes from the OMA database11, we reconstructed ancestral gene orders for 1,133 fully browsable ancestral genomes spanning all three domains of life. This revealed a significant association between gene order and function in the last eukaryotic common ancestor (LECA), as well as intriguing patterns of chromosomal evolution, such as conserved histone gene clusters in metazoans and younger gene adjacencies on sex chromosomes across various species. EdgeHOG is available as an open-source standalone tool to process arbitrary datasets (https://github.com/DessimozLab/edgehog). In combination with the recently released FastOMA15 software, it enables the reconstruction of ancestral gene orders for thousands of genomes (including entirely eukaryotic datasets) within a few days.

Fig. 1: HOGs explicitly model gene lineage, thus ancestral gene contents.
figure 1

Left: the species tree displays the evolutionary relationships between four extant species (S1, S2, S3 and S4). Its internal nodes are proxies for ancestors (A1,2, A1,2,3 and A1,2,3,4). The central gene tree illustrates an example of inferred evolutionary history for a gene family (coined X) present in all species. The HOG object of the family X can be computed from the reconciled gene tree of X as in EggNOG or using fast graph-based methods as in OMA. Once pyHAM processes the HOG of family X, the graph on the right is created: each node corresponds to an extant or ancestral gene and each edge links a descendant gene to its parent gene in its most direct ancestor. When paralogues are connected to the same parental gene, it models an event of gene duplication (happening here between ancestor A1,2,3,4 and descendant A1,2,3). The collection of gene nodes of all HOGs’ graphs at a given ancestral level eventually serves as proxies for its ancestral gene content.

Full size image

Results

We describe edgeHOG’s algorithmic principles, validation on simulated and empirical datasets and scalability and demonstrate how it enables biological insight into chromosomal evolution, including functional clustering in the LECA and patterns of gene adjacency retention across extant species and sex chromosomes.

Algorithm overview

EdgeHOG requires a rooted species tree, the coordinates of the genes on the chromosomes or contigs (as GFF files) and the HOGs of the genomes, which can be downloaded in OrthoXML format from various orthology databases or computed from proteomes using software such as OMA Standalone16 or FastOMA15.

Ancestral gene repertoire reconstruction

HOGs can be thought of as ancestral genes, as they encompass orthologues and paralogues descending from a common ancestral gene at a specific taxonomic level (that is, internal node of the species tree)17,18,19. HOGs at a lower taxonomic level are nested within HOGs of a higher level, thereby modelling the lineage of genes, assuming strict vertical inheritance along the species tree. When distinct HOGs defined at the same taxonomic level are nested in a higher level HOG, they can be thought of as ancestral in-paralogues (Fig. 1).

Bottom-up propagation of gene adjacencies

Using descendant-to-parent gene links within gene lineages, observed or predicted adjacencies between two genes at a given phylogenetic level are mapped to their corresponding parental genes in the upper taxonomic level. If a gene has no parent but its flanking neighbours have one, an edge is created between these two neighbours and propagated to the upper taxonomic level, thereby modelling gene emergence and insertion events between two older genes. This process ultimately constructs a network at each level of the phylogeny, where nodes represent ancestral genes and edges link genes inferred to be of closest proximity. The weight of each edge indicates the number of propagations from descendant extant genomes (Fig. 2a).

Fig. 2: EdgeHOG’s algorithm.
figure 2

a, The bottom-up phase. Traversing the guide species tree from leaves to root, an adjacency between two genes is propagated to a direct ancestor as long as it is inferred to have the two ancestral genes. The inferred gene gain (in grey) is accounted for by propagating an adjacency to the parental level only between the two flanking neighbours. All edges propagated to the parental level are summarized in a ‘propagated graph’, and the propagated form of each real edge is stored. Note that if duplicated genes or edges exist as distinct entities after the duplication event, they merge into a single gene or edge at the point of duplication and before it. b, The top-down phase. Traversing the tree from root to leaves, any adjacency not supported by parsimony is removed, that is, essentially any edge supported by only one child and not by the parent (hence wrongly propagated before the last common ancestor in which the edge emerged). c, The linearization phase. The linearization step flags conflicting genes (those having more than two neighbours) and removes edges until the number of neighbours of a conflicting gene is no more than two. Bottom: the order in which conflicting nodes are resolved and the hierarchy of neighbours for a conflicting node are explained. The linear path weight of a neighbour is the sum of weights of all edges in the path passing by the neighbour and ending at the first node encountered with less or more than two neighbours. For each ancestor, the linearized graph constitutes edgeHOG’s main output.

Full size image

Top-down removal of edges not explained by parsimony

A drawback of the bottom-up phase is that when a novel adjacency between two old genes has arisen through genomic rearrangement, propagating the adjacency to the ancestral level is essentially a mistake. Therefore, the top-down phase removes any edge propagated in ancestral synteny networks that is not supported by parsimony. This means an edge is removed if it was propagated before the last common ancestor in which the adjacency is inferred to have emerged (see Fig. 2b for details). Because the criterion of edge removal does not consider edge weights, the top-down phase is not affected by any potential tree imbalance.

Linearization of synteny networks

After edge removals, some ancestral genes may still have more than two neighbours due to orthology/paralogy misinferences, incorrect species tree topologies or convergent/reticulate evolution of gene adjacencies. The linearization step ‘resolves’ conflicting genes having more than two neighbours by selecting their two most likely flanking genes—those of maximal support (details in Fig. 2c). This results in linear ancestral contigs at each phylogenetic level, which collectively form an ancestral genome. Of note, the heuristics for determining the final linearized genome are applied independently at each internal node of the species tree, without influence from linearization choices made at other nodes.

Beyond reconstructing ancestral adjacencies, edgeHOG offers several additional features, such as predicting the orientation of genes in ancestral adjacencies, dating the age of gene adjacencies in both ancestral and extant genomes (using the information of the last common ancestor in which the adjacency is inferred to have emerged) or performing a phylostratigraphy of gene adjacency retentions, gains, losses or duplications at each internal node of the species tree.

Extensive benchmarking on simulated and empirical data

To evaluate edgeHOG’s performance in ancestral gene order reconstruction, we benchmarked it against AGORA using various simulated and real datasets (see the Supplementary Information for detailed results). In a simulation of 100 ancestral genomes, edgeHOG showed high accuracy, achieving a harmonic mean precision of 98.9% (percentage of predicted adjacencies being correct) and a recall of 96.8% (percentage of real adjacencies being predicted). This outperformed AGORA, which showed 96.0% precision and 94.9% recall (Fig. 3a). Moreover, edgeHOG’s performance was relatively consistent across all levels of the phylogenetic tree, while AGORA’s accuracy slightly declined for more recent ancestors due to a bias in its weighting strategy (see the Supplementary Information for details). In a more challenging simulation with particularly high rates of genomic rearrangement, edgeHOG outperformed AGORA more markedly, achieving a harmonic mean precision of 40.3% and recall of 18.8%, compared with AGORA’s 13.9% precision and 3.8% recall (Extended Data Fig. 1 and Supplementary Information).

Fig. 3: Benchmarks.
figure 3

a, The simulated genome evolution benchmark. Each dot represents one of the 99 ancestral levels in a species tree with 100 extant genomes. The x axis gives the relative evolutionary divergence44 from root of an ancestral node (0 for the root, near to 1 close to the leaves). The top row’s y axis gives the precision of each algorithmic step of edgeHOG and of Agora at each ancestral level, measuring the proportion of predicted edges that are true edges in the simulated ancestral genome. The bottom row’s y axis shows recall, that is, the proportion of true edges predicted by each method. b, The YGOB benchmark. The species tree depicts the phylogeny of the 20 yeast genomes in YGOB. The star indicates a whole genome duplication (WGD) event in the last common ancestor of 12 yeasts. The pink circle indicates the tree root, where YGOB-curated ancestral gene order is available. This curated order is compared with predictions by edgeHOG and AGORA, allowing evaluation of precision and recall. c, The masked extant gene orders benchmark. The species tree on the left shows the phylogeny of 50 vertebrate genomes from the OMA database, sampled to represent clade diversity. The ten coloured extant genomes correspond to those whose gene order is masked (and is to be inferred). The ten coloured internal levels correspond to the most direct ancestor of each masked species. Scatterplots compare edgeHOG and AGORA in inferring masked edges using a projection of each edge between two ancestral genes at the parental level onto their corresponding single-copy descendant genes in the extant masked species. Note the near 0% recall for Salmo trutta due to a whole gene duplication on its terminal branch, leaving no information to ‘phase’ the duplicated genes. Panel c silhouettes adapted from PhyloPic under Creative Commons licenses: Anguilla anguilla, Ingo Braasch (CC0 1.0); Salmo trutta, Carlos Cano-Barbacil (CC0 1.0); Cirripectes variolosus, Mykle Hoban (CC BY-SA 3.0); Latimeria chalumnae, Chuanixn Yu (CC0 1.0); Podarcis muralis, Titouan Montessuit (Attribution 4.0 International); Crocodylus, Becky Barnes (CC0 1.0); Nestor notabilis, Matías Muñoz (CC0 1.0); Tachyglossus aculeatus, Becky Barnes (PDM 1.0); Manis culionensis, Steven Traver (CC0 1.0); Tupaia javanica, Margot Michaud (CC0 1.0).

Full size image

To benchmark on empirical data, we took advantage of the expert and thorough work performed by the Yeast Gene Order Browser (YGOB) to manually curate the likely gene order in the last common ancestor of a clade of 20 yeast species20. Comparing predicted gene adjacencies with that annotated by YGOB, we found that edgeHOG’s precision and recall reached 91.7% and 77.5%, respectively, while AGORA’s reached 90.6% and 79.2% (Fig. 3b). Notably, both tools correctly predicted gene orientations over 99% of the time.

Since YGOB may favour the method that has the most in common with its inference process, we designed an additional empirical benchmark in which we masked the gene order of ten Vertebrata species, that is, treating each gene as if it were on its own contig, and inferred them using the gene orders of 40 other Vertebrata species (see Extended Data Fig. 2 for the sampling of the 50 genomes). Specifically, we inferred the gene adjacencies of each masked genome by mapping the predicted adjacencies of its most direct ancestor onto the corresponding descendant genes, with the rationale that the number of accurate predictions projected from ancestral edges can serve as a proxy for the quality of the ancestral gene order inference. EdgeHOG again showed slightly better performance, with an average precision and recall improvement of +1.5% and +0.4%, respectively, over AGORA (Fig. 3c and see Extended Data Fig. 3 for a detailed comparison of performance across species in relation to characteristics of each masked genome (for example, contiguity level) and its corresponding ancestor (for example, phylogenetic depth)). We also found that increasing the number of genomes to 156 instead of 50 improved recall (+2.1%) (Extended Data Fig. 2) but slightly reduced precision (−0.8%) (Extended Data Fig. 4), while increasing the number of gene adjacencies in ancestral genomes (Extended Data Fig. 5). For instance, edgeHOG inferred 11,051 adjacencies for the last common ancestor of Gnathostomata with 156 species, versus 8,193 with 50 species (Extended Data Fig. 5). This demonstrates that comparing more genomes and thus handling large datasets has the potential to improve the resolution of ancestral genomes.

Scalability

To evaluate how efficiently both tools handle large datasets, we measured their runtime (Fig. 4) and RAM usage (Extended Data Fig. 6) across eukaryotic phylogenies of increasing size. RAM usage scaled similarly for both tools, though AGORA uses ~29% less memory on average. However, differences in runtime were more pronounced. EdgeHOG’s runtime scaled linearly, benefiting from its tree traversal-based edge handling, whereas AGORA’s runtime inflated with larger phylogenies, due to its reliance on gene order comparisons that increase quadratically. In practical terms, edgeHOG took 1 h and 20 min to infer ancestral gene orders at each internal node of a phylogeny of 791 Eukaryotic genomes, while AGORA took 43 h and 19 min for the same task. As a result, edgeHOG currently stands as the only scalable software solution capable of reconstructing ancestral gene orders for datasets comprising thousands of eukaryotic genomes. For instance, a linear model fitted to the runtime data estimates that edgeHOG would process 10,000 eukaryotic genomes in approximately 17 h and 30 min.

Fig. 4: Runtime on one processor as a function of the size of the input phylogeny.
figure 4

Each dot corresponds to a clade of eukaryotic genomes from the OMA database.

Full size image

Ancestral genome orders across the three domains of life

EdgeHOG’s scalability made it possible for us to process all 2,845 extant genomes in the OMA database (1,965 bacteria, 173 archaea and 707 eukaryotes), in under 3 h on a single processor. To our knowledge, this represents the largest single-run inference of ancestral gene orders using genomes across all three domains of life. The resulting collection of 1,133 ancestral genomes represents a unique resource to study ancestral synteny across key clades of the tree of life. Details for browsing this resource are outlined in the latest OMA paper11 and a summary of the number of genes, gene adjacencies, contigs and contiguity levels for all extant and ancestral genomes of the OMA database is in Supplementary Table 1. In these resources, ancestral genomes include only genes on reconstructed contigs, excluding singleton genes (Discussion).

Reconstruction of the LECA

The unprecedented phylogenetic depth of the analysis enabled us to reconstruct ancestral gene order in the LECA (Fig. 5; see Extended Data Fig. 7a,b for the guide species tree). EdgeHOG inferred 1,009 ancestral contigs in LECA, with lengths ranging from 2 to 19 genes (Fig. 5a). The functional similarity among genes within contigs supports the inference, consistent with the link between gene linkage and functional association21 (Supplementary Table 2). The Gene Ontology (GO) enrichment analysis of contigs (GO terms of genes of a contig as foreground, GO terms of genes of the ancestral genome as background) confirms this trend by highlighting 194 contigs enriched in ancestral genes contributing to the same biological process (Fisher’s exact test, Bonferroni-corrected P value <0.05). As a sanity check, we repeated the analysis after randomizing gene order (preserving contig size), and the number of functionally enriched contigs was indeed much lower (mean of 14.6 and standard deviation of 7.9) (Fig. 5b). The tendency of neighbouring genes to be functionally related is unlikely biased by an over-representation of a gene family in multiple copies within contigs, as enriched contigs do not contain more ancestral in-paralogues than non-enriched ones (Mann–Whitney test, P = 0.99) (Extended Data Fig. 7c). Remarkably, reconstructed contigs contain genes that capture core pathways, with primary metabolism, translation, DNA repair and stress responses being the most represented categories of functions (Fig. 5a and Supplementary Table 2).

Fig. 5: Functional analysis of the 1,009 ancestral contiguous regions inferred by edgeHOG in the LECA.
figure 5

a, The landscape of ancestral contigs. Each dot represents an ancestral gene in a reconstructed contiguous region. The edges link the genes inferred to be mutually closest among reconstructible genes. Edge thickness is proportional to the square root of its weight (1–1,746), representing how often the edge was propagated from descendant genomes and thus the conservation level of genomic neighbourhoods. Within contigs, the black dots indicate genes whose descendant genes’ GO terms contribute to enriched pathways detected in the contig. A contig with only black dots means all its genes participate in the same biological pathway. The blue dots indicate gene families with uncharacterized functions. Enriched pathways are labelled only for contigs at the figure’s right border for clarity. AA, amino acids. b, The number of inferred contigs with enriched pathways in LECA (first column). The graph was randomized 100 times by gene swapping, and the number of contigs with GO enrichment in each randomized graph is shown (second column). c, The proportion of LECA adjacencies conserved in extant eukaryotic genomes. Each dot represents an extant genome from the OMA database. The x axis groups genomes as animals (n = 275; Metazoa), fungi (n = 224), plants (n = 86; Viridiplantae or Rhodophyta) or protists (n = 87; other eukaryotes). The y axis shows the proportion of LECA’s adjacencies conserved in each extant genome. The genomes are coloured by gene content completeness, assessed by OMArk. The boxplots show medians, quartiles and whiskers extending to 1.5× interquartile range beyond hinges.

Full size image

We computed for each LECA ancestral gene the fraction of descendants found on extant mitochondrial or chloroplast contigs (Supplementary Table 3). This showed that the long contig annotated as ‘ATP synthesis’ in Fig. 5a capture ancestral mitochondrion features as it contains consecutive gene adjacencies of the respiratory chain pathway (Supplementary Tables 2 and 3). However, a few contigs are erroneous based on our knowledge of eukaryotic evolution, such as contigs containing genes involved in photosynthesis, as chloroplasts emerged from the endosymbiosis of cyanobacteria after LECA in plants, let alone the cases of secondary endosymbiosis. For instance, we found that 17 gene adjacencies in LECA were probably induced by the cyanobacterial ancestry of choloroplasts, as these edges were shared only between Cyanobacteria and Chloroplast-containing eukaryotic lineages (Supplementary Table 3). This highlights potential for future algorithmic improvements, notably accounting for reticulated evolution (Discussion).

To assess LECA’s gene adjacency conservation in extant eukaryotes, we calculated the percentage of LECA edges retained per species (Supplementary Table 4, sheet 1) and the proportion of species retaining each adjacency (Supplementary Table 4, sheet 2). Extant genomes preserve 0–7.7% (1.73% average) of LECA’s adjacencies (Fig. 5c), with the histone 2A–2B adjacency being the most conserved (retained in 66% of species).

Dating gene adjacencies with edgeHOG

One novel feature of edgeHOG is the ability to assess the age of gene adjacencies of extant and ancestral genomes, that is, indicating the last common ancestor in which each adjacency is inferred to have emerged. It enables identification of conserved and divergent patterns in chromosomal organization over time. We inferred the last common ancestor (clade of origin) of all adjacencies for all eukaryotic genomes in our dataset and dated these adjacencies based on the estimated age of their common ancestor using the TimeTree22 resource (Fig. 6 and Supplementary Data 1). We observed remarkable patterns of chromosomal evolution in metazoan genomes.

Fig. 6: Estimated age of gene adjacencies within chromosomes of Homo sapiens, Gallus gallus, Danio rerio and Papilio machaon.
figure 6

Each subfigure corresponds to the karyotype of a given species. The y axis gives the name of the chromosome while the x axis the number of gene adjacencies. Each chromosome is represented as a stack of gene adjacencies, each coloured according to its estimated age in millions of years. The dark blue arrows indicate blocks of adjacencies dated to the eukaryotic common ancestor and enriched in adjacencies between histone genes. Squared in blue in the subfigure of G. gallus is a zoomed view on the histone cluster in chromosome 1: older adjacencies (dated to the eukaryotic common ancestor) are between different histone subunits while more recent adjacencies are between paralogous subunits (indicated by black triangles). The light teal arrows indicate sexual chromosomes, essentially composed of very recent adjacencies, particularly the heteromorphic chromosome (for example, Y human chromosome).

Full size image

First, a common pattern in metazoan genomes is the presence of sometimes large synteny blocks in which most adjacencies date from around 1.5 billion years (Fig. 6, blue arrows). The synteny blocks mainly comprise genes of the four subunits of the histone octamers (H2A, H2B, H3 and H4) and histone linkers (H1/H5). While paralogue adjacencies (for example, H3–H3) appear more recent, adjacencies between different histone gene families (hereafter referred to as ‘histone adjacencies’) are dated back to LECA (Fig. 6, Gallus gallus). Essentially, edgeHOG dates adjacencies between any representative of distinct gene families (HOGs) to the first occurrence of an adjacency between their common ancestors. Hence, though old adjacencies might be in multiple copies resulting from more recent tandem duplication, they are estimated as descending from ancestral single-copy adjacencies in LECA (contig with the strongest edge supports annotated as ‘chromatin organization’ in Fig. 5a). Histone adjacencies are common across eukaryotes, but metazoans are the only species exhibiting such clusters of histone gene adjacencies containing several copies of each subunit (Extended Data Fig. 8c) and have a higher proportion of histone adjacencies than the other clades (Extended Data Fig. 8a). This may be in part explained by metazoans having, overall, more copies of histone genes than most other eukaryotes, although this relationship is not observed in plants despite some of them having many histone copies as well (Extended Data Fig. 8b). Cluster number and size vary by species—from 12 clusters averaging 14.75 adjacencies in Bufo bufo to one cluster of 109 adjacencies in Drosophila melanogaster (Extended Data Fig. 8). Overall, our results highlight that the very old colocalization of histone subunits on the same contig in LECA (Fig. 5a) adopt a specific organization in animals where they still colocalize in the same locus but in many copies of each subunit, probably as a result of more recent tandem duplications.

Another notable pattern in adjacency ages involves sex chromosomes (teal arrows in Fig. 6). Heteromorphic sex chromosomes are pairs of homologous chromosomes that are morphologically distinct from one another, with one of them carrying a sex determination locus. They are traditionally called X and Y in species where males are heterogametic (XY) and females homogametic (XX) and Z and W when the opposite occurs (ZZ males and ZW females). These systems have been independently acquired multiple times23. In our dataset, heteromorphic sex chromosomes stand out as having younger adjacencies than other chromosomes (Mann–Whitney U test, adjacency ages on sex chromosomes versus other chromosomes in each species, alternative hypothesis: sex chromosome adjacencies are younger). As controls, we performed a similar analysis using each autosome as the focus instead. All comparison results (differences and test statistics) are given in Supplementary Table 5. We confirmed a significant trend regarding the X/Y system: both X and Y chromosomes had significantly younger adjacencies than the rest of the genome in all of the tested mammals and Diptera (chromosome X: n = 27, 23 mammals and 4 Diptera; chromosome Y: n = 12, 11 mammals and 1 Diptera), except Caenorhabditis elegans, where no differences could be detected between the X chromosome and the autosomes. In W/Z systems, the W chromosome had significantly younger adjacencies than the rest of the genomes in all considered species (n = 5: 3 birds, 1 Lepidoptera and 1 Neopterygii). However, no clear pattern emerged for the Z chromosome (n = 8: 6 birds, 1 Lepidoptera and 1 Neopterygii), as it harboured significantly younger adjacencies only in two birds. In addition, younger adjacencies appeared on the right arm of chromosome 4 in zebrafish, where sex determination occurs, but not in other fish with homomorphic sex chromosomes. However, younger adjacencies were not unique to heterochromosomes; one or more autosomes in each species also showed significantly younger adjacencies (Supplementary Table 5), such as bird microchromosomes, primate chromosome 19 and Drosophila chromosome 4.

Discussion

EdgeHOG unlocks key applications in comparative genomics, including tracking genomic rearrangements along a species phylogeny, identifying conserved gene clusters and improving genome assembly using gene order knowledge from other species (an example showing how ancestral gene adjacencies can expand contigs in a fragmented genome is available as a Jupyter notebook in the Figshare repository).

The ability to infer gene order conservation at large scale also facilitates the comparative genomics of fast-evolving intergenic regions, potentially identifying orthologous regulatory elements using syntenic genes to bracket non-coding regions. Likewise, it can help detect highly divergent orthologs using syntenic context, holding potential to enhance HOGs quality. Inferring HOGs, especially at deep evolutionary nodes, is challenging, often yielding more HOGs than expected ancestral genes11. This led us to introduce the HOG Completeness Score in the OMA browser—defined as the fraction of species in the clade represented in a HOG (ranging from 0 to 1). Low scores may indicate dubious HOGs with many inferred losses, while reliable HOGs typically score above 0.2. In the LECA reconstruction, we observed that low-score HOGs often remain as singletons, whereas high-score HOGs integrate into contigs (Extended Data Fig. 7d). Hence, edgeHOG may be useful not only to refine orthology inference but also to filter out dubious HOGs (typically excluded from reconstructed contigs).

As a powerful application illustrated in our analyses of LECA’s ancestral contigs, edgeHOG can identify conserved gene clusters and highlight potential new targets for functional studies, as genes located within the same neighbourhood can be coregulated or functionally related. Moreover, tracking gene cluster evolution can offer insights into how biological functions have maintained or adapted throughout evolution.

EdgeHOG’s unique option to pinpoint the clade of emergence for any gene adjacency is useful to detect evolutionary patterns of genome organization or rearrangement. For instance, it led us to recover two well-documented, outstanding patterns: histone clusters in Metazoa and relatively younger adjacencies of sex chromosomes. Histone clusters in metazoan chromosomes comprise ‘blocks’ of adjacencies dated by edgeHOG to LECA, consistent with the conservation of histone genes as quartet or quintet across Eukaryotes24 and in support with the most recent suggestions of acquisitions of the histone genes as a single unit from a viral precursor25. Multiple clusters of histone quartets/quintets in succession is a specific feature of metazoa, probably originating from a complex history of tandem duplication and believed to be tied to the mechanisms of histone regulation in animals24. Dating also revealed that sex chromosomes, particularly Y, W and X, tend to have younger gene adjacencies than autosomes, reflecting known features such as higher gene turnover, gene duplication and repetitive element expansion rates in sex chromosomes than in autosomes26, structural instability in X-specific repetitive elements27 and rapid degeneration in Y and W due to the lack of recombination with a chromosome counterpart28,29,30,31. While Z chromosomes showed no clear pattern, regions such as the right arm of zebrafish chromosome 4 (rich in recent genes, pseudogenes and duplications32) also displayed younger adjacencies, though the link to sex determination remains unclear33. Younger adjacencies were also found in autosomes, including chromosome 4 in Drosophila (a reverted sex chromosome34), chromosome 19 in primates (notable for high gene density, repeats and GC content35) and 15 chicken microchromosomes, many with elevated repeat and GC content36. The ability of edgeHOG to flag the ancient origin/specific organization of histone clusters in Metazoa and chromosomal regions enriched in younger gene adjacencies highlights its potential for unravelling uncommon gene order trajectories and exploring the origin of genome architectures.

In terms of limitations, edgeHOG assumes that shared gene adjacencies across genomes are inherited from their last common ancestor, although such adjacencies can arise from horizontal gene transfer or convergent rearrangements, potentially leading to incorrect inferences, for example, photosynthetic gene contigs in LECA due to primary or secondary chloroplast endosymbiosis (Supplementary Table 3). Results in clades with reticulate evolution should thus be interpreted cautiously. Mitigation strategies include filtering for high-confidence HOGs with high completeness scores (excluding HOGs suggesting excessive gene loss) or removing edges with low weights. On another note, edgeHOG’s linearized genomes prioritize microsynteny and precision and tend to have a lower contiguity level than AGORA’s (Extended Data Fig. 9). While both tools can propose reconstructions of contigs of thousands to dozens of genes in ancestral species, a single missing adjacency can prevent identifying two neighbouring contigs as part of the same chromosome. For now, tools such as DESCHRAMBLER6, which, unlike edgeHOG and AGORA, primarily optimize for contiguity may be more effective for ancestral karyotype reconstruction. Rather than bridging contigs together at the expanse of precision, future extensions of edgeHOG could group microsynteny contigs into a higher hierarchical level, for example, that of the ancestral chromosome.

Benchmarks consistently show that edgeHOG matches and even slightly exceeds AGORA in recall and precision, with better linearization near the leaves and the ability to model the emergence of a gene through dynamical reconnection of its flanking genes in the parent graph. Most and foremost, edgeHOG scales far more efficiently than AGORA (Fig. 5). The linear scalability of edgeHOG breaks new ground and its ability to process phylogenies of thousands of genomes makes it uniquely suited to keep up with today’s and tomorrow’s massive sequencing projects and unlock their potential in comparative genomics.

Combining the temporal dimension of gene repertoire evolution with the spatial dimension of gene order evolution provides a comprehensive understanding of genome organization and evolutionary dynamics. Hence, our software solution opens up varied applications and advances our knowledge of genome evolution.

Methods

Algorithm

The detailed algorithm of edgeHOG is included in the Supplementary Information.

Benchmarking (preparation of input data)

Input data for benchmarking are available in Supplementary Data 1. The 100 simulated lineages datasets were generated with ALF (alfsim binary version 4.0)37, with a low mutational rate (mutRate:= 30) to facilitate downstream detection of orthologs and minimize biases inherent to orthology misinferences. Parameters regarding genome compositions and rates of gene duplications, losses and rearrangements are in Supplementary Data 1. The YGOB dataset (v7-Aug2012)20 was downloaded from http://ygob.ucd.ie/. For the OMA Vertebrata dataset, a pruned OMA’s species tree of the 50 chosen genomes was used. HOGs were derived from the tree and the all-vs-all of the 50 genomes, exported directly from the OMA browser. For all datasets, preprocessing followed the same steps. OMA Standalone inferred HOGs from the guide species tree and extant proteomes16. HOGs were converted to reconciled gene trees, the input format for AGORA. Gene order data (GFF files for edgeHOG, ordered gene lists for AGORA) were generated from the known extant genome structures. For ALF’s output and YGOB, orders were known from metadata files. For Vertebrata, gene orders were loaded from the OMA’s HDF5 file (for the 10 masked species, each gene was considered as a singleton).

Benchmarking

EdgeHOG and AGORA (version 3.1, basic workflow3) were run with default parameters on all datasets. For the genome simulation benchmark, inferred adjacencies at each internal level of the species tree were compared with true adjacencies in the corresponding known ancestral genome output by ALF. For the YGOB benchmark, inferred adjacencies at the root of the species tree were compared with YGOB-curated adjacencies in this ancestor. For the masked Vertebrata species benchmarks, any adjacency between two genes in the direct ancestor of a masked extant species were propagated in the masked genome only if the two ancestral genes had each a unique descending gene in the masked genome (no descending paralogues). Projected adjacencies were then compared with real, unmasked adjacencies. For both simulated and YGOB datasets, comparing ancestral adjacencies required to perform a mapping of a modelled ancestral gene (HOG_id in edgeHOG, family_id in Agora) to the corresponding ‘real’ ancestral gene disclosed by ALF and YGOB. This mapping was done based on the maximal number of descending extant genes in common. For each benchmark, the recall score was computed as 100 × TP/(TP + FN) and the precision as 100 × TP/(TP + FP), where TP is the number of correctly inferred adjacencies, FN is the number of missed adjacencies and FP is the number of misinferred adjacencies.

Functional analysis of LECA contigs

The LECA genome was inferred from the Nov2022 OMA release. For each Eukaryota-level HOG, ancestral GO terms were assigned as the union of its extant descendants’ terms. A Gene Ontology Enrichment Analysis was performed using goatools version 1.3.1 (ref. 38), with contig HOGs as the foreground and all Eukaryota-level HOGs as background. Enriched terms were those with Bonferroni-adjusted P < 0.05 (Fisher’s exact test). Randomized graphs were generated by swapping HOGs among the collection of contigs, which affected only the gene content of contigs and not their topology. Contigs were visualized with Cytoscape version 3.10.0 (ref. 39).

LECA’s adjacencies conservation in eukaryotes

Using pyHAM 1.2.0 (ref. 40), we retrieved all descendant genes per species for each HOG on LECA’s contigs. For each ancestral adjacency, we checked extant synteny graphs in species where both ancestral genes had descendants. If an adjacency existed between any descendant genes, the extant adjacency was considered conserved in that species.

Dating gene adjacencies

The taxon of origin for gene adjacencies was determined using EdgeHOG’s date_edges option. Taxon ages were obtained from TimeTree22 (https://timetree.org) by uploading species lists from the OMA Database. Since TimeTree lacks some OMA species, we first considered a reduced OMA Taxonomy containing only species shared with TimeTree and attributed an age of all non-conflicting internal nodes between OMA and TimeTree. Finally, we attributed an age of 0 to any leaf in the OMA Taxonomy. Any node left with no age at this point was assigned the average age of its most recent ancestor and its oldest child with age info. A companion script for dating gene adjacencies with TimeTree is available in the EdgeHOG GitHub repository. Histone adjacency clusters were defined as groups with over four adjacencies between histone genes from distinct HOGs and fewer than ten genes separating these adjacencies. The cluster size equals the number of histone adjacencies within the cluster. For the sex chromosomes, we selected genomes with clearly identified sex chromosomes (X, Y, Z and W) or numbered chromosomes from the OMA Database, excluding fungi with Roman numeral chromosomes. Only canonical chromosomes (numbers or letters) were considered, excluding incomplete contigs and scaffolds. Comparisons were made between each sex chromosome and all other complete chromosomes, as well as each autosome against the other chromosomes. One-sided Mann–Whitney tests assessed whether the distribution of adjacency ages was similar between the sex and the other chromosomes, with the alternative hypothesis being the sex chromosome having younger adjacencies than the others. The P values were adjusted for multiple testing based on the number of chromosomes per species.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.