We implemented content-based synteny comparisons between each amniote vertebrate and multiple invertebrate outgroup genomes. Initial ohnolog candidates were identified, in each vertebrate genome, using a window-based approach to detect putative synteny blocks between each vertebrate and the six outgroup genomes (outgroup comparison, Fig 1J), extending earlier similar approaches [17, 30, 31]. Additional synteny block candidates were also identified by comparing each vertebrate genome to itself (self comparison, Fig 1I) [3, 32] and ohnolog pair candidates were further restricted to paralogous pairs duplicated at the base of vertebrates according to Ensembl compara [33–35] (see S1 Text, Supplementary Materials and Methods). S1 Fig lists the numbers of human ohnolog pair candidates identified by each invertebrate outgroup and human-human synteny comparison, before applying any filtering on the statistical support of candidate synteny blocks. We identified a total of 15,107 such putative ohnolog pair candidates, including 11,428 identified with at least one outgroup and 15,054 identified by self comparison alone.

To narrow down this initial list of ohnolog candidates, we developed a quantitative approach to assess the statistical confidence of each ohnolog pair candidate. This quantitative approach and corresponding ‘q-score’, ranging from 0 to 1, estimates the probability that each ohnolog pair is simply identified by chance. Hence, lower q-scores imply more statistically significant ohnolog pairs (see S1 Text). Finally, we integrated q-scores for outgroup-comparison and self-comparison from all vertebrates, and filtered the ohnolog pairs based on the resulting combined q-scores. A flowchart summarizing our algorithmic approach is depicted in Fig 2. The pipeline of the approach is outlined below with methodological details described in Supplementary Materials and Methods (S1 Text).

The strict, intermediate and relaxed criteria lead to three sets of ohnolog pairs in the human genome with decreasing statistical confidence levels: 2,695 ohnolog pairs with very high confidence, 4,827 with high confidence and 8,178 with medium confidence, respectively (Table 1). These predicted ohnolog pairs are also significantly different from ohnolog pairs reported in earlier studies [3, 4], Table 1. In particular, 617 (23%) of the 2,695 strict ohnologs pairs from our analysis are not identified in [3]. For example, the strict ohnolog pairs between the transcription factors SOX11 and SOX12 or between the microtubule-associated proteins MAP2, MAP4 and MAPT are missing in [3]. Conversely, 3,695 (44%) of the 8,383 ohnolog pairs reported in [3] are excluded by the present analysis. More precisely, we found that 1,853 (50%) of these 3,695 ohnolog pairs ruled out by our analysis have not been duplicated at the base of vertebrates according to Ensembl compara, while 813 (22%) discarded ohnolog pairs are not supported by our quantitative multi-genome synteny comparison and the remaining 1,029 (28%) are excluded by both duplication timing and quantitative multi-genome synteny assessment. For example, the 3-oxoacid CoA-transferase genes OXCT1 and OXCT2, previously reported as ohnologs [3], have in fact been duplicated more recently than the 2R-WGD (i.e. in mammals according to Ensembl compara). By contrast, the signaling genes WNT1 and WNT3, also reported as an ohnolog pair [3] are not supported by our quantitative multi-genome synteny criteria and have also been duplicated earlier than the 2R-WGD (i.e. in bilateria or coelomata according to Ensembl compara).

The distribution of our ohnolog pairs with respect to all six outgroups is depicted on a six way Venn diagram in Fig 3 (percentages) and S8 Fig (numbers). Ohnolog pairs range from 1,416 with sea urchin comparison to a maximum of 5,994 using Drosophila melanogaster as outgroup. There are only 3.8% (293) ohnolog pairs identified by all outgroups, while each outgroup combination shaded in green in Fig 3 contributes to more than 2% of the total number of ohnolog pairs. This illustrates that many ohnologs would not be identified using just a single outgroup genome owing to lineage specific rearrangements in the outgroup genomes, limitations of genome assembly/annotation or homology criteria. In particular, while 90% (6,943) ohnolog pairs in human are identified by at least one chordate outgroup genome, 10% (772) ohnolog pairs are only identified by synteny comparison with non-chordate genomes. For example, the homeobox protein ohnolog pair VAX1/VAX2 and the nuclear receptor co-repressor ohnolog pair LCOR/LCORL are only identified by synteny comparison with D. melanogaster and C. elegans.

The final human ohnolog counts for strict, intermediate and relaxed criteria are respectively, 3,544 ohnologs (Strict Criteria); 5,504 ohnologs (Intermediate Criteria) and 7,831 ohnologs (Relaxed Criteria), Table 1. This is also to be contrasted with the results of previous studies that used either content-based synteny comparison with a single outgroup [17, 31] or only self comparison [3, 4, 32] without statistical significance criteria to filter out spurious synteny block conservation. We found that the available sets of human ohnologs from these early studies also present significant differences from our results. For instance, the set of 7,075 ohnolog genes from [3] shows significant differences from ours (S9 Fig), as 14%, 18% and 23% of our human ohnologs for strict, intermediate and relaxed criteria, respectively, have not been identified in [3]. Conversely, 57%, 33% and 15% of this early ohnolog data set are excluded from our strict, intermediate and relaxed human ohnolog sets, respectively (S9 Fig). As discussed above, this is due to inconsistent duplication times, according to Ensembl Compara, and/or limited statistical supports for each confidence criteria.

We then reconstructed ohnolog families from ohnolog pairs using a depth first search algorithm [36] (S1 Text). The resulting ohnolog families also contain paralogs which are small scale duplicates with respect to each other but form ohnolog pairs with a third gene of the family. Accounting for such small scale duplicates, eventually lead to ohnolog families with an expected maximum of four ohnologs retained from the two rounds of WGD in early vertebrates. However, as most genes lose their duplicates after WGD, most ohnolog families are expected to be of size two or three.

We obtained 1,381, 2,024 and 2,642 ohnolog families using strict, intermediate and relaxed criteria, respectively, for the human genome. Most remarkably, for almost all of these families, the size never exceeds four ohnologs, as expected for two rounds of WGD. As depicted in Table 1, all but 7 ohnolog families (99.5%) have a size smaller or equal to four for the strict criteria. Even with the most relaxed criteria, 96.7% of ohnolog families are consistent with a maximum family size of four ohnologs. Furthermore, a sharp decline in the number of families was observed beyond size four, suggesting a limited number of false positive ohnologs incompatible with two rounds of genome duplications. Interestingly, however, many three- or four-ohnolog families could not be identified independently in individual amniote genomes, but only by integrating synteny information from different amniote genomes, such as the four-ohnolog family ERAS/HRAS/KRAS/NRAS (relaxed criteria).

We also applied the same approach to generate ohnolog families from the ohnolog pairs provided by [3] and [4]. 95.1% of ohnolog families from [3] are consistent with two rounds of WGD and only 85.4% of ohnolog families from [4] have sizes up to four ohnologs. Clearly families exceeding four ohnologs must result either from the erroneous concatenation of distinct ohnolog families or include non-ohnolog genes. For instance, the ohnolog status of TRPV5 and TRPV6 [3] from the large family of six ion channels (TRPV1-6) are not supported by our quantitative assessment of self- and outgroup synteny. Conversely, we could also identified previously overlooked ohnologs, through high confidence assessment of self- and outgroup synteny. For instance, the guanine exchange factor RGL2 was found to be part of a four-ohnolog family with strict criteria, RGL1/RGL2/RGL3/RALGDS, RGL4 (with RGL4 a small scale duplicate of RALGDS).

In addition to the human genome, our synteny comparison approach across multiple genomes also identified ohnologs in five other amniote genomes: four mammals (mouse, rat, pig and dog) and one bird (chicken). Starting from ohnolog pairs in each species, the same approach was used to generate ohnolog families. A summary of individual ohnologs, ohnolog pairs and ohnolog families for these genomes is given in S2 Fig for strict, intermediate and relaxed quantitative criteria.

The level of annotation of these genomes is variable and the number of annotated protein coding genes range from 15,310 for chicken to 22,865 for the rat genome (S3 Fig). Using the relaxed criteria, a minimum of 4,282 to a maximum of 9,708 ohnolog pairs could be identified for chicken and rat, respectively. The six way Venn diagram in Fig 4 summarizes the fractions of retention versus lineage specific loss of ohnologs in the analyzed amniote genomes for the relaxed criteria (see S10 Fig for ohnolog numbers). Statistics for the strict criteria are given in S11 Fig. The identification of consensus ohnologs in this context implies that we are able to detect their ohnolog status through self- and outgroup synteny comparison or, alternatively, through orthology with bona fide ohnologs in other amniotes (see S1 Text). Indeed, ohnologs that are no longer in significant synteny in a particular vertebrate genome can still be identified, as long as their ortholog status can be unequivocally established with proper ohnologs in other vertebrates. This enables to circumvent strict synteny conditions in a specific genome.

By contrast to the small fraction of ohnolog genes identified by the six outgroups (i.e. 3.8%, Fig 4), 36.6% of predicted ohnologs are shared by all six amniotes, 53.9% by the five mammals and 74.3% by human, mouse and rat, while only a few other combinations of specific amniotes contribute to more than 2% of all ohnologs (see sectors shaded in red in Fig 4). This illustrates that the ohnologs have been largely conserved in mammals and to a lesser extent across amniotes. Likewise, ohnolog family sizes in each amniote genome consistently follow similar distributions as observed in human (Table 1) with a sharp decline in the number of families beyond the maximum size of four ohnologs (S2 Fig). In fact, the numbers of ohnologs in each family are most often the same in human and other mammals (in particular mouse) with occasional differences, typically missing ohnologs, in chicken which has significantly fewer genes (including ohnologs) than other amniotes considered in this study. For example, chicken has lost a number of adipokine genes [37] such as SERPINE1, which is part of a four-ohnolog family in mammals, SERPINE1/SERPINE2/SERPINE3/SERPINI1|SERPINI2 (where SERPINI1 and SERPINI2 are small scale duplicates). Similarly, all three ohnolog genes in the family of DNA binding Forkhead box protein A, i.e. FOXA1/FOXA2/FOXA3, are missing in the annotated chicken genome. Hence, differences in the shared ohnologs in Fig 4 arise due to lineage specific ohnolog loss or, possibly, due to missing annotations of genes and/or orthologs in these genomes.

We have so far restricted our synteny conservation analysis across multiple genomes to selected amniote genomes. In particular, amphibians and fishes have not been included in the analysis. This is because assembled chromosomal scaffolds of available amphibians (e.g. Xenopus) and non-teleost fishes (e.g. elephant shark and coelacanth) do not contain enough genes to be included in a content-based synteny conservation analysis (e.g. 81% of X. tropicalis genes are on chromosomal scaffolds with fewer than 50 genes). As for teleost fish genomes, they experienced a third more recent (3R) WGD, about 300 MY ago [38] in addition to the two rounds of (2R) WGD common to all vertebrates. This additional 3R WGD implies methodological issues specific to teleost fish genomes, which will be addressed in a forthcoming extension of our computational approach to identify ohnologs through multiple genome synteny comparison.

As outlined in the introduction, ohnologs have been reported to be preferentially retained in functional categories associated with development, signaling and gene regulation in the human genome [3, 7–10]. We performed a Gene Ontology (GO) enrichment analysis on four amniote vertebrates using DAVID [39] and observed the same general trend across these amniote genomes (Fig 5A). This confirms that ohnologs are associated with similar functional categories in different vertebrates.

In addition, ohnologs have also been associated with disease mutations [5, 12–14], in particular with dominant deleterious mutations frequently implicated in cancers and dominant genetic diseases [5, 6, 15]. Fig 5B confirms such cancer and genetic disease associations for all three ohnolog confidence criteria adopted in this study. This is particularly significant for core cancer genes [5, 40] (amounting for just 8.3% of non-ohnologs but up to 21.6–26% of ohnologs, i.e. a 2.6–3.1 fold increase, p = 3.4 × 10⁻¹⁵³ Fisher Exact Test) and autosomal dominant diseases (amounting for just 2.1% of non-ohnologs but up to 5.4–5.9% of ohnologs, i.e. a 2.6–2.8 fold increase, p = 3.4 × 10⁻²⁷ Fisher Exact Test) in agreement with earlier reports [5, 6] and evolutionary models [15]. We also analyzed the enrichment of ohnologs in genes with autoinhibitory protein folds, which are prone to dominant deleterious mutations. To this end, we collected genes with autoinhibitory protein folds either from careful literature curation [5] or based on the annotation of structural domains frequently associated with autoinhibition (i.e. SH3, DH, PH, CH, Drf and Eth domains), identified using Hidden Markov Model (HMM) search [41] against the PFAM database [42] (see Supplementary Methods). We observed that the ohnologs are particularly enriched in genes with autoinhibitory protein folds (amounting for just 1.4% of non-ohnologs but up to 9–12% of ohnologs, i.e. a 6.4–8.6 fold increase, p = 4.4 × 10⁻¹⁵⁰ Fisher Exact Test) [5].

The data of all the ohnolog pairs and families for the six vertebrate genomes is accessible through the ‘Ohnologs’ server at http://ohnologs.curie.fr/. There, users can i) search for a particular gene, ii) browse pre-compiled ohnolog families and ohnolog pairs or iii) generate ohnolog families based on their own, user-defined, quantitative filters. The server is implemented in Perl-CGI and is hosted on a virtual machine at Institut Curie.

On the Search page (S12 Fig), the user can search for a gene of interest in any of the six available vertebrates using either Ensembl Id, gene symbol or any desired keywords. Search by functional categories is also possible using Gene Ontology Id or term. If a keyword search does not match any gene directly, we display all the genes matching that keyword in gene symbol, text description or GO term. A hyperlink from this page directs to the details on its ohnolog families and its possible association with human diseases points to Genecards [43] and Cosmic [44] databases. This page also contain links to details in UniProt and Entrez databases if available. If the gene exists in our analysis, and is an ohnolog, users are directed to the details about ohnolog families for each statistical confidence levels (i.e., strict, intermediate and relaxed criteria), S13 Fig.

Alternatively, users can also generate ohnolog families using our multi genome comparison analysis, for any of the six available vertebrate genomes using an arbitrary, user-defined, quantitative criteria for the outgroup and self comparisons. The default values correspond to the strict criteria. The result pages display all the pre-calculated or custom generated families, which can also be downloaded.

In the light of the importance of ohnologs in the evolution of vertebrates and their enhanced association with diseases, our analysis provides a useful resource to gain further insights on the impact of WGD in extant vertebrates.