SATURN: OrthoFinder orthogroups, and how six gene groupings place species together

Analysis·Project Mnemiopsis Phagocytes·Tools OrthoFinder 3.1.0 SATURN ESM-1b ESMC-600M (FANTASIA v4) eggNOG-mapper Python

ImportantSummary

• OrthoFinder placed 83.3% of the 91,104 proteins (one per gene, four species) in 12,254 orthogroups; 4,987 contain all four species and 2,577 of those are single-copy. Zebrafish genes often come in pairs relative to the other species (teleost genome duplication), so zebrafish has the fewest 1:1 orthologs. OrthoFinder did not write Metazoa-level hierarchical orthogroups (section 1).
• The two orthology references agree on how much cross-species signal there is. Only about a third (Mnemiopsis, oyster) to a half (Ciona, zebrafish) of SATURN genes share a group with another species’ SATURN gene, whether eggNOG or OrthoFinder is used (Table 8).
• Macrogenes put almost everything with another species: 84–92% of invertebrate and Ciona genes, and 51–71% of zebrafish genes, sit in multi-species macrogenes (Figure 2).
• Macrogenes over-group the three non-vertebrates and set zebrafish apart. Mnemiopsis + oyster + Ciona groups hold 12–17% of genes in macrogenes vs 3.6–5.3% in orthogroups, while Ciona + zebrafish groups, the closest pair by orthology, are under-represented (Figure 3, Table 9). This is strongest with ESMC.
• Agreement on gene pairs is low everywhere. The two orthology references share 33–49% of their cross-species pairs; ESM-1b macrogenes recover 30–43% of orthology pairs (ESMC 15–21%); and two ESM-1b runs on different species sets share only ~42% of their pairs (Figure 4).
• How orthogroups sit in macrogenes does not depend on the reference: ESM-1b mostly merges whole orthogroups or cuts across them (5% match, 29–33% merged); ESMC cuts across them for three-quarters of genes (Figure 5).

This entry follows Oct 9 (section 6 of which summarizes the OrthoFinder benchmark scores). Here: what OrthoFinder produced, then a side-by-side of six ways of grouping the same SATURN genes — two orthology references and four macrogene sets — asking which species end up together and how orthogroups split or merge.

Key terms

Table 1: Terms used in this entry.

Term	Meaning here
Orthogroup (OG)	Genes descended from one gene in the last common ancestor of the species analysed. eggNOG Metazoa OGs are precomputed (each gene assigned by its best database hit); OrthoFinder OGs are inferred de novo from our four proteomes.
HOG (hierarchical orthogroup)	An OrthoFinder orthogroup defined at one node of the species tree (N0 = root, N1, N2 …), using gene trees.
Grouping	Any partition of genes: an orthology reference or a set of macrogenes. Six here (Table 2).
SATURN genes	The 31,702 genes present in all four macrogene runs (8,000 highly variable genes per species, minus a few absent from some runs). All comparisons use these genes.
Species span	The number of species (1–4) with at least one SATURN gene in a group.
Cross-species pair	Two SATURN genes from different species placed in the same group.
match / OGs merged / OG subdivided / cross-cutting	For a gene in a reference orthogroup of ≥ 2 genes: its OG is kept whole in a group of its own / kept whole but shares the group with other OGs / split, with each piece in a group of its own / both. As on Oct 9.

Table 2: The six groupings (all macrogene sets: K = 3,000).

Grouping	What it is
eggNOG Metazoa OGs	eggNOG-mapper 2.1.13, DB 5.0.2, Metazoa level; primary OG per gene. 21,674 of the SATURN genes have one
OrthoFinder orthogroups	This entry; every gene has a group (singletons included)
Canonical 5sp ESM-1b	Macrogenes from the canonical 5-species run (urchin genes dropped)
ESM-1b 4sp	Macrogene-only run, 4 species, ESM-1b
ESMC raw 4sp	Same, ESMC-600M layer_0 (pre-norm) vectors
ESMC rescaled 4sp	Same, ESMC vectors rescaled to ESM-1b’s scale

1. The OrthoFinder run and its outputs

OrthoFinder 3.1.0 on the four proteomes (longest protein per gene, same protein → gene bridges as the SATURN embeddings; 06a), with MSA gene trees (-M msa -A mafft -T fasttree) and the fixed species tree (Mlei,(Cgig,(Crob,Drer))). 70 min on 16 cores, 15.9 GB peak memory (Pegasus, normal queue; environment rebuilt the same day, see that entry).

Table 3: OrthoFinder overall statistics (Comparative_Genomics_Statistics/Statistics_Overall.tsv).

Statistic	Value
Genes (proteins)	91,104 (Mnemiopsis 15,545; oyster 31,371; Ciona 18,788; zebrafish 25,400)
Genes in orthogroups	75,861 (83.3%)
Unassigned genes	15,243 (16.7%)
Orthogroups	12,254 (mean 6.2 genes, median 4)
Species-specific orthogroups	3,561, holding 22,576 genes (24.8%)
Orthogroups with all 4 species	4,987, of which 2,577 single-copy in every species
G50 / O50	5 genes / 3,263 orthogroups

Table 4: OrthoFinder per-species statistics (Statistics_PerSpecies.tsv).

Statistic	Mnemiopsis	oyster	Ciona	zebrafish
Genes	15,545	31,371	18,788	25,400
% of genes in orthogroups	75.8	88.1	75.1	88.0
Unassigned genes	3,765	3,748	4,685	3,045
Orthogroups containing the species (% of all)	6,660 (54.3)	9,682 (79.0)	7,938 (64.8)	8,800 (71.8)
Species-specific orthogroups	642	1,530	504	885
% of genes in species-specific orthogroups	21.9	34.9	17.6	19.3

Table 5: Pairwise orthologs from OrthoFinder (OrthologuesStats_*.tsv; genes counted in the first species of each pair). All 12 directions in D0_orthofinder_orthologues.csv.

Species pair	1:1 orthologs	Genes of the first species with several orthologs in the second (one-to-many)	Genes of the first species sharing one ortholog in the second (many-to-one)
oyster – Ciona	5,737	388	4,366
Mnemiopsis – Ciona	4,454	417	1,118
oyster – zebrafish	4,393	2,327	2,628
Mnemiopsis – oyster	4,322	827	1,206
Ciona – zebrafish	4,270	2,463	1,215
Mnemiopsis – zebrafish	3,237	1,756	710
zebrafish – Ciona	4,270	267	8,141

• Oyster has the most lineage-specific genes: 35% of its genes are in oyster-only orthogroups, consistent with its many lineage-specific gene-family expansions. Mnemiopsis and Ciona have the most genes with no orthogroup at all (~24–25%).
• Zebrafish genes usually come in pairs relative to the other species. 8,141 zebrafish genes share a single Ciona ortholog with another zebrafish gene (many-to-one), and 2,463 Ciona genes have several zebrafish orthologs, as expected after the teleost genome duplication. As a result, zebrafish has fewer 1:1 orthologs with Ciona (4,270) than oyster does (5,737), and the fewest of all with Mnemiopsis (3,237). This duplication may also contribute to the zebrafish-only macrogenes in section 3 (paralog pairs grouped together); not tested here.

What the results folder contains, and what we use:

Table 6: The OrthoFinder results folder.

Folder / file	Contents	Used here
Orthogroups/Orthogroups.tsv (+ _UnassignedGenes.tsv, .GeneCount.tsv)	Orthogroup membership per species	Yes (06c → ALL4_gene_orthofinder_OG.csv)
Phylogenetic_Hierarchical_Orthogroups/N1.tsv, N2.tsv	HOGs for oyster + Ciona + zebrafish (N1) and Ciona + zebrafish (N2)	No: no root (N0) file
Orthologues/	Pairwise orthologs per species pair (1:1, 1:many …)	Not yet
Resolved_Gene_Trees/, Gene_Duplication_Events/	Rooted gene trees; duplications mapped to species-tree nodes	Not yet
Species_Tree/	The fixed tree with node labels (N0 root, N1, N2)	Interpreting HOG files
Comparative_Genomics_Statistics/	Overall and per-species statistics	Table 3

No Metazoa-level HOGs. OrthoFinder wrote HOGs only for N1 and N2, not for the root, the only node that includes Mnemiopsis; a 3-species test behaved the same way. The comparisons below therefore use OrthoFinder orthogroups. Adding a choanoflagellate outgroup in the planned 6-species run should make the animal node a non-root node.

Coverage of SATURN genes. Every SATURN gene is in the OrthoFinder tables (the gene bridges work). Genome-wide, the share of SATURN genes whose orthogroup includes another species follows the divergence gradient: Mnemiopsis 43.9%, oyster 46.6%, Ciona 53.2%, zebrafish 78.0%. Within the SATURN gene set (only other species’ SATURN genes counted), the shares are lower (Table 8).

2. Species span: how many species share each group

Table 7: Species span of each grouping on the SATURN genes (eggNOG: its 21,674 assigned genes). Effective number of groups = exp(entropy of group sizes).

Grouping	Genes	Groups	Median size	Effective no. of groups	% of genes in groups spanning 1 / 2 / 3 / 4 species	4-species groups
eggNOG Metazoa OGs	21,674	10,443	1	7,155	42 / 31 / 18 / 10	220
OrthoFinder orthogroups	31,702	15,944	1	9,756	59 / 21 / 13 / 7	251
Canonical 5sp ESM-1b	31,702	2,963	8	2,138	15 / 22 / 34 / 30	581
ESM-1b 4sp	31,702	2,997	7	2,167	15 / 21 / 36 / 28	552
ESMC raw 4sp	31,702	2,994	8	2,106	22 / 20 / 34 / 25	412
ESMC rescaled 4sp	31,702	2,989	7	1,936	21 / 19 / 34 / 27	391

Figure 1: Genes by the number of species in their group. Left: all SATURN genes (eggNOG on its assigned genes); right: the 21,674 eggNOG-annotated genes, where all six groupings are defined on the same genes.

• Orthogroups are small and mostly single-species within the SATURN genes: 42% (eggNOG) and 59% (OrthoFinder) of genes have no other species’ SATURN gene in their group. OrthoFinder’s higher share comes from the ~10,000 genes eggNOG cannot annotate; on the same genes it is 49% (Figure 1, right).
• Macrogenes are the reverse: 78–85% of genes sit in groups spanning ≥ 2 species, and 25–30% in groups with all four. ESMC leaves more genes in single-species macrogenes (21–22% vs 15%), as on Oct 8.

3. Which species end up together

Table 8: % of each species’ SATURN genes in a group with at least one other species’ SATURN gene. eggNOG values are converted to % of all SATURN genes (assigned-gene shares 66.3 / 64.1 / 68.8 / 44.0 × the share of genes with an OG).

Grouping	Mnemiopsis	oyster	Ciona	zebrafish
eggNOG Metazoa OGs (as % of all SATURN genes)	33.3	40.1	43.2	43.4
OrthoFinder orthogroups	33.6	39.5	41.7	47.8
Canonical 5sp ESM-1b	92.0	88.1	90.1	70.9
ESM-1b 4sp	92.4	89.6	89.1	68.3
ESMC raw 4sp	87.6	84.3	84.1	56.7
ESMC rescaled 4sp	90.6	84.3	89.9	50.8

Figure 2: % of each species’ genes in a multi-species group. Left: all SATURN genes (eggNOG as % of its assigned genes); right: eggNOG-annotated genes.

• The two orthology references give nearly the same per-species picture once eggNOG’s unannotated genes are counted: a third of Mnemiopsis genes, and under half for every species, have a SATURN-gene ortholog in another species.
• Zebrafish is the exception for macrogenes: it is the species most often in a multi-species orthogroup (OrthoFinder) but the least often in a multi-species macrogene, especially with ESMC (51–57%).

Table 9: Selected species pairs (all SATURN genes; *eggNOG as % of its assigned genes). Full 4 × 4 matrices for all six groupings in D3_species_pair_cooccurrence.csv.

% of species A’s genes whose group contains species B	eggNOG*	OrthoFinder	ESM-1b 4sp	ESMC raw	ESMC rescaled
Mnemiopsis → oyster	33.8	16.2	76.0	72.9	78.1
Mnemiopsis → Ciona	40.9	18.8	79.8	74.2	80.5
Mnemiopsis → zebrafish	41.3	20.8	42.1	30.8	37.5
oyster → Ciona	42.9	23.9	75.5	72.2	71.6
Ciona → zebrafish	42.4	27.4	48.5	36.9	41.5
zebrafish → Ciona	27.5	30.8	49.2	39.0	35.5

Figure 3: % of genes in groups with each exact species combination, on the 21,674 eggNOG-annotated genes.

• By orthology, Ciona’s most frequent partner is zebrafish (OrthoFinder: 27% of Ciona genes; Ciona + zebrafish groups hold 8.5–8.7% of genes), as expected for two chordates.
• Macrogenes instead group Ciona with Mnemiopsis and oyster (70–80% of Ciona or Mnemiopsis genes share a macrogene with the other), and Mnemiopsis + oyster + Ciona groups hold 12–17% of genes vs 3.6–5.3% in orthogroups (Figure 3). Groups pairing zebrafish with one other species are under-represented (Mnemiopsis + zebrafish: 0.8–0.9% with ESMC vs 5% in orthogroups).
• Reading: in embedding space, zebrafish genes sit apart from the other three species. This fits the species signal seen in the embeddings on Oct 6, and is strongest for ESMC. Whether it reflects real divergence, uneven training data (vertebrates are better sampled) or an embedding artefact is open; a fifth species (human or fly) will help separate these.

4. Agreement on gene pairs

Figure 4: Agreement on cross-species gene pairs (eggNOG-annotated genes): % of the row grouping’s pairs also placed together by the column grouping.

Table 10: Cross-species pairs and agreement with each orthology reference (eggNOG-annotated genes).

Grouping	Cross-species pairs (annotated genes)	% recovered of eggNOG pairs	% recovered of OrthoFinder pairs	% of own pairs that are eggNOG / OrthoFinder pairs
eggNOG Metazoa OGs	26,985	—	49.0	— / 32.5
OrthoFinder orthogroups	17,917	32.5	—	49.0 / —
Canonical 5sp ESM-1b	59,881	29.1	40.2	13.1 / 12.0
ESM-1b 4sp	61,412	30.2	42.9	13.3 / 12.5
ESMC raw 4sp	75,563	15.1	19.7	5.4 / 4.7
ESMC rescaled 4sp	74,519	15.9	20.6	5.7 / 5.0

• The references themselves agree only partly: half of OrthoFinder’s pairs are eggNOG pairs, but only a third of eggNOG’s are OrthoFinder pairs (eggNOG groups more genes together).
• Macrogenes make 2–4× more cross-species pairs than orthogroups, and 87–95% of them are not orthology pairs (or involve genes whose orthologs are not among the SATURN genes).
• Macrogene sets agree poorly with each other. The two ESM-1b runs, same model with a different species set and training run, share ~42% of pairs; ESMC raw vs rescaled ~21%; ESM-1b vs ESMC 12–16%. This matches Oct 9: aggregate scores are reproducible, individual macrogene memberships are not.

5. How orthogroups split and merge across groupings

Figure 5: How reference orthogroups relate to each grouping (eggNOG-annotated genes). Each bar: % of genes in reference orthogroups of ≥ 2 genes.

Table 11: Nesting of reference orthogroups within each grouping (eggNOG-annotated genes; categories as % of genes in reference OGs with ≥ 2 genes). Canonical 5sp ESM-1b is within 1 point of ESM-1b 4sp throughout (D5_nesting_vs_reference.csv).

Grouping	Reference	Homogeneity	Completeness	ARI	match	OGs merged	OG subdivided	cross-cutting
OrthoFinder orthogroups	eggNOG	0.96	0.95	0.41	34.1	14.9	29.8	21.2
ESM-1b 4sp	eggNOG	0.81	0.94	0.21	4.3	28.4	4.2	63.1
ESMC raw 4sp	eggNOG	0.76	0.91	0.11	1.6	12.0	4.4	82.0
ESMC rescaled 4sp	eggNOG	0.77	0.91	0.12	2.7	12.8	6.1	78.3
ESM-1b 4sp	OrthoFinder	0.82	0.95	0.25	5.0	32.9	5.3	56.7
ESMC raw 4sp	OrthoFinder	0.76	0.92	0.13	2.2	14.7	5.9	77.2
ESMC rescaled 4sp	OrthoFinder	0.78	0.92	0.14	3.6	15.8	8.4	72.3

• eggNOG and OrthoFinder have similar granularity but cut genes differently: a third of genes are in identical groups, and OrthoFinder splits eggNOG OGs about twice as often as it merges them.
• Macrogenes relate to both references the same way. ESM-1b keeps orthogroups whole (match + merged) for 33–38% of genes; ESMC for 14–19%. Nearly all the rest are cross-cutting: their orthogroup is split across macrogenes that also contain other orthogroups.

Interpretation

1. Orthology and embeddings disagree about how much is shared between species. By sequence orthology, at most half of the SATURN genes have a cross-species partner among the other species’ SATURN genes; macrogenes give a partner to 84–92% of non-vertebrate genes. The difference is the set where embeddings are expected to add something (remote homologs), and also where they can add noise.
2. The pattern of who groups with whom is not phylogenetic. Macrogenes join Ciona with the ctenophore and oyster more than with zebrafish. Together with the species signal in the embeddings, this suggests a zebrafish (vertebrate) offset in embedding space rather than a biological grouping, and argues for checking species offsets before reading cross-species macrogenes as homology.
3. The two orthology references are interchangeable for these questions, which makes the ESM-1b vs ESMC conclusions robust to the choice of reference (as in Oct 9 section 6).

Caveats

• Everything is counted within the SATURN gene set. A gene can look single-species because its ortholog was not a highly variable gene in the other species; “1:1” means 1:1 among SATURN genes.
• eggNOG is defined only on 68% of genes, so its numbers come from a different (better-conserved) gene set unless the annotated-genes universe is used.
• No Metazoa-level HOGs: OrthoFinder orthogroups lump all duplications since the last common ancestor; HOGs would separate older from newer paralogs.
• One macrogene run per model at K = 3,000; Oct 9 shows the same model ranking at every K.

Possible next steps

• Copy Comparative_Genomics_Statistics/ to the Mac, rerun 06d, and add per-species and ortholog-type statistics (D0) to section 1.
• Check whether zebrafish-only macrogenes are enriched for zebrafish paralog pairs (genes sharing one ortholog in Ciona; OrthoFinder Orthologues/).
• Test the zebrafish offset directly: per-species centring of embeddings before KMeans, and whether Ciona then moves towards zebrafish.
• 6-species OrthoFinder (+ human, fly, choanoflagellate outgroup) for Metazoa-level HOGs.
• Use OrthoFinder Orthologues/ (pairwise 1:1) as a stricter ortholog set for recall.

Scripts

All in pipelines/scripts/saturn_embedding_comparison/.

Table 12: Scripts used in this entry.

Script	Where	What it does
06a_orthofinder_prep.py	Pegasus	longest protein per gene, same bridges as SATURN; fixed species tree
06b_orthofinder_4sp.sh	Pegasus	OrthoFinder 3.1.0 (micromamba env, FastTree preflight)
06c_orthofinder_to_reference.py	Pegasus	orthogroups → gene-level reference table + coverage of SATURN genes
06d_grouping_species_overlap.py	Mac	six groupings: species span, species pairs and combinations, pair agreement, nesting vs each reference (D0–D5)

References

• Emms D.M., Kelly S. (2019) OrthoFinder: phylogenetic orthology inference for comparative genomics. Genome Biology 20:238. https://doi.org/10.1186/s13059-019-1832-y
• Emms D.M., Liu Y., Belcher L., Holmes J., Kelly S. (2025) OrthoFinder 3 (bioRxiv; cited by the program).
• Huerta-Cepas J. et al. (2019) eggNOG 5.0. Nucleic Acids Research 47:D309–D314. https://doi.org/10.1093/nar/gky1085
• Cantalapiedra C.P. et al. (2021) eggNOG-mapper v2. Molecular Biology and Evolution 38:5825–5829.
• Rosen Y. et al. (2024) SATURN. Nature Methods 21:1492–1500. https://doi.org/10.1038/s41592-024-02191-z
• Rosenberg A., Hirschberg J. (2007) V-measure. EMNLP-CoNLL 2007.
• Hubert L., Arabie P. (1985) Comparing partitions. Journal of Classification 2:193–218.