henhaobio
Sign in

bioRxivBioinformaticsPreprint

OrthoFinder: scalable phylogenetic orthology inference for comparative genomics

David EmmsiD, Yi LiuiD, Laurence J. BelcheriD, Jonathan HolmesiD, Steven L. KellyiD

PDFDOICode

In-depth summary

Claude Opus 5.5 · xhigh effort · 7 Nov 2026 · from the full text of version 1 (16 Jul 2025) · CC BY 4.0

OrthoFinder v3 splits orthogroups at duplications predating the species-tree root, raising OrthoBench accuracy, and assigns extra species to a core run, finishing 1024 species in 128 hours.

Most orthology methods start from all-vs-all similarity searches, which scale with the square of species number. That is a bottleneck now that the Darwin Tree of Life and Earth BioGenome Project aim to cover all eukaryotes; OrthoFinder v2 defined orthogroups by MCL clustering of similarity scores alone, while v3 refines them with gene-tree reconciliation and extends the authors' earlier single-gene SHOOT placement to add whole species sets to an existing run.

Key results

  1. ShownSplitting MCL orthogroups at gene duplications that predate the species-tree root improves orthogroup accuracy. OrthoFinder v3 ranked first on an overall score across seven OrthoBench measures, ahead of v2 and seven other tools.

    OF3 had the highest recall, the fewest missing genes, the lowest entropy, and a 5-7% higher f-score than OF2 with identical settings (abstract: 7% relative gain). SonicParanoid2 (default) had 3.7% higher precision than OF3_Align_ST, but OF3 recall was 19% higher. FastOMA fused 10% of reference orthogroups vs 2.85%. · Abstract; Improved phylogenetic delineation of orthogroups

  2. ShownIn its core-plus-assign mode, OrthoFinder v3 was the only tested tool to finish a 1024-species Ensembl set within the 7-day limit. Runtime grew close to linearly from 128 to 1024 species.

    OF3 (64-species core, all other species assigned) finished 1024 species in 128 hours, with 32 threads and up to 200 GB RAM. Only SonicParanoid2 (fast mode) and FastOMA also finished 512 species. Above 64 species, OF3 was faster than every competitor and x8 faster than OF2. · Achieving enhanced scalability without…; Assessing scalability performance

  3. ShownOn the Quest for Orthologs benchmark, OrthoFinder v3 was on the Pareto frontier for species-tree discordance (eukaryotes and bacteria), enzyme classification and three human-curated reference sets. Its recall was well above FastOMA's.

    78 proteomes; 14 other tools compared. Vs FastOMA: eukaryote Robinson-Foulds 0.06 vs 0.05 but recall 15721 vs 8686 (80% higher); bacteria Robinson-Foulds 0.590 vs 0.587 with 23% higher recall; Enzyme Commission precision 0.933 vs 0.928, recall 183368 vs 157049. Highest recall on VGNC and SwissTree; TreeFam recall 0.72 vs 0.74 for OF2. · OrthoFinder 3 performs accurate ortholog…; Assessing ortholog inference accuracy

  4. ShownThe scalable Linear mode keeps most of the accuracy gain. On OrthoBench it fell slightly below full OrthoFinder v3 on some measures, but still beat every OrthoFinder v2 implementation and all other tools tested.

    OF3_Linear was scored in the same seven-measure OrthoBench comparison (Figure 2). The main text gives this ordering only in words; per-method scores are in Supplemental File 1 and the f-scores in Supplementary S1 Table 1. · Achieving enhanced scalability without…; Assessing orthogroup inference accuracy

  5. ShownAbove 128 species, the assignment mode uses less peak RAM than OrthoFinder v2, and the saving grows as more species are added.

    At 256 species, v3 Linear needed 3.4-fold less RAM than v2 run with DendroBLAST. At the biggest sizes any competitor still finished, v3's peak memory was about 4-fold lower (Figure 4B). · Achieving enhanced scalability without…

Methods, data and reproducibility

OrthoFinder v3 builds a gene tree per MCL orthogroup, reconciles it with the species tree and splits it at duplications predating the root. A scalable mode runs OrthoFinder on a core species set, then assigns other species' genes to core orthogroups using SHOOT profiles and DIAMOND. It was tested on OrthoBench, Quest for Orthologs and Ensembl subsets.

  • OrthoBench expert-curated reference orthogroups
  • Ensembl rapid release proteomes (accessed 29 August 2024, longest variant per gene), subsampled with PDA into nested sets of 2-1024 species · multiple species · n = 1789 species
  • Quest for Orthologs 2022 reference proteomes · 48 eukaryotes, 23 bacteria (called prokaryotes in Methods), 7 archaea · n = 78 proteomes
  • Published Quest for Orthologs results for OrthoFinder v2 and 13 other orthology tools · n = 14 tools
Methods & tools
OrthoFinder, SHOOT, DIAMOND, MCL, FAMSA, MAFFT, FastTree, ASTRAL-Pro, DendroBLAST, Phylogenetic Diversity Analyzer (PDA), BUSCO, FastOMA
Assay conditions & compute
Scalability tests ran on a Linux server with AMD processors. Each tool had 32 threads (16 cores), up to 200 GB RAM and a 7-day timeout; OrthoFinder v3 took 128 hours for 1024 species. A core set under 100 species is recommended on conventional hardware.
Reproducibility gaps
Benchmark details, per-method scores, output files and runtime data are in Supplemental Files 1-3, which are not linked in the text. Ensembl and Quest for Orthologs inputs are named without URLs or accessions, and no dependency versions or environment files are given.
Limitations (6)
  • On OrthoBench, OrthoFinder v3 has lower precision than several other methods. SonicParanoid2 (default) had 3.7% higher precision than OF3_Align_ST; the authors point to OF3's higher recall as the offset. · stated by the authors
  • The scalable Linear mode scores slightly below non-scalable OrthoFinder v3 on some OrthoBench measures. · stated by the authors
  • The scalable workflow adds a step for the user (splitting input into core and assign sets), and the core set should stay under 100 species on conventional computing resources. · stated by the authors
  • For scale, the defaults use FAMSA and FastTree '-fastest'. Users who substitute costlier alignment or tree programs through config.json are told to check their compute resources first. · stated by the authors
  • The main-text Quest for Orthologs numbers do not say whether they come from the Align or the Linear run. The bacterial Robinson-Foulds comparison (0.590 vs 0.587) is called marginally better, although in the eukaryote comparison the lower value is the better one. · noted in this summary
  • Runtime and memory figures come from one server configuration, with no replicate runs reported. Above 64 species, v3 always used the same 64-species core, and the text does not test how core size or composition affects results. · noted in this summary

Read this if you run orthology inference or comparative genomics across hundreds to thousands of proteomes and want to know what OrthoFinder v3's core-and-assign mode costs in accuracy and saves in runtime and memory.

Citations
44

Abstract

Here, we present a major advance of the OrthoFinder method. This extends OrthoFinders high accuracy comparative genomic framework to provide substantially enhanced scalability and accuracy. Specifically, we show that enhanced phylogenetic delineation of orthogroups provides a 7% relative increase in orthogroup inference accuracy. We further demonstrate that a new gene assignment method substantially reduces overall runtime RAM usage without compromising accuracy. The latest version of OrthoFinder is available at https://github.com/OrthoFinder/OrthoFinder.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISEmms '19Emms '20Steenwyk '24Derelle '20Emms '22Hu '19
87% semantic matchBioinformatics
David Emms, Steven L. Kelly·Genome biology·Nov 14, 2019· 7,675 citations