henhaobio
Sign in

Genome biologyBioinformaticsPeer-reviewed journal article

OrthoFinder: phylogenetic orthology inference for comparative genomics

David EmmsiD, Steven L. KellyiD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Here, we present a major advance of the OrthoFinder method. This extends OrthoFinders high accuracy orthogroup inference to provide phylogenetic inference of orthologs, rooted genes trees, gene duplication events, the rooted species tree, and comparative genomic statistics. Each output is benchmarked on appropriate real or simulated datasets and, where comparable methods exist, OrthoFinder is equivalent to or outperforms these methods. Furthermore, OrthoFinder is the most accurate ortholog inference method on the Quest for Orthologs benchmark test. Finally, OrthoFinders comprehensive phylogenetic analysis is achieved with equivalent speed and scalability to the fastest, score-based heuristic methods. OrthoFinder is available at https://github.com/davidemms/OrthoFinder.

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 2 (24 Apr 2019) · CC BY 4.0

OrthoFinder now infers orthogroups, rooted gene and species trees, duplications and orthologs from proteomes alone, topping Quest for Orthologs accuracy tests while rivaling heuristic tools' speed.

Score-based ortholog tools infer gene relationships from pairwise similarity, which uneven evolutionary rates can mislead, while tree-based curated databases cover only provider-chosen species. OrthoFinder supplies the missing automated pieces (rooted gene trees from sequence alone, species tree inference and rooting, and scalable ortholog calling tolerant of incomplete lineage sorting and tree error) at speeds competitive with score-based tools.

Key results

  1. ShownThe default, fastest configuration (DIAMOND search, DendroBLAST trees) gave the highest ortholog F-score of all methods on the Quest for Orthologs SwissTree and TreeFam-A gold-standard tree tests; slower BLAST and alignment-based configurations were marginally better still.

    Default OrthoFinder was 3%-17% (SwissTree) and 2%-28% (TreeFam-A) more accurate than every other method; '-S blast' and '-S blast -M msa' (MAFFT, FastTree) runs added a further 1-2%; no single competitor was consistently runner-up. · OrthoFinder has the highest orthology inference…; Ortholog Benchmarking

  2. ShownAll OrthoFinder configurations matched or beat every competitor on the Standard and Generalised Species Tree Discordance Tests (STDT, GSTDT) pseudo F-score, and across all four tests recovered more orthologs than any other method at similar precision.

    Pseudo F-score (ortholog-set recovery as recall, 1 minus normalised Robinson-Foulds distance as precision): default 0%-41% (STDT) and 11%-60% (GSTDT) above competitors, other versions a further 1%-6%; default recall 0.2% to 77.8% higher, precision from 3.4% lower to 18% higher than competitors. · OrthoFinder has the highest orthology inference…; Ortholog Benchmarking

  3. ShownOrthoFinder scaled to hundreds of fungal genomes as the second-fastest of all user-runnable Quest for Orthologs methods, and was the only one to also return gene trees and gene duplication events.

    With 16 threads, default OrthoFinder took 192s for 4 species and 1.8 days for 256 species, vs 1.2 days for SonicParanoid and 4.1 days for the next fastest; at 64 species, the largest set every method finished within the 120 hour cut-off, the slowest took 200 times longer than OrthoFinder. · OrthoFinder is fast and scales well to hundreds…; Performance Testing

  4. ShownA hybrid duplication-loss-coalescent (DLC) algorithm, reconciling only nodes with overlapping species sets via pre-computed parsimonious sub-case solutions, improved duplication (hence ortholog) calling over the species-overlap method while staying fast enough for genome-scale tree sets.

    On simulated gene trees, F-score rose from 72% (species-overlap alone) to 80% (hybrid); only DLCpar (full) was more accurate (reported F-score 91.8%, against 75.5% for species-overlap in that comparison). The hybrid handled 18651 trees (948449 genes) from 128 fungi in 141s, 8 and 187 times faster than Notung and DLCpar (search); DLCpar (full) did not finish a 4-species set in 120 hours. · OrthoFinder implements a novel…; Ortholog inference and identification of gene…; Simulation Tests of OrthoFinder Gene Duplication…

  5. ShownAlthough fully automated, OrthoFinder outscored static ortholog databases built with varying degrees of human curation when scores were averaged over the four Quest for Orthologs tests.

    Every OrthoFinder configuration had a higher mean F-score/pseudo F-score than each database; the default was 2% to 17% higher and the BLAST plus alignment-based tree version 5% to 20% above the databases (Supplementary Fig. 2). · OrthoFinder has the highest orthology inference…

Methods, data and reproducibility

Pipeline: DIAMOND all-vs-all search and orthogroup inference; DendroBLAST (or alignment plus maximum-likelihood) trees per orthogroup; STAG species tree from all orthogroup trees, rooted by STRIDE; gene tree rooting by outgroup and ancient-duplication bipartition scores; then a hybrid species-overlap/DLC pass calling orthologs and duplications.

  • Quest for Orthologs benchmark reference proteomes across eukaryotes, bacteria and archaea; orthologs scored by SwissTree, TreeFam-A, STDT and GSTDT tests · multiple species (Eukarya, Bacteria, Archaea) · n = 64 reference proteomes (Fig. 4 legend: 66 species) · 10.5281/zenodo.1481147
  • Fungal proteomes from Ensembl Genomes (one species per genus), randomly subsampled into 4- to 256-species sets with the same last common ancestor, for runtime tests · fungi (multiple species) · n = 272 species; subsets of 4, 8, 16, 32, 64, 128 and 256 · 10.5281/zenodo.1481147
  • OrthoFinder orthogroup trees from the 4- to 128-species fungal sets, used to time ortholog and duplication inference methods · fungi (multiple species) · n = 18651 trees, 948449 genes (128-species set)
  • Simulated 'flies' gene trees from a known species tree with divergence, duplication, loss and population parameters modelling incomplete lineage sorting; rooted trees supplied · simulated flies · n = 12,000 trees; 12 species; 12,032 duplication events
  • Simulated 'primates' gene trees from a known species tree, varying effective population size and duplication rate; rooted trees supplied · simulated primates · n = 7,500 trees; 17 species; 16,066 duplication events
  • Simulated 'metazoa' alignments with heterogeneous duplication/loss rates and complex sequence evolution; trees inferred with FastTree, rooted with reconroot · simulated metazoa · n = 2,000 gene trees; 40 species; 4,967 duplication events
  • Ten chordate proteomes (Ciona intestinalis to Homo sapiens) run with default settings to illustrate standard OrthoFinder outputs · chordates (e.g., Homo sapiens, Mus musculus, Danio rerio) · n = 10 species · 10.5281/zenodo.1481147
Methods & tools
OrthoFinder (original orthogroup algorithm plus new tree-based steps), DIAMOND (default all-versus-all sequence search), BLAST and MMseqs2 (supported alternative searches), DendroBLAST (default orthogroup gene trees), MAFFT and FastTree (optional alignment and maximum-likelihood trees), STAG (Species Tree from All Genes), STRIDE (species tree root inference from duplication events), DLCpar (full and search), Notung and GSDI Forester (duplication-loss reconciliation), Species-overlap ortholog method, SonicParanoid and other user-runnable Quest for Orthologs methods, Quest for Orthologs benchmark service (SwissTree, TreeFam-A, STDT, GSTDT; RAxML)
Software & environment
OrthoFinder options: default (DIAMOND, DendroBLAST), '-S blast', '-S blast -M msa' (MAFFT, FastTree); runtime runs '-S diamond -t 16 -a 16'; JSON config adds search/alignment/tree tools; DLC tests via scripts/resolve.py ('--no_resolve' = species-overlap).
Assay conditions & compute
Runtime benchmarks on a single Intel E5-2640v3 Haswell node (16 cores, 16 threads) of Oxford University's ARCUS-B server, with a 120 hour per-run cut-off; reconciliation timings run in parallel on 16 threads via GNU Parallel.
Reproducibility gaps
Exact OrthoFinder release and third-party software versions used for benchmarks are not stated; benchmark data tables (Supplementary Tables 1 and 2) and supplementary figures are not part of the provided text.
Limitations (6)
  • The authors concede DLCpar (full) was the most accurate duplication caller on simulated data (F-score 91.8%) and built the hybrid because DLCpar (full) could not finish a 4-species dataset within 120 hours, so the hybrid is most accurate only among scalable methods. · stated by the authors
  • Discordance benchmarks count any gene tree/species tree conflict as an error, so OrthoFinder can be penalized when DLC parsimony calls discordant genes orthologs; the authors argue this likely penalizes every method for some correct predictions. · stated by the authors
  • Only three configurations (DIAMOND/DendroBLAST, BLAST/DendroBLAST, BLAST with MAFFT/FastTree) were benchmarked; the authors note testing every alignment and tree-inference combination is infeasible, and the default is the fastest but least accurate variant. · stated by the authors
  • Gene tree rooting relies on the inferred rooted species tree, and correct rooting is critical for ortholog calls; the authors recommend supplying a known rooted species tree (Newick) to remove species tree inference error. · stated by the authors
  • Some reported values conflict: the species-overlap method is assigned F-scores of 75.5% and 72% without clarifying the datasets behind each, and the Quest for Orthologs set is 64 proteomes in Methods but 66 species in the Fig. 4 legend. · noted in this summary
  • STAG and STRIDE accuracy is cited from separate studies, and duplication simulations supplied pre-rooted gene trees, so the species tree and gene tree rooting steps are not separately quantified here; runtime scaling was tested only on fungal proteomes. · noted in this summary

Read this if you need genome-scale orthologs, gene trees, a rooted species tree or duplication events from proteomes, or are weighing phylogenetic against similarity-score orthology tools.

Citations
7,675
FWCI
168.51
Field percentile
top 1%

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 '25Emms '20Emms '22Steenwyk '24Tsai '26Rapoport '17
87% semantic matchBioinformatics
David Emms, Yi Liu, Laurence J. Belcher+2 more·bioRxiv·Jul 15, 2025· 44 citations