OrthoFinder: improved phylogenetic orthology inference with enhanced accuracy and scalability

The original OrthoFinder algorithm applied a per-species-pair gene-length normalization that accounted for both evolutionary distance and gene-length dependence in pairwise bit scores between sequences.

“The original OrthoFinder algorithm applied a per-species-pair gene-length normalization that accounted for both evolutionary distance and gene-length dependence in pairwise bit scores between sequences12.”

• cites [12] Emms, D. M. & Kelly, S. OrthoFinder: phylogenetic orthology inference for comparative genomics. Genome Biol. 20, 238 (2019). sourcePartly supported

  The Background and Results ('OrthoFinder algorithm overview and summary of results files'; page unavailable) say the original implementation inferred orthogroups using sequence similarity scores. They do not describe per-species-pair gene-length normalization, bit-score dependence, or adjustment for evolutionary distance; searches for those terms found no supporting passage. (LLM, conf=0.79)

  Experimental — accuracy not yet measured

The OrthoBench dataset consists of expert-curated reference orthogroups.

“we compared this new version of OrthoFinder with a set of alternative methods on the OrthoBench dataset of expert-curated reference orthogroups20,21.”

• cites [20] Trachana, K. et al. Orthology prediction methods: a quality assessment using curated protein families. BioEssays 33, 769–780 (2011). sourceSupported
• cites [21] Emms, D. M. & Kelly, S. Benchmarking orthogroup inference accuracy: revisiting OrthoBench. Genome Biol. Evol. 12, 2258–2266 (2020). sourceSupported

The SHOOT profile algorithm enables rapid addition of new species to existing OrthoFinder runs.

“This is achieved through adaptation and further development of the SHOOT profile algorithm19, which enables rapid addition of new species to existing OrthoFinder runs.”

• cites [19] Emms, D. M. & Kelly, S. SHOOT: phylogenetic gene search and ortholog inference. Genome Biol. 23, 85 (2022). sourcePartly supported

  The Materials and methods sections on database preparation and search describe SHOOT profiles assigning novel gene sequences to precomputed trees for rapid phylogenetic placement. They do not say that the algorithm adds new species to existing OrthoFinder runs; OrthoFinder results are instead described as input to SHOOT database preparation. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

Achieving high orthology inference accuracy at scale is a major challenge facing method development in orthology inference.

“one of the major challenges facing method development in this field is how to achieve high orthology inference accuracy at scale7,8.”

• cites [7] Cosentino, S., Sriswasdi, S. & Iwasaki, W. SonicParanoid2: fast, accurate, and comprehensive orthology inference with machine learning and language models. Genome Biol. 25, 195 (2024). sourceSupported
• cites [8] Majidian, S. et al. Orthology inference at scale with FastOMA. Nat. Methods 22, 269–272 (2025). sourceSupported

In previous versions of OrthoFinder, orthogroups were defined solely based on MCL clustering of sequence similarity search results.

“In previous versions of OrthoFinder, orthologs were defined phylogenetically, whereas orthogroups were defined solely based on Markov Cluster Algorithm (MCL) clustering of sequence similarity search results12,16.”

• cites [12] Emms, D. M. & Kelly, S. OrthoFinder: phylogenetic orthology inference for comparative genomics. Genome Biol. 20, 238 (2019). sourcePartly supported

  The Results say sequence-similarity scores provide the raw data for orthogroup inference, and the Methods say the original OrthoFinder algorithm is used for that step (no page numbers available). Our copy does not say that the algorithm uses MCL clustering, so it does not support the claim's more specific description. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [16] Emms, D. M. & Kelly, S. OrthoFinder: solving fundamental biases in whole genome comparisons dramatically improves orthogroup inference accuracy. Genome Biol. 16, 157 (2015). sourceSupported

The entry point to the majority of orthology inference methods is an all-versus-all sequence similarity search.

“the entry point to the majority of inference methods is an all-versus-all sequence similarity search from which orthology relationships can be inferred12.”

• cites [12] Emms, D. M. & Kelly, S. OrthoFinder: phylogenetic orthology inference for comparative genomics. Genome Biol. 20, 238 (2019). sourceSupported

Widespread genome sequencing efforts are creating an unprecedented scalability challenge for automated orthology inference methods.

“These pioneering efforts, combined with widespread sequencing efforts that are distributed across many biological communities, are creating an unprecedented scalability challenge for automated orthology inference methods4,11.”

• cites [4] Linard, B. et al. Ten years of collaborative progress in the Quest for Orthologs. Mol. Biol. Evol. 38, 3033–3045 (2021). sourceSupported
• cites [11] Langschied, F. et al. Quest for Orthologs in the era of biodiversity genomics. Genome Biol. Evol. 16, evae224 (2024). sourceSupported

OrthoFinder v3 completed the 1,024-species run in 128 h.

“OrthoFinder v3 completed this run in 128 h.”

OrthoFinder v3 linear shows a 3.4-fold decrease in RAM consumption for 256 species compared with OrthoFinder v2 DendroBlast.

“For example, v3 linear shows a 3.4-fold decrease in RAM consumption for 256 species compared with v2 DendroBlast.”

The updated OrthoFinder method is capable of analyzing thousands of species using a two-step process.

“The updated OrthoFinder method is now capable of analyzing thousands of species using a two-step process.”

OrthoFinder v3 is on the Pareto frontier for species tree discordance tests in both eukaryotes and bacteria.

“OrthoFinder v3 is on the Pareto frontier for species tree discordance tests in both eukaryotes (Fig. 5a), and bacteria (Fig. 5b), demonstrating its ability to perform accurate orthology inference across different domains of life.”

Enhanced phylogenetic delineation of orthogroups provides a 5-7% relative increase in orthogroup inference accuracy compared with previous OrthoFinder.

“Specifically, we show that enhanced phylogenetic delineation of orthogroups provides a 7% relative increase in orthogroup inference accuracy.”

• Also stated in: § Improved phylogenetic delineation of orthogroups · plain

  “OrthoFinder v3 had 5–7% higher accuracy (f-score harmonic mean of recall and precision; Supplementary Table 1.1) than the previous version of OrthoFinder run with identical settings (OF2_Align_ST and OF2_DB; Fig. 2).”

The latest version of OrthoFinder is scalable to thousands of genomes on conventional computing resources.

“Finally, we show that the latest version of OrthoFinder is scalable to thousands of genomes on conventional computing resources.”

• AssumptionRests on an assumption: The Linux server used for benchmarking (AMD processors, 32 threads, 16 cores, up to 200 GB RAM) is representative of conventional computing resources.

On the TreeFam dataset, OrthoFinder v3 was marginally outperformed in recall by OrthoFinder v2 (0.72 versus 0.74).

“was only marginally outperformed in recall on the TreeFam dataset by OrthoFinder v2 (0.72 versus 0.74)”

Using the accelerated search functionality does not compromise orthology inference accuracy.

“We demonstrate that using such an accelerated search functionality does not compromise orthology inference accuracy.”

In the analysis presented here, OrthoFinder and FastOMA were the only orthology inference tools capable of analyzing datasets comprising thousands of genomes.

“In the analysis presented here, OrthoFinder and FastOMA were the only orthology inference tools capable of analyzing datasets comprising thousands of genomes.”

The spread of reference orthogroup fission events varied across tools, with low-precision tools such as Broccoli and OrthoHMM making relatively few RefOG fissions.

“The spread of reference orthogroup fissions events across the tools varied, with low-precision tools such as Broccoli and OrthoHMM making relatively few RefOG fissions.”

OrthoFinder v3 is on the Pareto frontier for the enzyme classification test.

“OrthoFinder v3 is on the Pareto frontier for the enzyme classification test (Fig. 5c), which measures how well predicted orthologs match to Enzyme Commission numbers (which are given depending on the chemical reaction they catalyze).”

OF3_Linear and FastOMA exhibited comparable trends in runtime reduction and memory scaling as thread counts increased.

“Both tools exhibited comparable trends in runtime reduction and memory scaling as thread counts increased.”

OrthoFinder v3 (in either implementation) outperformed all other orthogroup inference methods, including OrthoFinder v2, on the OrthoBench dataset.

“This revealed that OrthoFinder v3 (in either implementation) outperformed all other orthogroup inference methods including OrthoFinder v2 (Fig. 2).”

It is possible to rapidly assign gene sets from large numbers of species to a phylogenetically partitioned and structured database of biological sequences in near-linear time.

“We show that it is possible to rapidly assign gene sets from large numbers of species to a phylogenetically partitioned and structured database of biological sequences in near-linear time.”

Runtime decreased as threads increased up to 32, beyond which performance gains plateaued.

“Runtime decreased as threads increased up to 32, beyond which performance gains plateaued.”

OrthoFinder v3 followed a near linear runtime trend for 128-1,024 species.

“OrthoFinder v3 followed a near linear runtime trend for 128–1,024 species.”

The new OrthoFinder v3 workflow for high scalability does not compromise accuracy.

“Analysis of the resulting scores revealed that the new OrthoFinder v3 workflow for high scalability does not compromise accuracy (Fig. 5).”

Using phylogenetic delimitation improves the inference of orthogroups.

“We also show that using phylogenetic delimitation improves the inference of orthogroups”

It is possible to rapidly search and assign sequence sets from large numbers of species to a phylogenetically partitioned and structured database of biological sequences in near linear time.

“We show that it is possible to rapidly search and assign sequence sets from large numbers of species to a phylogenetically partitioned and structured database of biological sequences in near linear time.”

OrthoFinder v3 Linear only slightly underperformed compared with the nonscalable version, with slight decreases in some measures.

“OrthoFinder v3 Linear only slightly underperformed compared with the nonscalable version, with slight decreases in some measures observed.”

On the largest datasets that other methods could complete, OrthoFinder v3 used approximately 4-fold lower peak memory consumption.

“On the largest datasets that other methods could complete, OrthoFinder v3 used approximately 4-fold lower peak memory consumption (Fig. 4b).”

FastOMA erroneously fused 3.5 times more reference orthogroups than other methods (10% compared with 2.85%).

“Almost all tools were identical in the production of fused reference orthogroups (RefOG), with the exception of FastOMA, which erroneously fused 3.5 times more reference orthogroups than other methods (10% compared with 2.85%).”

OrthoFinder v3 achieved the highest recall of all methods on both the VGNC and SwissTree human-curated reference sets.

“OrthoFinder v3 achieved the highest recall of all methods on both the Vertebrate Gene Nomenclature Committee (VGNC) and SwissTree human-curated reference sets”

OrthoFinder v3 Linear was more accurate than any implementation of OrthoFinder v2 and any other competitor methods.

“However, this scalable version was more accurate than any implementation of OrthoFinder v2 and any other competitor methods (Fig. 2), supporting the use of this accelerated implementation of OrthoFinder in comparative genomic analyses.”

SonicParanoid2 (fast mode) and FastOMA were the only two methods able to run 512 proteomes within the cutoff time.

“SonicParanoid2 (fast mode) and FastOMA were the only two methods able to run 512 proteomes within the cutoff time.”

OrthoFinder tends to produce orthogroups whose sizes and membership most accurately recapture the reference dataset.

“Analysis of the entropy scores, which are similar to a weighted combination of fission and fusion events, revealed that OrthoFinder tends to produce orthogroups whose sizes and membership most accurately recapture the reference dataset (Fig. 2).”

Errors in orthogroup delineation propagate through all subsequent steps of the analysis, causing collateral errors in orthology inference.

“Errors in orthogroup delineation at this stage propagate through all subsequent steps of the analysis, causing collateral errors in orthology inference.”

The original per-species-pair gene-length normalization requires a large number of pairwise hits between the genes in each species to fit the required correction functions.

“However, this normalization step requires a large number of pairwise hits between the genes in each species to fit the required correction functions.”

OrthoFinder v3 exhibits high accuracy ortholog inference across multiple benchmark tests.

“In summary, similar to the high accuracy of its orthogroup inference, OrthoFinder v3 also exhibits high accuracy ortholog inference across multiple benchmark tests.”

Some genes fail to be assigned to any core orthogroups during the initial assignment step because newly added species sometimes contain new genes that arose de novo and are not found in the core species set.

“This occurs because the new species (or sets of species) being added will sometimes contain new genes that have arisen de novo and are not found in the core species set.”