Pierre-Alain ChaumeilpersName, Aaron J. MussigpersName, Philip HugenholtzpersName, Donovan H. ParkspersName (2022date). “GTDB-Tk v2: memory friendly classification with the genome taxonomy database.” Bioinformaticstitle·j.
Object of study —
Research question —
Contribution —
phenomenon · claim-type · cross-domain-transfer · mechanism · claim-relation · replication-status · limitation — grounded in the full text| phenomenon | substantial memory requirements of the GTDB-Tk classification toolkit caused by the growing GTDB bacterial reference tree — “However, the growing size of the GTDB bacterial reference tree has resulted in GTDB-Tk requiring substantial amounts of memory (∼320 GB) which limits its adoption and ease of use.” | |
| claim-type | method — “Here, we present an update to GTDB-Tk that uses a divide-and-conquer approach where user genomes are initially placed into a bacterial reference tree with family-level representatives followed by placement into an appropriate class-level subtree comprising species representatives.” | |
| contribution — evidence | — “Here, we present an update to GTDB-Tk that uses a divide-and-conquer approach where user genomes are initially placed into a bacterial reference tree with family-level representatives followed by placement into an appropriate class-level subtree comprising species representatives.” | |
| extends | GTDB-Tk (previous version) — “Here, we present an update to GTDB-Tk that uses a divide-and-conquer approach where user genomes are initially placed into a bacterial reference tree with family-level representatives followed by placement into an appropriate class-level subtree comprising species representatives.” | claim_relations[0] |
| sense | backbone tree — a tree containing one genome representative for each family — “First, a genome is placed into a backbone tree consisting of a single genome representative for each family (see Supplementary Methods ).” | senses[0] |
| effect direction | divide-and-conquer approach —causes→ memory requirements — reduces — “The divide-and-conquer approach implemented in GTDB-Tk v2 reduced the maximum memory requirements from ∼320 GB to <55 GB when run with the GTDB R07-RS207 reference trees.” — “reduced the maximum memory requirements” | concept_edges[0] |
| note | divide-and-conquer approach —causes→ memory requirements — The sentence measures the reduction from approximately 320 GB to below 55 GB under the divide-and-conquer approach; the large quantified effect sets 0.96. — “The divide-and-conquer approach implemented in GTDB-Tk v2 reduced the maximum memory requirements from ∼320 GB to <55 GB when run with the GTDB R07-RS207 reference trees.” | concept_edges[0] |
| weight | divide-and-conquer approach —causes→ memory requirements — 0.96 — “The divide-and-conquer approach implemented in GTDB-Tk v2 reduced the maximum memory requirements from ∼320 GB to <55 GB when run with the GTDB R07-RS207 reference trees.” | concept_edges[0] |
| note | class-level subtrees —part-of→ bacterial reference tree — The sentence says the bacterial reference tree is divided into class-level subtrees, directly licensing part-of; 0.95 reflects categorical structure without using a perfect weight. — “Here, we show that the memory requirements of GTDB-Tk are reduced by dividing the GTDB bacterial reference tree into class-level subtrees and demonstrate that taxonomic classifications are largely unimpacted by this change.” | concept_edges[1] |
| weight | class-level subtrees —part-of→ bacterial reference tree — 0.95 — “Here, we show that the memory requirements of GTDB-Tk are reduced by dividing the GTDB bacterial reference tree into class-level subtrees and demonstrate that taxonomic classifications are largely unimpacted by this change.” | concept_edges[1] |
| audit | maximum likelihood (ML) placement tool pplacer —causes→ memory requirements — Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span writes 'memory-intensive operation requiring ~320 GB of RAM', not 'memory requirements'. — “Unfortunately, ML placement with pplacer is a memory-intensive operation requiring ∼320 GB of RAM when using the GTDB R07-RS207 bacterial reference tree comprised of 62 291 genomes.” | concept_edges[2] |
| note | maximum likelihood (ML) placement tool pplacer —causes→ memory requirements — The sentence calls pplacer placement memory-intensive and measures approximately 320 GB of RAM, directly setting a high causal weight of 0.97. — “Unfortunately, ML placement with pplacer is a memory-intensive operation requiring ∼320 GB of RAM when using the GTDB R07-RS207 bacterial reference tree comprised of 62 291 genomes.” | concept_edges[2] |
| weight | maximum likelihood (ML) placement tool pplacer —causes→ memory requirements — 0.97 — “Unfortunately, ML placement with pplacer is a memory-intensive operation requiring ∼320 GB of RAM when using the GTDB R07-RS207 bacterial reference tree comprised of 62 291 genomes.” | concept_edges[2] |
| audit | GTDB-Tk —operationalised-as→ divide-and-conquer approach — Audit 2026-09-28: DOUBTFUL - read against its span and kept, but GTDB-Tk USES a divide-and-conquer approach; operationalised-as is the wrong type. — “Here, we present an update to GTDB-Tk that uses a divide-and-conquer approach where user genomes are initially placed into a bacterial reference tree with family-level representatives followed by placement into an appropriate class-level subtree comprising species representatives.” | concept_edges[3] |
| note | GTDB-Tk —operationalised-as→ divide-and-conquer approach — The contribution states that the GTDB-Tk update uses the divide-and-conquer approach and describes its two placements; this explicit implementation sets 0.97. — “Here, we present an update to GTDB-Tk that uses a divide-and-conquer approach where user genomes are initially placed into a bacterial reference tree with family-level representatives followed by placement into an appropriate class-level subtree comprising species representatives.” | concept_edges[3] |
| weight | GTDB-Tk —operationalised-as→ divide-and-conquer approach — 0.97 — “Here, we present an update to GTDB-Tk that uses a divide-and-conquer approach where user genomes are initially placed into a bacterial reference tree with family-level representatives followed by placement into an appropriate class-level subtree comprising species representatives.” | concept_edges[3] |
finding · quantity · variable · definition · key-term — grounded in the full text| finding | Here, we show that the memory requirements of GTDB-Tk are reduced by dividing the GTDB bacterial reference tree into class-level subtrees and demonstrate that taxonomic classifications are largely unimpacted by this change. — “Here, we show that the memory requirements of GTDB-Tk are reduced by dividing the GTDB bacterial reference tree into class-level subtrees and demonstrate that taxonomic classifications are largely unimpacted by this change.” | content_tags.findings[0] |
| finding | This substantially reduces the memory requirements of GTDB-Tk while having minimal impact on classification. — “This substantially reduces the memory requirements of GTDB-Tk while having minimal impact on classification.” | content_tags.findings[1] |
| finding | 3.2 Reduced memory requirements The divide-and-conquer approach implemented in GTDB-Tk v2 reduced the maximum memory requirements from ∼320 GB to <55 GB when run with the GTDB R07-RS207 reference trees. — “The divide-and-conquer approach implemented in GTDB-Tk v2 reduced the maximum memory requirements from ∼320 GB to <55 GB when run with the GTDB R07-RS207 reference trees.” | content_tags.findings[2] |
| finding | If the genome is assigned to a class within this backbone tree, it is then placed into a class-level subtree to obtain a more refined placement for the genome. — “If the genome is assigned to a class within this backbone tree, it is then placed into a class-level subtree to obtain a more refined placement for the genome.” | content_tags.findings[3] |
| finding | A genome placed into a class-level subtree and assigned to a phylum belonging to one of the classes contained in the subtree is assigned the classification determined in the subtree. — “A genome placed into a class-level subtree and assigned to a phylum belonging to one of the classes contained in the subtree is assigned the classification determined in the subtree.” | content_tags.findings[4] |
| finding | Only 12 genomes (0.07%) did not have identical classifications between GTDB-Tk v1 and the divide-and-conquer approach used in GTDB-Tk v2 ( Supplementary Table S1 ). — “Only 12 genomes (0.07%) did not have identical classifications between GTDB-Tk v1 and the divide-and-conquer approach used in GTDB-Tk v2 ( Supplementary Table S1 ).” | content_tags.findings[5] |
| finding | 4 Summary GTDB-Tk v2 requires only a sixth of the memory of GTDB-Tk v1 while providing almost identical classifications. — “4 Summary GTDB-Tk v2 requires only a sixth of the memory of GTDB-Tk v1 while providing almost identical classifications.” | content_tags.findings[6] |
| quantity | ~320 GB RAM; reference tree of 62,291 genomes (printed '62 291') — reported sample, observation, or event count — “Unfortunately, ML placement with pplacer is a memory-intensive operation requiring ∼320 GB of RAM when using the GTDB R07-RS207 bacterial reference tree comprised of 62 291 genomes.” | content_tags.quantities[0] |
| quantity | at most 10,540 genomes (printed '10 540'); 10% more than the Gammaproteobacteria — reported sample, observation, or event count — “Each class-level subtree was formed by selecting the largest class in the reference tree and traversing towards the root until the subtree contained at most 10 540 genomes (10% more than the Gammaproteobacteria ).” | content_tags.quantities[1] |
| quantity | ~60,000 genomes (printed '60 000') dereplicated to 23,548 (printed '23 548') — reported sample, observation, or event count — “GTDB-Tk v1 and v2 classifications were further evaluated by dereplicating the ∼60 000 genomes introduced in GTDB R07-RS207 to 23 548 genomes by randomly selecting a single genome per species.” | content_tags.quantities[2] |
| quantity | 13 genomes (0.06%): 5 conflicting, 5 overclassified, 3 underclassified — reported sample, observation, or event count — “Only 13 genomes (0.06%) had different GTDB-Tk v1 and GTDB-Tk v2 classifications with 5 having conflicting assignments, 5 being overclassified and 3 being underclassified ( Supplementary Table S3 ).” | content_tags.quantities[3] |
| quantity | 22–35%; 1000; 1–64; >40%; 5000 — reported sample, observation, or event count — “GTDB-Tk v2 also ran 22–35% faster when processing 1000 genomes with 1–64 CPUs ( Supplementary Fig. S1A ) and was >40% faster when processing 5000 genomes using 32 CPUs ( Supplementary Fig. S1B ).” | content_tags.quantities[4] |
| concept | GTDB-Tk — “The Genome Taxonomy Database (GTDB) and associated taxonomic classification toolkit (GTDB-Tk) have been widely adopted by the microbiology community.” | concepts[0] |
| concept | divide-and-conquer approach — “3 Results Here, we demonstrate that the taxonomic classifications produced by the divide-and-conquer approach implemented in GTDB-Tk v2 are nearly equivalent to those produced by GTDB v1 while providing a substantial reduction in required memory.” | concepts[1] |
| concept | class-level subtrees — “The class-level subtrees were constructed in a greedy manner with the maximum size of a class-level subtree being set based on the number of species representatives in the largest class ( Gammaproteobacteria with 9582 genomes in GTDB R07-RS207).” | concepts[2] |
| concept | backbone tree — “First, a genome is placed into a backbone tree consisting of a single genome representative for each family (see Supplementary Methods ).” | concepts[3] |
| concept | lowest common ancestor — “Otherwise, the genome is classified by taking the lowest common ancestor between the backbone and class-level subtree.” | concepts[4] |
| concept | bacterial reference tree — “However, the growing size of the GTDB bacterial reference tree has resulted in GTDB-Tk requiring substantial amounts of memory (∼320 GB) which limits its adoption and ease of use.” | concepts[5] |
| concept | relative evolutionary divergence (RED) — “These classifications are consistent with the GTDB framework and based on the same relative evolutionary divergence (RED) and average nucleotide identity (ANI) criteria for circumscribing taxa ( Parks et al. , 2020 , 2022 ).” | concepts[6] |
| concept | maximum likelihood (ML) placement tool pplacer — “A primary step in assigning classifications is placing genomes into the GTDB bacterial or archaeal reference trees using the maximum likelihood (ML) placement tool pplacer ( Matsen et al. 2010 ).” | concepts[7] |
| concept | memory requirements — “Unfortunately, the size of this tree results in the memory requirements of GTDB-Tk being impractical.” | concepts[8] |
| concept | query genomes — “Each class-level subtree was then expanded to contain a single genome from each phylum in order to allow query genomes to be placed as the most basal member of a class.” | concepts[9] |
| concept | taxonomic rank — “The majority of incongruence was due to genomes being over- (six genomes) or under-classified (four genomes) by a single taxonomic rank.” | concepts[10] |
| concept | GTDB reference packages — “These 23 548 genomes were then classified using the GTDB-Tk R06-RS202 reference package to further evaluate classifications on genomes with varying degrees of taxonomic novelty and to ensure results were robust with different GTDB reference packages ( Supplementary Table S2 ).” | concepts[11] |
| concept | computing hardware — “This ensures GTDB-Tk can continue to be used on readily available computing hardware even as the size of the GTDB bacterial reference tree increases.” | concepts[12] |
Full text — tagged
The reference list is not part of the text we hold for this paper. Our copy was extracted from the publisher's page and the bibliography did not survive that extraction, so the article below ends before its references. Nothing has been removed by us.
Abstract. Abstract Summary The Genome Taxonomy Database (GTDB) and associated taxonomic classification toolkit (GTDB-Tkterm·mth) have been widely adopted by the microbiology community. However, the growing size of the GTDB bacterial reference treeterm·mth has resulted in GTDB-Tkterm·mth requiring substantial amounts of memory (∼320 GB) which limits its adoption and ease of use. Here, we present an update to GTDB-Tkterm·mth that uses a divide-and-conquer approachterm·mth where user genomes are initially placed into a bacterial reference treeterm·mth with family-level representatives followed by placement into an appropriate class-level subtreeterm·mth comprising species representatives. This substantially reduces the memory requirementsterm·mth of GTDB-Tkterm·mth while having minimal impact on classification. Availability and implementation GTDB-Tkterm·mth is implemented in Python and licenced under the GNU General Public Licence v3.0. Source code and documentation are available at: https://github.com/ecogenomics/gtdbtkterm·mth . Supplementary information Supplementary data are available at Bioinformatics online. 1 Introduction The Genome Taxonomy Database (GTDB) and associated taxonomic classification toolkit (GTDB-Tkterm·mth) has been used to assign taxonomic classifications to tens of thousands of bacterial and archaeal isolate genomes and metagenome-assemble genomes recovered from environmental and human-associated samples ( Almeida et al. , 2021 ; ChaumeilpersName et al. , 2019 ; Nayfach et al. , 2021 ). These classifications are consistent with the GTDB framework and based on the same relative evolutionary divergence (RED)term·mth and average nucleotide identity (ANI) criteria for circumscribing taxa ( ParkspersName et al. , 2020 , 2022 ). A primary step in assigning classifications is placing genomes into the GTDB bacterial or archaeal reference trees using the maximum likelihood (ML) placement tool pplacerterm·mth ( Matsen et al. 2010 ). Unfortunately, ML placement with pplacer is a memory-intensive operation requiring ∼320 GB of RAM when using the GTDB R07-RS207 bacterial reference treeterm·mth comprised of 62 291 genomes. Adding to this challenge is the lack of favourable alternatives to pplacer as the EPA-ng ML placement method requires more memory than pplacer and distance-based methods such as APPLES-2 have inferior performance ( Balaban et al. , 2022 ; Barbera et al. , 2019 ; Koning et al. , 2021). The GTDB bacterial reference treeterm·mth has been growing rapidly in size with each GTDB release and is expected to grow by upwards of 30% per year for the next few years ( ParkspersName et al. , 2022 ). Unfortunately, the size of this tree results in the memory requirementsterm·mth of GTDB-Tkterm·mth being impractical. Here, we show that the memory requirementsterm·mth of GTDB-Tkterm·mth are reduced by dividing the GTDB bacterial reference treeterm·mth into class-level subtreesterm·mth and demonstrate that taxonomic classifications are largely unimpacted by this change. 2 Materials and methods GTDB-Tkterm·mth v2 divides the GTDB bacterial reference treeterm·mth into class-level subtreesterm·mth to reduce memory requirementsterm·mth. Placement of a genome with pplacer now consists of two steps. First, a genome is placed into a backbone treeterm·mth consisting of a single genome representative for each family (see Supplementary Methods ). If the genome is assigned to a class within this backbone treeterm·mth, it is then placed into a class-level subtreeterm·mth to obtain a more refined placement for the genome. The class-level subtreesterm·mth were constructed in a greedy manner with the maximum size of a class-level subtreeterm·mth being set based on the number of species representatives in the largest class ( Gammaproteobacteria with 9582 genomes in GTDB R07-RS207). Each class-level subtreeterm·mth was formed by selecting the largest class in the reference tree and traversing towards the root until the subtree contained at most 10 540 genomes (10% more than the Gammaproteobacteria ). This subtree was then pruned from the reference tree and the procedure repeated until all classes were assigned to a class-level subtreeterm·mth. For GTDB R07-RS207, this resulted in seven class-level trees. Each class-level subtreeterm·mth was then expanded to contain a single genome from each phylum in order to allow query genomesterm·mth to be placed as the most basal member of a class. Final taxonomic classifications use the same RED and ANI criterion as GTDB-Tkterm·mth v1 (ChaumeilpersName et al. , 2020) with the following additional rules: A genome not placed into a class-level subtreeterm·mth is assigned the classification determined in the backbone treeterm·mth. A genome placed into a class-level subtreeterm·mth and assigned to a phylum belonging to one of the classes contained in the subtree is assigned the classification determined in the subtree. Otherwise, the genome is classified by taking the lowest common ancestorterm·mth between the backbone and class-level subtreeterm·mth. 3 Results Here, we demonstrate that the taxonomic classifications produced by the divide-and-conquer approachterm·mth implemented in GTDB-Tkterm·mth v2 are nearly equivalent to those produced by GTDB v1 while providing a substantial reduction in required memory. 3.1 Similarity of classifications on diverse sets of genomes The concordance between GTDB-Tkterm·mth v1 and v2 classifications was first assessed using 16 710 bacterial genomes from the GEMs dataset ( Nayfach et al. , 2021 ) that represent novel taxa relative to GTDB R07-RS207 ( Table 1 ). Only 12 genomes (0.07%) did not have identical classifications between GTDB-Tkterm·mth v1 and the divide-and-conquer approachterm·mth used in GTDB-Tkterm·mth v2 ( Supplementary Table S1 ). The majority of incongruence was due to genomes being over- (six genomes) or under-classified (four genomes) by a single taxonomic rankterm·mth. Only two genomes had conflicting taxonomic assignments, and these were both relatively poor-quality genomes assigned as new classes in alternative phyla ( Supplementary Table S1 ). Table 1. Novelty of GEM genomes relative to GTDB R07-RS207 based on GTDB-Tkterm·mth v1 classifications GTDB-Tkterm·mth v2 classifications relative to GTDB-Tkterm·mth v1 classifications Taxon novelty No. genomes Congruent Conflict a Underclassified b Overclassified c Novel phylum 3 2 0 0 1 Novel class 42 36 2 2 2 Novel order 144 143 0 0 1 Novel family 543 540 0 1 2 Novel genus 3222 3221 0 1 0 Novel species 12 756 12 756 0 0 0 Note : GTDB-Tkterm·mth v2 predicts a different taxon (a), or less (b) or more (c) resolved classifications than GTDB-Tkterm·mth v1 (see Supplementary Table S1 ). GTDB-Tkterm·mth v1 and v2 classifications were further evaluated by dereplicating the ∼60 000 genomes introduced in GTDB R07-RS207 to 23 548 genomes by randomly selecting a single genome per species. These 23 548 genomes were then classified using the GTDB-Tkterm·mth R06-RS202 reference package to further evaluate classifications on genomes with varying degrees of taxonomic novelty and to ensure results were robust with different GTDB reference packagesterm·mth ( Supplementary Table S2 ). Only 13 genomes (0.06%) had different GTDB-Tkterm·mth v1 and GTDB-Tkterm·mth v2 classifications with 5 having conflicting assignments, 5 being overclassified and 3 being underclassified ( Supplementary Table S3 ). 3.2 Reduced memory requirementsterm·mth The divide-and-conquer approachterm·mth implemented in GTDB-Tkterm·mth v2 reduced the maximum memory requirementsterm·mth from ∼320 GB to <55 GB when run with the GTDB R07-RS207 reference trees. GTDB-Tkterm·mth v2 also ran 22–35% faster when processing 1000 genomes with 1–64 CPUs ( Supplementary Fig. S1A ) and was >40% faster when processing 5000 genomes using 32 CPUs ( Supplementary Fig. S1B ). 4 Summary GTDB-Tkterm·mth v2 requires only a sixth of the memory of GTDB-Tkterm·mth v1 while providing almost identical classifications. More importantly, the divide-and-conquer approachterm·mth used in GTDB-Tkterm·mth v2 allows memory requirementsterm·mth to be controlled by tailoring the size of the largest subtree. This ensures GTDB-Tkterm·mth can continue to be used on readily available computing hardwareterm·mth even as the size of the GTDB bacterial reference treeterm·mth increases. Supplementary Material btac672_Supplementary_Data Click here for additional data file. Acknowledgements We thank Morgan Price for insightful advice on using FastTree with pplacer, Brian Kemish for his help in maintaining our computing infrastructure, Maria Chuvochina and Christian Rinke for helpful discussions, and the GTDB-Tkterm·mth community for their bug reports and suggestions on features to improve GTDB-Tkterm·mth. Funding This work was supported by UQ Strategic Funding and Australian Research Council Laureate Fellowship [FL150100038]. Conflict of Interest : none declared. Contributor Information Pierre-Alain ChaumeilpersName, Australian Centre for Ecogenomics, School of Chemistry and Molecular Biosciences, The University of Queensland, St Lucia, QLD 4072, Australia; Research Computing Center, The University of Queensland, St Lucia, QLD 4072, Australia. Aaron J MussigpersName, Australian Centre for Ecogenomics, School of Chemistry and Molecular Biosciences, The University of Queensland, St Lucia, QLD 4072, Australia. Philip HugenholtzpersName, Australian Centre for Ecogenomics, School of Chemistry and Molecular Biosciences, The University of Queensland, St Lucia, QLD 4072, Australia. Donovan H ParkspersName, Australian Centre for Ecogenomics, School of Chemistry and Molecular Biosciences, The University of Queensland, St Lucia, QLD 4072, Australia.