henhaobio
Sign in

Nature CommunicationsBioinformaticsPeer-reviewed journal article

High-throughput ANI Analysis of 90K Prokaryotic Genomes Reveals Clear Species Boundaries

Chirag JainiD, Luis M. Rodriguez‐RiD, Adam M. PhillippyiD, Konstantinos T. KonstantinidisiD, Srinivas AluruiD

PDFDOIPreprintCodeData

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 1 (27 Nov 2017) · CC BY 4.0

FastANI accurately estimates average nucleotide identity (ANI) up to 782-fold faster than BLAST; across 90K prokaryotic genomes, an 83-95% ANI gap supports a 95% species boundary.

Whether prokaryotes form discrete species or a genetic continuum had been argued from small, cultivation-biased genome sets, and testing it at scale was blocked because BLAST-based ANI scales poorly to the ~90K available genomes while MinHash-based Mash is inaccurate for draft genomes and at 80-90% ANI. FastANI makes all-versus-all ANI over the NCBI prokaryotic collection tractable, enabling a database-scale test of the 95% ANI species threshold.

Key results

  1. ShownAcross essentially all NCBI prokaryotic genomes, pairwise ANI values cluster above 95% or below 83%, leaving a wide, sparsely populated 83-95% interval that separates within-species from between-species comparisons.

    8.01 billion comparisons among 89,499 genomes (91,761 assemblies minus 2,262 with N50 < 10 Kbp); 679,765,100 pairs gave 76-100% ANI, of which only 17,132,536 (2.5%) fell at 83-95%, equal to 0.21% of all pairs; overall 99.8% of pairs were above 95% or below 83%. · Abstract; Large-scale Pairwise Comparison Indicates…; Figure legends

  2. ShownA 95% ANI cutoff reproduces existing species names with high recall and precision, and most disagreements reflect documented taxonomic conflicts, chiefly Escherichia coli versus Shigella, rather than ANI misestimation.

    Recall 98.5% and precision 93.1% for same named species; of the 6.9% of ≥95% pairs spanning different names, 5.6% were E. coli vs. Shigella spp., 0.5% Mycobacterium (M. tuberculosis complex) and 0.2% N. gonorrhoeae vs. N. meningitidis (interquartile range 94.9-95.2%); excluding E. coli/Shigella, precision was 98.7%. Species labels: 9% from NCBI taxonomy, 78% from GenBank names. · Large-scale Pairwise Comparison Indicates…; Figure legends

  3. ShownFastANI closely tracks BLASTn-based ANI (ANIb) on complete and draft genomes in the 80-100% range, whereas Mash, even at a sketch size of 100,000, departs from ANIb for divergent genomes, fragmented drafts and near-identical strains.

    Pearson r with ANIb > 0.99 on D1-D3 (RefSeq closed genomes; B. cereus and E. coli drafts); on D4 (B. anthracis, ANIb > 99.9 for all pairs) r was 0.681, rising to 0.944 after removing two misassembled genomes; Mash diverged on D1, D3 and D4 even at sketch size 100,000; FastANI-ANIb differences showed no trend with N50, genome length or reciprocal fragment count. · Accuracy; Figure legends; Tables

  4. ShownFastANI cuts ANI runtime by two to three orders of magnitude relative to ANIb and parallelizes by splitting the reference set into parts, making all-versus-all ANI across roughly 90K genomes computationally feasible.

    Single-thread speedup over ANIb (vs. FastANI compute time; indexing done once) ranged from 50x (D3, E. coli) to 782x (D1, RefSeq closed genomes); 80 parallel processes on a 40-core node cut compute time to 2, 8, 46 and 6 s for D1-D4; the 8.01 billion-pair NCBI run took 77K CPU hours vs. 190K CPU hours for the largest prior ANI study (86 million comparisons). · Computational Speedup; Large-scale Pairwise Comparison Indicates…; Figure legends; Tables

  5. ShownThe 83-95% ANI gap is stable through the database's history, persisting as sequencing priorities shifted from phylogenetically distant organisms toward closely related strains for microdiversity and epidemiological work.

    Density curves for four time periods of the NCBI collection each show two peaks on opposite sides of 95% ANI despite changing genus composition; a resampling that equalized expected representation of heavily and sparsely sampled named species also kept the gap, but is reported as data not shown. · Large-scale Pairwise Comparison Indicates…; Figure legends

Methods, data and reproducibility

FastANI splits the query into non-overlapping 3 Kbp fragments, maps them to the reference with Mashmap's winnowed-MinHash identity estimates (80% cutoff), and averages identities of reciprocal best matches (minimum 50). Benchmarked against BLASTn-based ANI and Mash, it was run all-versus-all on NCBI prokaryotic assemblies and compared with named species.

  • D1: closed prokaryotic genomes from NCBI RefSeq; one genome queried against every genome in the set · multiple prokaryotic species · n = 76 FastANI-reported matches to the query
  • D2: Bacillus cereus draft assemblies from the NCBI Genome prokaryote section, assemblies with N50 < 10 Kbp removed; single query genome · Bacillus cereus · n = 570 FastANI-reported matches to the query
  • D3: Escherichia coli draft assemblies from the NCBI Genome prokaryote section; single query genome; largest benchmark set (peak 48 GB memory) · Escherichia coli · n = 4,271 FastANI-reported matches to the query
  • D4: Bacillus anthracis draft assemblies of closely related strains (ANIb > 99.9 for all pairs); single query genome · Bacillus anthracis · n = 464 genomes vs. the query; two judged misassembled
  • NCBI Genome prokaryotic assemblies available on March 15, 2017, minus 2,262 with N50 < 10 Kbp, compared all-versus-all in both query/reference orders · prokaryotes (multiple species) · n = 89,499 of 91,761 assemblies; 8.01 billion comparisons
  • Subset of the 89,499 genomes with an assignable species name (NCBI taxonomy link or parsed GenBank organism name, ambiguous names excluded) for testing the 95% cutoff · prokaryotes (multiple species) · n = 9% via taxonomy links, 78% via names; 13% discarded
  • Genome pairs from the NCBI run plotted for both query/reference orientations to test order dependence; outliers checked for contamination with MyTaxaScan · prokaryotes (multiple species) · n = 451M pairs; 4,966 outlier pairs (difference ≥ 2)
  • Bartonella quintana (query) and Bartonella henselae (reference) genomes used to illustrate FastANI's visualization of reciprocal mappings with genoPlotR · Bartonella quintana; Bartonella henselae · NC_018533.1 (query); NC_005956.1 (reference)
Methods & tools
FastANI (alignment-free ANI estimator), Mashmap (winnowed-MinHash approximate sequence mapper), ANIb (BLASTn-based ANI, reference method), Mash (MinHash ANI; sketch sizes 1,000 to 100,000), Mauve (genome synteny alignment), genoPlotR (plots of reciprocal mappings), MyTaxaScan (assembly contamination check), Krona (taxonomic composition charts), NCBI RefSeq, NCBI Genome database (prokaryote assemblies), NCBI Taxonomy
Software & environment
FastANI built on Mashmap; compared against BLASTn-based ANIb and Mash at sketch sizes 1,000, 10,000 and 100,000; Mauve, genoPlotR, MyTaxaScan and Krona for diagnostics; FastANI settings I0 = 80%, l = 3 Kbp, τ = 50.
Assay conditions & compute
Single node with two Intel Xeon E5-2698 v4 20-core processors (40 physical cores), up to 80 parallel FastANI processes; peak memory 48 GB on D3; all-versus-all NCBI analysis took 77K CPU hours for 8.01 billion comparisons.
Reproducibility gaps
No software versions or containers are given for FastANI, Mash or BLAST; the species-balanced resampling analysis behind the claim of robustness to oversampled species is reported only as data not shown.
Limitations (6)
  • Scope is limited to 80-100% ANI: mapping uses an 80% identity cutoff and at least 50 reciprocal 3 Kbp fragments (at least 150 Kbp shared) must be found, so pairs far below 80% ANI go unscored; longer fragments cut runtime and memory but lower accuracy. · stated by the authors
  • Assembly errors and contamination distort estimates: two of 464 B. anthracis genomes in D4 diverged from ANIb owing to misassembly seen in Mauve, and 4,966 of 451M pairs showed an ANI difference of at least 2 when query and reference were swapped, traced to contaminated drafts. · stated by the authors
  • ANI does not strictly track core-genome evolutionary relatedness, since shared orthologs vary across pairs, and nomenclature conflicts (E. coli vs. Shigella, the M. tuberculosis complex, N. gonorrhoeae vs. N. meningitidis) prevent 95% ANI from matching every named species. · stated by the authors
  • Mechanisms producing the discontinuity are left unresolved: a steep fall in recombination near 90-95% ANI and competition-driven ecological sweeps are proposed as explanations but not tested. · stated by the authors
  • Accuracy versus ANIb was benchmarked with one query genome per dataset (76, 570, 4,271 and 464 FastANI-reported matches across D1-D4), so agreement for other query genomes and in all-versus-all mode is inferred rather than directly measured. · noted in this summary
  • The headline 99.8% uses all 8.01 billion pairs as denominator, most of which have no ANI value in the 76-100% window; among the 679,765,100 pairs that do, 2.5% (17,132,536) sit at 83-95%, so the gap is pronounced but not empty. · noted in this summary

Read this if you need fast, accurate ANI for thousands of complete or draft prokaryotic genomes, or want database-scale evidence on whether 95% ANI marks prokaryotic species boundaries.

Citations
5,748
FWCI
96.06
Field percentile
top 1%

Abstract

A fundamental question in microbiology is whether there is a continuum of genetic diversity among genomes or clear species boundaries prevail instead. Answering this question requires robust measurement of whole-genome relatedness among thousands of genomes and from diverge phylogenetic lineages. Whole-genome similarity metrics such as Average Nucleotide Identity (ANI) can provide the resolution needed for this task, overcoming several limitations of traditional techniques used for the same purposes. Although the number of genomes currently available may be adequate, the associated bioinformatics tools for analysis are lagging behind these developments and cannot scale to large datasets. Here, we present a new method, FastANI, to compute ANI using alignment-free approximate sequence mapping. Our analyses demonstrate that FastANI produces an accurate ANI estimate and is up to three orders of magnitude faster when compared to an alignment (e.g., BLAST)-based approach. We leverage FastANI to compute pairwise ANI values among all prokaryotic genomes available in the NCBI database. Our results reveal a clear genetic discontinuity among the database genomes, with 99.8% of the total 8 billion genome pairs analyzed showing either >95% intra-species ANI or <83% inter-species ANI values. We further show that this discontinuity is recovered with or without the most frequently represented species in the database and is robust to historic additions in the public genome databases. Therefore, 95% ANI represents an accurate threshold for demarcating almost all currently named prokaryotic species, and wide species boundaries may exist for prokaryotes.

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
THISMurray '20Larralde '25Rodriguez-R '24Ussery '26Hernández-S '23Gosselin '20
74% semantic matchMicrobiology
C. S. Murray, Y. Gao, M. Wu·bioRxiv·Jul 27, 2020