henhaobio
Sign in

bioRxivBioinformaticsPreprint

Extending and improving metagenomic taxonomic profiling with uncharacterized species with MetaPhlAn 4

Aitor Blanco‐MígueziD, Francesco BeghiniiD, Fabio CumboiD, Lauren J. McIveriD, Kelsey N. ThompsoniD, Moreno ZolfoiD, Paolo ManghiiD, Léonard DuboisiD, Kun D. HuangiD, Andrew Maltez ThomasiD, Gianmarco PiccinnoiD, Elisa PiperniiD, Michal PunčochářiD, Mireia Valles‐ColomeriD, Adrian TettiD, Francesca GiordanoiD, Richard DaviesiD, Jonathan WolfiD, Sarah E. BerryiD, Tim D. SpectoriD, Eric A. FranzosaiD, Edoardo PasolliiD, Francesco AsnicariD, Curtis HuttenhoweriD, Nicola SegataiD

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

Abstract

Metagenomic assembly enables novel organism discovery from microbial communities, but from most metagenomes it can only capture few abundant organisms. Here, we present a method - MetaPhlAn 4 - to integrate information from both metagenome assemblies and microbial isolate genomes for improved and more comprehensive metagenomic taxonomic profiling. From a curated collection of 1.01M prokaryotic reference and metagenome-assembled genomes, we defined unique marker genes for 26,970 species-level genome bins, 4,992 of them taxonomically unidentified at the species level. MetaPhlAn 4 explains [~]20% more reads in most international human gut microbiomes and >40% in less-characterized environments such as the rumen microbiome, and proved more accurate than available alternatives on synthetic evaluations while also reliably quantifying organisms with no cultured isolates. Application of the method to >24,500 metagenomes highlighted previously undetected species to be strong biomarkers for host conditions and lifestyles in human and mice microbiomes, and showed that even previously uncharacterized species can be genetically profiled at the resolution of single microbial strains. MetaPhlAn 4 thus integrates the novelty of metagenomic assemblies with the sensitivity and fidelity of reference-based analyses, providing efficient metagenomic profiling of uncharacterized species and enabling deeper and more comprehensive microbiome biomarker detection.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (22 Aug 2022) · CC BY 4.0

Marker genes for 26,970 species-level genome bins, 4,992 defined only by metagenome-assembled genomes, let MetaPhlAn 4 quantify uncharacterized species and beat MetaPhlAn 3, mOTUs and Bracken on synthetic benchmarks.

Reference-based profilers detect only species already catalogued in their databases, while metagenomic assembly finds new organisms but in complex communities recovers mostly the abundant ones. MetaPhlAn 4 combines the two by grouping 1.01M isolate genomes and metagenome-assembled genomes (MAGs) into species-level genome bins and mapping reads to bin-specific marker genes, so species without isolates become quantifiable.

Key results

  1. ShownOn 123 synthetic metagenomes scored with OPAL against NCBI taxonomy, MetaPhlAn 4 had higher species-level F1 and lower Bray-Curtis dissimilarity than MetaPhlAn 3, mOTUs 2.6 and Bracken 2.5. Compared with MetaPhlAn 3, it detected more true species with a modest rise in false positives.

    118 CAMI II plus 5 SynPhlAn-nonhuman samples (4.1B reads). Mean Bray-Curtis to gold standard 0.14 ± 0.07 vs 0.21 ± 0.12 for MetaPhlAn 3; true positives 83.02 ± 47.38 vs 70.62 ± 35.85; false positives 11.61 ± 8.4 vs 8.69 ± 4.89, 84.6% of them from SGB species-group labels. With an SGB-based gold standard: F1 0.96 ± 0.024, Bray-Curtis 0.034 ± 0.022. · MetaPhlAn 4 improves the performance of…; MetaPhlAn 4 synthetic evaluation

  2. SuggestedIn simulations, bins defined only by MAGs (uSGBs) were detected and quantified about as accurately as known-species bins (kSGBs). This held when the simulated genomes were absent from the database.

    65 synthetic metagenomes covering 12 human body sites, animal hosts and non-host environments, mixing kSGBs and uSGBs: uSGB mean F1 0.97 ± 0.02, on par with the kSGB mean F1 of 0.96 ± 0.024 from the SGB-based evaluation of the CAMI II/SynPhlAn samples. 5 further metagenomes built from MAGs and reference genomes not in the database: mean F1 0.98 ± 0.006. · MetaPhlAn 4 improves the performance of…; MetaPhlAn 4 synthetic evaluation

  3. ShownAcross 24.5k public metagenomes, MetaPhlAn 4 explained more reads than MetaPhlAn 3 in every environment. Gains were largest in non-human gut microbiomes, where detected uSGBs outnumbered known species except in non-human primates.

    Abstract: about 20% more reads explained in most human gut microbiomes and >40% in rumen. Mean fold increase in explained reads: 1.95 airways, 3.26 wild mice, up to 14.15 rumen, 6.65 ocean. 11,132 SGBs reached ≥1% prevalence in some environment, 3,527 (31.68%) of them uSGBs; uSGBs averaged 23.13% (s.d. 17.89%) of profile richness. · Abstract; MetaPhlAn 4 expands the profiled fraction of…; MetaPhlAn 4 application to human and non-human…

  4. SuggestedProfiling bins without isolates changed host-association results. Most diet biomarkers in mice were uSGBs, and MetaPhlAn 4 profiles improved random-forest prediction of most diet and cardiometabolic markers in ZOE PREDICT 1.

    Mouse cohort: 632 SGBs (60.8% uSGBs) vs 108 species with MetaPhlAn 3; mixed models found 18 diet biomarkers at FDR < 0.2, 13 (72%) of them uSGBs. ZOE PREDICT 1 (1,001 people): all but two of 19 markers better predicted, median AUC 0.74 (4.84% improvement); 10-year atherosclerotic cardiovascular disease risk +0.106 AUC; Healthy Eating Index +0.072 AUC. · Biomarkers of diet in mice are dominated by…; Stronger links between the human gut microbiome…; Differential abundance analysis of diet-related…; Associations between microbiome and host diet…

  5. SuggestedStrainPhlAn 4 extends strain-level phylogenies to uSGBs. For the health-associated Lachnospiraceae SGB4894, it resolved subclades linked to lifestyle and geography and found little mother-infant or household strain sharing.

    Vs StrainPhlAn 3: mean 1.33% higher correlation with MAG-based trees (100 samples, 3 kSGBs). SGB4894 tree: 1,683 samples, 37 markers; geographic vs genetic distance Spearman 0.505; no sharing in 21 mother-infant cases; 5.63% of positive household adults shared a strain. Higher in healthy controls than in Crohn's disease and ulcerative colitis in three cohorts each (e.g. p=2.50e-28, 1.85e-22). · Stronger links between the human gut microbiome…; StrainPhlAn 4 reconstructs large phylogenies of…; StrainPhlAn 4 evaluation; Figure legends

Methods, data and reproducibility

Isolate genomes from NCBI and MAGs from human, animal and environmental samples were filtered with CheckM, dereplicated and Mash-clustered at 5% genetic distance into species-level genome bins (SGBs). Core genes from UniRef90/Uniclust90 pangenomes were screened for SGB-specificity, giving 5.1M markers. Reads are mapped with Bowtie 2; abundance is a trimmed mean marker coverage, normalized within each sample.

  • Bacterial and archaeal isolate reference genomes from NCBI (the Pasolli et al. catalog plus GenBank genomes added through November 2020) · bacteria and archaea · n = 236,620 (169,111 after quality control and dereplication)
  • Metagenome-assembled genomes from human body sites, animal hosts (including 22 non-human primate species), soil, fresh water and ocean · bacteria and archaea · n = 771,528 (560,084 after quality control and dereplication)
  • CAMI II taxonomic profiling challenge synthetic metagenomes (five human body sites and murine gut) · simulated human-associated and mouse gut communities · n = 118
  • Author-built synthetic metagenomes: SynPhlAn-nonhuman, kSGB/uSGB sets across 12 environments, and sets from genomes outside the database · microbial (simulated) · n = 5 SynPhlAn-nonhuman; 65 kSGB/uSGB; 5 with held-out genomes
  • Public shotgun metagenomes from human body sites, non-human primates, mice, ruminants, soil, freshwater and ocean · multiple hosts and environments · n = 24,515 samples from 145 datasets
  • Human gut metagenomes with age metadata, plus disease case/control datasets, via curatedMetagenomicData · human · n = 19,468 samples from 86 datasets; 21 case/control datasets
  • Public mouse gut microbiome cohort, high-fat vs normal chow diet, 8 genetic backgrounds and 6 vendors · mouse · n = 184 (181 in Methods; 43 high-fat and 88 chow mice modeled)
  • ZOE PREDICT 1 gut metagenomes with diet and cardiometabolic phenotypes, UK · human · n = 1,001 individuals
Methods & tools
MetaPhlAn 4, StrainPhlAn 4, Bowtie 2, CheckM 1.1.4, Mash, PhyloPhlAn 3, Prokka 1.14, DIAMOND 0.9.24, MMseqs2, UniRef90, OPAL, curatedMetagenomicData
Software & environment
MetaPhlAn 4.beta.1 with the vJan21 marker database, compared against MetaPhlAn 3.0.12. Versions are given for most dependencies (e.g. CheckM 1.1.4, Bowtie 2 2.3.5.1, DIAMOND 0.9.24, Prokka 1.14, scikit-learn 0.22.2). Distributed via Bioconda and PIP.
Reproducibility gaps
No runtime or memory figures are reported. Public metagenomes are cited through the publications in Table S8 rather than accession numbers, and per-sample results are in supplementary tables not included in this text.
Limitations (6)
  • Environmental and other under-studied communities still contain many sequences outside current uSGBs. Soil is especially poorly covered (2,495 MAGs defining 26 uSGBs). · stated by the authors
  • Viral and eukaryotic microbes are largely not represented, because their genome architectures and quality-control needs differ from those of bacteria and archaea. · stated by the authors
  • Because SGBs are whole-genome OTU-like clusters, the trade-off between sensitivity and specificity when applying quality-control filters to identify real but rare taxa remains unresolved. · stated by the authors
  • OPAL scoring against NCBI taxonomy cannot credit merged SGB species groups, which penalizes MetaPhlAn 4 in the standard comparison; 84.6% of its false positives came from such labels. · stated by the authors
  • Accuracy for uSGBs comes only from synthetic metagenomes the authors built, mostly from genomes in their own catalog. Only 5 simulated samples used genomes outside the database. · noted in this summary
  • The mouse diet and ZOE PREDICT 1 findings each come from a single cohort, and the mouse biomarkers and ZOE uSGB correlations are called at a permissive FDR < 0.2. The mouse cohort is given as 184 samples in Results but 181 in Methods and Fig. 3. · noted in this summary

Read this if you run taxonomic profiling on shotgun metagenomes and want abundances, or strain trees, for species known only from MAGs, especially in mouse, rumen or other hosts with few cultured isolates.

Citations
89

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
THISRuscheweyh '21Deng '25Woodcroft '24Fritscher '26Li '22Pita '24
75% semantic matchBioinformatics
Hans‐Joachim Ruscheweyh, Alessio Milanese, Lucas Paoli+8 more·bioRxiv·Apr 20, 2021· 9 citations