ArticlePDF Available

Identification and assembly of genomes and genetic elements in complex metagenomic samples without using reference genomes

August 2014
Nature Biotechnology 32(8):822 - 828

DOI:10.1038/nbt.2939

Authors:

Mathieu Almeida

French National Institute for Agriculture, Food, and Environment (INRAE)

Agnieszka Sierakowska Juncker

Technical University of Denmark

Simon Rasmussen

University of Copenhagen

Show all 54 authorsHide

Most current approaches for analyzing metagenomic data rely on comparisons to reference genomes, but the microbial diversity of many environments extends far beyond what is covered by reference databases. De novo segregation of complex metagenomic data into specific biological entities, such as particular bacterial strains or viruses, remains a largely unsolved problem. Here we present a method, based on binning co-abundant genes across a series of metagenomic samples, that enables comprehensive discovery of new microbial organisms, viruses and co-inherited genetic entities and aids assembly of microbial genomes without the need for reference sequences. We demonstrate the method on data from 396 human gut microbiome samples and identify 7,381 co-abundance gene groups (CAGs), including 741 metagenomic species (MGS). We use these to assemble 238 high-quality microbial genomes and identify affiliations between MGS and hundreds of viruses or genetic entities. Our method provides the means for comprehensive profiling of the diversity within complex metagenomic samples.

Content uploaded by Joel Dore

Content may be subject to copyright.

nature biotechnology ADVANCE ONLINE PUBLICATION 1

ARTICLES

Natural microbial communities typically contain a wide diversity of

organisms, viruses, and other chromosomal and extra-chromosomal

genetic elements. The microbiome of the human distal gut is among

the most complex communities ever studied, with an estimated 1,000

different microbial species across human populations1 and millions

of different genes2. Current computational analysis strategies for

Most current approaches for analyzing metagenomic data rely on comparisons to reference genomes, but the microbial

diversity of many environments extends far beyond what is covered by reference databases. De novo segregation of complex

metagenomic data into specific biological entities, such as particular bacterial strains or viruses, remains a largely unsolved

problem. Here we present a method, based on binning co-abundant genes across a series of metagenomic samples, that enables

comprehensive discovery of new microbial organisms, viruses and co-inherited genetic entities and aids assembly of microbial

genomes without the need for reference sequences. We demonstrate the method on data from 396 human gut microbiome

samples and identify 7,381 co-abundance gene groups (CAGs), including 741 metagenomic species (MGS). We use these

to assemble 238 high-quality microbial genomes and identify affiliations between MGS and hundreds of viruses or genetic

entities. Our method provides the means for comprehensive profiling of the diversity within complex metagenomic samples.

Identification and assembly of genomes and genetic

elements in complex metagenomic samples without

using reference genomes

H Bjørn Nielsen

1,2,32

, Mathieu Almeida

3–5,32

, Agnieszka Sierakowska Juncker

1,2

, Simon Rasmussen

, Junhua Li

6–8

Shinichi Sunagawa

, Damian R Plichta

, Laurent Gautier

, Anders G Pedersen

, Emmanuelle Le Chatelier

3,4

Eric Pelletier

10–12

, Ida Bonde

1,2

, Trine Nielsen

, Chaysavanh Manichanh

, Manimozhiyan Arumugam

7,9,13

Jean-Michel Batto

3,4

, Marcelo B Quintanilha dos Santos

, Nikolaj Blom

, Natalia Borruel

, Kristoffer S Burgdorf

Fouad Boumezbeur

3,4

, Francesc Casellas

, Joël Doré

3,4

, Piotr Dworzynski

, Francisco Guarner

Torben Hansen

13,15

, Falk Hildebrand

16,17

, Rolf S Kaas

, Sean Kennedy

3,4

, Karsten Kristiansen

7,19

, Jens Roat Kultima

Pierre Léonard

3,4

, Florence Levenez

3,4

, Ole Lund

, Bouziane Moumen

3,4

, Denis Le Paslier

10–12

, Nicolas Pons

3,4

Oluf Pedersen

13,20–22

, Edi Prifti

3,4

, Junjie Qin

6,7

, Jeroen Raes

17,23,24

, Søren Sørensen

, Julien Tap

Sebastian Tims

, David W Ussery

, Takuji Yamada

9,27

, MetaHIT Consortium

, Pierre Renault

Thomas Sicheritz-Ponten

1,2

, Peer Bork

9,29

, Jun Wang

7,13,19,30

, Søren Brunak

1,2

& S Dusko Ehrlich

3,4,31

1Center for Biological Sequence Analysis, Technical University of Denmark, Kongens Lyngby, Denmark. 2Novo Nordisk Foundation Center for Biosustainability,

Technical University of Denmark, Kongens Lyngby, Denmark. 3INRA, Institut National de la Recherche Agronomique, UMR 14121 MICALIS, Jouy en Josas,

France. 4INRA, Institut National de la Recherche Agronomique, US 1367 Metagenopolis, Jouy en Josas, France. 5Department of Computer Science, Center

for Bioinformatics and Computational Biology, University of Maryland, USA. 6BGI Hong Kong Research Institute, Hong Kong, China. 7BGI-Shenzhen, Shenzhen,

China. 8School of Bioscience and Biotechnology, South China University of Technology, Guangzhou, China. 9European Molecular Biology Laboratory, Heidelberg,

Germany. 10Commissariat à l’Énergie Atomique et aux Énergies Alternatives, Institut de Génomique, Évry, France. 11Centre National de la Recherche Scientifique,

Évry, France. 12Université d’Évry Val d’Essonne, Évry, France. 13The Novo Nordisk Foundation Center for Basic Metabolic Research, University of Copenhagen,

Copenhagen, Denmark. 14Digestive System Research Unit, University Hospital Vall d’Hebron, Ciberehd, Barcelona, Spain. 15Faculty of Health Sciences, University

of Southern Denmark, Odense, Denmark. 16Department of Structural Biology, VIB, Brussels, Belgium. 17Department of Bioscience Engineering, Vrije Universiteit,

Brussels, Belgium. 18National Food Institute, Division for Epidemiology and Microbial Genomics, Technical University of Denmark, Kongens Lyngby, Denmark.

19Department of Biology, University of Copenhagen, Copenhagen, Denmark. 20Hagedorn Research Institute, Gentofte, Denmark. 21Institute of Biomedical Science,

Faculty of Health and Medical Sciences, University of Copenhagen, Copenhagen, Denmark. 22Faculty of Health, Aarhus University, Aarhus, Denmark. 23Department of

Microbiology and Immunology, Rega Institute, KU Leuven, Belgium. 24VIB Center for the Biology of Disease, Leuven, Belgium. 25Section of Microbiology, Department

of Biology, University of Copenhagen, Copenhagen, Denmark. 26Laboratory of Microbiology, Wageningen University, Wageningen, The Netherlands. 27Department

of Biological Information, Tokyo Institute of Technology, Yokohama, Japan. 28A full list of members and affiliations appears at the end of the paper. 29Max Delbrück

Centre for Molecular Medicine, Berlin, Germany. 30Princess Al Jawhara Center of Excellence in the Research of Hereditary Disorders, King Abdulaziz University,

Jeddah, Saudi Arabia. 31King’s College London, Centre for Host-Microbiome Interactions, Dental Institute Central Office, Guy’s Hospital, United Kingdom.

32These authors contributed equally to this work. Correspondence should be addressed to S.B. (brunak@cbs.dtu.dk) or S.D.E. (dusko.ehrlich@jouy.inra.fr).

Received 12 February; accepted 22 May; published online 6 July 2014; doi:10.1038/nbt.2939

metagenomic data rely largely on comparisons to reference genomes

from cultivated microbes. However, these reference genomes

represent only a fraction of the species and viruses present. Moreover,

bacterial genomes from different isolates of the same species usually

show considerable genetic heterogeneity when compared3,4. This

variation may be the result of clonal differences, environmental

2 ADVANCE ONLINE PUBLICATION nature biotechnology

ARTICLES

adaptation or possibly artifacts from the cultivation process.

Therefore, reference genomes represent only a small proportion of the

biological diversity of microbial systems, and thus methods relying

on them place limitations on structuring and analyzing metagenomic

data. In particular, they limit our ability to segregate metagenomic

data into coherent biological entities and fail to describe previously

unknown species, phages and modules of genetic variation within

microbial species.

De novo assembly of genomes from complex metagenomic data is

inherently difficult due to the many sequence ambiguities that confuse

the assembly process5. Hence, a typical metagenomic assembly will

result in a large set of independent contigs that are not easily aggregated

into biological entities6. Previous methods have addressed the assembly

problem from different angles. Iverson et al.7 used tetranucleotide-

based binning to assist the assembly of species with a distinct and

consistent skew in base composition. In complex communities this

approach works only for a few organisms with extreme base compo-

sitions. Moreover, some known genomes have inconsistent tetranu-

cleotide distributions, which would compromise any assembly using

this approach. An alternative is to use differences in the abundance

of genetic sequences measured in a single biological sample to sepa-

rate organisms8,9. Such methods rely on the notion that abundance

is constant across genetic entities (i.e., each gene on a specific bacte-

rial chromosome will be found in a sample in the same abundance

as any other gene on that chromosome). These abundance-based

methods have been used to segregate the most abundant species in

waste water9, which is a community with limited complexity, or to

further segregate a subset of sequence reads with similarity to known

reference genomes8. However, both Albertsen et al.9 and Wang et al.8

acknowledge that their methods cannot segregate taxonomically

related species or genera, respectively; also these methods cannot

give comprehensive segregation of all entities in complex samples.

Proper structuring of the complete metagenomic composition is

important not only for understanding the microbial communities10,

but also for making statistical associations between the metagenomic

data and descriptors of the system. In the case of the human microbi-

ome such descriptors include clinical parameters of the human host.

For example, we have previously used co-abundance profiles to bin the

2% of a gene catalog with strongest correlation to the human type 2

diabetes phenotype11. This was manageable with 2% of the genes but

such clustering using distance matrices is not possible for an entire

microbiome (as calculating a 3.9 million × 3.9 million gene distance

matrix is impractical even for large supercomputers).

The method we present here is based on segregating biological enti-

ties by co-abundance but overcomes the limited resolution of previous

methods (e.g., not being able to segregate related organisms) and

enables the complete segregation of complex metagenomic samples.

The increased resolution is achieved by using co-abundance profiles

across many samples. Here we use 396 samples, but we also show that

species can be segregated accurately using as few as 18 samples. The

computational problem of generating a distance matrix for a complete

metagenome is overcome by using a method that extracts groups of

genes that correlate (in terms of abundance) to randomly picked seed

genes. This clustering approach can be done in hours on a powerful

desktop computer.

Segregating a metagenome into groups of genes that have similar

abundance (CAGs) allows identification of biological entities like

species and phages, as well as small genetic entities representing

co-inherited clonal heterogeneity. Phage sequences have previously

been identified in complex communities by sequence similarity or

size separation before sequencing12,13, but a general method that can

identify novel species, phages and genetic heterogeneity from generic

metagenomics data has been lacking. In addition, small biological

entities have, with a few exceptions14,15, not been affiliated to specific

microbial species within the community.

Here we assign genes from 396 human gut microbiome samples

into 7,381 CAGs, and define subsets of these as MGS and phage-like

CAGs. From the MGS, we assembled 238 unique genomes that meet

the high-quality draft genome standard of the Human Microbiome

Project16. We found that a large set of smaller CAGs could be affili-

ated to the MGS by dependency associations, and that the persistence

of some MGS is related to the occurrence of specific dependency-

associated CAGs.

RESULTS

Comprehensive co-abundance gene segregation

Our method for co-abundance segregation uses a metagenomic data

set consisting of a number of samples of the same type. In this study

we use a deeply sequenced data set of 396 human stool samples from

Spanish and Danish individuals, including 124 samples from a pre-

vious study2 (see Online Methods and Supplementary Data 1 for

details). Seventy-seven of the Spanish individuals were sampled twice,

with an average of 6 months between the samplings. The first step is

the assembly of the sequencing reads from each sample into genes,

which are then pooled into a nonredundant gene catalog (Fig. 1).

Our gene catalog contained 3,871,657 genes (Supplementary Fig. 1);

only 10% of this gene catalog could be assigned to a taxonomic group

at the species level (Supplementary Note 1 and Online Methods).

We picked a gene at random as a ‘seed’ and defined this and other

genes with similar abundance profiles as a ‘canopy’, which was defined

as those genes with Pearson correlation coefficient (PCC) > 0.9 to

the seed gene profile. A canopy profile was then determined as the

median abundance profile of the gene group and was used itera-

tively for recapturing the canopy until the canopy profile stabilizes

(Supplementary Fig. 2). New random seed genes were picked until all

genes were assigned to a canopy. Canopies that passed a canopy rejec-

tion criteria—canopies must contain more than two genes and 90% of

the canopy profile signal must originate from more than three sam-

ples—were identified as CAGs. It should be noted that large canopies

tend to be determined quickly, thereby reducing the computational

cost of subsequent canopies. The method depends on two parameters:

the gene inclusion criterion and the canopy rejection criteria.

Clustering of our data set binned 1.53 million genes (represent-

ing 68% of the mapped sequence reads; the remaining genes were

in canopies that did not pass the rejection criteria) into 7,381 CAGs,

which ranged in size from 3 to 6,319 genes. The size distribution of

the groups, in terms of numbers of genes, was bimodal with peaks at

~50 genes and ~1,700 genes, respectively (Fig. 2a).

Most complete genomes of bacteria or archaea contain >700 genes

and the 741 largest CAGs had >700 genes (Supplementary Fig. 3 and

Supplementary Note 2). The genes in these 741 CAGs were highly

consistent in base composition, had highly correlated abundance pro-

files in an independent set of 115 samples17 (Supplementary Note 3)

and had consistent taxonomical annotation. For 115 of these CAGs,

>95% of the taxonomically annotated genes were annotated to the

same species (Supplementary Figs. 4–6 and Supplementary Data 2)

and most were similar to a reference sequence from a specific strain

within a species. We refer to these CAGs with >700 genes as MGS. In

nine cases we identified several distinct MGS from the same species

(e.g., three from Fecalibacterium prausnitzii).

The individual gene-abundance profiles of all MGS were highly

coherent and most were distinct from genes not included in the MGS.

nature biotechnology ADVANCE ONLINE PUBLICATION 3

ARTICLES

Consequently, changing the gene inclusion criterion to PCC > 0.8

or PCC > 0.95 increased or reduced the size of the average MGS

by only 5% and 17%, respectively (Supplementary Figs. 7 and 8a

and Supplementary Note 4). In contrast, when we defined gene sets

by sequence similarity to reference genomes, the abundance profiles

were inconsistent (Supplementary Fig. 8b). Importantly, clustering

of randomly permuted abundance profiles only resulted in very few

canopies that pass the canopy rejection criterion and all of these were

small (Supplementary Fig. 9).

Nineteen of the individuals sampled consumed a defined fermented

milk product containing the previously sequenced Bifidobacterium

animalis subsp. lactis CNCM I-2494 (ref. 18), and we used this spe-

cies as a benchmark to assess the ability of our method to identify

and segregate a particular species. Although on average only 0.3%

of the sequence reads in the 19 samples originated from B. animalis,

we were able to capture 95% of the B. animalis reference genes into

one MGS (MGS:337). Subsampling of the data showed that the

B. animalis MGS can be segregated using as little as 700 K sequence

reads per sample or from a much smaller sample set consisting of

only 18 samples (Fig. 3).

Together the MGS and CAGs provide a detailed overview of the

microbial community and precise estimates of the species richness

that are in strong agreement with the observed gene richness (PCC =

0.96; Supplementary Note 5 and Supplementary Fig. 10).

Genome assembly

Grouping genes into an MGS using our co-abundance method also

assists de novo genome assembly by providing the basis for segregating

sequence reads into those that derive from a distinct biological entity

(Fig. 1). That is, for an individual sample, reads can be selected that

map to a particular MGS; these are species-specific subsets of the full

set of reads from that metagenomic sample. We used standard genome

assembly tools to assemble these subsets of reads for all MGS. Of

these 238 unique MGS genomes—including 181 new genomes from

previously unsequenced species—met the Human Microbiome

Project high-quality draft genome standard (Supplementary Figs. 11

and 12 and Supplementary Data 3). We refer to this strategy as

MGS-augmented assembly. The MGS-augmented assembly of

MGS:337 covered 95% of the reference genome of the benchmark

species B. animalis subsp. lactis CNCM I-2494 with 99.9% identity

(Fig. 4). In addition, 44 of the MGS-augmented assemblies were

closely related to known reference genomes and covered an average

of 78% of these genomes with an average sequence identity of 98.4%

(Supplementary Data 4).

To reduce the possibility of making chimeric assemblies of closely

related strains, we used only sequence reads from a single sample

Deep sequencing

and de novo

gene assembly

Abundance

profiles of gene

catalog

Nonredundant

gene catalog

Canopy

clustering by

co-abundance

MGS

Random

seed gene

MGS

+ +

CAG

MGS-specific reads

MGS augmented assembly

Sequence reads

from sample x

HQ genome

Standard genome

assembly

Matching reads

to MGS

Figure 1 Overview of co-abundance clustering and the MGS-augmented

assembly. DNA from a series of independent biological samples from

microbial communities, here originating from the human gut microbiome,

is extracted and shotgun sequenced. Genes assembled and identified

in individual samples are then integrated to form a cross-sample,

nonredundant gene catalog. The abundance profile of each gene in the

catalog is assessed by counting the matching sequence reads in each

sample. To facilitate co-abundance clustering of large gene catalogs, we

used random seed genes as ‘baits’ for identifying groups of genes that

correlate (PCC > 0.9, gray dashed circle) to the abundance profile of the

bait genes. The fixed PCC distance threshold is called a canopy (dashed

circles). To center the canopy on a co-abundance gene group (CAG), the

median gene abundance profile of the genes within the original seed

canopy (or subsequent canopies, symbolized as +) is used iteratively

to recapture a new canopy until it settles on a particular profile (off-set

circles). The gene content of a settled canopy (black dashed circles) is

named a metagenomic species (MGS) if it contains 700 or more genes.

The smaller groups remain referred to as CAGs. Sequence reads from

individual samples that map to the MGS genes and their contigs are then

extracted and used to assembly a draft genome sequence for an MGS;

we refer to this process as MGS-augmented genome assembly. The use

of sample-specific sequence reads in the assemblies helps discriminate

between closely related strains.

600

400

200

03 50 200 700 1,800 2,800 5,6004,5003,5001,100

Genes in CAG

CAGs

a800

100

1,000

5,000

CAG size (genes, log)

Essential genes

Phage-like CAGs

Transposases,

integrases

& recombinases

CRISPR

Restriction

endonucleases

DNA methylation

Glycosyltransferases

Dependency-associated

Figure 2 Size distributions of co-abundance gene groups (CAGs).

(a) Histogram showing the CAG size distribution in terms of gene content.

The scale is logarithmic as indicated by the bar widths. (b) Bee swarm

plot showing CAGs that are significantly enriched (Fishers exact test,

P < 0.001) for the indicated gene annotation, as well as phage-like CAGs

and dependency-associated CAGs, plotted against the number of genes

contained in the CAGs. Here every point represents an enriched CAG or

MGS and the width of the swarms shows the distribution. The dashed line

marks the 700-gene threshold separating small CAGs from MGS.

4 ADVANCE ONLINE PUBLICATION nature biotechnology

ARTICLES

for every such assembly. However, >100 MGS could be assembled

from multiple samples, hence, the total number of high-quality draft

genome assemblies was 360.

Functional characterization of small CAGs

The 6,640 CAGs with <700 genes (an average of 44 genes) showed

abundance correlations similar to those of the MGS and several lines

of evidence suggest that many of these CAGs represent biological

entities or clonal differences within species. Of the small CAGs,

848 were enriched for proteins characteristic of phages13 or for genes

with similarity to specific known phage taxa (Supplementary Data 5).

These phage-like CAGs have a median coding sequence length of 28kb,

which is substantially longer than previous phage assemblies from

complex communities (in which ~60% are <1 kb)13. An average of

113 (±37, s.d.) phage-like CAGs were identified per sample. Although

bacteriophage taxonomy is fairly limited, we observed consistent

species- or family-level taxonomical annotation in 35 and 172 phage-

like CAGs, respectively. As expected from the presence of phage-like

CAGs, transposase, integrase and recombinase encoding genes were

enriched in the smaller CAGs (Fig. 2b).

Another class of functions that were enriched in smaller CAGs are

functions that are important for biotic interactions. These include

clustered, regularly interspaced, short palindromic repeats (CRISPR)-

associated genes, which function in bacteria and archaea as a sequence-

dependent adaptive immune system against foreign nucleic acids19.

In addition to core CRISPR-associated genes, several CAGs were

enriched for specific subtypes of these genes (Supplementary Fig. 13).

Similarly, restriction endonucleases and DNA methylases, which are

part of the nonadaptive defense system, were enriched in 120 small

CAGs. Also, genes involved in modification of bacterial exterior

surfaces, which are important for bacterial identification and mask-

ing, were enriched in more than 400 small CAGs. These included

genes involved in modifications of the cell wall and, in particular,

glycosyltransferases.

Dependency associations affiliate small CAGs to MGS

The existence of small CAGs that represent phages and clonal

differences implies that such CAGs depend on cellular organisms

for their proliferation. In relationships that are nonpromiscuous,

a dependent CAG should therefore never occur independently of

the hosting microorganism. Indeed, we identified significant depend-

ency associations by comparing the absence-presence profiles

throughout all samples for all pairs of CAGs (including the MGS)

using Fisher’s exact test and excluding relationships where a potential

dependent CAG was observed independently of the potential hosting

CAG (Fig. 5a). In this way the network becomes directional from

the dependency-associated CAG to the hosting CAG (which may be

observed independently).

400 K 1,943,1131,600 K1,200 K800 K

1 M

1,860,762

Bidobacterium animalis subsp. lactis CNCM I-2494

chromosomal positions

MGS:337 augmented assembly

positions

Figure 4 Comparison of the MGS:337 augmented assembly and the

B. animalis reference genome. BLAST dot-plot comparing the MGS:337

augmented assembly (y axis) to the B. animalis subsp. lactis CNCM

I-2494 reference genome (x axis). The dot-blot shows the relative

chromosomal positions of matching sequence on the MGS-augmented

assembly and the B. animalis reference genome. The MGS-augmented

assembly covers 95% of the reference genome with 99.9% identity.

The plot shows an inversion in the assembly relative to the reference

genome around position 1,300 K.

100

Specificity (% B. animalis genes) Specificity (% B. animalis genes) Specificity (% B. animalis genes)

11 M 5 M 2.5 M 1 M 700 k 500 k 350 k 250 k 200 k 150 k 100 k

Genes captured

Sequencing reads per sample (downsized)

2,000

4,000

6,000

8,000

100

Genes captured

Number of samples containing B. animalis

19 15 10 8 6 4 3 12

2,000

4,000

100

Genes captured

Samples used

393 350 300 250 200 150 100 50 30 20 18 15 12 10 8

2,000

4,000

6,000

8,000

Figure 3 Benchmarking sensitivity and specificity of the co-abundance

clustering across a range of sequencing depths or sample numbers.

B. animalis subsp. lactis CNCM I-2494 was used as a benchmark

species because 19 samples originated from individuals who had

consumed a defined fermented milk product containing this strain.

(For each clustering, the size (number of genes captured) is shown

as bars (left axis); and the specificity (percentage of genes matching

the B. animalis reference genome with > 95% sequence identity over

100 bp or better captured in the MGS that is most similar to

B. animalis) is shown as a line (right axis).) (a) Co-abundance

clustering using reduced data sets to simulate the sequencing

depths indicated (x axis). At a sequence depth of 700 K reads,

97% of the B. animalis genes were captured, and at a depth of

200 K reads 98.6% of the captured genes were from B. animalis.

(b) Co-abundance clustering of random sample subsets containing the

indicated number of samples (x-axis) from individuals that consumed

the DFMP. Here the total sample size was kept constant at 375 samples.

the indicated size (x axis). These sample subsets included 19 samples

from individuals who had consumed the DFMP, except when they

contained <19 samples (i.e., 19-8 DFMP individuals per subset).

In b and c, samples were downsized to 11 million sequence reads per

sample. Error bars, ±1 s.d. from the mean (n = 5).

nature biotechnology ADVANCE ONLINE PUBLICATION 5

ARTICLES

The resulting network of the most significant (Fishers exact test:

P < 10−10, corrected) dependency associations is shown in

Supplementary Figure 14 and contains 882 relationships between 1,205

CAGs (Supplementary Data 6). As expected, the network is signifi-

cantly over-represented for small CAGs that associate to an MGS (odds

ratio 12.7, Fisher’s exact test: P < < 1 × 10−100) and many of the sub-

networks are centered on an MGS, but there are also nine small CAGs

that connect MGS pairs of the same genus. Biologically these may be

genetic elements or phages that are shared between related species.

For 413 of the associations, sequence contigs were found in

individual samples that bridged the dependent and hosting CAGs

(enrichment relative to all CAGs pairs: odds ratio 2,513, Fisher’s

exact test P < < 1 × 10−100). This indicates that many dependency-

associated CAGs are genomically integrated in the hosting CAG

in some samples.

The dependency associations connect CAGs into subnetworks,

which can be used to guide further study of the components. For

example, the subnetwork centered on Sutterella wadsworthensis

(MGS:135, Fig. 5b) contains eight dependency associations including

the phage-like CAG:3731 and a CAG containing CRISPR-associated

genes and a repeat region (CAG:4011). The sample-wise detection of

the CRISPR and phage-like CAGs were anti-correlated (Matthew’s

correlation coefficient −0.7) and one of the CRISPR spacers had a

15-bp sequence match to the phage, suggesting that the CRISPR

prevents the phage from infecting the bacterium19.

Sample-specific, MGS-augmented assemblies of the Escherichia

coli MGS:4 and its dependency-associated CAGs (Supplementary

Fig. 15 and Supplementary Data 7) demonstrated strong sequence

similarities throughout the majority of the chromosome, but had

some differences. In particular the largest of the E. coli–associated

CAGs (CAG:427, containing 345 genes) was absent in some sample-

specific assemblies (Fig. 5c). In the MGS:4 genome the integration of

CAG:427 was spread across several locations, suggesting that it repre-

sents several independent insertion and/or rearrangement events.

CAG:427

MGS:4

CRISPR

MGS:135

CAG:4065

CAG:4755

CAG:4703

CAG:5806

CAG:2350

CAG: 2543

CAG:3731

CAG:4011

MCC = – 0.7

Samples 1-318

Abundance

CAG:2350

MGS:135

Figure 5 Dependency associations among

MGS and CAGs. (a) A typical example of

a significant dependency association. The

abundance of the MGS:135 (S. wadsworthensis)

and the small CAG:2350 across 318 fecal

samples are shown as blue and red curves,

respectively (upper panel, logarithmic scale).

Below the sample-wise presence of the

two CAGs is shown as bars. CAG:2350 is

significantly co-occurring with MGS:135 and

never detected independently (Fishers exact

test, P = 9 × 10−74). The samples were sorted

according to the abundance of MGS:135.

(b) The dependency-association subnetwork

of CAGs associated to S. wadsworthensis

(MGS:135). Arrows show dependency

associations and solid arrows indicate that

co-assembly of the MGS and the CAG in one or more samples supported the association. Blue coloring indicates CAGs dominated by genes with

species level similarity to S. wadsworthensis. CAG:2543 and CAG:3731 are enriched for phage genes, and CAG:4011 contains a series of CRISPR-

associated genes and a CRISPR cluster. The CRISPR complex containing CAG:4011 and one of the phages-like CAG:3731 anti-correlate (Matthews

correlation coefficient = −0.7) and spacers of the CRISPR show sequence complementarity to the phage. (c) The E. coli (MGS:4) and its nine

dependency-associated CAGs were co-assembled to high-quality draft genomes in each of 11 samples. The outer black circle represents the

consensus assembly of the E. coli–centered agglomerate and each of the gray circles represents alignment of the assembly from a particular sample.

The positions and sequence coverage of CAG:427 are marked in red, across the assemblies.

Figure 6 Gut persistence probability for B. adolescentis. The gut

persistence of B. adolescentis (MGS:119) populations stratified by the

presence (red curves) or absence (black curves) of the dependency-

associated CAG:2298 observed across 54 human individuals who had

the bacterium in the first of two fecal samples. B. adolescentis had

substantially higher persistence probability with CAG:2298 present.

(a) Interval-censored Kaplan-Meier curves showing the cumulative loss

of populations of B. adolescentis over time across the cohort of human

individuals. Points (+) indicate time (in days) of the second of two

samplings from a human individual. The curve shows the “losses” when

they are registered at the second time point and not when the loss actually

happened (i.e., the data are interval-censored). (b) Model-based estimates

of annual gut persistence probability for B. adolescentis with or without

the dependency-associated CAG. Note that annual persistence probability

with the CAG (mean estimate = 88%) is much larger than without (mean

estimate = 18%). In the Bayesian logistic regression framework used

here, estimates are expressed as probability distributions over the possible

values for parameters of interest. We therefore obtain both an estimate of

a parameter, and quantification of how certain we are of the estimate. This

figure shows the posterior probability distribution over possible values for

the annual persistence probabilities; with the shaded areas indicating the

95% highest-density intervals (i.e., the parameter values with the most

support).

6 ADVANCE ONLINE PUBLICATION nature biotechnology

ARTICLES

MGS persistence is influenced by associated CAGs

To investigate the effect that dependency-associated CAGs may have

on their host MGS, we analyzed 73 of the individuals that were sam-

pled at two different time points (four of the original 77 sample pairs

were discarded). From each of these sample pairs we assessed if a

given MGS was present at the first time point and whether it was still

present at the second time point. Based on this information, it was

possible to estimate the persistence of an MGS across the cohort of

human individuals, and whether this persistence was influenced by

dependency-associated CAGs.

We found that there was a greater probability of B. adolescentis

(MGS:119) persisting at the second time point when its dependency-

associated CAG:2298 was also present (Fig. 6a). To further analyze this

phenomenon, we used logistic regression to infer the annual persistence

probabilities for MGS with or without their dependency-associated

CAGs (Supplementary Data 6). The credibility of these estimates was

quantified using Bayesian statistical methods20, which, in brief, output

a posterior probability distribution over the possible annual persistence

probabilities. From this analysis we identified 26 cases in which the

presence of a specific dependency-associated CAG correlated with a

substantially altered annual persistence probability of its host MGS. For

example, the annual persistence probability of B. adolescentis (MGS:119)

was estimated to be 88% in individuals where it was observed in asso-

ciation with CAG:2298, but 18% in individuals where CAG:2298 was

absent (Fig. 6b; posterior probability that the CAG:2298 effect is larger

than zero = 99.94%). We observed similar, positive effects for CAGs

dependency-associated with Prevotella copri, E. coli, F. prausnitzii and

12 other MGS (Supplementary Fig. 15 and Supplementary Data 6). In

addition, ten dependency-associated CAGs had a substantial negative

effect on the persistence probability of their hosting MGS.

The dependency-associated CAGs that increased the MGS persistence

probability contained a range of gene sets, including CRISPR-associated

genes (CAG:2720), collagen adhesion protein and gram-positive

anchor proteins (CAG:2888), and thioredoxin family proteins that

might be important for the tolerance of reactive oxygen species (ROS).

This is in line with our observation that the most common species in

the human gut microbiome have genes that mediate ROS tolerance

(Supplementar y Note 6 and Supplementary Data 8). Among the

dependency-associated CAGs that contributed negatively to the MGS

persistence probability, we observed three phage-like CAGs.

DISCUSSION

The method presented here allows complete co-abundance cluster-

ing of microbiomes. The resulting CAGs provide insight into the

microbial species present in metagenomic samples and their genetic

makeup, and thereby provide details important for understanding

the content of a microbial community. Clustering is purely data

driven and therefore circumvents the need for reference genomes,

presequencing filtering and cultivation of microbial species. The

ability of our method to discriminate between strains of the same

species indicates the power of co-abundance to segregate closely

related biological entities, in contrast to the findings from gene sets

that are defined by sequence similarity to known reference genomes

(Supplementary Fig. 8). Inaccurate discrimination among closely

related species potentially leads to false associations between clinical

conditions and putative species. Although we identify a few cases of

chimeric assemblies (Supplementary Note 7), we have no indication

of CAGs constituting multiple species; however, such entities could

in principle exist in very close co-abundance.

The method should be generally applicable to sets of deep-

sequenced shotgun metagenomics samples. The exact number of

samples and sequencing depth needed depends on the complexity

of the microbial community and the abundance of the microbes.

However, our benchmarking using B. animalis suggests that the

number of samples used is critical (here 18 samples), whereas the

necessary sequencing depth (here 0.7 M reads) is easily reached with

current sequencing technologies (Fig. 3). This emphasis on sample

number over sequencing depth is likely to hold across different types

of microbial communities.

Interestingly, we found that most genes involved in resistance to

antibiotics, except vancomycin resistance genes, were not found as

members of any CAG. This is in line with the fact that most antibiotic

genes, except vancomycin resistance genes, are known to act alone to

provide antibiotic resistance (Supplementary Note 8).

Although, many of the CAGs and their dependency association are

not understood at present, our findings suggest that even small CAGs

represent biologically meaningful entities, either in the form of phages

or clonal differences of microbial species. This is consistent with pre-

vious findings that the genetic differences that make the E. coli O104:

H4 strain a cause of severe food poisoning are just a few virulence

factors, including a Shiga toxin 2–encoding prophage21. Therefore,

thorough descriptions of genetic heterogeneity, which are enabled

by our co-abundance method, are likely to be important for disease

association studies. Moreover, although the relationships we observed

between specific smaller CAGs and their host MGS are only associa-

tions, they do suggest functionally and/or evolutionarily important

relationships that may prove critical for future understanding and

engineering of microbial communities. Furthermore the approach

that allowed us to determine conditional persistence probabilities for

microbial species in a community is likely to be important for reveal-

ing relationships or conditions that are critical for the persistence or

elimination of specific microbes in future studies. Here the func-

tions of genes within the CAGs associated with differential persistence

suggest that tolerating ROS and anchoring to the intestinal epithelia

are important for microbial persistence in the gut. Our findings also

suggest specific phage-microbe relationships that reduce persist-

ence. Therefore, our co-abundance–based method should facilitate

advances in understanding microbial biology as well as enabling

de novo genome assembly and comprehensive characterization of

complex metagenomic samples.

METHODS

Methods and any associated references are available in the online

version of the paper.

Accession codes. Sequence data were deposited at EBI with the

accession code ERP002061; and MGS-augmented assemblies

were deposited at EBI under (PRJEB674 to PRJEB1046). 454

sequencing reads were added to the NCBI BioProjectID 32811.

The 3.9 M gene catalog and the CAGs are available for download

from https://www.cbs.dtu.dk/projects/CAG/. Source code for the

MGS canopy algorithm is available as Supplementary Software and

from: http://git.dworzynski.eu/mgs-canopy-algorithm.

Note: Any Supplementary Information and Source Data files are available in the

online version of the paper.

ACKNOWLEDGMENTS

The research leading to these results has received funding from the European

Community’s Seventh Framework Programme FP7- HEALTH-F4-2007-201052:

Metagenomics of the Human Intestinal Tract (MetaHIT) and FP7-HEALTH-2010-

261376: International Human Microbiome Standards, as well as the Novo Nordisk

Foundation Center for Biosustainability. Work on the clustering concept has been

supported by the OpenGPU FUI collaborative research projects, with funding

nature biotechnology ADVANCE ONLINE PUBLICATION 7

ARTICLES

from DGCIS. Researchers on the project were granted access to the HPC resources

of CCRT under the allocation 2011-036707 made by GENCI (Grand Equipement

National de Calcul Intensif). The company Alliance Services Plus (AS+) has

provided help to scale up the process, especially, V. Arslan, D. Tello, V. Ducrot,

T. Saidani and S. Monot. The authors affiliated with MGP are funded, in part, by

the Metagenopolis ANR-11-DPBS-0001 grant. Ciberehd is funded by the Instituto

de Salud Carlos III (Spain). M.A. was supported by a grant from the Ministère de la

Recherche et de l’Education Nationale (France).

AUTHOR CONTRIBUTIONS

All authors are members of the Metagenomics of the Human Intestinal

Tract (MetaHIT) Consortium. S.D.E. and S.B. managed the project. F.C., N.B.,

F.G., T.H., K.S.B. and T.N. performed clinical sampling. F.L. and C.M. performed

DNA extraction. J.L., E.P. and D.L.P. performed sequencing. S.D.E., H.B.N.,

M.A., A.S.J., S.R., P.R. and P.B. designed the analyses. H.B.N., A.S.J., S.R., M.A.,

A.G.P., D.R.P., L.G., I.B., M.B., M.B.Q.d.S., M.A., J.L., J.T., S.S., T.Y., E.P., D.L.P. and

R.S.K. performed the data analyses. H.B.N., S.B., A.S.J., S.R., A.G.P. and M.A.

wrote the manuscript. H.B.N., S.B., S.D.E., D.R.P., I.B., P.B., E.P., O.P. and D.W.U.

revised the manuscript. The MetaHIT Consortium members contributed to the

design and execution of the study.

COMPETING FINANCIAL INTERESTS

The authors declare no competing financial interests.

Reprints and permissions information is available online at http://www.nature.com/

reprints/index.html.

1. Fodor, A.A. et al. The “most wanted” taxa from the human microbiome for whole

genome sequencing. PLoS ONE 7, e41294 (2012).

2. Qin, J. et al. A human gut microbial gene catalogue established by metagenomic

sequencing. Nature 464, 59–65 (2010).

3. Lukjancenko, O., Wassenaar, T.M. & Ussery, D.W. Comparison of 61 sequenced

Escherichia coli genomes. Microb. Ecol. 60, 708–720 (2010).

4. Fitzsimons, M.S. et al. Nearly ﬁnished genomes produced using gel microdroplet

culturing reveal substantial intraspecies genomic diversity within the human

microbiome. Genome Res. 23, 878–888 (2013).

5. Pop, M. Genome assembly reborn: recent computational challenges. Brief. Bioinform.

10, 354–366 (2009).

6. Wooley, J.C., Godzik, A. & Friedberg, I. A primer on metagenomics. PLOS Comput.

Biol. 6, e1000667 (2010).

7. Iverson, V. et al. Untangling genomes from metagenomes: revealing an uncultured

class of marine Euryarchaeota. Science 335, 587–590 (2012).

8. Wang, Y., Leung, H.C.M., Yiu, S.M. & Chin, F.Y.L. MetaCluster 5.0: a two-round

binning approach for metagenomic data for low-abundance species in a noisy

sample. Bioinformatics 28, i356–i362 (2012).

9. Albertsen, M. et al. Genome sequences of rare, uncultured bacteria obtained

by differential coverage binning of multiple metagenomes. Nat. Biotechnol. 31,

533–538 (2013).

10. Raes, J. & Bork, P. Molecular eco-systems biology: towards an understanding of

community function. Nat. Rev. Microbiol. 6, 693–699 (2008).

11. Qin, J. et al. A metagenome-wide association study of gut microbiota in type 2

diabetes. Nature 490, 55–60 (2012).

12. Reyes, A. et al. Viruses in the faecal microbiota of monozygotic twins and their

mothers. Nature 466, 334–338 (2010).

13. Minot, S. et al. The human gut virome: inter-individual variation and dynamic

response to diet. Genome Res. 21, 1616–1625 (2011).

14. Stern, A., Mick, E., Tirosh, I., Sagy, O. & Sorek, R. CRISPR targeting reveals a

reservoir of common phages associated with the human gut microbiome. Genome

Res. 22, 1985–1994 (2012).

15. Zhang, Q., Rho, M., Tang, H., Doak, T.G. & Ye, Y. CRISPR-Cas systems target a

diverse collection of invasive mobile genetic elements in human microbiomes.

Genome Biol. 14, R40 (2013).

16. Chain, P.S.G. et al. Genomics. Genome project standards in a new era of sequencing.

Science 326, 236–237 (2009).

17. Le Chatelier, E. et al. Richness of human gut microbiome correlates with metabolic

markers. Nature 500, 541–546 (2013).

18. Chervaux, C. et al. Genome sequence of the probiotic strain Biﬁdobacterium

animalis subsp. lactis CNCM I-2494. J. Bacteriol. 193, 5560–5561 (2011).

19. Terns, M.P. & Terns, R.M. CRISPR-based adaptive immune systems. Curr. Opin.

Microbiol. 14, 321–327 (2011).

20. Kruschke, J.K. Bayesian data analysis. Wiley Interdiscip. Rev. Cogn. Sci. 1,

658–676 (2010).

21. Karch, H. et al. The enemy within us: lessons from the 2011 European Escherichia

coli O104:H4 outbreak. EMBO Mol. Med. 4, 841–848 (2012).

MetaHIT Consortium:

H Bjørn Nielsen1,2,32, Mathieu Almeida3–5,32, Agnieszka S Juncker1,2, Simon Rasmussen1, Junhua Li6–8,

Shinichi Sunagawa9, Damian R Plichta1, Laurent Gautier1, Anders G Pedersen1, Emmanuelle Le Chatelier3,4,

Eric Pelletier10–12, Ida Bonde1,2, Trine Nielsen13, Chaysavanh Manichanh14, Manimozhiyan Arumugam9,

Jean-Michel Batto3,4, Marcelo B. Quintanilha dos Santos1, Nikolaj Blom2, Natalia Borruel14, Kristoffer S Burgdorf13,

Fouad Boumezbeur3,4, Francesc Casellas14, Joël Doré3,4, Piotr Dworzynski1, Francisco Guarner14,

Torben Hansen13,15, Falk Hildebrand16,17, Rolf S Kaas18, Sean Kennedy3,4, Karsten Kristiansen19, Jens Roat Kultima9,

Pierre Leonard3, Florence Levenez3,4, Ole Lund1, Bouziane Moumen3,4, Denis Le Paslier10–12, Nicolas Pons3,4,

Oluf Pedersen13,20–22, Edi Prifti3,4, Junjie Qin6,7, Jeroen Raes17,23,24, Søren Sørensen25, Julien Tap9,

Sebastian Tims26, David W Ussery1, Takuji Yamada9,27, Pierre Renault3, Thomas Sicheritz-Ponten1,2,

Peer Bork9,29, Jun Wang7,13,19,30, Søren Brunak1,2, S Dusko Ehrlich3,4,31, Alexandre Jamet3, Alexandre Mérieux33,

Antonella Cultrone3, Antonio Torrejon14, Benoit Quinquis4, Christian Brechot33, Christine Delorme3,

Christine M’Rini33, Willem M de Vos26, Emmanuelle Maguin3, Encarna Varela14, Eric Guedon3, Falony Gwen16,

Florence Haimet4, François Artiguenave10, Gaetana Vandemeulebrouck3, Gérard Denariaz34, Ghalia Khaci3,

Hervé Blottière4, Jan Knol35, Jean Weissenbach10, Johan E T van Hylckama Vlieg34, Jørgensen Torben26,

Julian Parkhill36, Keith Turner36, Maarten van de Guchte3, Maria Antolin14, Maria Rescigno37,

Michiel Kleerebezem26, Muriel Derrien34, Nathalie Galleron4, Nicolas Sanchez3, Niels Grarup21, Patrick Veiga34,

Raish Oozeer35, Rozenn Dervyn3, Séverine Layec3, Thomas Bruls10, Yohanan Winogradski3 &

Zoetendal Erwin G26

33Institut Mérieux, Lyon, France. 34Danone Research, Palaiseau, France. 35Gut Biology & Microbiology, Danone Research, Center for Specialized Nutrition,

Wageningen, the Netherlands. 36The Wellcome Trust Sanger Institute, Hinxton, Cambridge, U.K. 37Istituto Europeo di Oncologia, Milan, Italy.

nature biotechnology doi:10.1038/nbt.2939

ONLINE METHODS

Sample description. 396 stool samples from 177 Danish and 141 Spanish

human individuals were collected (Supplementary Data 1). 124 of the samples

were sequenced and used previously2. The Spanish samples included

13 individuals with Crohn’s disease and 69 with ulcerative colitis. 77 of the

Spanish individuals were sampled twice with, on average, 6 months between

the samplings. The Danish samples include healthy individuals ranging in

body mass index from 18 to 42. All were subjected to Illumina deep sequencing

resulting in 4.5 Gb sequence per sample on average, and a total of 23.2 billion

high-quality sequencing reads with an average length of 77 bp. The study was

approved by the local Ethical Committees of the Capital Region of Denmark

(HC-2008-017) and Clinical Research Ethics Committees (Comités de Ética en

Investigación Clínica, CEIC, Spain) and informed consent was obtained.

Construction of a nonredundant metagenomic gene catalog. Illumina raw

sequencing reads from 396 metagenomic samples (Supplementary Data 1)

were processed using the MOCAT software package22. In brief, >23.2 bil-

lion raw sequencing reads were filtered using the FastX software (http://

hannonlab.cshl.edu/fastx_toolkit) with a quality cutoff of 20 and reads shorter

than 30 bp discarded. High-quality reads (92% of raw reads) were assembled

into scaftigs (i.e., continuous sequences within scaffolds) using SOAPdenovo

(version 1.05)23. Genes were predicted on 18.5 M scaftigs longer than 500 bp

(35 Gbp in total) using MetaGeneMark24. Predicted genes from all samples

(45.4 M in total) were clustered using BLAT by single linkage. Any two genes

with greater than 95% identity and covering more than 90% of the shorter gene

were clustered together. Finally, cluster representatives shorter than 100 bp

were discarded resulting in a set of 4,201,877 nonredundant genes. From this

set, we removed genes that were considered spurious or likely originated from

human, animals or plants to yield a final set of 3,871,657 genes that formed

the reference gene catalog. For a comparison to our previous gene catalog12,

see Supplementary Data 9.

Quantification of reference gene abundances. High-quality reads were

mapped to the reference gene catalog using the screen function in MOCAT22.

Briefly, reads were mapped with SOAPaligner (version 2.21)25 with options:

–M 4 (find best hits), –l 30 (seed length), –r 1 (random assignment of multiple

hits) and –v 5 (maximum number of mismatches). Mapped reads were subse-

quently filtered using a 30-bp length and 95% identity cutoff and gene-length

normalized base counts were calculated using the soap.coverage script (avail-

able at: http://soap.genomics.org.cn/down/soap.coverage.tar.gz). For samples

where 11 M or more sequence reads were obtained (n = 393), 11 M sequence

reads were drawn randomly (without replacement). These randomly drawn

reads were mapped to the gene catalog and the number of reads counted to

form a downsized depth or abundance matrix. The 11 M downsized depth

matrix was used to estimate CAG abundances, gene and MGS richness. Similar

downsizings were done for the reduced sampling depths (Fig. 3).

Taxonomical annotation. Catalog genes were assigned taxonomical annota-

tion by sequence similarity to a database of 3,048 reference genomes (ftp://ftp.

ncbi.nlm.nih.gov/genomes/Bacteria/ and ftp://ftp.ncbi.nlm.nih.gov/genomes/

Bacteria_DRAFT/, July 2012), using BLASTN, only accepting alignments with

100 bp or longer. Sequence similarity of 95%, 85% and 75% or better was

used for species, genus and phylum level taxonomical annotation, respectively.

MGS were assigned a species level annotation if more than 50% of the genes

comprising the CAG were assigned a given species level taxonomy (including

genes with no match). MGS were described to have ‘clear and unambiguous

similarity to a known species’ when 90% or more of the genes were annotated

to the same species. Selected CAGs that appear in figures and could not be

assigned to a genus or species by DNA similarity (MGS:11, MGS:17, MGS:124

and MGS:225) were in addition taxonomically annotated by similarity to the

UniProt database (BLASTP, best hit, E < 0.001) to get an approximate taxo-

nomical annotation.

Phage definition and taxonomy annotation. A CAG was called phage-like

if it passed one of two criteria. (i) If a CAG contained a minimum of ten

phage-taxonomy annotated genes and 80% of these were consistent at the

species, genus or family level. Here phage-taxonomy annotated genes were

defined as genes with a top-3 blastp hit (E < 0.001, against the combined

NCBI nr Sept. 2013 and ACLAME26 0.4 database) to a viral organism listed in

the International Committee on Taxonomy of Viruses (ICTV) master species

list (release 2012). (ii) If a CAG-encoded five or more distinct characteristic

phage functions and ≥40% of the CAG genes were most similar to known

phage genes. Phage-functional classes were defined as proteins with a best-hit

(hmmscan27, domE < 0.001, against Pfam-A28 27.0) to one of 16 phage-specific

Pfam functions defined by Minot et al.13 or as proteins matching the

corresponding set of functions identified among phage orthologous

groups (blastp, E < 0.001, against POG VQ29). A characteristic phage

function was counted only once per CAG. Furthermore, a gene most similar

to known phage genes was defined as a gene with a best hit (blastp, E < 0.001,

against the combined NCBI nr and the ACLAME 0.4 database) to a viral organ-

ism. All phage-like CAGs were taxonomically annotated to species, genus or

family level using a 50% consistency criteria across ICTV annotated genes

(top-3 blastp hits, E < 0.001, against the combined NCBI nr and ACLAME26

database). Interestingly, the functions “tail,” “portal,” “terminase” and “capsid”

were each found in ≥70% of all phage-like CAGs and on average in only 5%

of other small CAGs.

Gene annotations and enrichment analysis. Functional annotation (includ-

ing CRISPR-associated genes) of the gene catalog was obtained by aligning

predicted proteins to the UniProt database using BLASTP (best hit with

E < 0.001) and proteins from the eggNOG (v3) database30 using BLASTP

(WU-BLAST 2.0, default parameters except E = 1 × 10−5, B = 10,000) and were

assigned to an orthologous group as described elsewhere31.

Genes of MGS:11, CAG:4957, MGS:17 and MGS:124 (appearing in

Supplementary Fig. 16c) was aligned to proteins listed by Roessner et al.32

as experimentally verified and strictly anaerobe corrin ring biosynthesis

proteins (60 coverage, 40% identity). CRISPR repeat-spacer segments were

identified with CRISPR-recognition tool (CRT, ver. 1.2)33 in selected CAG

assemblies. Genes were annotated as virulence or antibiotic resistance genes

when BLASTP alignments exceeded 80% identity over 80% of the length of

protein in the Virulence Factor Database (VFDB, February 2012 version) or

ResFinder34 (version 1.2) database, respectively.

From 271 essential genes from the genome of Bacillus subtilis strain 168

(ref. 35), 252 clusters of orthologous genes (COGs) were deduced (ftp://ftp.

ncbi.nih.gov/genomes/Bacteria/Bacillus_subtilis_168_uid57675/NC_000964.

ptt manually curated, see Supplementary Data 10). Genes aligning to these

COGs were termed essential genes.

CAGs significantly enriched for a specific annotation were identified using

Fisher’s exact test (P < 0.001 for Fig. 2b). Significant biases in eggNOG30 anno-

tation, as a function of the MGS observation frequency across the samples,

were identified using Wilcoxon rank sum test (P < 1 × 10−15; Supplementary

Data 8).

Co-abundance clustering. The canopy-based clustering of the gene catalog

was performed by iteratively picking a seed gene among the not yet clustered

genes and aggregate genes with abundance profiles within a fixed distance

from the seed gene abundance profile (Pearson correlation coefficient >

0.9 and Spearman’s rank correlation coefficient > 0.6) into the seed canopy.

Canopies with median abundance profiles within a distance of 0.97 PCC

from one another were merged. Canopies with 2 or less genes (such cano-

pies included a total of 1.7 M genes), or for which the canopy abundance

signal from any three samples constituted 90% or more of the total signal

across all samples, for which the median profile was detected in less than

four samples, or for which one sample made up 90% of the total signal (1.1 M

genes) were discarded for having insufficient supporting evidence (based

on Monte Carlo simulation; Supplementary Fig. 9). Canopies that passed

these criteria were called CAGs. CAGs with more than 700 genes are also

referred to as MGS or just species. Note that the number of clusters was not

predefined for the canopy-based clustering. CAG abundance profiles were

calculated as the sample-wise median gene depth signal (downsized). A CAG

was considered observed in a sample when its abundance profile exceeded

zero in that sample.

Source code for the canopy algorithm is freely available as Supplementary

Software and at http://git.dworzynski.eu/mgs-canopy-algorithm.

nature biotechnology

doi:10.1038/nbt.2939

MGS-augmented assembly. For each of the 741 MGS we performed a de novo

MGS-augmented assembly, using the subset of sequence reads that mapped

to the contigs from which the MGS genes originated. For each MGS, we per-

formed independent and sample-specific augmented assemblies on the two

samples from which most sequence reads mapped to the MGS and the sample

from which most of the MGS gene containing de novo contigs were derived.

For a given sample, the reads were aligned using Burrows-Wheeler Aligner36

(bwa-0.5.9) to the MGS specific scaffolds and the mapped reads, including

unmapped mates, were extracted. These reads were then corrected by Quake37

using k = 15. The reads were then de novo assembled with Velvet (1.2.01) using

k-mers from 21 to 45 and the parameters ‘-cov_cutoff auto’ and ‘-exp_cov auto’.

As several samples were used for assembly of each MGS, the best assembly

was selected based on ranking of contig N50 and the number of contigs in the

assemblies38. Contigs with read depth of less than half the average depth of all

contigs were removed from the assemblies38,39. The contigs and scaffolds were

then filtered to 100- and 500-bp minimum lengths, respectively, and gaps in

scaffolds were filled using SOAPdenovo GapCloser (1.10).

Assembly statistics. General assembly statistics were calculated using

assemblathon_stats.pl40 and coverage was calculated by aligning reads to the

contigs using bwa (0.5.9)36 and BEDtools. To assess the quality of the assem-

blies, we adopted the six high-quality draft assembly criteria from the Human

Microbiome Project (HMP)16. Five of these criteria address the contiguity of

the assembly, and one criterion, genome completeness, by counting core genes

contained in the assembly. The criteria are (i) 90% of the genome assembly must

be included in contigs >500 bp, (ii) 90% of the assembled bases must be at >

5× read coverage, (iii) The contig N50 must be >5 kb, (iv) scaffold N50 must

be >20 kb, (v) average contig length must be >5 kb and (vi) >90% of the core

genes must be present in the assembly. The core gene ratios were determined

using HMP standard operating procedure for both bacteria and archaea. In

short, blastx was used to identify core genes from the scaffolds and proteins

with at least 30% identity and 30% coverage for Bacteria and 50% identity and

70% coverage for Archaea were considered a core gene hit. The ratio of core

genes identified was then calculated using get_coregroups_coverage.pl (HMP

tools and protocols). In total 360 sample-specific MGS-augmented assemblies

from 247 unique MGS passed all six criteria (Supplementary Data 3). In addi-

tion, 139 unique assemblies passed five criteria.

We determined the number of novel species by aligning all proteins

to Uniprot using blastp and converted taxids from strain to species level

using NCBI taxonomy. An assembly was considered previously unsequenced

if less than 10% of the genes could be aligned with a minimum of 95%

identity over 33aa to genes from a species. 181 of the 238 high-quality

assembled draft genomes plus 83 assemblies passing 5 criteria were identified

as novel species.

Screening for chimeric assemblies. Because the HMP criteria were created

for single genome assembly, we applied three additional metrics to account

for putative chimeric assemblies arising from metagenomic data, (i) uniform-

ity of the contig read depth distribution, (ii) identification of multiple copies

of conserved 40 COGs30 and (iii) inter-assembly tetranucleotide frequency

(TNF) consistency.

Because assemblies consisting of genomic regions from different

organisms are likely to have multimodal coverage distributions, we performed

peak detection on the contig read coverage distributions for all assem-

blies passing 4-6 HMP criteria and assemblies with more than 1 peak were

manually inspected. From the presence of multiple copies of COGs, we were

able to identify three assemblies as chimeric. Of the 247 unique high-quality

draft assemblies, 9 (3.6%) were identified as potentially chimeric, and for the

additional 139 assemblies that passed five criteria, we identified 3 potential

chimeric assemblies (2.3%) and 1 without any core genes (MGS:3246).

The remaining assemblies have been deposited at the European Nucleotide

Archive (ENA).

Furthermore, tetranucleotide frequencies z-scores were calculated for all

assemblies and HMP reference assemblies as described by Teeling et al.41. For

each assembly the frequencies were calculated in windows of 5 kb to avoid

biases introduced by different scaffold lengths. If a scaffold was shorter than

the window size it was still included in the calculations. Within each assembly a

median tetranucleotide frequency z-profile was created and the tetranucleotide

frequency z-scores of each 5-kb window were correlated to this median profile

using PCC. The resulting high-quality draft genomes showed comparable TNF

correlations to the single organism HMP reference genomes indicating a low

rate of chimeric assemblies (Supplementary Fig. 12).

Comparison of MGS-augmented assemblies and reference genomes. To

estimate the completion level of the MGS-augmented assemblies, 299 draft

reference genomes from the human intestinal tract HMP DACC database and

the NCBI collection of complete reference genomes (both version updated

from 2012/04) were used as a reference set for a blast comparison procedure.

44 of the assemblies that passed 5 or more of the 6 HMP criteria (including

the bacteria/archaea core ratio criteria) were similar to a reference genome.

The contigs and scaffolds of these assemblies were projected on their closest

reference genomes using the GAGE pipeline for assembly quality evaluation42.

First nucmer (default parameters) was used to align the contigs/scaffolds to

the reference genome. Then delta-filter was used to remove low identity match

(parameters: -I 95, -o 80). Finally dnadiff was used to compare the assemblies

and the closest reference genome and estimate the mean identity and cover-

age of each contigs and scaffolds (Supplementar y Data 4). Additionally, the

MGS:337 assembly, which did not meet the six criteria, was 99.9% identical

to B. animalis subsp. lactis CNCM I-2494 (ref. 18) and covered 95% of this

reference genome (Fig. 4).

To search for potential contaminants, unaligned scaffold fragments

were blasted to the complete reference genome set, and the best hit (with

identity and coverage threshold of ≥95% and ≥80%, respectively) was

extracted. Scaffolds that matched to a different genus were considered

potential contaminants. Of the 44 MGS-augmented assemblies, only 16

contained any scaffolds with similarity to an alternative genus. In general

these scaffolds were small with an average size of only 2,721 bp. If we con-

sider unaligned scaffold with similarity to an alternative genus as a potential

contaminant, the mean contamination rate was estimated to 1.00 scaffold

per HQ assembly.

MGS-augmented assembly gap closure using Sanger sequence data. To

further experimentally validate the coherence of the sample-specific MGS-

augmented assemblies, we used Sanger sequence data from eight samples2.

Fecal microbial DNA from those individuals was used to construct plasmid-

based (pCNS) clone libraries of 3 kb long inserts, containing 250,000

clones each. Clone insert ends were sequenced using the Sanger technology

(ABI3730XL). Sequences were subsequently subjected to vector cleaning

and quality trimming, generating on average 230,468 (±5,145) reads per

sample. The same DNA was used for pyrosequencing (454GSFlx-Titanium),

resulting in 2,362,978 reads on average per sample (±3,245,603). For

each reference subject, Sanger and 454 reads as well as Velvet contigs gener-

ated from Illumina sequencing of the same DNA were combined for assembly

using the 454-Newbler software (v2.3). CAGs detected in a given reference

subject were compared with Sanger reads from that individual using blastn.

High-scoring segment pair (HSPs) covering at least 90% of the length of the

smallest read or velvet contig with at least 90% identity were selected, and

corresponding reads extracted. Scaffolding of the CAG contigs with paired

Sanger reads was then achieved using the bambus software43. Only assem-

blies with >1× coverage were kept, and used to assess the rate of gap closure

(Supplementar y Data 3). On average 64% of the assembly gaps were closed,

and in particular, the MGS:710 assembly was closed to only 3 scaffolds from

an initial 32 scaffolds.

Phylogeny of the MGS assemblies. We used all nonchimeric assemblies pass-

ing 5 and 6 HMP criteria (139 and 247 assemblies, respectively) and 296 HMP

gut microbiome reference genomes (HMPDACC) and 1,506 reference genomes

to construct a phylogeny based on 40 phylogenetic marker proteins (COGs)44.

For each assembly, proteins were predicted using Prodigal and aligned using

blastp to the individual COG proteins, and the best hits were selected requir-

ing at least 50% id over 50% of the COG sequence. For each COG the MGS

assembly and reference proteins were aligned using muscle and here joined to

a single alignment file for each COG using muscle-profile. The 40 individual

protein alignments were concatenated to a single alignment for each reference

nature biotechnology doi:10.1038/nbt.2939

genome/MGS assembly and alignments containing less than 35 COGs

were removed from further analysis, resulting in 337 MGS assemblies for

the final tree. The phylogenetic tree was constructed with FastTree using the

JTT substitution matrix with the parameters “-gamma -pseudo -spr 4 -mlacc

3 -slownni” and visualized using ITOL45.

Co-assembly of E . coli and dependency-associated CAGs. A pool of the

main E. coli (MGS:4) and its nine dependency-associated CAGs (CAG:427,

CAG:1345, CAG:2136, CAG:2318, CAG:2530, CAG:2610, CAG:3070,

CAG:3196 and CAG:5108) were used for recruiting 1,708 contigs for a

pooled assembly, across 247 selected samples. Subsequently, de novo assembly

(as described above) from 13 of these samples passed five or more HMP

criteria (Supplementary Data 7). A consensus assembly was generated

from the contigs of these assemblies using minimus2, where each assembly

was joined to the consensus in separate steps46. The consensus assembly

contained 4.3 Mb sequences in 45 contigs with a contig N50 of 129 kb.

Subsequently all the individual assemblies were aligned with blastn to the

consensus assembly, and contigs without a significant hit were pooled and

clustered using cd-hit-est with the parameters “-c 0.8 -n 7”47. To further reduce

redundancy of the extra contigs they were cut into 500-bp ‘reads’ with 250-bp

overlap and reassembled using Newbler 2.6. The resulting 157 contigs were

then added to the consensus assembly obtained from minimus2 to a final

assembly of 4.91 Mb in 202 contigs.

Dependency associations. A CAG was considered dependency associated on

another CAG when the sample-wise overlapping detections of the CAG pair

were statistically significantly over-represented (Fisher’s exact test, upper tail,

Bonferroni corrected P < 1 × 10−10) and the dependency-associated CAG was

not detected independently.

Smaller CAGs (<700 genes) were considered ‘co-existence associated’ when

their detections were significantly enriched (Fisher’s exact test, Bonferroni

corrected P < 0.05) in samples where an MGS pair was co-observed, and never

occurred independently of one of the two MGS (the host). Here an MGS pair

consisted of a host MGS and a companion MGS. An MGS was considered a

potential companion if it co-existed with a potential host species in samples

from ten or more individuals and was found independently of the host species

in samples from ten or more human individuals. For the co-existence-

associated relationships where the small CAG was not observed independently

of any of the two MGS, the host species were determined as the MGS with

the strongest abundance correlation to the small CAG across samples where

both were detected, and by the sample specific co-assembly. No inconsistency

between these measures was found.

Dependency-associated small CAGs were considered significantly

absent in samples where a specific companion species was found, when it was

significantly enriched in samples where the companion species was absent

compared to samples where the host MGS was found (Fisher’s exact test,

Bonferroni corrected P < 0.05). Furthermore, the small CAG could never be

observed independent of the hosting species. Again, an MGS was considered

as a potential companion species if it co-existed with a host species in samples

from ten or more individuals.

For all types of dependency associations a CAG was considered detected in

a sample if the CAG abundance profile exceeded zero. Furthermore, only

CAGs detected in ≥10 and ≤308 samples were considered. To ensure

independence between the observations only one sample per individual

was used (n = 318). Dependency-associated, ‘co-existence–associated’ and

‘co-existence–absent’ CAGs showed correlation to the species richness

comparable to that of all CAGs.

Estimation of CAG persistence probabilities. Data from 73 human individu-

als, which were sampled twice, were used to estimate the annual persistence

probabilities of MGS with or without dependency-associated CAGs (Fig. 6,

Supplementary Fig. 17 and Supplementary Data 6). All of the 2 × 73 stool

samples included in this analysis resulted in at least 11 M sequence reads, and

samples yielding more than this were downsized to 11 M reads. Furthermore,

all included sample pairs exhibited strong stability between the samplings,

in that they were more similar to each other than to 99% of the other

samples in the cohort (using the Spearman correlation coefficient of the MGS

abundances as similarity measure). Four of the original 77 sample pairs did

not pass these criteria.

The main idea in this analysis was the following: for a fraction of the

73 sample pairs, a given MGS is present in the sample obtained at time point 1.

For a subset of these sample pairs, the same MGS was also present at time point 2.

Based on these, data logistic regression can be used to estimate an annual

persistence probability for the MGS. The predictor variable (time between

two consecutive samples) is continuous, whereas the outcome variable (pres-

ence or absence of an MGS) is binary. Logistic regression is used to estimate

how the probability of an MGS still being present depends on the amount of

time passed.

This computation is based on the assumption that an MGS has a typical

probability per time unit of persisting in the gut of an individual. Thus the

likelihood of observing an MGS at time point 2 is expected to be smaller

the more time that has passed between the two samplings. Specifically, this

decline is assumed to be exponential; thus if the probability that a given MGS

will persist for a year is P(1) = 0.7, then the probability that it will persist for

two years is P(2) = 0.72 = 0.49, etc. This assumption seems to fit well with

Kaplan-Meier cur ves constructed from these data (Fig. 6a). Of course, the

persistence of a given MGS in any individual is likely to depend on the specific

conditions in that individual. We simply assume that there is a typical overall

annual persistence probability associated with the MGS (on average, a given

MGS has a typical tendency to persist in the gut of any individual), and real

data will be scattered around this average according to unidentified covariates

and stochastic effects.

Annual persistence probabilities were estimated in a probabilistic (Bayesian)

model–based framework that explicitly accounts for time dependence. In this

approach we assume that the annual persistence probability for an MGS is

determined by the inherent resilience of the MGS itself, in combination with

possible additional effects (positive or negative) caused by a set of dependency-

associated CAGs. Specifically, we assume that the annual persistence

probability, P, for a given MGS, depends on the effects of a set of dependency-

associated CAGs in the form of a logistic regression model: ln(P/[1-P]) =

logit(P) = b0 + Sum[biXi] or, equivalently: P = expit(b0 + Sum[biXi]). Here, the

regression coefficient b0 corresponds to the inherent persistence tendency of

the MGS itself, bi corresponds to the effect of dependency-associated CAG

number i and Xi is a binary variable indicating whether dependency-associated

CAG number i is present or absent for a given sample. “Expit” is the sigmoidal,

logistic function (the inverse of the logit function). The index, i, runs over all

the dependency-associated CAGs for a given MGS.

The probability that a CAG will survive for t days, P(t), can be found from its

annual persistence probability, P, in the following way: P(t) = P[t/365]. The like-

lihood for a data point where the MGS survives (i.e., where it is still present at

the second sample, after t days have elapsed) is therefore given by the following

expression: L = P[t/365] = [expit(b0 + Sum[biXi])] [t/365]. For data points where a

CAG does not survive, the likelihood is simply: L = 1 − P[t/365] = 1 − [expit(b0 +

Sum[biXi])] [t/365]. As recommended in Gelman et al.48 the priors for all b0

regression coefficients (which correspond to the inherent persistence of all

MGS) are t-distributions with µ = 0, d.f. = 1, and rate = 0.1 (corresponding to

scale = 10). The priors for all bi regression coefficients (corresponding to the

effects on persistence of the dependency-associated CAGs) are t-distributions

with µ = 0, d.f. = 1 and rate = 0.4 (corresponding to scale = 2.5). These are

conservative priors that help keep the correlation coefficients close to zero.

Given these expressions for priors and likelihoods, it is possible to perform a

Bayesian analysis of the model, resulting in estimates of the above-mentioned

regression coefficients. However, since the regression coefficients themselves

can be difficult to interpret, we instead report the following derived measures:

(i) the annual persistence probability for each MGS. This can be computed as:

P = expit(b0). (ii) The annual persistence probability for a specific MGS when

together with a given dependency-associated CAG. This can be computed as:

P = expit(b0 + bj), where j refers to the specific dependency-associated CAG.

(iii) The effect of the dependency-associated CAG. We have chosen to simply

express this as the absolute difference between the above two measures. (For

instance, if the annual persistence probability of an MGS, together with a

specific dependency-associated CAG, is 0.75, and the annual persistence prob-

ability of the MGS alone is 0.5, then the effect of the dependency-associated

CAG is reported as 0.75 – 0.5 = 0.25).

nature biotechnology

doi:10.1038/nbt.2939

For the analysis of coexistence between a pair of MGS and an associated

CAG (Supplementary Note 9, Supplementary Fig. 16 and Supplementary

Data 11), there were insufficient data to obtain estimates for each individual

CAG. We therefore pooled all data points for CAGs having a positive effect on

the persistence of their MGS host, and estimated an overall effect for these.

Note that, in the Bayesian framework, estimates are expressed as prob-

ability distributions over the possible values for parameters of interest20. We

therefore obtain both an estimate of a parameter, and quantification of how

certain we are of the estimate. To declare an effect to be substantially different

from zero, we require that its 95% highest posterior density interval (the “95%

HDI”) should be located entirely outside of a “region of practical equivalence”

to 0 (a “ROPE”). In this analysis the ROPE was defined to be [–0.02, 0.02].

The 95% HDI is the narrowest interval that includes 95% of the probability.

By design, all parameter values inside a 95% HDI will be more likely than

all values outside. In this work we have identified 26 dependency-associated

CAGs where we are more than 95% certain that they have a nonzero effect on

the persistence probability of an MGS (Supplementary Data 6).

The model was implemented and analyzed in a Bayesian framework by

Markov chain Monte Carlo (MCMC) using the JAGS package49. Convergence

of MCMC was checked by running two independent chains and verifying that

they arrived at similar posterior distributions. In particular it was checked that

the potential scale reduction factor (“R-hat”) for each estimated parameter

was <1.1 (ref. 50).

22. Kultima, J.R. et al. MOCAT: a metagenomics assembly and gene prediction toolkit.

PLOS ONE 7, e47656 (2012).

23. Li, R. et al. De novo assembly of human genomes with massively parallel short

read sequencing. Genome Res. 20, 265–272 (2010).

24. Zhu, W., Lomsadze, A. & Borodovsky, M. Ab initio gene identiﬁcation in metagenomic

sequences. Nucleic Acids Res. 38, e132 (2010).

25. Li, R. et al. SOAP2: an improved ultrafast tool for short read alignment.

Bioinformatics 25, 1966–1967 (2009).

26. Leplae, R., Lima-Mendez, G. & Toussaint, A. ACLAME: a classiﬁcation of mobile

genetic elements, update 2010. Nucleic Acids Res. 38, D57–D61 (2010).

27. Finn, R.D., Clements, J. & Eddy, S.R. HMMER web server: interactive sequence

similarity searching. Nucleic Acids Res. 39, W29–37 (2011).

28. Punta, M. et al. The Pfam protein families database. Nucleic Acids Res. 40,

D290–D301 (2012).

29. Kristensen, D.M., Cai, X. & Mushegian, A. Evolutionarily conserved orthologous

families in phages are relatively rare in their prokaryotic hosts. J. Bacteriol. 193,

1806–1814 (2011).

30. Powell, S. et al. eggNOG v3.0: orthologous groups covering 1133 organisms at 41

different taxonomic ranges. Nucleic Acids Res. 40, D284–D289 (2012).

31. Tringe, S.G. et al. Comparative metagenomics of microbial communities. Science

308, 554–557 (2005).

32. Roessner, C.A. & Scott, A.I. Fine-tuning our knowledge of the anaerobic route to

cobalamin (vitamin B12). J. Bacteriol. 188, 7331–7334 (2006).

33. Bland, C. et al. CRISPR recognition tool (CRT): a tool for automatic detection of

clustered regularly interspaced palindromic repeats. BMC Bioinformatics 8, 209

(2007).

34. Zankari, E. et al. Identiﬁcation of acquired antimicrobial resistance genes.

J. Antimicrob. Chemother. 67, 2640–2644 (2012).

35. Kobayashi, K. et al. Essential Bacillus subtilis genes. Proc. Natl. Acad. Sci. USA

100, 4678–4683 (2003).

36. Li, H. & Durbin, R. Fast and accurate short read alignment with Burrows-Wheeler

transform. Bioinformatics 25, 1754–1760 (2009).

37. Kelley, D.R., Schatz, M.C. & Salzberg, S.L. Quake: quality-aware detection and

correction of sequencing errors. Genome Biol. 11, R116 (2010).

38. Zerbino, D.R. & Birney, E. Velvet: algorithms for de novo short read assembly using

de Bruijn graphs. Genome Res. 18, 821–829 (2008).

39. Mavromatis, K. et al. Use of simulated data sets to evaluate the ﬁdelity of

metagenomic processing methods. Nat. Methods 4, 495–500 (2007).

40. Earl, D. et al. Assemblathon 1: a competitive assessment of de novo short read

assembly methods. Genome Res. 21, 2224–2241 (2011).

41. Teeling, H., Meyerdierks, A., Bauer, M., Amann, R. & Glöckner, F.O. Application of

tetranucleotide frequencies for the assignment of genomic fragments. Environ.

Microbiol. 6, 938–947 (2004).

42. Salzberg, S.L. et al. GAGE: a critical evaluation of genome assemblies and assembly

algorithms. Genome Res. 22, 557–567 (2012).

43. Koren, S., Treangen, T.J. & Pop, M. Bambus 2: scaffolding metagenomes.

Bioinformatics 27, 2964–2971 (2011).

44. Ciccarelli, F.D. et al. Toward automatic reconstruction of a highly resolved tree of

life. Science 311, 1283–1287 (2006).

45. Letunic, I. & Bork, P. Interactive Tree Of Life v2: online annotation and display of

phylogenetic trees made easy. Nucleic Acids Res. 39, W475–W478 (2011).

46. Treangen, T.J., Sommer, D.D., Angly, F.E., Koren, S. & Pop, M. Next generation

sequence assembly with AMOS. Curr. Protoc. Bioinformatics Chapter 11, Unit 11.8

(2011).

47. Li, W. & Godzik, A. Cd-hit: a fast program for clustering and comparing large

sets of protein or nucleotide sequences. Bioinformatics 22, 1658–1659

(2006).

48. Gelman, A., Jakulin, A., Pittau, M.G. & Su, Y. A weakly informative default prior

distribution for logistic and other regression models. Ann. Appl. Stat. 2, 1360–1383

(2008).

49. Plummer, M. JAGS: a program for analysis of Bayesian graphical models using

Gibbs sampling. in Proc. 3rd Int. Work. Distrib. Stat. Comput. March, 20–22

(2003).

50. Gelman, A. & Rubin, D. Inference from iterative simulation using multiple sequences.

Stat. Sci. 7, 457–511 (1992).

Multi-biome analysis identifies distinct gut microbial signatures and their crosstalk in ulcerative colitis and Crohn’s disease

Article

Full-text available

Nov 2024

The integrative multi-kingdom interaction of the gut microbiome in ulcerative colitis (UC) and Crohn’s disease (CD) remains underinvestigated. Here, we perform shotgun metagenomic sequencing of feces from patients with UC and CD, and healthy controls in the Japanese 4D cohort, profiling bacterial taxa, gene functions, and antibacterial genes, bacteriophages, and fungi. External metagenomic datasets from the US, Spain, the Netherlands, and China were analyzed to validate our multi-biome findings. We found that Enterococcus faecium and Bifidobacterium spp. were enriched in both diseases. Enriched Escherichia coli was characteristic of CD and was linked to numerous antibiotic resistance genes involved in efflux pumps and adherent-invasive Escherichia coli virulence factors. Virome changes correlated with shifts in the bacteriome, including increased abundances of phages encoding pathogenic genes. Saccharomyces paradoxus and Saccharomyces cerevisiae were enriched in UC and CD, respectively. Saccharomyces cerevisiae and Escherichia coli had negative associations with short-chain fatty acid (SCFA)-producing bacteria in CD. Multi-biome signatures and their interactions in UC and CD showed high similarities between Japan and other countries. Since bacteria, phages, and fungi formed multiple hubs of intra- or trans-kingdom networks with SCFA producers and pathobionts in UC and CD, an approach targeting the interaction network may hold therapeutic promise.

Микробиом кишечника человека: научная парадигма и клиническая практика: The human gut microbiome: scientific paradigm shifts and clinical practice today and tomorrow

Article

Jun 2024

Исследования микробиома человека, начатые 15 лет назад и активно продолжающиеся в настоящее время, позволили во многом пересмотреть наши взгляды о глубинных механизмах взаимодействия организма человека и населяющих его микроорганизмов. Представшая картина оказалась много сложнее, чем ожидалось, особенно это относится к микробиому кишечника. В настоящее время научные концепции того, что из себя представляет микробиом кишечника, продолжают активно разрабатываться. В основе их лежит, конечно, редукционный подход. Так, предлагается теория «корового микро�биома» и сопряженная с ней концепция «здорового микробиома». Еще более упрощенно подходит к решению проблемы многообразия и вариабельности кишечных микробиомов теория «энтеротипов». Она допускает существование предпочтительных (устойчивых) вариантов микробиомов, структурно организованных вокруг определенных групп микроорганиз�мов. Если структурообразующими являются анаэробные грамотрицательные палочки, то такие микробиомы отличаются низким видовым разнообразием, доминированием небольшой группы микроорганизмов и низкой стрессоустойчивостью. Именно такие типы микробиомов обнаруживают у пациентов с различными патологиями, в том числе, внекишечной локализации. Поэтому предполагается, что выделение энтеротипов имеет практический смысл. В настоящее время научная концепция микробиомов имеет мало выходов в медицинскую практику. Но уже имеющийся практический опыт применения про- и синбиотиков, препаратов бактериофагов и фекальной трансплантации свидетельствует о реальной возможно�сти изменения микробиомов для профилактики и лечения различных заболеваний человека.

A Swedish dietary guideline index, gut microbial α-diversity and prevalence of metabolic syndrome – observations in the Swedish CArdioPulmonary bioImage Study (SCAPIS)

Article

Full-text available

Nov 2024

Background: Metabolic syndrome (MetS) is characterized by coexisting risk factors for type 2 diabetes and cardiovascular disease. Diet is of importance in their aetiology, and gut microbiota (GM) may constitute a link between diet and metabolic health. Understanding the interplay between diet and GM could contribute novel insights for future dietary guidelines, and aid in preventive actions to motivate adherence to dietary guidelines. Objective: We intended to create a Swedish dietary guideline index (SweDGI) measuring adherence to 12 Swedish dietary guidelines and examine whether SweDGI and its components are associated with GM α-diversity (Shannon index) and prevalent MetS, and if the association between the Shannon index and MetS differs depending on SweDGI. Design: SweDGI was based on food-frequency data assessed 2014–2018 in 10,396 diabetes-free participants from the Malmö and Uppsala-sites of the Swedish CArdioPulmonary bioImage Study (SCAPIS) (50–64 y, 53% women). We estimated the Shannon index from shotgun metagenomic sequencing-data to assess microbial richness and evenness. We used a general linear model to examine cross-sectional SweDGI-Shannon associations and logistic regression for associations with MetS. Results: Most guidelines were followed by less than half of the participants. Men showed poorer adherence. Higher SweDGI was linked to higher Shannon index (P-trend across five SweDGI-groups = 1.7 × 10-12). Most guidelines contributed to this observation. Higher SweDGI and Shannon index were associated with lower MetS-prevalence, where the lowest prevalence was observed among those with both high SweDGI and high Shannon index (odds ratio:0.43; 95% confidence interval:0.35, 0.52). Both the Shannon index and SweDGI were associated with MetS, independently of the level of the other factor (P-interaction = 0.82). Conclusions: We created a new index to comprehensively reflect adherence to the Swedish dietary guidelines in sub-cohorts within the large multicentre SCAPIS study. Better adherence was associated with a richer and more even GM and lower prevalence of MetS. The inverse association between the Shannon index and MetS was consistent at different levels of adherence to dietary guidelines

MOSAIC: A Pipeline for MicrobiOme Studies Analytical Integration and Correction

Preprint

Nov 2024

Large-scale and consortium microbiome studies have enabled identification of reliable population-level biomedical signals, wherein integration is essential to eliminate unwanted variations between batches or studies and retain biological signals. Many strategies, each with distinct advantages and limitations, have been adapted or developed for microbiome data. The optimal strategy for a given study needs to be determined on a data-specific, case-by-case basis. Here, we develop the first-of-its-kind MicrobiOme Studies Analytical Integration and Correction (MOSAIC) pipeline to enable a convenient, fair, and comprehensive comparison of integration strategies. It includes modules for pre-processing, integration, and evaluation of artifact removal and signal preservation, using metrics relevant to common microbiome analyses, including alpha and beta diversities, disease prediction, and differential abundance analysis. We applied MOSAIC to extensive real-world and simulated data and found that though no single strategy excels in all aspects, yet certain strategies, the ComBat and ConQuR families, perform better overall.

The human gut metacommunity as a conceptual aid in the development of precision medicine

Article

Full-text available

Oct 2024

Gerald W Tannock

Human gut microbiomes (microbiotas) are highly individualistic in taxonomic composition but nevertheless are functionally similar. Thus, collectively, they comprise a “metacommunity.” In ecological terminology, the assembly of human gut microbiomes is influenced by four processes: selection, speciation, drift, and dispersal. As a result of fortuitous events associated with these processes, individual microbiomes are taxonomically “tailor-made” for each host. However, functionally they are “off-the-shelf” because of similar functional outputs resulting from metabolic redundancy developed in host-microbe symbiosis. Because of this, future microbiological and molecular studies of microbiomes should emphasize the metabolic interplay that drives the human gut metacommunity and that results in these similar functional outputs. This knowledge will support the development of remedies for specific functional dysbioses and hence provide practical examples of precision medicine.

Effect of Micro-granular Activated Carbon on Bacteriophage MS2 Removal and Fouling Control in Flat-Plate MBR

Article

Nov 2024

GutMetaNet: an integrated database for exploring horizontal gene transfer and functional redundancy in the human gut microbiome

Article

Full-text available

Nov 2024

Metagenomic studies have revealed the critical roles of complex microbial interactions, including horizontal gene transfer (HGT) and functional redundancy (FR), in shaping the gut microbiome’s functional capacity and resilience. However, the lack of comprehensive data integration and systematic analysis approaches has limited the in-depth exploration of HGT and FR dynamics across large-scale gut microbiome datasets. To address this gap, we present GutMetaNet (https://gutmetanet.deepomics.org/), a first-of-its-kind database integrating extensive human gut microbiome data with comprehensive HGT and FR analyses. GutMetaNet contains 21 567 human gut metagenome samples with whole-genome shotgun sequencing data related to various health conditions. Through systematic analysis, we have characterized the taxonomic profiles and FR profiles, and identified 14 636 HGT events using a shared reference genome database across the collected samples. These HGT events have been curated into 8049 clusters, which are annotated with categorized mobile genetic elements, including transposons, prophages, integrative mobilizable elements, genomic islands, integrative conjugative elements and group II introns. Additionally, GutMetaNet incorporates automated analyses and visualizations for the HGT events and FR, serving as an efficient platform for in-depth exploration of the interactions among gut microbiome taxa and their implications for human health.

Analyses of the gut microbial composition of domestic pig louse Haematopinus suis

Article

Nov 2024
MICROB PATHOGENESIS

Haematopinus suis is an obligatory ectoparasite of the domestic pig, serving as a vector of several swine pathogens and posing great threats to the pig industry. The gut microbiome of lice is thought of an important mediator of their healthy physiology. However, there is a great paucity of lice-associated microbial communities’ structure and function. The current study aimed to profile the gut microbiome and to understand the microbial functions of swine lice by metagenomic sequencing and bioinformatics analyses. In total, 102,358 (77.2 %) nonredundant genes were cataloged, by contrast, only a small proportion of genes were assigned to microbial taxa and functional assemblages. Bacteria of known or potential public health significance such as Anaplasma phagocytophilum, Chlamydia trachomatis, Waddlia chondrophila, Bacillus cereus, and Leptotrichia goodfellowii were observed in all samples. The integrated microbial profile further illustrated the evolutionary relevance of endosymbionts and detailed the functional composition, and findings suggested H. suis may acquire adenosylcobalamin by feeding due to an adenosylcobalamin synthesis defect and a lack of complete synthases of endosymbionts. Sucking lice contained fewer functional genes compared with ticks and fleas probably because of the obligate host specificity of parasitic lice. In addition, the genes from the intestines contained encompassed most of the microbial functional genes in sucking lice. A wide range of unknown taxonomic and functional assemblages were discovered, which improves our understanding related to microbial features and physiological activities of sucking lice. In general, this study increases the characterization of the microbiota of lice and offers clues for preventing and controlling lice infestation in swine production in the future.

Concomitant gut dysbiosis and defective gut barrier serve as the bridges between hypercortisolism and chronic systemic inflammation in Cushing disease

Article

Oct 2024
EUR J ENDOCRINOL

Objective The aim of this study was to investigate the gut microbial signatures and related pathophysiological implications in patients with Cushing’s disease (CD). Design and methods 27 patients with CD and 45 healthy controls were enrolled. Based on obtained metagenomics data, we performed correlation, network study and genome interaction group (GIG) analysis. Fecal metabolomics and serum ELISA analysis were conducted in dichotomized CD patients. Caco-2 cells were incubated with gradient concentrations of cortisol for subsequent transepithelial electrical resistance (TEER) measurement, FITC-dextran transwell permeability assay, qPCR and western blot analysis. Results Gut microbial composition in patients with CD was notably different from that in healthy controls. Network analysis revealed that Eubacterium siraeum might serve as the core specie in the gut microbial system of CD patients. Subsequent GIG analysis identified the positive correlations between GIG9 and UFC. Further serum ELISA and fecal metabolomics uncovered that CD patients with elevated UFC levels were characterized with increased lipopolysaccharide binding protein (LBP). Moreover, remarkable positive association was found between LBP level and relative abundance of Eubacterium siraeum. TEER and FITC-Dextran transwell assays demonstrated that hypercortisolism induced increased gut permeability. Further qPCR and western blot analysis suggested that dysregulated AhR/Claudin 2 axis might be involved in the development of hypercortisolism-induced defective gut barrier function Conclusions Disease activity associated dysbiosis and defective gut barrier might jointly facilitate the development of systemic inflammation in patients with CD.

MicroHDF: predicting host phenotypes with metagenomic data using a deep forest-based framework

Article

Oct 2024

The gut microbiota plays a vital role in human health, and significant effort has been made to predict human phenotypes, especially diseases, with the microbiota as a promising indicator or predictor with machine learning (ML) methods. However, the accuracy is impacted by a lot of factors when predicting host phenotypes with the metagenomic data, e.g. small sample size, class imbalance, high-dimensional features, etc. To address these challenges, we propose MicroHDF, an interpretable deep learning framework to predict host phenotypes, where a cascade layers of deep forest units is designed for handling sample class imbalance and high dimensional features. The experimental results show that the performance of MicroHDF is competitive with that of existing state-of-the-art methods on 13 publicly available datasets of six different diseases. In particular, it performs best with the area under the receiver operating characteristic curve of 0.9182 ± 0.0098 and 0.9469 ± 0.0076 for inflammatory bowel disease (IBD) and liver cirrhosis, respectively. Our MicroHDF also shows better performance and robustness in cross-study validation. Furthermore, MicroHDF is applied to two high-risk diseases, IBD and autism spectrum disorder, as case studies to identify potential biomarkers. In conclusion, our method provides an effective and reliable prediction of the host phenotype and discovers informative features with biological insights.

Application of tetranucleotide frequencies for the assignment of genomic fragments

Article

Full-text available

Jan 2004

A metagenome-wide association study of gut microbiota in type 2 diabetes

Article

Full-text available

Jan 2012

Assessment and characterization of gut microbiota has become a major research area in human disease, including type 2 diabetes, the most prevalent endocrine disease worldwide. To carry out analysis on gut microbial content in patients with type 2 diabetes, we developed a protocol for a metagenome-wide association study (MGWAS) and undertook a two-stage MGWAS based on deep shotgun sequencing of the gut microbial DNA from 345 Chinese individuals. We identified and validated approximately 60,000 type-2-diabetes-associated markers and established the concept of a metagenomic linkage group, enabling taxonomic species-level analyses. MGWAS analysis showed that patients with type 2 diabetes were characterized by a moderate degree of gut microbial dysbiosis, a decrease in the abundance of some universal butyrate-producing bacteria and an increase in various opportunistic pathogens, as well as an enrichment of other microbial functions conferring sulphate reduction and oxidative stress resistance. An analysis of 23 additional individuals demonstrated that these gut microbial markers might be useful for classifying type 2 diabetes.

Richness of human gut microbiome correlates with metabolic markers

Article

Full-text available

Aug 2013
NATURE

We are facing a global metabolic health crisis provoked by an obesity epidemic. Here we report the human gut microbial composition in a population sample of 123 non-obese and 169 obese Danish individuals. We find two groups of individuals that differ by the number of gut microbial genes and thus gut bacterial richness. They contain known and previously unknown bacterial species at different proportions; individuals with a low bacterial richness (23% of the population) are characterized by more marked overall adiposity, insulin resistance and dyslipidaemia and a more pronounced inflammatory phenotype when compared with high bacterial richness individuals. The obese individuals among the lower bacterial richness group also gain more weight over time. Only a few bacterial species are sufficient to distinguish between individuals with high and low bacterial richness, and even between lean and obese participants. Our classifications based on variation in the gut microbiome identify subsets of individuals in the general white adult population who may be at increased risk of progressing to adiposity-associated co-morbidities.

Bayesian data analysis for newcomers

Article

Apr 2017

This article explains the foundational concepts of Bayesian data analysis using virtually no mathematical notation. Bayesian ideas already match your intuitions from everyday reasoning and from traditional data analysis. Simple examples of Bayesian data analysis are presented that illustrate how the information delivered by a Bayesian analysis can be directly interpreted. Bayesian approaches to null-value assessment are discussed. The article clarifies misconceptions about Bayesian methods that newcomers might have acquired elsewhere. We discuss prior distributions and explain how they are not a liability but an important asset. We discuss the relation of Bayesian data analysis to Bayesian models of mind, and we briefly discuss what methodological problems Bayesian data analysis is not meant to solve. After you have read this article, you should have a clear sense of how Bayesian data analysis works and the sort of information it delivers, and why that information is so intuitive and useful for drawing conclusions from data.

HMMER web server: Interactive sequence similarity searching

Article

Jan 2011

Interactive Tree of Life v2: online annotation and display of phylogenetic trees made easy

Article

Jan 2011

Toward automatic reconstruction of a highly resolved tree of life (vol 311, pg 1283, 2006)

Article

May 2006

Francesca D Ciccarelli

A Human Gut Microbial Gene Catalogue Established by Metagenomic Sequencing

Article

Mar 2010
NATURE

To understand the impact of gut microbes on human health and well-being it is crucial to assess their genetic potential. Here we describe the Illumina-based metagenomic sequencing, assembly and characterization of 3.3 million non-redundant microbial genes, derived from 576.7 gigabases of sequence, from faecal samples of 124 European individuals. The gene set, approximately 150 times larger than the human gene complement, contains an overwhelming majority of the prevalent (more frequent) microbial genes of the cohort and probably includes a large proportion of the prevalent human intestinal microbial genes. The genes are largely shared among individuals of the cohort. Over 99% of the genes are bacterial, indicating that the entire cohort harbours between 1,000 and 1,150 prevalent bacterial species and each individual at least 160 such species, which are also largely shared. We define and describe the minimal gut metagenome and the minimal gut bacterial genome in terms of functions present in all individuals and most bacteria, respectively.

A weakly informative default prior distribution for logistic and other regression models

Article

Aug 2007
ANN APPL STAT

We propose a new prior distribution for classical (non-hierarchical) logistic regres- sion models, constructed by first scaling all nonbinary variables to have mean 0 and standard deviation 0.5, and then placing independent Student-t prior distributions on the coefficients. As a default choice, we recommend the Cauchydistribution with center 0 and scale 2.5, which in the simplest setting is a longer-tailed version of the distribu- tion attained by assuming one-half additional success and one-half additional failure in a logistic regression. Cross-validation on a corpus of datasets shows the Cauchy class of prior distributions to outperform existing implementations of Gaussian and Laplace priors. We recommend this prior distribution as a default choice for routine applied use. It has the advantage of always giving answers, even when there is complete separation in logistic regression (a common problem, even when the sample size is large and the number of predictors is small) and also automatically applying more shrinkage to higher- order interactions. This can be useful in routine data analysis as well as in automated procedures such as chained equations for missing-data imputation. We implement a procedure to fit generalized linear models in R with the Student-t prior distribution by incorporating an approximate EM algorithm into the usual itera- tively weighted least squares. We illustrate with several examples, including a series of logistic regressions predicting voting preferences, a small bioassay experiment, and an imputation model for a public health data set.

Defining Genome Project Standards in a New Era of Sequencing (GSC8 Meeting)

Article

Oct 2009

Chain

The Genomic Standards Consortium was formed in September 2005. It is an international, open-membership working body which promotes standardization in the description of genomes and the exchange and integration of genomic data. The 2009 meeting was an activity of a five-year funding "Research Coordination Network" from the National Science Foundation and was organized held at the DOE Joint Genome Institute with organizational support provided by the JGI and by the University of California - San Diego.