ArticlePDF Available

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

Authors:

Abstract

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.
© 2014 Nature America, Inc. All rights reserved.
nature biotechnologyADVANCE 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
1
, Junhua Li
6–8
,
Shinichi Sunagawa
9
, Damian R Plichta
1
, Laurent Gautier
1
, Anders G Pedersen
1
, Emmanuelle Le Chatelier
3,4
,
Eric Pelletier
10–12
, Ida Bonde
1,2
, Trine Nielsen
13
, Chaysavanh Manichanh
14
, Manimozhiyan Arumugam
7,9,13
,
Jean-Michel Batto
3,4
, Marcelo B Quintanilha dos Santos
1
, Nikolaj Blom
2
, Natalia Borruel
14
, Kristoffer S Burgdorf
13
,
Fouad Boumezbeur
3,4
, Francesc Casellas
14
, Joël Doré
3,4
, Piotr Dworzynski
1
, Francisco Guarner
14
,
Torben Hansen
13,15
, Falk Hildebrand
16,17
, Rolf S Kaas
18
, Sean Kennedy
3,4
, Karsten Kristiansen
7,19
, Jens Roat Kultima
9
,
Pierre Léonard
3,4
, Florence Levenez
3,4
, Ole Lund
1
, 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
25
, Julien Tap
9
,
Sebastian Tims
26
, David W Ussery
1
, Takuji Yamada
9,27
, MetaHIT Consortium
28
, Pierre Renault
3
,
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
© 2014 Nature America, Inc. All rights reserved.
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. 46 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.
© 2014 Nature America, Inc. All rights reserved.
nature biotechnologyADVANCE 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
+
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
10
100
1,000
5,000
CAG size (genes, log)
b
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.
© 2014 Nature America, Inc. All rights reserved.
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
80
60
40
20
0
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
0
100
80
60
40
20
0
Genes captured
Number of samples containing B. animalis
19 15 10 8 6 4 3 12
2,000
4,000
0
100
80
60
40
20
0
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
0
a
b
c
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.
(c) Co-abundance clusterings on a series of random sample subsets of
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).
© 2014 Nature America, Inc. All rights reserved.
nature biotechnologyADVANCE 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. coliassociated
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
c
CRISPR
MGS:135
CAG:4065
CAG:4755
CAG:4703
CAG:5806
CAG:2350
CAG: 2543
CAG:3731
CAG:4011
MCC = – 0.7
b
a
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).
© 2014 Nature America, Inc. All rights reserved.
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
© 2014 Nature America, Inc. All rights reserved.
nature biotechnologyADVANCE 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.
© 2014 Nature America, Inc. All rights reserved.
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.
© 2014 Nature America, Inc. All rights reserved.
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
© 2014 Nature America, Inc. All rights reserved.
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).
© 2014 Nature America, Inc. All rights reserved.
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).
... Уже первые результаты, полученные на весьма ограниченной в численном отношении группе здоровых добровольцев, показали не только невероятное разнообразие видов микроорганизмов у каждого конкретного индивидуума (α-разнообразие) и сложную индивидуальную динамику, но и еще больший разброс между отдельными участниками групп (β-разнообразие). Довольно быстро стало понятно, что подходы классической экологии, да и микробной экологии, базирующиеся на них, слабо применимы к такой сложной системе [6,7]. Большие надежды возлагались на молекулярно-биологические методы, арсенал которых весьма обширен и продолжает постоянно пополняться. ...
... Получено огромное количество метагеномов, но их расшифровка значительно отстает. Для преодоления этих трудностей при метагеномных исследованиях нередко оперируют условным понятием «метагеномные виды» («metagenomic species, MGS») [6], или ведется поиск маркерных генов, кото-рые в достаточной мере характеризуют виды микроорганизмов («metagenomic operational taxonomic units, mOTUs») [9]. Попытки решения возникающих трудностей путем совершенствования биоинформационных программных алгоритмов пока не привели к ожидаемым результатам: новые алгоритмы не позволяют адекватно описывать структуру микробных сообществ человека [10]. ...
Article
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.
... The microbiome has been shown to play critical roles in human health and disease, such as type II diabetes (T2D) [1][2][3] , inflammatory bowel disease (IBD) [4][5][6][7] , colorectal cancer (CRC) [8][9][10][11] , among others. Given the importance of microbiome and the reduced cost of profiling technologies, large-scale and consortium microbiome studies, which involve multiple batches or studies on the same disease, have emerged and become the trend 5,12,13 . ...
Preprint
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.
... Mutation of genes occurs commonly; de novo mutations are estimated at 2 × 10 9 to 6 × 10 12 single nucleotide polymorphisms per microbiota per day. Gene loss and gene gain (by horizontal gene transfer) in bacterial species reveals a genetically mutable community in real time (Nielsen et al., 2014;Zhao et al., 2019). Consequently, strains of the same species differ in genetic characteristics where "core" genes common to all strains plus "dispensable" genes (variable presence) together comprising the pangenome, can be recognized (Medini et al., 2005;Truong et al., 2017). ...
Article
Full-text available
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.
... Metagenomic species were defined by binning of co-abundance genes, as previously described. 33 Alpha diversity, a metric of the diversity of species within a sample, was estimated using the Shannon diversity index, inverse Simpson index, and richness (number of species). Beta diversity, a metric of the composition dissimilarity between samples, was estimated using the Bray-Curtis dissimilarity. ...
Conference Paper
Background Previous population-based studies investigating the relationship between physical activity and the gut microbiota have relied on self-reported activity, prone to reporting bias. Here, we investigated the associations of accelerometer-based sedentary (SED), moderate-intensity (MPA), and vigorous-intensity (VPA) physical activity with the gut microbiota using cross-sectional data from the Swedish CArdioPulmonary bioImage Study.
... and mapped using BWA mem v0.7.16a52 to a gene catalog of > 14 million non-redundant microbial genes (version "HG3A") as earlier described 20 . Metagenomic species (MGS) were defined by the concept of binning of co-abundant genes as previously described 21 . Using this approach altogether 1985 MGS were identified of which 881, present in at least 5% of the participants, were included in the analyses. ...
Article
Full-text available
Medication can affect the gut microbiota composition and function. The aim of this study was to investigate connections between use of common non-antibiotic medicines and the gut microbiota composition and function in a large Swedish cohort (N = 2223). Use of 67 medications and polypharmacy (≥ 5 medications), based on self-reported and prescription registry data, were associated with the relative abundance of 881 gut metagenomic species (> 5% prevalence) and 103 gut metabolic modules (GMMs). Altogether, 97 associations of 26 medications with 40 species and of four medications with five GMMs were observed (false discovery rate < 5%). Several earlier findings were replicated like the positive associations of proton pump inhibitors (PPIs) with numerous oral species, and those of metformin with Escherichia species and with lactate consumption I and arginine degradation II. Several new associations were observed between, among others, use of antidepressants, beta-blockers, nonsteroidal anti-inflammatory drugs and calcium channel blockers, and specific species. Polypharmacy was positively associated with Enterococcus faecalis, Bacteroides uniformis, Rothia mucilaginosa, Escherichia coli and Limosilactobacillus vaginalis, and with 13 GMMs. We confirmed several previous findings and identified numerous new associations between use of medications/polypharmacy and the gut microbiota composition and functional potential. Further studies are needed to confirm the new findings.
Article
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.
Article
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.
Article
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.
Article
The dissemination of antibiotic resistance genes (ARGs) through plasmids is a major mechanism for the development of bacterial antimicrobial resistance. The adaptation and evolution mechanisms of multidrug-resistant (MDR) plasmids with their hosts are not fully understood. Herein, we conducted experimental evolution of a 244 kb MDR plasmid (pJXP9) under various conditions including no antibiotics and mono- or combinational drug treatments of colistin (CS), cefotaxime (CTX), and ciprofloxacin (CIP). Our results showed that long-term with or without positive selections for pJXP9, spanning approximately 600 generations, led to modifications of the plasmid-encoded MDR and conjugative transfer regions. These modifications could mitigate the fitness cost of plasmid carriage and enhance plasmid maintenance. The extent of plasmid modifications and the evolution of plasmid-encoded antibiotic resistance depended on treatment type, particularly the drug class and duration of exposure. Interestingly, prolonged exposure to mono- and combinational drugs of CS and CIP resulted in a substantial loss of the plasmid-encoded MDR region and antibiotic resistance, comparable to the selection condition without antibiotic. By contrast, combinational treatment with CTX contributed to the maintenance of the MDR region over a long period of time. Furthermore, drug selection was able to maintain and even amplify the corresponding plasmid-encoded ARGs, with co-selection of ARGs in the adjacent regions. In addition, parallel mutations in chromosomal arcA were also found to be associated with pJXP9 plasmid carriage among endpoint-evolved clones from diverse treatments. Meanwhile, arcA deletion improved the persistence of pJXP9 plasmid without drugs. Overall, our findings indicated that plasmid-borne MDR region deletion and chromosomal arcA inactivation mutation jointly contributed to co-adaptation and co-evolution between MDR IncHI2 plasmid and Salmonella Typhimurium under different drug selection pressure. IMPORTANCE The plasmid-mediated dissemination of antibiotic resistance genes has become a significant concern for human health, even though the carriage of multidrug-resistant (MDR) plasmids is frequently associated with fitness costs for the bacterial host. However, the mechanisms by which MDR plasmids and bacterial pairs evolve plasmid-mediated antibiotic resistance in the presence of antibiotic selections are not fully understood. Herein, we conducted an experimental evolution of a large multidrug-resistant plasmid in a Salmonella enterica Typhimurium host under single and combinatorial drug selection pressures. Our results show the adaptive evolution of plasmid-encoded antibiotic resistance through alterations of the MDR region in the plasmid, in particular substantial loss of the MDR region, in response to different positive selections, especially mono- and combinational drugs of colistin and ciprofloxacin. In addition, strong parallel mutations in chromosomal arcA were associated with pJXP9 carriage in Salmonella Typhimurium from diverse treatments. Our results thus highlight promoting the loss of the plasmid’s MDR region could offer an alternative approach for combating plasmid-encoded antibiotic resistance.
Article
Full-text available
Background The human gut microbiome develops rapidly during infancy, a key window of development coinciding with the maturation of the adaptive immune system. However, little is known about the microbiome growth dynamics over the first few months of life and whether there are any generalizable patterns across human populations. We performed metagenomic sequencing on stool samples (n = 94) from a cohort of infants (n = 15) at monthly intervals in the first 6 months of life, augmenting our dataset with seven published studies for a total of 4,441 metagenomes from 1,162 infants. Results Strain-level de novo analysis was used to identify 592 of the most abundant organisms in the infant gut microbiome. Previously unrecognized consortia were identified which exhibited highly correlated abundances across samples and were composed of diverse species spanning multiple genera. Analysis of a published cohort of infants with cystic fibrosis identified one such novel consortium of diverse Enterobacterales which was positively correlated with weight gain. While all studies showed an increased community stability during the first year of life, microbial dynamics varied widely in the first few months of life, both by study and by individual. Conclusion By augmenting published metagenomic datasets with data from a newly established cohort, we were able to identify novel groups of organisms that are correlated with measures of robust human development. We hypothesize that the presence of these groups may impact human health in aggregate in ways that individual species may not in isolation.
Article
Full-text available
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.
Article
Full-text available
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.
Article
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.
Article
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.
Article
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.
Article
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.