RESEARCH ARTICLE SUMMARY

MICROBIOTA

A sparse covarying unit thatdescribes healthy and impairedhuman gut microbiota developmentArjun S Raman Jeanette L Gehrig Siddarth Venkatesh Hao-Wei Chang Matthew C HibberdSathish Subramanian Gagandeep Kang Pascal O Bessong Aldo AM LimaMargaret N Kosek William A Petri Jr Dmitry A Rodionov Aleksandr A ArzamasovSemen A Leyn Andrei L Osterman Sayeeda Huq Ishita Mostafa Munirul IslamMustafa Mahfuz Rashidul Haque Tahmeed Ahmed Michael J Barratt Jeffrey I Gordon

INTRODUCTION Ecosystems such as the hu-man gut microbiota are typically described bya ldquoparts listrdquo with enumeration of componentmembers Accordingly the abundances of com-munity components are commonly used as ametric for relating its configuration to featuresof its habitat and to the biological state of thehost Although this approach has providedmuch insight the structure and function ofbiological systems are emergent arising fromthe collective action of constituent parts ratherthan each part acting in isolation This char-acteristic demands a different approach to de-

scribing the form of a microbiotamdashone thattakes into consideration the abundances aswell as the interactions between members

RATIONALE Borrowing from the fields ofeconophysics and protein evolution whereidentification of conserved covariation hasprovided insights about the organization ofcomplex dynamic systems we searched for fea-tures amidst the seemingly intractable complex-ity of human gut microbial communities thatcould serve as a framework for understandinghow they assemble and function

RESULTS A statistical workflow was devel-oped to identify conserved bacterial taxon-taxon covariance in the gut communities ofhealthy members of a Bangladeshi birth co-hort who provided fecal samples monthly frompostnatal months 1 to 60 The results revealedan ldquoecogrouprdquo of 15 bacterial taxa that togetherexhibited consistent covariation by 20 monthsof age and beyond Ecogroup taxa also describedgut microbiota development in healthy mem-bers of birth cohorts residing in BangladeshIndia and Peru to an extent comparable to

what is achieved whenconsidering all detectedbacterial taxa this findingsuggests that the ecogroupnetwork isaconservedgen-eral feature of microbiotaorganization Moreover

the ecogroup provided a framework for char-acterizing the state of perturbed microbiota de-velopment in Bangladeshi children with severeacute malnutrition (SAM) and moderate acutemalnutrition (MAM) as well as a quantitativemetric for defining the efficacy of standard ver-sus microbiota-directed therapeutic foods in re-configuring their gut communities toward astate seen in age-matched healthy children liv-ing in the same locale These results highlightthe importance of the ecogroup as a descriptorboth for fundamental and practical uses A con-sortium of cultured ecogroup taxa introducedinto gnotobiotic piglets reenacted changes intheir relative abundances that were observed inhuman communities as the animals transitionedfrom exclusive milk feeding to a fully weanedstate consuming a prototypic Bangladeshi dietThis pattern of change correlated with the rep-resentation of a sparse set of metabolic path-ways in the genomes of these organisms and inthe fully weaned state with their expression

CONCLUSIONThe ecogroup represents a sim-plified feature of community organization andcomponents that could play key roles in commu-nity assembly and function As the gutmicrobiotaconstantly faces environmental challenges ldquoem-beddingrdquo a sparse network of covarying taxa in alarger framework of independently varying orga-nisms could represent an elegant architecturalsolution developed by nature tomaintain robust-ness while enabling adaptation The approachused to identify and characterize the sparse net-work of covarying ecogroup taxa is in principlegeneralizable to a wide variety of ecosystems

RESEARCH

Raman et al Science 365 140 (2019) 12 July 2019 1 of 1

The list of author affiliations is available in the full article onlineCorresponding author Email jgordonwustleduThis is an open-access article distributed under the terms ofthe Creative Commons Attribution license (httpcreative-commonsorglicensesby40) which permits unrestricteduse distribution and reproduction in any medium providedthe original work is properly citedCite this article as A S Raman et al Science 365 eaau4735(2019) DOI 101126scienceaau4735

Ecogroup as a concisedescription of microbiotaform (Top) Network diagramof covarying taxa where node(taxon) color indicatesecogroup (green) or non-ecogroup (gray) node sizeindicates number of mutuallycovarying taxa and connectionbetween nodes indicatescovariance between two taxa(Bottom) Measuring therepresentation of ecogrouptaxa reveals that children withSAM treated with standardtherapeutic foods have anecogroup profile similar to thatof children with untreatedMAM indicating persistentperturbations in their gutcommunity relative to healthychildren In contrast childrenwith MAM treated with atherapeutic food designed totarget the microbiota (MDCF-2)have an ecogroup profile thatoverlaps nearly entirely withthat of healthy children

-003

-002

002

-001

0

0

0-001

001

-002

002

-002-003

-004-004 PC1 (31)PC2 (25)

PC3

(15

)

Healthy

MAMMDCF2

MAMMDCF3 MAM

RUSFMAMMDCF1

MAMuntreated

SAM untreated

SAM at discharge

SAM 1 mo post-discharge

SAM 12 mo post-dischargeSAM 6 mo post-discharge

Ecogroup taxaIndependentlyvarying taxa

E hallii

Blautia

Rtorques

Cmitsuokai

Ruminococcaceae

Bbifidum

Enterobacteriaceae

Blongum

Bifidobacterium

PrevotellaPcopri

(840914)

Prevotella

Dialister

S thermophilusF prausnitzii (514940)

F prausnitzii(851865)

Ecoli

Erectale

Lruminis

Efaecalis

Pcopri(588929)

Sgallolyticus

ClostridialesSventriculi

Prevotellaceae

Lgarvieae

ON OUR WEBSITE

Read the full articleat httpdxdoiorg101126scienceaau4735

on February 4 2021

httpsciencesciencem

agorgD

ownloaded from

RESEARCH ARTICLE

MICROBIOTA

A sparse covarying unit thatdescribes healthy and impairedhuman gut microbiota developmentArjun S Raman12 Jeanette L Gehrig12 Siddarth Venkatesh12 Hao-Wei Chang12Matthew C Hibberd12 Sathish Subramanian12 Gagandeep Kang3 Pascal O Bessong4Aldo AM Lima5 Margaret N Kosek67dagger William A Petri Jr8 Dmitry A Rodionov910Aleksandr A Arzamasov910 Semen A Leyn910 Andrei L Osterman10 Sayeeda Huq11Ishita Mostafa11 Munirul Islam11 Mustafa Mahfuz11 Rashidul Haque11Tahmeed Ahmed11 Michael J Barratt12 Jeffrey I Gordon12Dagger

Characterizing the organization of the human gut microbiota is a formidable challengegiven the number of possible interactions between its components Using a statisticalapproach initially applied to financial markets we measured temporally conservedcovariance among bacterial taxa in the microbiota of healthy members of a Bangladeshibirth cohort sampled from 1 to 60 months of age The results revealed an ldquoecogrouprdquoof 15 covarying bacterial taxa that provide a concise description of microbiotadevelopment in healthy children from this and other low-income countries and a means formonitoring community repair in undernourished children treated with therapeutic foodsFeatures of ecogroup population dynamics were recapitulated in gnotobiotic piglets asthey transitioned from exclusive milk feeding to a fully weaned state consuming arepresentative Bangladeshi diet

Innumerable studies of the functioning ofbiological systems have underscored theimportance of characterizing interactionsbetween their component parts (1ndash5) De-fining microbial communities in this way

can present a seemingly intractable challenge(1ndash3 6) For example the gastrointestinal tractof a healthy adult human harbors multiplespecies with multiple strain-level variants of a

given species that can engage in higher-orderinteractions with other community membersUsing a conservative species count of 100 thenumber of terms needed to mathematically rep-resent all possible species-species interactions(pairwise and higher-order) is ~1030 A centralquestion is how biologically important inter-actions between component members can beidentified so as to reduce the number of fea-tures necessary for characterization of microbialcommunity properties such as assembly duringthe postnatal period or temporal responses tovarious perturbationsCo-occurrence analysis has been used to de-

scribe community organization but is limitedin its ability to describe interactions betweenmicrobes (7 8) Recently developed approacheshave focused on defining microbe-microbe inter-actions using cross-sectional data (9 10) althoughthese methods were not explicitly designed toaddress the temporal conservation of these in-teractions in for example longitudinal studiesTherefore we turned to approaches developedin the fields of econophysics and protein evo-lution Applying the concept of statistical co-variance coupled with analytical techniquesof matrix decomposition has identified co-fluctuating economic sectors and cooperativeamino acid networks of functional relevance(11ndash13) The underlying presumption is that co-variation that is conserved is covariation thatmay be informative about the organization ofcomplex dynamic systems

In this spirit we have developed a computa-tionalworkflow to calculate temporally conservedcovariance of gut bacterial taxa over time inmembers of a healthy Bangladeshi birth cohortsampledmonthly for the first five postnatal yearsThe results revealed a network of 15 covaryingbacteria that we term an ldquoecogrouprdquo Ecogrouptaxa not only describe healthy gut microbial de-velopment in children residing in Bangladeshas well as several other low- and middle-incomecountries they also distinguish the microbiotaof Bangladeshi children with untreated moder-ate and severe acutemalnutrition and the degreeto which these communities are reconfiguredtoward a healthy state in response to severaltherapeutic food interventions Colonizing germ-free piglets with a consortium of ecogroup taxaand following them during the transition fromexclusive milk feeding through weaning onto arepresentative diet consumed by Bangladeshichildren recapitulates features of healthy com-munity development and reveals microbial ge-nomic features and expressed metabolic attributesimportant for fitness during succession

Identifying the ecogroup

Thirty-six members of a birth cohort with con-sistently healthy anthropometric scores livingwithin the Mirpur district of Dhaka Bangladeshunderwent monthly fecal sampling for the first60 months of postnatal life [height-for-age Zscore (HAZ) ndash092 plusmn 119 (mean plusmn SD) weight-for-height Z score (WHZ) ndash048 plusmn 133 n =1961 fecal samples 55 plusmn 4 samples collectedper individual table S1] In Bangladesh themedian duration of breastfeeding is 4 monthswhereas the weaning process is long with amedian of 25 months (14) Samples collectedless frequently or only after 36 months from19 other children from Mirpur were also in-cluded in our analysis (HAZ ndash058 plusmn 112 WHZndash025 plusmn 096 n = 257 plusmn 105 samples per child)Amplicons generated from variable region 4 (V4)of bacterial 16S rRNA genes present in these2455 fecal samples were sequenced and theresulting reads were assigned to operationaltaxonomic units with ge97 nucleotide sequenceidentity (97ID OTUs) (15 16) (fig S1) In total118 97ID OTUs were represented at a relativefractional abundance of at least 0001 (01)in at least two of the samples collected overthe 60-month periodAn initial broad description of microbiota

development in this cohort was obtained byapplying unweighted and weighted UniFracto compute overall phylogenetic dissimilaritybetween gut communities from the 36 childrensampled monthly from 1 to 60 months and 49fecal samples collected in a previous study from12 unrelated adults aged 23 to 41 years livingin Mirpur (17) This metric indicated that themean ldquoinfantchild-to-adultrdquo distance decreasesto ldquoadult-to-adultrdquo by 3 years of age (fig S2 Aand B) Alpha diversity also increased to adult-like levels during this time period (fig S2 C andD) As an additional description of communitydevelopment we used the 16S rDNA dataset to

RESEARCH

Raman et al Science 365 eaau4735 (2019) 12 July 2019 1 of 11

1Edison Family Center for Genome Sciences and SystemsBiology Washington University School of Medicine St LouisMO 63110 USA 2Center for Gut Microbiome and NutritionResearch Washington University School of MedicineSt Louis MO 63110 USA 3Translational Health Science andTechnology Institute Faridabad Haryana India 4HIVAIDSand Global Health Research Programme Department ofMicrobiology University of Venda Thohoyandou 0950 SouthAfrica 5Center for Global Health Department of Physiologyand Pharmacology Clinical Research Unit and Institute ofBiomedicine School of Medicine Federal University of CearaacuteFortaleza CE 60430270 Brazil 6Department of InternationalHealth Bloomberg School of Public Health Johns HopkinsUniversity Baltimore MD 21205 USA 7AB PRISMA RamirezHurtado 622 Iquitos Peru 8Departments of MedicineMicrobiology and Pathology University of Virginia School ofMedicine Charlottesville VA 22908 USA 9A A KharkevichInstitute for Information Transmission Problems RussianAcademy of Sciences Moscow 127994 Russia 10Infectiousand Inflammatory Disease Center Sanford Burnham PrebysMedical Discovery Institute La Jolla CA 92037 USA11International Centre for Diarrhoeal Disease ResearchBangladesh Dhaka 1212 BangladeshPresent address Department of Medicine Massachusetts GeneralHospital Boston MA 02114 USA daggerPresent address Departmentof Medicine University of Virginia School of Medicine CharlottesvilleVA 22908 USADaggerCorresponding author Email jgordonwustledu

on February 4 2021

httpsciencesciencem

agorgD

ownloaded from

construct a sparse Random Forests (RF)ndashderivedmodel comprising age-discriminatory taxa (fig S3A to E) Microbiota ldquoagerdquo can be computed bynoting the fractional abundances of these age-discriminatory taxa in a given sample obtainedat a given time point (14) Applying the RF-generated model disclosed a high degree ofcorrelation between microbiota age and chron-ologic age (R2 = 08) (fig S3C) Although theseapproaches provide measures of communitydevelopment they do not characterize inter-actions between community members duringthis processPrincipal components analysis (PCA) applied

to taxa present in monthly fecal samples offersa way to mathematically characterize gut micro-biota organization by defining principal com-ponents (eigenvectors) The result of PCA is aranked list of principal components (principalcomponent spectrum or ldquoeigenspectrumrdquo) whereeach principal component carries a percent-age of data variance Tracking the principalcomponent spectrum through time offers adescription of the evolving temporal organiza-tion of the gut microbiota The approach weused iterative PCA (iPCA) is described in fig S4AFor each month we created a matrix where therows were fecal samples and the columns com-prised the 118 taxa described above In the ex-ample shown time point 1 considers monthlyfractional abundance data from month 1 and areference time point The dissimilarity betweenthe two time points is reflected in the primaryprincipal component (PC1) The system is con-sidered to be ldquostablerdquo at the time point whereadding subsequent monthsrsquo data negligibly con-tributes to variance mathematically this is whenthe eigenvalue of PC1 reaches an asymptoteWe performed iPCA on sequentially joined

monthly data with month 36 taken as a refer-ence (fig S4B) Month 36was chosen on the basisof the results of phylogenetic dissimilarity anddiversitymeasurements presented in fig S2 [notethat previous cross-sectional studies using thesemetrics had also indicated that an adult-like con-figuration was achieved by this time point eg(18)] iPCA revealed that month 20 and beyondsignify a time period of minimal structural var-iation in the gut microbiota (fig S4B) This con-clusion was supported by using the very lasttime point in the 5-year longitudinal study asthe reference (fig S4C) Therefore we were ableto design a workflow to compute reproduciblecovariance (covariance conserved across time ina mature community assemblage as opposed totransient covariance that may occur during com-munity assembly) usingmonths 20 to 60withouthaving to make any a priori assumptions aboutthe importance of any taxon For each monthspanning postnatal months 20 to 60 we calcu-lated the covariance between the 118 taxa over allindividuals to generate monthly taxon-taxon co-variance matrices (19) (see Fig 1A fig S5 andtable S2A) The matrices were averaged to asingle taxon-taxon matrix (hCi j

binit ) that repre-sented a definition of consistent covariancewherei and j are bacterial taxa and t designates the

month (Fig 1B and table S2B) PCA performedon this matrix revealed that PC1 encompassed80 of the data variance (Fig 1C see supple-mentary text for a sensitivity analysis of theworkflow) A group of 15 covarying taxa repre-sented the top 20 of all taxa projections alongPC1 (Fig 1C see table S3 for different thresh-old cutoffs) They include OTUs assigned toBifidobacterium longum another member ofBifidobacterium Faecalibacterium prausnitziia member of Clostridiales Prevotella copriStreptococcus thermophilus and Lactobacillusruminis all of which are age-discriminatorybacterial strains identified from RF-based anal-ysis of bacterial 16S rDNA datasets generatedfrom healthy members of this Bangladeshi co-hort (fig S3D)The results of PCA performed on data gen-

erated from 478 samples collected from childrensampled at postnatal months 50 to 60 providean illustration of statistical covariation betweenthese taxa PC1 reveals that B longum (OTU559527) and L ruminis (OTU 1107027) positivelycovary with one another across samples and neg-atively covary with two P copri strains (OTUs840914 and 588929) PC2 documents how twoF prausnitzii OTUs (514940 and 851865) posi-tively covary with each other and negativelycovary with S gallolyticus (OTU 349024) andE coli (OTU 1111294) PC3 discloses that thetwo P copri OTUs negatively covary with thetwo F prausnitzii OTUs (Fig 1D)Figure 1E provides a graphical depiction of

this network of covarying taxa Each green noderepresents one of the 15 OTUs that manifest ahigh degree of conserved covariance betweenmonths 20 and 60 Two nodes are connected byan edge if their temporally averaged covariancevalue (hCi j

binit from Fig 1B) is within the top 20of all such values Node size is proportional tothe number of connections (edges) present Thegreen nodes collectively covarywith one anotherIn contrast gray nodes depict taxa that covarywith green nodes but not with one another(Fig 1E) The green nodes constitute an ldquoinsu-latedrdquo ecostructure its members exhibit signif-icant intragroup covariation (fig S6 and tableS2C) We chose the term ldquoecogrouprdquo to reflectthe conserved collective statistical covariationof this sparse network of 15 organisms

Microbiota development in otherbirth cohorts

We asked whether components of the ecogroupprovide a concise description of postnatal devel-opment of the microbiota in healthy membersof the Bangladeshi cohort and if so whetherchanges in the representation of these taxa fol-low a pattern that is shared across other healthybirth cohorts representing distinct geographiclocales and anthropologic features (20) Moreoverwe postulated that if ecogroup taxa are informa-tive biomarkers of normal community develop-ment these taxamight be useful for characterizingimpaired development andor the extent to whichcommunity repair is achieved as a function ofvarious therapeutic interventions (21)

Three different matrices were created whereeach row was a fecal sample collected from anindividual at a particular month in the healthyBangladeshi cohort and columns were either (i)all 118 taxa (ii) the 15 ecogroup taxa or (iii) theremaining 103 non-ecogroup taxa PCA wasperformed on the rows of these matrices fecalsamples were plotted on the first three principalcomponents The left panel of Fig 2A shows theresults obtained when considering the fractionalrepresentation of all 118 taxa in fecal samplescollected at postnatal months 4 10 and 20There is substantial interpersonal variation ingut community structure at postnatal month 1as evidenced by the broad distribution alongPC1 but this variation converges by month 4(Fig 2A fig S7 A and B and table S2D) There-after changes in the structure of the fecal micro-biota are depicted by right-to-left movementalong PC1 with minor variance observed alongPC2 and PC3 Minimal movement along PC1 isobserved after month 20 (fig S7C) consistentwith the results of iPCA in fig S4 B and C (No-tably children in this cohort had completedweaning bymonth 23 see fig S8 for a descriptionof the nature and timing of their dietary tran-sitions) Ecogroup taxa recapitulate the variancedepicted by PC1 PC2 and PC3 Moreover theecogroup taxa capture (i) the significant inter-personal variation observed at postnatal month 1(ii) the subsequent convergence to a B longumndashpredominant microbiota at postnatal month 4and (iii) temporal changes noted at postnatalmonths 10 and 20 (Fig 2 A andB fig S7 A andBmiddle panels and table S2E) In contrast theremaining 103 non-ecogroup taxa provide aless informative representation of developmen-tal changes in the microbiota as exemplified bythe fact that PC1 PC2 and PC3 each capturele10 of the variance (Fig 2A and fig S7 A andB right panels) The importance of taxa withlow average fractional abundances and largestandard deviations such as P copri (Fig 2Binset) is often overlooked when they are con-sidered in isolation However analysis of taxon-taxon covariation can reveal relationships betweenmember species as illustrated by P copri andB longum (Fig 1D blue box)To determine the extent to which the eco-

group is a generalizable descriptor of the micro-biota in infants and children with healthy growthphenotypes we turned to the MAL-ED networkof study sites located in low- and middle-incomecountries (20 21) Fecal samples had been col-lected monthly for the first two postnatal yearsallowing sparse 30-taxon RF-generated models ofnormal community development to be generatedfrommembers of birth cohorts residing in LoretoPeru (periurban area) and Vellore India (urbanarea) (supplementary text fig S9 and table S4)Our ability to identify a network of covaryingtaxa in the Mirpur cohort depended on a high-resolution time-series study that extended wellbeyond the month at which the microbiota wasdetermined to be ldquostablerdquo (month 20) This du-ration of sampling did not occur at these otherMAL-ED sites obviating our ability to identify

Raman et al Science 365 eaau4735 (2019) 12 July 2019 2 of 11

RESEARCH | RESEARCH ARTICLEon F

ebruary 4 2021

httpsciencesciencemagorg

Dow

nloaded from

Raman et al Science 365 eaau4735 (2019) 12 July 2019 3 of 11

Fig 1 Defining a sparse consistently covarying network of bacterialtaxa (ldquoecogrouprdquo) in healthy Bangladeshi children (A) WorkflowLeft 16S rDNA sequencing of fecal microbiota samples collected monthlyfrom healthy members of the birth cohort from postnatal months20 to 60 For each month a matrix is created where rows are taxa andcolumns are fecal samples of individuals Center Taxon-taxon covariancematrices for each month are calculated Right Monthly taxon-taxoncovariance matrices are normalized relative to the maximum monthlycovariance value If a normalized monthly covariance value for a given (i j)taxon-taxon pair is within the top or bottom 10 of all monthly covariancevalues it is converted to a ldquo1rdquo otherwise it is assigned a ldquo0rdquo This binarizedcovariance matrix is defined as Cij

bin Concatenating Cijbin for all months

creates a three-dimensional matrix ethCi jbinTHORNt (B) Temporally conserved taxon-

taxon covariance matrix The binarized covariance values for each(i j) pair of taxa in ethCij

binTHORNt are averaged over all months to give a temporallyweighted covariance value for each taxon-taxon pair (hCi j

binit) In the limitthat two taxa always covary with each other hCij

binit = 1 If two taxa never

covary with each other hCi jbinit = 0 The matrix shown illustrates sparse

temporally conserved coupling with many taxa showing no consistentcovariance (hCij

binit asymp 0 white pixels) but a few exhibiting a high degreeof conserved covariance (hCi j

binit ge 05 deep red pixels) (C) Eigende-composition of temporally conserved covariance matrix Note that 80 of

the data variance in hCijbinit can be represented by a single principal

component The histogram shows projections of taxa along PC1 data arefit to a generalized extreme value distribution (red line) Applying a 20threshold to this distribution identifies 15 taxa that reproducibly covaryover time (D) Fecal samples from postnatal months 50 to 60 shown on aPCA space ordinated by the 15 taxa in (C) Heat maps illustrate thefractional abundance of taxa responsible for the variance along each

principal component The blue box shown in the left portion of theprojection along PC1 highlights the subset of healthy children who have ahigh representation of P copri relative to B longum (E) Graphicalrepresentation of the sparse covarying network of 15 taxa (greennodes) See text for details

RESEARCH | RESEARCH ARTICLEon F

ebruary 4 2021

httpsciencesciencemagorg

Dow

nloaded from

conserved covariance among taxa However totest how well the 15 ecogroup taxa identified inthe Mirpur cohort could characterize the devel-oping microbiota of children living in thesecountries we created two matrices where eachrow was a fecal microbiota sample from theIndian or Peruvian cohorts and columns wereeither all taxa identified in the Peruvian andIndian samples or just the 15 ecogroup taxa

identified from the Bangladeshi birth cohort PCAwas performed on the rows of these matrices andthe same analysis performed as described forthe healthy Bangladeshi birth cohort The re-sults show that the ecogroup taxa identified inmembers of the healthy Bangladeshi cohortalso provide a concise description of commu-nity development in healthy members of theseother two birth cohorts that is (i) they capture

the variance depicted by PC1 PC2 and PC3 ascompared to considering all taxa and (ii) changesin their fractional abundances followed tempo-ral patterns similar to those documented in theBangladeshi cohort (fig S10 and table S2F)

Ecogroup configuration in acutemalnutrition before and after treatment

Bangladeshi children with acute malnutritionhave perturbed microbiota development theirgut communities appear younger than those ofchronologically age-matched individuals (14 21)We examined whether ecogroup taxa providea useful way to characterize the microbiota ofchildren with moderate or severe acute mal-nutrition (MAM and SAM respectively) priorto and after food-based therapeutic interventionsIn the accompanying paper Gehrig et al describe63 children from Mirpur diagnosed with MAMaged 12 to 18 months who were enrolled in adouble-blind randomized controlled feedingtrial of different microbiota-directed comple-mentary foods (MDCFs) (21) Fecal sampleswere collected for 9 weeks at weekly intervalsThe first 2 weeks comprised a pretreatment ob-servation period Over the next 4 weeks chil-dren received either one of three MDCFs or aready-to-use supplementary food (RUSF) rep-resenting a form of conventional therapy thatunlike the MDCFs was not designed to targetspecific members of the gut microbiota and re-pair community immaturity The last 2 weeksrepresented the post-treatment observation pe-riod In total we identified 945 97ID OTUsthat had a fractional abundance of at least 0001(01) in at least two fecal samples collected fromone or more participants prior to during andafter treatment (n = 531 samples) Gehrig et al(21) also describe another trial involving 54 hos-pitalized Bangladeshi children with SAM aged6 to 36 months where each participant wastreated with one of three standard therapeuticfoods and then followed over a 12-month periodafter discharge In total we identified 944 97ID OTUs that had a fractional abundance of atleast 0001 in at least two fecal samples collectedfrom one or more participants in this trial (n =618 samples)Amatrix was created that included (i) all fecal

samples from the SAM trial (ii) pretreatmentsamples from childrenwithMAMenrolled in allfour arms of the MDCF trial (iii) MAM samplesobtained 2 weeks after treatment with one ofthe three MDCFs or the RUSF and (iv) fecalsamples from age-matched healthy Bangladeshichildren (table S5) Each row of thematrix was afecal sample each columnwas an ecogroup taxonand each element in the matrix was the frac-tional abundance of an ecogroup taxon within aparticular fecal sample PCA was performedon the rows of this matrix Centroids for eachcohort were computed and plotted on the PCAspace (Fig 3A) At the time of discharge afterreceiving standard therapeutic foods the mi-crobiota of children with SAM remained in anincompletely repaired state Although there wassome improvement at 1 month after discharge

Raman et al Science 365 eaau4735 (2019) 12 July 2019 4 of 11

A

0 08

Average fractionalabundance

B

1

Ave

rage

frac

tiona

l abu

ndan

ce

23

45

1020

4060

Month

Month 10

Month 20

All taxa (118) Ecogrouptaxa (15)

Non-ecogrouptaxa (103)

PC

2 (9)

PC

3(8

)

PC

2 (9)

PC

3(8

)

-01

008

004

0080006004002-004 0

01

PC1 (50)

PC

2 (12)

PC

3(9

) 01

00 -002

-004-006

005 -008

01

-005

PC1 (55)

PC2 (11

)

PC1 (10)

PC

3(1

0)

-04

02

01015

0100050-01

-005

0

-01

008

004

0080006004002-004 0

01

PC1 (50)

PC

2 (12)

PC

3(9

)

01

00 -002

-004-006

005 -008

01

-005

PC1 (55)

PC2 (11

)

PC1 (10)

PC

3(1

0)

-04

02

01015

010005

0-01-005

0

-01008

004

-004

00800060040020

01

PC1 (50)

PC

2 (12)

PC

3(9

)

01

00 -002

-004-006005 -008

01

-005

PC1 (55)

PC2 (11

)

PC

3(1

0)

-04

02

015010

0050-01-005

0

Month 4

PC

2 (9)

PC

3(8

)

PC1 (10)

01

008

004

Month

01 2 3 4 5 10 20 40 60

Ave

rage

frac

tiona

lab

unda

nce

P c

opri

0

05

Blongum

1

Bifidobacte

rium

Sgallolyt

icus

Lruminis

Ecoli

Fprausnitz

ii (514940)

Clostridiales

Pcopri (

588929)

Fprausnitz

ii (851865)

Erecta

le

Pcopri (

840914)

Prevotella

Stherm

ophilus

Efaeca

lis

Dialister

Fig 2 Characterizing healthy gut microbiota development in the Bangladeshi birth cohort(A) PCA spaces were created Each point in the spaces represents a fecal sample described byeither all taxa present at a fractional abundance greater than 0001 (01) (118 taxa) ecogroup taxa(15) or non-ecogroup taxa (103) The spatial distribution of fecal samples in each PCA space isshown for the indicated postnatal months (B) Bar graph illustrating average fractional abundanceof ecogroup taxa as a function of postnatal month (see table S2E) Inset Average fractionalabundance (plusmnSD) of P copri as a function of time

RESEARCH | RESEARCH ARTICLEon F

ebruary 4 2021

httpsciencesciencemagorg

Dow

nloaded from

there was minimal additional improvement evi-dent at 6 or 12 months at which times theirmicrobiota resembled that of untreated chil-dren with MAM (Fig 3A) The microbiota ofchildren with MAM that were treated withMDCF-1 MDCF-3 and RUSF clustered together

whereas the microbiota of those treated withMDCF-2 closely resembled that of healthy chil-dren Notably MDCF-2 was also distinct amongthe four treatment types in eliciting changes inthe plasma proteome indicative of improvedhealth status including changes in biomarkers

and mediators of metabolism bone growth cen-tral nervous system development and immunefunction [see (21) for details]PCA measures the effect of treatment on the

gut microbiota by considering a constellationof changes in fractional abundance of ecogroup

Raman et al Science 365 eaau4735 (2019) 12 July 2019 5 of 11

Fig 3 Ecogroup taxa define the response of the microbiota of children with SAM and MAM to various nutritional interventions (A) Centroidsof each indicated cohort are plotted on a PCA space Arrows indicate the temporal progression of microbiota reconfiguration for children with SAMtreated with conventional therapy and children with MAM treated with a RUSF or a MDCF (B) Matrix decomposition of the axes shown in (A) highlightsthe taxa that are important for fecal sample variance observed along each principal component (C and D) Average fractional abundance of ecogrouptaxa identified in (B) in the fecal microbiota of members of the SAM and MAM cohorts as a function of treatment (see table S2G)

RESEARCH | RESEARCH ARTICLEon F

ebruary 4 2021

httpsciencesciencemagorg

Dow

nloaded from

taxa with the premise that the fractional abun-dances as well as the covariation of these taxaare important for characterizing community con-figuration The left panel of Fig 3B shows thatthe relationship between the fractional represen-tations of B longum (OTU 559527) and E coli(OTU 1111294) determines microbiota positionalong PC1 in Fig 3A The center and right panelsof Fig 3B show that the relationship between thefractional representations of B longum E coliS gallolyticus (OTU 349024) and P copri (OTUs588929 and 840914) determines position alongPC2 whereas position along PC3 reflectsthe relationship between the abundances ofS gallolyticus and the two P copri OTUs B longumS gallolyticus and E coli are the predominantecogroup taxa represented in the microbiota ofchildren with untreated SAM (Fig 3C and tableS2G) Treatment results in movement of theirmicrobiota along PC1 and PC3 in Fig 3Athis movement is associated with a decrease inB longum S gallolyticus and E coli (Fig 3C andtable S2G) Differences between the microbiotaof healthy children and those with SAM priorto and during the 12 months after treatmentwith standard therapeutic foods are manifest bydifferences in their respective positions alongPC1 and PC3 (Fig 3A) These differences sig-nify incomplete repair to a ldquohealthyrdquo state andhighlight the need to achieve further decreasesin the fractional abundance of B longum (asso-ciated with movement to the right of PC1) alongwith further decreases in the fractional abun-dance of S gallolyticus and increases in P copri(associated with positive movement alongPC3) The representation of B longum P copriS gallolyticus and E coli in the microbiota of12- to 18-month-old children with untreated MAMaccounts for their positive projection along PC1and PC3 relative to the microbiota of childrenwith untreated SAM (Fig 3A) Among the testedtherapeutic foods MDCF-2 was uniquely asso-ciated with a positive movement along PC1 (Fig3A) this corresponds to decreased fractionalabundance of B longum (Fig 3D and table S2G)and more complete community repairTwo other methods SparCC and SPIEC-EASI

have been used to describe microbiota organi-zation (9 10) As these methods were designedfor cross-sectional studies we adapted them(see supplementary text) so we could comparetheir ability to identify (i) temporally conservedaspects of community organization and (ii) thedegree to which SAM and MAM microbiota arerepaired with different food-based interventionswith the approach we had used to identify theecogroup SparCC identifies a subset of eco-group taxa that describe healthy gut micro-biota development in members of the 5-yearhealthy Bangladeshi cohort study (fig S11 Aand B) SparCC clearly separates the microbiotaof children with untreated SAM from healthycontrols and shows that treatment with standardtherapeutic foods fails to repair their microbiotato a healthy state or even to a state seen inchildren with untreated MAM Compared to theapproach described in Fig 1A SparCC does not

as clearly separate MAM from healthy or (byextension) the differential effects of MDCFtreatment although it does place MDCF-2ndashtreated microbiota closest to that of healthychildren (fig S11C) One explanation is thatP copri does not contribute as prominently to thecollective group of correlated taxa identified bySparCC (fig S11 and table S6 A and B) SPIEC-EASI identifies P copri and other PrevotellaOTUs as key microbes (fig S12 A and B and tableS6 C to E) However SPIEC-EASI does not pro-vide as informative a description of the temporalpattern of healthy gut microbial developmentas does the ecogroup taxa [note the relative lackof movement over time of community configu-ration from right to left along PC1 in fig S12Ccompared to Fig 2A (ecogroup taxa) and figS11B (SparCC)] The 15 interacting taxa iden-tified by SPIEC-EASI separate untreated andtreated SAM and MAM microbiota from oneanother and from healthy (fig S12D) As withthe two other approaches although less clearlythan with the ecogroup taxa SPIEC-EASI showsthat MDCF-2 is most effective in changing theconfiguration of the MAM-associated micro-biota toward a healthy state relative to MDCF-1MDCF-3 and RUSF Together these findings pro-vide support for considering temporally conservedtaxon-taxon covariance when characterizing themicrobiota of children with undernutrition priorto and after various therapeutic interventions

Ecogroup taxa in a gnotobioticpiglet model of postnatal Bangladeshidietary transitions

Our observations raise questions about thenature of the interactions among B longumP copri and other ecogroup taxa during post-natal development as a function of the dietarytransitions that occur when children progressfrom exclusive milk feeding to complementaryfeeding to a fully weaned state To address thisissue we colonized germ-free piglets withecogroup taxa and tracked the dynamics ofconsortium members over time We turned tognotobiotic piglets rather than mice becausethe former have physiologic and metabolic qual-ities more similar to that of humans (22) Pigletswere derived as germ-free at birth and were fedan irradiated sowrsquos-milk replacement (Soweena)for the first four postnatal days (fig S13A) Piglets(n = 5) were then colonized by oral gavagewith a consortium of seven cultured sequencedB longum strains recovered from the fecal mi-crobiota of children living in Mirpur Bangladeshas well as three other countries (Peru Malawiand the United States) (fig S13A) On the basisof their genome sequences (table S7) six strainswere classified as B longum subspecies infantisand one as B longum subspecies longum The ga-vage mixture also contained two Bifidobacteriumbreve strains which we used as comparators todelineate factors that contribute to the fitnessof the B longum strains given the phylogeneticsimilarity of their genomes Beginning on post-natal day 4 a diet representative of that con-sumed by 18-month-old children living in Mirpur

[Mirpur-18 (21)] was added to food bowls con-taining Soweena On postnatal day 7 pigletswere gavaged with a second consortium con-sisting of 16 additional cultured sequenced eco-group taxa (fig S13A) representing 13 of the 15species shown in Fig 1C During postnatal days5 to 22 the amount of Mirpur-18 added to foodbowls was progressively increased while theamount of Soweena was decreased once a fullyweaned state was achieved on day 22 animalswere monotonously fed the Mirpur-18 diet un-til they were euthanized on postnatal day 29Piglets increased their weight by 185 plusmn 31(mean plusmn SD) between postnatal days 7 and 29To define features in ecogroup strains that

relate to their fitness during the series of dietarytransitions that mimic those experienced bychildren living in Mirpur we performed short-read shotgun sequencing of community DNAprepared from rectal swabs obtained at 11 timepoints spanning experimental days 5 to 29 (figS13A) and along the length of the gut at thetime of euthanasia The results are presented inFig 4A and table S2H After gavage of remain-ing ecogroup members the representation ofall B longum strains diminished rapidly Frompostnatal day 8 to day 22 as the animals werebeing weaned S gallolyticus E coli E aviumL salivarius and P copri exhibited distinctpatterns of temporal change in their represen-tation After the animals were fully weaned therewas a pronounced increase in P copri which be-came the dominant member of the cecal colonicand fecal microbiota (Fig 4A and fig S13B) Therelationship between the abundances of P copriand B longum is comparable in these piglets tothat observed in the healthy Bangladeshi chil-dren who were used to evaluate the microbiotaconfigurations of untreated and treated childrenwith MAM and SAM (Fig 3 C and D)The representations of 81 mcSEED metabolic

modules (see methods) in strain genomes wereused to make in silico predictions about theircapacity to synthesize amino acids and B vita-mins utilize a variety of carbohydrates andgenerate short-chain fatty acids Predicted pheno-types were scored as either a ldquo1rdquo or a ldquo0rdquo sig-nifying auxotrophy or prototrophy in the case ofamino acid and B-vitamin biosynthesis or theability or inability to utilize various carbohydrates(table S8) PCA of a ldquobinary phenotype matrixrdquo ofall strains present at a fractional representationof ge0001 in fecal samples collected from post-natal day 8 to day 18 identified 14 carbohydrateutilization pathways plus the capacity to synthe-size cysteine folate and pantothenate as genomicfeatures that distinguish these strains from eachother (table S9) Hierarchical clustering by thesepredicted metabolic phenotypes also groupedthese strains by their fitness (Fig 4 B and C)We performed microbial RNA-seq using cecal

contents to characterize the expression of genesencoding components of mcSEEDmetabolic mod-ules presentwithin the ecogroup strains [The frac-tional representations of these strains in the cecumand feces at the time of euthanasia were highlycorrelated (r2 = 098 table S10)] Figure S14A

Raman et al Science 365 eaau4735 (2019) 12 July 2019 6 of 11

RESEARCH | RESEARCH ARTICLEon F

ebruary 4 2021

httpsciencesciencemagorg

Dow

nloaded from

illustrates the workflow used to generate amcSEED ldquoenrichment matrixrdquo (ME) that signifiesthe extent towhich the aggregate transcript levelsof components of a given mcSEED metabolicmodule in a given bacterial strain quantitativelydiffer from that of a reference strain BecauseP copri had the highest fractional representa-tion on postnatal day 29 it was used as thereference (fig S14B and table S2I) PCA wasperformed on the mcSEED enrichment matrix(Fig 5A and table S11A) The results revealed thatthe transcriptomes of Bifidobacterium strainscluster together and are distinct from those ofP copri E coli B luti and E avium Moreoverthe distribution of strains along PC1 based on

their mcSEED enrichment profiles correlatedwith their fractional representation (fitness) inthe cecal and fecal microbiota (Fig 5A inset)To identify which expressed components of

mcSEED metabolic modules contribute to thedifferences in the fractional representation werequired a way to relate the principal compo-nents of the rows (metabolic modules) and col-umns (strains) of the mcSEED enrichment matrixTo do so we used singular value decomposition(SVD fig S14 C and D) Relative to P copri themost distinguishing features of the Bifidobacteriumtranscriptomes were markedly reduced or absentexpression of pathways involved in (i) biosynthesisof cysteine tyrosine tryptophan and asparagine

(ii) utilization of several carbohydrates (xyloseand b-xylosides plus galacturonateglucuronateglucuronide) (iii) biosynthesis of queuosine and(iv) uptake of cobalt related to cobalamin bio-synthesis (Fig 5B and tables S2J and S11B)Moreover expression of four of these pathways(cysteine and asparagine biosynthesis xyloseb-xyloside and galacturonateglucuronateglucuronide utilization) exclusively differentiateP copri B luti E coli and E avium from allnine Bifidobacterium species and the other fivestrains whose transcripts were represented inthe community metatranscriptome (Fig 5B)The biological significance of expression of

these distinguishingmcSEEDmetabolic modules

Raman et al Science 365 eaau4735 (2019) 12 July 2019 7 of 11

Fig 4 Distinguishing genomic features related to the fitnesslandscape of ecogroup strains in gnotobiotic piglets (A) Averagefractional abundances of strains plotted over time (see table S10)The summary of the experimental design shows when the various taxawere first introduced by gavage and how the diet changed over time Seefig S13A for complete strain designations (B) Genome features thatdistinguish among strains whose average fractional abundances in thefecal microbiota of piglets was ge0001 between postnatal days 8 and 22These distinguishing features are mcSEED metabolic phenotypes color-

coded according to whether they are predicted to endow the hoststrain with prototrophy for amino acids and B vitamins or the capacityto utilize the indicated carbohydrate Strains are hierarchicallyclustered according to the representation of these metabolic pathways(C) Heat map depicting the fractional representation of the strains shownin (B) at the indicated time points Strains are hierarchically clusteredaccording to the mcSEED metabolic phenotypes in (B) Note that thepattern of clustering defined by phenotypes also clusters strains bytheir fitness

RESEARCH | RESEARCH ARTICLEon F

ebruary 4 2021

httpsciencesciencemagorg

Dow

nloaded from

demanded a further contextualization basedon whether these systems were complete orincompletely represented in the strain genomesFigure 5C shows that all of the Bifidobacteriumstrains contain complete metabolic pathwaysfor tyrosine asparagine and tryptophan biosyn-thesis but do not contain complete metabolicpathways for cysteine biosynthesis utilizationpathways for galactose xylose and glucuronidesand B-vitamin synthetic pathways for queuosineand cobalamin In contrast E coli and B luti

have mcSEED binary phenotype profiles similarto that of P copri and contain complete meta-bolic pathways for cysteine biosynthesis andxylose utilization (table S2J) These results in-dicate that genomic features of the Bifidobac-terium strains examined limit their ability tothrive in the context of the Mirpur-18 diet anda community that contains the other ecogroupstrains In contrast the fact that P copri andother ecogroup strains contain and expressthese metabolic pathways provides support for

their importance in maintaining their fitnessunder these conditions As such the feature-reduction approachusedhere provides a rationalefor testing nutritional interventions that targetthese pathways in ecogroup members in chil-dren at risk for or who already have perturbedmicrobiota development

Conclusions

We have developed a statistical approach toidentify a group of 15 covarying bacterial taxathat we term an ecogroup We found that theecogroup is a conserved structural feature ofthe developing gut microbiota of healthy mem-bers of several birth cohorts residing in dif-ferent countries Moreover the ecogroup canbe used to distinguish the microbiota of chil-dren with different degrees of undernutrition(SAM MAM) and to quantify the ability of theirgut communities to be reconfigured toward ahealthy state with a MDCF Studies of gnoto-biotic piglets subjected to a set of dietary tran-sitions designed to model those experiencedby members of the Bangladeshi healthy birthcohort demonstrate that temporal changes inthe fitness of ecogroup taxa can occur in theabsence of other gut communitymembers Theseobservations suggest that the approach used toidentify the ecogroup may be useful in charac-terizing microbial community organization inmembers of other longitudinally sampled (hu-man) cohortsA critical feature of biological systems is that

they function reliably yet adapt when faced withenvironmental fluctuations (23 24) An architec-ture of sparse but tight coupling enables rapidevolution to new functions in proteins (25 26)Studies ofmacro-ecosystems such as ant colonieshave argued that adaptive behaviors are depen-dent on proper network organization (27) Thegut microbiota must satisfy the constraints ofsurvival namely withstanding insult and main-taining functionality (robustness) while stillhaving the capacity for plasticity ldquoEmbeddingrdquoa sparse network of covarying taxa in a largerframework of independently varying organ-isms could represent an elegant architecturalsolution developed by nature to maintain ro-bustness while enabling adaptation

MethodsHuman studies

A previously completed NIH birth cohort study(ldquoField Studies of Amebiasis in BangladeshrdquoClinicalTrialsgov identifier NCT02734264) wasconducted at the International Centre for Diar-rhoeal Disease Research Bangladesh (icddrb)Anthropometric data and fecal samples werecollected monthly from enrollment throughpostnatal month 60 Informed consent was ob-tained from the mother or guardian of eachchild The research protocol was approved by theinstitutional review boards of the icddrb and theUniversity of Virginia CharlottesvilleIn the case of the MAL-ED birth cohort study

(ldquoInteractions of Enteric Infections and Mal-nutrition and the Consequences for Child Health

Raman et al Science 365 eaau4735 (2019) 12 July 2019 8 of 11

Fig 5 Distinguishing features of mcSEED metabolic module expression related to the fitnessof ecogroup strains in weaned gnotobiotic piglets See fig S13A for full strain designations(A) The transcriptomes of cecal community members were classified on the basis of gene assignmentsto 81 mcSEED metabolic modules (see count matrix in fig S14B) Each strain is plotted on the firsttwo principal components of the enrichment matrix in fig S14B The inset shows that fractionalrepresentation (fitness) of strains correlates with their expression profiles as judged by positionalong PC1 (B) Singular value decomposition (SVD fig S14C) identifies which among the 81expressed metabolic modules most distinguish the indicated strains in the cecal community andMirpur-18 diet contexts (fig S14D) (C) Expressed discriminatory metabolic modules identified bySVD in (B) are shown as complete or incompletely represented in the genomes of the indicatedstrains by red pixels (predicted prototrophy for the amino acid or the ability to utilize thecarbohydrate shown) or by white pixels (auxotrophy or the inability to utilize the carbohydrate)Strains and metabolic modules are hierarchically clustered

RESEARCH | RESEARCH ARTICLEon F

ebruary 4 2021

httpsciencesciencemagorg

Dow

nloaded from

and Developmentrdquo ClinicalTrialsgov identifierNCT02441426) anthropometric data and fecalsamples were collected every month from enroll-ment to 24 months of age The study protocolwas approved by institutional review boards ateach of the study sitesThe accompanying paper by Gehrig et al (21)

describes studies that enrolled (i) Bangladeshichildren with MAM in a double-blind random-ized four-group parallel assignment inter-ventional trial study of microbiota-directedcomplementary food (MDCF) prototypes con-ducted in Dhaka Bangladesh (ClinicalTrialsgovidentifier NCT03084731) (ii) a reference cohortof age-matched healthy children from the samecommunity and (iii) a subcohort of 54 childrenwith SAM who were treated with one of three dif-ferent therapeutic foods and followed for 12monthsafter discharge with serial anthropometry andbiospecimen collection (ldquoDevelopment and FieldTesting of Ready-to-Use Therapeutic Foods Madeof Local Ingredients in Bangladesh for the Treat-ment of Children with SAMrdquo ClinicalTrialsgovidentifier NCT01889329) The research protocolsfor these studies were approved by the EthicalReview Committee at the icddrb Informed con-sent was obtained from the motherguardian ofeach child Use of biospecimens and metadatafrom each of the human studies for the analysesdescribed in this report was approved by theWashington University Human Research Protec-tion Office (HRPO)

Collection and storage of fecal samplesand clinical metadata

Fecal samples were placed in a cold box with icepacks within 1 hour of production by the donorand collected by field workers for transport backto the lab (NIH Birth Cohort MAL-ED study)For the ldquoDevelopment and Field Testing of Ready-to-Use Therapeutic Foods Made of Local In-gredients in Bangladesh for the Treatment ofChildren with SAMrdquo study the healthy referencecohort and the MDCF trial samples were flash-frozen in liquid nitrogenndashcharged dry shippers(CX-100 Taylor-Wharton Cryogenics) shortly aftertheir production by the infant or child Biospeci-mens were subsequently transported to the locallaboratory and transferred to ndash80degC freezerswithin 8 hours of collection Sampleswere shippedon dry ice to Washington University and archivedin a biospecimen repository at ndash80degC

Sequencing bacterial V4-16S rDNAamplicons and assigning taxonomy

Methods used for isolation of DNA from fro-zen fecal samples generation of V4-16S rDNAamplicons sequencing of these amplicons cluster-ing of sequencing reads into 97 ID OTUs and as-signing taxonomy are described in Gehrig et al (21)

Generation of RF-derived models of gutmicrobiota development

We produced RF-derived models of gut micro-biota development from the Peruvian Indianand ldquoaggregaterdquoV4-16S rDNAdatasets generatedfrom 22 14 and 28 healthy participants respec-

tively (see supplementary text for a description ofthe aggregate dataset) Model building for eachbirth cohort was initiated by regressing the re-lative abundance values of all identified 97IDOTUs in all fecal samples against the chronologicage of each donor at the time each sample wasprocured (R package ldquorandomForestrdquo ntree =10000) For each country site OTUswere rankedon the basis of their feature importance scorescalculated from the observed increases in meansquare error (MSE) when values for that OTUwere randomized Feature importance scoresweredetermined over 100 iterations of the algorithmTo determine how many OTUs were required tocreate a RF-based model comparable in accuracyto a model comprising all OTUs we performedan internal 100-fold cross-validation where mod-els with sequentially fewer input OTUs werecompared to one another Limiting the country-specific models to the top 30 ranked OTUs hadonly minimal impact on accuracy (within 1 ofthe MSE obtained with all OTUs) In additionto calculating the R2 of the chronological ageversus predicted microbiota age for reciprocalcross-validation of the RF-derived models wealso calculated the mean absolute error (MAE)and root mean square error (RMSE) for the ap-plication of each model to each dataset to fur-ther assess model quality (table S12)

Comparing OTUs with DADA2 ampliconsequence variants (ASVs) (fig S1)

Each OTU in the ecogroup and each OTU in thesparse RF-derived models that had 100 se-quence identity to an ASV was identified eachof these OTUs was defined as a ldquoprimary OTUsequencerdquo and the ASV as the ldquocorrect ASV se-quencerdquo The primary OTU sequence was thenmutated according to the maximum sequencevariance accepted by QIIME for a ge97ID OTU(ie le3) to create a library of 1000 derivativesequences Each sequence in the librarywas thencompared to a database of all ASVs producedfrom DADA2 analysis (28) of all 16S rDNA data-sets generated from all birth cohorts described inthis report and in Gehrig et al (21) The ASVwiththe maximum sequence identity to each mem-ber of each library of 1000 derivative sequenceswas noted If this ASVmatched the correct ASVsequence the OTU derivative sequence in thelibrary was assigned a ldquo1rdquo otherwise it was as-signed a ldquo0rdquo An average over all 1000 derivativesequences in a given library was then calculatedThis process was iterated 10 separate timescreating 10 trials of 1000 derived sequences foreach OTU An average over all 10 trials wasthen calculated thereby defining the prob-ability of an OTU being ascribed to the correctASV given the accepted sequence ldquoentropyrdquo ofQIIME (15) The results demonstrated that V4-16S rDNA sequences comprising a 97ID OTUgenerated by QIIME map directly to the singleASV sequence deduced by DADA2

Studies of gnotobiotic piglets

Experiments involving gnotobiotic piglets wereperformed under the supervision of a veterinar-

ian using protocols approved by the WashingtonUniversity Animal Studies Committee

Diets

Piglets were initially bottle-fed with an irradiatedsowrsquos milk replacement (Soweena Litter LifeMerrick catalog number C30287N) Soweenapowder (120-g aliquots in vacuum-sealed steri-lized packets) was gamma-irradiated (gt20 Gy)and reconstituted as a liquid solution in the gnoto-biotic isolator (120 g per liter of autoclavedwater) The procedure for producing Mirpur-18is detailed in Gehrig et al (21)

Husbandry

Feeding The protocol used for generating germ-free piglets was based on our previous publica-tion (29) with modifications (21) Piglets werefed at 3-hour intervals for the first 3 postnataldays at 4-hour intervals from postnatal days4 to 8 and at 6-hour intervals from postnatalday 9 to the end of the experiment Introduc-tion of solid foods began on postnatal day 4and weaning was accomplished by day 22 Eachgnotobiotic isolator was equipped with fourstainless steel bowls and one 2-gallon waterereach 2-gallon waterer (Valley Vet MaryvilleKS catalog number 17544) was equipped withtwo 05-inch nipples (Valley Vet catalog num-ber 17352) During the first 3 days after birthall four bowls were filled with Soweena Fromdays 4 to 12 at each feeding one bowl was filledwith Mirpur-18 while the remaining three bowlswere filled with Soweena On day 12 one bowl ofmilk was replaced with a bowl of water Fromday 15 to day 19 each daytime feeding consistedof placement of two bowls of water and twobowls of Mirpur-18 In nighttime one bowl ofwater was replaced with Soweena (ie each iso-lator at each feeding had two bowls ofMirpur-18one bowl of water and one bowl of Soweena)From postnatal days 20 and 21 only one bowlwas provided with Soweena and the amount ofmilk added was reduced by one half each dayduring this period On day 22 the last bowl ofmilk was replaced with a bowl of water therebycompleting the weaning process After weaningtwo bowls of fresh sterilizedwater and two bowlsof fresh Mirpur-18 were introduced into each iso-lator every 6 hours to enable ad libitum feedingThe 2-gallon waterer was replenished with freshsterilized water every 2 to 3 days Mirpur-18 con-sumption was monitored by noting the amountof input food required to maintain a filled bowlduring a 24-hour period Piglets were weigheddaily using a sling (catalog number 887600 Pre-mier Inc Charlotte NC) Environmental enrich-ment was provided within the isolators includingplastic balls for ldquorootingrdquo activity and rubber hosesand stainless steel toys for chewing and manipu-lating The behavior and health status of the pig-lets weremonitored every 3 to 4 hours throughoutthe day andnight during the first 13 postnatal daysand then every 6 hours until the time of eutha-nasia on day 29Bacterial genome assembly annotation

in silico metabolic reconstructions and phenotype

Raman et al Science 365 eaau4735 (2019) 12 July 2019 9 of 11

RESEARCH | RESEARCH ARTICLEon F

ebruary 4 2021

httpsciencesciencemagorg

Dow

nloaded from

predictions Barcoded paired-end genomic libra-ries were prepared for each bacterial isolate andthe libraries were sequenced (Illumina MiSeqinstrument paired-end 150- or 250-nt reads)Reads were demultiplexed and assembled con-tigs with greater than 10times coverage were initiallyannotated using Prokka (30) followed by anno-tation at various levels by mapping protein se-quences to the Prokaryotic Peptide Sequencedatabase of the Kyoto Encyclopedia of GenesandGenomes (KEGG) as described inGehrig et al(21) Additional annotations were based on SEEDa genomic integration platform that includes agrowing collection of complete and nearly com-plete microbial genomes with draft annotationsperformed by the RAST server (31) SEED con-tains a set of tools for comparative genomicanalysis annotation curation and in silico re-construction of microbial metabolism MicrobialCommunity SEED (mcSEED) is an application ofthe SEED platform thatwe have used formanualcuration of a large and growing set of bacterialgenomes representing members of the humangut microbiota (currently ~2600) mcSEED sub-systems (32) are user-curated liststables ofspecific functions (enzymes transporters tran-scriptional regulators) that capture current (andever-expanding) knowledge of specific metabolicpathways or groups of pathways projected ontothis set of ~2600 genomes mcSEED pathwaysare lists of genes comprising a particular meta-bolic pathway ormodule theymay bemore gran-ular than a subsystem splitting it into certainaspects (eg uptake of a nutrient separately fromitsmetabolism) mcSEED pathways are presentedas lists of assigned genes and their annotations intable S7 As detailed in Gehrig et al (21) predictedphenotypes are generated from the collection ofmcSEED subsystems represented in a microbialgenome and the results described in the form ofa binary phenotypematrix (BPM prototrophy orauxotrophy for an amino acid or B vitamin theability to utilize specific carbohydrates andorgenerate short-chain fatty acid products of fer-mentation) Table S7 presents the supportingevidence for assigning a given phenotype to anorganismColonization Bacterial strains were cultured

under anaerobic conditions in pre-reducedWilkins-Chalgren anaerobe broth (Oxoid Inc)or MegaMedium (21 33) Methods used forsequencing assembling and annotating bac-terial genomes are described in Gehrig et al(21) An equivalent mixture of each B longumstrain or additional ecogroup strain was preparedby adjusting the volumes of each culture based onoptical density (OD600) readings An equal volumeof pre-reduced PBS containing 30 glycerol wasadded to the mixture and aliquots were frozenand stored at ndash80degC until use Each piglet re-ceived an intragastric gavage (Kendall Kangaroo27 mm diameter feeding tube catalog number8888260406 Covidien Minneapolis MN) of11 ml of a solution containing the bacterial con-sortia listed in fig S13A and Soweena (110 vv)The fecal microbiota was sampled using rectalswabs on the days indicated in fig S13A

Euthanasia and assessment of communitycomposition along the length of the intestineEuthanasia was performed on experimentalday 29 according to American Veterinary Med-ical Association (AVMA) guidelines The smallintestine was divided into 20 sections of equallength the first 1 cm of the 1st 5th 10th 15thand 20th sections were opened with an incisionand luminal contents were harvested with sterilecell scraper (Falcon catalog number 353085)Luminal contents were also harvested from thececum proximal colon (10 cm of the mid-spiralregion) and distal colon (10 cm from the anus)Methods for isolation of DNA from luminal andfecal samples and short-read shotgun sequenc-ing of community DNA samples (COPRO-seq)are all detailed in Gehrig et al (21)Microbial RNA-seq Isolation of RNA from

cecal contents harvested from piglets at thetime of euthanasia depletion of ribosomal rRNA(Ribo-Zero Kit Illumina) and bacterial RNA pu-rificationwere performed (21) Double-strandedcomplementary DNA and indexed Illumina li-brarieswerepreparedusing theSMARTerStrandedRNA-seq kit (Takara Bio USA) Libraries wereanalyzedwith aBioanalyzer (Agilent) to determinefragment size distribution and then sequenced[Illumina NextSeq platform 75-nt unidirectionalreads 369 (plusmn54) times 106 reads per sample (mean plusmnSD) n = 5 samples] Fluorescence was not mea-sured from the first four cycles of sequencing asthis library preparation strategy introduces threenontemplated deoxyguanines Transcripts werequantified (34) normalized (transcripts per kilo-base per million reads TPM) and then aggre-gated according to their representation in mcSEEDand KEGG subsystemspathway modules (21)

REFERENCES AND NOTES

1 W Z Lidicker Jr A clarification of interactions inecological systems Bioscience 29 375ndash377 (1979)doi 1023071307540

2 K Faust J Raes Microbial interactions From networks tomodels Nat Rev Microbiol 10 538ndash550 (2012) doi 101038nrmicro2832 pmid 22796884

3 M Layeghifard D M Hwang D S Guttman Disentanglinginteractions in the microbiome A network perspectiveTrends Microbiol 25 217ndash228 (2017) doi 101016jtim201611008 pmid 27916383

4 A R Ives B Dennis K L Cottingham S R CarpenterEstimating community stability and ecological interactionsfrom time-series data Ecol Monogr 73 301ndash330 (2003)doi 1018900012-9615(2003)073[0301ECSAEI]20CO2

5 D R Hekstra S Leibler Contingency and statistical laws inreplicate microbial closed ecosystems Cell 149 1164ndash1173(2012) doi 101016jcell201203040 pmid 22632978

6 S Weiss et al Correlation detection strategies in microbialdata sets vary widely in sensitivity and precision ISME J10 1669ndash1681 (2016) doi 101038ismej2015235pmid 26905627

7 K Faust et al Microbial co-occurrence relationships in thehuman microbiome PLOS Comput Biol 8 e1002606 (2012)doi 101371journalpcbi1002606 pmid 22807668

8 A Zelezniak et al Metabolic dependencies drive speciesco-occurrence in diverse microbial communities Proc NatlAcad Sci USA 112 6449ndash6454 (2015) doi 101073pnas1421834112 pmid 25941371

9 J Friedman E J Alm Inferring correlation networks fromgenomic survey data PLOS Comput Biol 8 e1002687 (2012)doi 101371journalpcbi1002687 pmid 23028285

10 Z D Kurtz et al Sparse and compositionally robust inferenceof microbial ecological networks PLOS Comput Biol 11e1004226 (2015) doi 101371journalpcbi1004226pmid 25950956

11 V Plerou et al Random matrix approach to cross correlationsin financial data Phys Rev E 65 066126 (2002) doi 101103PhysRevE65066126 pmid 12188802

12 S W Lockless R Ranganathan Evolutionarily conservedpathways of energetic connectivity in protein families Science286 295ndash299 (1999) doi 101126science2865438295pmid 10514373

13 N Halabi O Rivoire S Leibler R Ranganathan Proteinsectors Evolutionary units of three-dimensional structureCell 138 774ndash786 (2009) doi 101016jcell200907038pmid 19703402

14 S Subramanian et al Persistent gut microbiota immaturity inmalnourished Bangladeshi children Nature 510 417ndash421(2014) doi 101038nature13421 pmid 24896187

15 J G Caporaso et al QIIME allows analysis of high-throughputcommunity sequencing data Nat Methods 7 335ndash336 (2010)doi 101038nmethf303 pmid 20383131

16 A direct comparison of these OTUs and amplicon sequencevariants (ASVs) identified using a bioinformatic pipelinedesigned to reduce sequencing errors disclosed good agree-ment between the two methods (fig S1 and methods)Therefore we retained OTU designations for this study

17 A Hsiao et al Members of the human gut microbiota involvedin recovery from Vibrio cholerae infection Nature 515423ndash426 (2014) doi 101038nature13738 pmid 25231861

18 T Yatsunenko et al Human gut microbiome viewedacross age and geography Nature 486 222ndash227 (2012)doi 101038nature11053 pmid 22699611

19 Each monthly covariance matrix was normalized against thehighest covariance value for that month (see fig S5 A to Dand table S2A for the example of month 60) Because sometaxon-taxon covariance values are zero as a result of theabsence of a taxon (eg fig S5C) fitting a probabilitydistribution over all of the covariance values becomes apractical constraint Therefore we retained the nonzero valuesacross months 20 to 60 yielding 80 of the original 118 taxaValues in the normalized covariance matrix for each monthwere then fit to a t-location scale probability distributionbecause the monthly normalized covariance histograms weresignificantly heavy-tailed (eg fig S5D) Given our desire toidentify which taxon-taxon covariance values were consistentlyin the tails of these probability distributions over time theelements in each monthly covariance matrix were binarized toa ldquo1rdquo if they fell within the top or bottom 10 and a ldquo0rdquo if theirvalues were within the remaining 80 of the probabilitydistribution this isolated the most covarying taxon-taxon pairs[ethCij

binTHORNt where i and j are bacterial taxa and t designates themonth] Monthly binarized covariance matrices were thenaveraged over time to create an 80 times 80 covariance matrixthat signifies temporally conserved taxon-taxon covariation(hCij

binit Fig 1B)20 MAL-ED Network Investigators The MAL-ED study A

multinational and multidisciplinary approach to understand therelationship between enteric pathogens malnutrition gutphysiology physical growth cognitive development andimmune responses in infants and children up to 2 years of agein resource-poor environments Clin Infect Dis 59S193ndashS206 (2014) pmid 25305287

21 J L Gehrig et al Effects of microbiota-directed foods ingnotobiotic animals and undernourished children Science 365eaau4732 (2019)

22 E Miller D Ullrey The pig as a model for human nutritionAnnu Rev Nutr 7 361ndash382 (1987)

23 J A Draghi T L Parsons G P Wagner J B PlotkinMutational robustness can facilitate adaptation Nature 463353ndash355 (2010) doi 101038nature08694 pmid 20090752

24 M Kirschner J Gerhart Evolvability Proc Natl AcadSci USA 95 8420ndash8427 (1998) doi 101073pnas95158420 pmid 9671692

25 R N McLaughlin Jr F J Poelwijk A Raman W S GosalR Ranganathan The spatial architecture of protein functionand adaptation Nature 491 138ndash142 (2012) doi 101038nature11500 pmid 23041932

26 A S Raman K I White R Ranganathan Origins of allosteryand evolvability in proteins A case study Cell 166 468ndash480(2016) doi 101016jcell201605047 pmid 27321669

27 D M Gordon The ecology of collective behavior PLOS Biol12 e1001805 (2014) doi 101371journalpbio1001805pmid 24618695

28 B J Callahan et al DADA2 High-resolution sample inferencefrom Illumina amplicon data Nat Methods 13 581ndash583 (2016)doi 101038nmeth3869 pmid 27214047

29 M R Charbonneau et al Sialylated milk oligosaccharidespromote microbiota-dependent growth in models of infant

Raman et al Science 365 eaau4735 (2019) 12 July 2019 10 of 11

RESEARCH | RESEARCH ARTICLEon F

ebruary 4 2021

httpsciencesciencemagorg

Dow

nloaded from

undernutrition Cell 164 859ndash871 (2016) doi 101016jcell201601024 pmid 26898329

30 T Seemann Prokka Rapid prokaryotic genome annotationBioinformatics 30 2068ndash2069 (2014) doi 101093bioinformaticsbtu153 pmid 24642063

31 R Overbeek et al The SEED and the Rapid Annotation ofmicrobial genomes using Subsystems Technology (RAST)Nucleic Acids Res 42 D206ndashD214 (2014) doi 101093nargkt1226 pmid 24293654

32 R Overbeek et al The subsystems approach to genomeannotation and its use in the project to annotate 1000 genomesNucleic Acids Res 33 5691ndash5702 (2005) doi 101093nargki866 pmid 16214803

33 A L Goodman et al Extensive personal human gutmicrobiota culture collections characterized andmanipulated in gnotobiotic mice Proc Natl AcadSci USA 108 6252ndash6257 (2011) doi 101073pnas1102938108 pmid 21436049

34 M C Hibberd et al The effects of micronutrient deficiencieson bacterial species from the human gut microbiotaSci Transl Med 9 eaal4069 (2017) doi 101126scitranslmedaal4069 pmid 28515336

35 Github deposition of code Zenodo doi 105281zenodo3255003Also available for download at githubcomarjunsramanRaman_et_al_Science_2019

ACKNOWLEDGMENTS

We are indebted to the families of study subjects for their activeparticipation and assistance We thank the staff and investigators aticddrb for their contributions to the recruitment and enrollment ofparticipants in the 5-year Bangladeshi birth cohort study plus theinterventional studies of children with SAM and MAM as well as thecollection of biospecimens and data We also thank the study teammembers and health care workers involved in the MAL-ED birthcohort studies M Gottlieb D Lang K Tountas and M McGrath whoprovided invaluable assistance in coordinating the MAL-ED

collaboration and providing access to key clinical datasets M MeierS Deng and J Hoisington-Loacutepez for superb technical assistanceD OrsquoDonnell J Serugo and M Talcott for their indispensable helpwith gnotobiotic piglet husbandry and R Olson for technical supportwith the mcSEED-based genome analysis and subsystem curationFunding Supported by the Bill amp Melinda Gates Foundation as part ofthe Breast Milk Gut Microbiome and Immunity (BMMI) ProjectThe 5-year birth cohort study of Bangladeshi children was funded byNIH grant AI043596 (WAP) ASR is a postdoctoral fellowsupported by Washington University School of Medicine PhysicianScientist Training Program and in part by NIH grant DK30292 DARAAA and SAL were supported by Russian Science Foundationgrant 19-14-00305 JIG is the recipient of a Thought Leader awardfrom Agilent Technologies Author contributions RH and WAPdesigned and oversaw the 5-year birth cohort study they togetherwith TA were responsible for coordinating various aspects ofbiospecimen and metadata collection SH MM RH WAP andTA (Bangladesh) MNK (Peru) GK (India) POB (South Africa) andAAML (Brazil) oversaw the MAL-ED studies SH IM MI MMand TA were responsible for studies involving the SAM and MAMcohorts JLG and SS generated 16S rDNA datasets from humanfecal samples MJB managed the repository of biospecimensand associated clinical metadata used for the studies describedabove H-WC performed the experiments with gnotobiotic pigletswith the assistance of ASR SV and MCH DAR AAA SALand ALO performed in silico metabolic reconstructions based on thegenome sequences of bacterial strains introduced into gnotobioticpiglets ASR conceived the mathematical approach and wrote all ofthe computational workflow for identifying ecogroup taxa performedthe sensitivity analysis of the workflow compared the SparCC andSPIEC-EASI algorithms with the workflow and undertook the analysesof gut microbial communities from subjects enrolled in the SAMMDCF Peruvian and Indian cohort studies as well as the gnotobioticpiglet experiment with JLG SV MJB and JIG contributing invarious supportive ways ASR and JIG wrote the paper Competinginterests JIG is a co-founder of Matatu Inc a company

characterizing the role of diet-by-microbiota interactions in animalhealth WAP serves as a consultant to TechLab Inc a company thatmakes diagnostic tests for enteric infections and has served as aconsultant for Perrigo Nutritionals LLC which produces infantformula Data and materials availability Bacterial V4-16S rDNAsequences in raw format (prior to postprocessing and data analysis)shotgun datasets generated from cultured bacterial strains andCOPRO-seq and microbial RNA-seq datasets obtained fromgnotobiotic piglets have been deposited at the European NucleotideArchive under study accession number PRJEB27068 Code has beenarchived at Zenodo (35) Fecal specimens from the MAL-ED birthcohorts in Bangladesh (icddrb Dhaka) Brazil (Federal University ofCearaacute Fortaleza) India (Christian Medical College Vellore) Peru(JHSPHAB PRISMA) South Africa (University of Venda) and fromthe NIH birth cohort and SAMMDCF studies at icddrb were providedto Washington University under material transfer agreementsThis work is licensed under a Creative Commons Attribution 40International (CC BY 40) license which permits unrestricted usedistribution and reproduction in any medium provided the originalwork is properly cited To view a copy of this license visit httpcreativecommonsorglicensesby40 This license does not applyto figuresphotosartwork or other content included in the articlethat is credited to a third party obtain authorization from the rightsholder before using such material

SUPPLEMENTARY MATERIALS

sciencesciencemagorgcontent3656449eaau4735supplDC1Supplementary TextFigs S1 to S16Tables S1 to S13References (36ndash40)

13 June 2018 resubmitted 24 April 2019Accepted 7 June 2019101126scienceaau4735

Raman et al Science 365 eaau4735 (2019) 12 July 2019 11 of 11

RESEARCH | RESEARCH ARTICLEon F

ebruary 4 2021

httpsciencesciencemagorg

Dow

nloaded from

developmentA sparse covarying unit that describes healthy and impaired human gut microbiota

Haque Tahmeed Ahmed Michael J Barratt and Jeffrey I GordonA Arzamasov Semen A Leyn Andrei L Osterman Sayeeda Huq Ishita Mostafa Munirul Islam Mustafa Mahfuz Rashidul

AleksandrGagandeep Kang Pascal O Bessong Aldo AM Lima Margaret N Kosek William A Petri Jr Dmitry A Rodionov Arjun S Raman Jeanette L Gehrig Siddarth Venkatesh Hao-Wei Chang Matthew C Hibberd Sathish Subramanian

DOI 101126scienceaau4735 (6449) eaau4735365Science

this issue p eaau4732 p eaau4735Sciencemetabolic and growth profiles on a healthier trajectoryage-characteristic gut microbiota The designed diets entrained maturation of the childrens microbiota and put theirstate that might be expected to support the growth of a child These were first tested in mice inoculated with recovery Diets were then designed using pig and mouse models to nudge the microbiota into a mature post-weaningmalnutrition The authors investigated the interactions between therapeutic diet microbiota development and growth

monitored metabolic parameters in healthy Bangladeshi children and those recovering from severe acuteet alRaman andet altherapeutic intervention with standard commercial complementary foods children may fail to thrive Gehrig

Childhood malnutrition is accompanied by growth stunting and immaturity of the gut microbiota Even afterMalnutrition and dietary repair

ARTICLE TOOLS httpsciencesciencemagorgcontent3656449eaau4735

MATERIALSSUPPLEMENTARY httpsciencesciencemagorgcontentsuppl201907103656449eaau4735DC1

CONTENTRELATED

httpstmsciencemagorgcontentscitransmed4137137rv6fullhttpstmsciencemagorgcontentscitransmed4137137rv7fullhttpstmsciencemagorgcontentscitransmed6220220ra11fullhttpstmsciencemagorgcontentscitransmed7276276ra24fullhttpstmsciencemagorgcontentscitransmed8366366ra164fullhttpsciencesciencemagorgcontentsci3656449109fullhttpsciencesciencemagorgcontentsci3656449eaau4732full

REFERENCES

httpsciencesciencemagorgcontent3656449eaau4735BIBLThis article cites 40 articles 10 of which you can access for free

PERMISSIONS httpwwwsciencemagorghelpreprints-and-permissions

Terms of ServiceUse of this article is subject to the

is a registered trademark of AAASScienceScience 1200 New York Avenue NW Washington DC 20005 The title (print ISSN 0036-8075 online ISSN 1095-9203) is published by the American Association for the Advancement ofScience

Copyright copy 2018 American Association for the Advancement of Science

on February 4 2021

httpsciencesciencem

agorgD

ownloaded from