A Microbiota-Directed Food Intervention for Undernourished Children

BACKGROUND More than 30 million children worldwide suffer from moderate acute malnutrition (MAM). Current treatments have limited effectiveness and much remains unknown about pathogenesis. Children with MAM exhibit perturbed development of their gut microbiota. METHODS Slum-dwelling Bangladeshi children, aged 12 to 18 months, with moderate acute malnutrition (n=124) received a microbiota-directed complementary food (MDCF-2) or an existing ready-to-use supplementary food (RUSF), twice daily for three months followed by a 1-month period of monitoring. We obtained weight-for-length, weight-for-age, and length-for-age Z-scores and mid-upper arm circumference at baseline and fortnightly, through four months. We compared the rate of change of these related phenotypes between baseline and three months, and between baseline and four months. We also measured levels of 4,977 proteins in plasma plus 209 bacterial taxa in fecal samples. RESULTS 118 children completed the intervention (n=59/arm). The rate of change in weight-for-length Z-score (β-WLZ), weight-for-age Z-score, and mid upper arm circumference is consistent with a benefit of MDCF-2 on growth over the course of the study including the one-month follow-up. Receipt of MDCF-2 was linked to the magnitude of change in levels of 70 β-WLZ-positively correlated plasma proteins including mediators of bone growth, neurodevelopment and inflammation (gene set enrichment analysis [GSEA];p<0.001) and the abundances of 23 WLZ-associated bacterial taxa (GSEA;p<0.001). CONCLUSIONS These findings provide support for further clinical investigation of MDCF-2 as a dietary supplement for young children with MAM and provide insight into mechanisms by which this targeted manipulation of microbiota components may be linked to growth. (Supported by the Bill and Melinda Gates Foundation and the NIH; ClinicalTrials.gov identifier:NCT04015999)


Nutritional supplementation
During the first month of the study, each child was brought to a study center twice daily (morning and afternoon). On each visit, mothers were provided 25 g of their assigned food supplement (MDCF-2 or RUSF) and asked to spoon feed their child, under the supervision of trained study personnel, until she/he refused to eat further. The amount of food consumed at each visit was recorded by subtracting that left over from the offered amount (pre-weighed napkins were used to collect any food regurgitated or spilled, which was deducted from the amount provided). In the second month, each child was provided 25 g of their assigned food supplement at the feeding center, and an additional 25 g was provided in a clean container to feed at home that afternoon. In the third month, two separate containers (for morning and afternoon feeding) containing 25 g of study diet were delivered each day to each enrolled child at her/his home. Any unconsumed diet from each feeding was retained in the container; the weight of study diet consumed was determined by weighing the food remaining in the container. Mothers were advised to avoid feeding their children during the 2-hour period before each visit but to otherwise continue their usual breastfeeding and complementary feeding practices throughout the day/study. Current recommendations for treating children with MAM are predicated on the principle that dietary management should be based on the optimal use of locally available nutrient-dense foods to improve their nutritional status and prevent them from becoming further malnourished (e.g., developing SAM). However, at present, there are no evidence-informed recommendations regarding the composition of supplementary foods used to treat children with MAM (1,2).
To develop MDCFs, we selected 12 locally available complementary food ingredients that are commonly consumed by the target population in Dhaka, Bangladesh. We designed 21 unique diets containing different combinations of these ingredients and tested them in gnotobiotic mice colonized with a consortium of 9 'target' bacterial strains; these strains were selected based on their underrepresentation in the fecal microbiota of Bangladeshi children with acute malnutrition when compared to the microbiota of age-matched healthy children from the same population. (Note that administration of a consortium of ONLINE SUPPLEMENT 7 five of these target strains had been previously shown to ameliorate growth faltering in gnotobiotic mice colonized with intact microbiota from a stunted, underweight child) (3). We identified ingredients (chickpea, peanut, banana and soyflour) that significantly promoted the representation of one or more of the target bacterial strains. Formulations containing these ingredients were subsequently tested in gnotobiotic piglets colonized with the bacterial targets to ascertain their effects on bacterial fitness as well as on host biology in a species that is physiologically and metabolically more similar to humans than mice. Based on these results and tests of organoleptic acceptability, lead MDCF prototypes were advanced to a 1-month controlled feeding study involving 15-17 Bangladeshi children with MAM/arm to determine their capacity to produce microbiota repair. This study led to the selection of MDCF-2 for the follow-up POC trial described here due to the superiority of its ability to promote microbiota repair and its beneficial effects on plasma biomarkers of healthy growth.
In addition to the inclusion of microbiota-directed ingredients, MDCF-2 was designed to approximate the recommended nutritional composition of specially formulated foods for the prevention and treatment of acute malnutrition in 12-18 month-old children (2); specifically, provision of ~250 kcal/day of supplemental energy, of which 45-50 percent should be from fat and 8-10 percent from protein. Culturally acceptable, ready-to-use supplementary foods (RUSF) composed of locally-available food ingredients (rice/lentils) have been previously demonstrated to provide modest but statistically significant improvements in linear growth (compared to nutritional counselling alone) after 12 months of treatment of rural Bangladeshi children during the weaning period (6-18 months) (4,5). A recent metaanalysis indicated that supplementary foods are often more effective at promoting anthropometric recovery in children with MAM than nutrition counseling, with or without the addition of micronutrient supplements. In addition, supplementation with high-quality protein and adequate micronutrient content, for 3 months, is recommended (6).

ONLINE SUPPLEMENT
8 Fecal samples were collected at participants' homes within 20 minutes of production by study personnel, transferred in 2 mL cryovials to Cryo Exchange vapor shippers (Taylor-Wharton/Worthington Industries, CX-100) and transported to the study center where they were recorded and stored at -80 o C. EDTAplasma was prepared from blood collected during scheduled visits to the study center as previously described (7) and stored at -80 o C. Coded biospecimens were shipped to Washington University on dry ice where they were stored at -80 o C, along with associated metadata, in a dedicated repository with approval from the Washington University Human Research Protection Office.

Immunization and breastfeeding status
In Bangladesh, the 'Expanded Programme on Immunization' (EPI) starts with Bacillus Calmette-Guerin (BCG) vaccine, pentavalent vaccine, inactivated polio vaccine (IPV) and bivalent oral poliovirus vaccine (bOPV), Pneumococcal conjugate vaccine (PCV) and measles and rubella (MR) virus vaccine to protect children from nine preventable diseases. We categorized immunization status as 'complete' (received all six vaccinations), 'partial' (completed less than six), and 'none' (not vaccinated at all). Breastfeeding status at enrollment was categorized as 'exclusively breast fed', 'partially breast fed' or 'never breast fed' since birth.

Preparation of MDCF-2 and RUSF
A food processing laboratory was established in the Mirpur area, in close proximity to the nutrition centers where the intervention was provided. All raw ingredients were purchased from a single local market in Dhaka. Each step of food preparation, including cleaning, roasting, particle size reduction, homogeneous blending, and supply to the nutrition centers was performed and monitored by icddr,b study investigators and field supervisors. Upon receiving the raw dry food ingredients (rice, lentils, chickpeas, soybeans, peanuts), any foreign material, grains or seeds were removed manually and by using a sieve.
Ingredients were roasted in an open pan at 120-130 °C for 8-10 minutes, then allowed to cool and subsequently ground. At this stage, peanut was ready for mixing. The other food ingredients were converted into fine particles by blending for 4 to 5 minutes and sieving. Sugar was ground and the resulting fine powder was mixed with the other ingredients. Unpeeled whole green bananas were placed ONLINE SUPPLEMENT 9 in a deep pan and boiled in water for 17-20 minutes until they were tender. The peel was removed and the fruit was grated into small pieces which, after cooling, were mashed with a potato masher. The weights of all the ingredients required for preparing MDCF-2 and RUSF were recorded, pre-weighed micronutrient premix powder was added, and the supplementary foods were produced in small batches by mixing all ingredients in an electric blender.
The MDCF-2 and RUSF formulations were prepared fresh daily, dispensed and fed to participants on the same day. Samples of the food were routinely cultured at the icddr,b Food Safety Laboratory; tests included scoring total aerobes, total coliforms, Escherichia coli, Enterobacteriaceae, Bacillus cereus, Salmonella spp, Shigella spp, Campylobacter spp, coagulase positive and other Staphylococci, as well as yeasts and molds. The nutritional composition (energy content, moisture, protein, total fat, total carbohydrate, dietary fiber, ash) of the ingredients was assessed at the Institute of Nutrition, Mahidol University, Thailand using standard procedures.

Processing of plasma samples
The aptamer based SomaScan 5K Proteomic Assay plasma/serum kit (SomaLogic) was used to quantify the abundances of 5,284 proteins in plasma samples collected from children. Plasma samples were processed according to manufacturer's instructions as previously described (8). Briefly, 50 L of plasma was incubated with NHS-biotin-tagged, protein-specific aptamer probes ('SOMAmers') to form protein-SOMAmer complexes that were immobilized on streptavidin beads. The complexes were subsequently cleaved, denatured, eluted, and hybridized to a custom Agilent DNA microarray. Arrays were scanned with an Agilent SureScan instrument at 5 µm resolution and the Cy3 fluorescence signal was quantified and processed using SomaLogic's SomaScan standardization procedures (8). SOMAmers that were not specific to human proteins, or that were marked by SOMAlogic as deprecated, were removed.
Additionally, SOMAmers were removed whose median fluorescence signal across all samples was within 4.9 median average deviances (MAD) from blanks, resulting in a total of 4,977 SOMAmers that passed quality control ( Figure S4). Protein abundances were log2-transformed and quantile-normalized prior to all downstream analyses (8).

Processing of fecal samples for microbiota analysis
Fecal samples were pulverized in liquid nitrogen. Methods for DNA extraction and purification have been described previously (9). Briefly, DNA was extracted from ~50 mg of pulverized material by beadbeating with 500 mL of 0.1 mm diameter zirconia/silica beads in a solution of 500 ml phenol:chloroform:isoamyl alcohol (25:24:1), 210 mL 20% SDS, and 500 mL buffer A (200 mM NaCl, 200 mM Trizma base, 20 mM EDTA), followed by purification (Qiaquick columns, Qiagen) and storage in Tris-EDTA (TE) buffer. Purified DNA was quantified (Quant-iT dsDNA broad range kit, Invitrogen), adjusted to 1 ng/uL concentration, and used to generate bacterial V4 16S rDNA amplicons and subsequent indexed Illumina libraries (9). Libraries were quantified, pooled, and sequenced using an Illumina MiSeq instrument to generate paired-end, 250 nt reads (3.29x10 4  9.93x10 3 reads/sample; mean  SD). Amplicon sequences were processed to trim adapter and primer sequences using bbtools (v37.02). DADA2 (10) was used to analyze preprocessed, paired-end sequence data to obtain and quantify errorcorrected amplicon sequence variants (ASVs) in R (v3.6.1). Taxonomic assignments were performed using the DADA2 implementation of the Ribosomal Database Project Naïve Bayesian Classifier (database v16) at a minimum bootstrap confidence of  80% (option 'minboot = 80'). Tables of ASV abundances (counts) for each sample were combined with sample metadata and taxonomic assignment into a phyloseq (v1.3.0) object in R. Samples with fewer than 2000 reads were excluded from further analysis. In addition to removing samples with insufficient absolute read depth, we also performed a rarefaction analysis based on the R package vegan (v2.5.6). Samples for which ASV richness reached saturation at their corresponding sampling depth were considered sufficiently sampled ( Figure S7).
Contaminating mitochondrial or chloroplast ASV sequences were removed, along with any bacterialorigin ASVs lacking phylum-level taxonomic classification. A count filter was applied to remove any ASVs present below five counts in fewer than 5% of samples, yielding a filtered table containing 209 ASVs across 939 samples. This filtered ASV table was adjusted for library size and normalized (variance stabilizing transformation) using DESeq2 (11).

Quantification of 23 enteropathogens by multiplex qPCR
Samples were prepared as described in (9). Briefly, DNA and RNA were extracted and purified from fecal samples. cDNAs were prepared using the High Capacity cDNA Reverse Transcription Kit (Applied Biosystems, Foster City, CA). The resulting cDNA products were amplified using Specific Target Amplification, hybridized with TaqMan primers and probes against 23 bacterial, viral, and protozoan pathogens (12) and quantified using the 96.96 Dynamic Array microfluidic digital PCR system (Fluidigm Corp. San Francisco, CA). The qPCR assay used in this study has been widely deployed in large multicenter cohort studies of undernourished children involving mixed-species microbial communities and is highly specific for the 23 target pathogens in these contexts (13,14).

Study design: power calculation
In our previous 1-month-long, pilot study of different MDCFs (9), the pretreatment WLZ score of children assigned to the MDCF-2 arm was -2.2. After one month of supplementation, WLZ was -1.7.
Considering a WLZ of -2 at baseline and -1.7 at the end of the intervention, and a pooled SD of 0.53, we calculated that a sample size of 49 in each arm would be necessary to achieve 80% power at a 5% level of significance. Furthermore, assuming a 20% rate of attrition, we determined that 62 children would be required per treatment group for the 4-month-long study described in the current report.

Analysis of enrollment and baseline characteristics
Comparisons of demographic, anthropometric, and environmental features at enrollment between children receiving MDCF-2 or RUSF were performed using two-sided unpaired t-tests for normally distributed features, Wilcoxon rank-sum tests for measurements with skewed distributions, or Chi-squared tests for categorical variables. The first day of intervention began on average 5.88 ± 0.14 (SEM) days after enrollment. For all analyses of anthropometric measurements, the first day of intervention was used as the baseline measurement. Comparisons of baseline anthropometric measures between children receiving ONLINE SUPPLEMENT 12 MDCF-2 and those receiving RUSF were performed using a linear model controlling for baseline age, gender, and any history of illness 7 days prior to enrollment.

Analysis of anthropometric responses to MDCF-2 or RUSF
Ponderal growth rate was calculated using a mixed-effects linear model that predicted WLZ from weeks in the intervention, controlling for baseline age, gender, any history of illness 7 days preceding enrollment, and a random intercept for each participant ID (PID). The model took the form: The rates of growth for children receiving MDCF-2 or RUSF reported in Table 2 are β1 in Equation (1) and represent the weekly increase in WLZ in a given treatment arm. The same equation was used to assess weekly improvements in WAZ, MUAC and LAZ.
A comparison of the effects of MDCF-2 and RUSF on growth rates was performed using a mixed effects linear model predicting WLZ from the interaction between weeks in the intervention and treatment, controlling for baseline age, gender, any history of illness 7 days preceding enrollment, weeks in the intervention, treatment, and a random intercept for each participant. The model took the form: The differential rate of ponderal growth as a function of treatment arm (MDCF-2 versus RUSF) reported in Table 2 is β1 in Equation (2) and represents how much more WLZ improved in children receiving MDCF-2 compared to those receiving RUSF per week. Equation (2) was also used to compare the effects ONLINE SUPPLEMENT 13 of MDCF-2 and RUSF on rates of change of WAZ, MUAC and LAZ by substituting WLZ for the anthropometric measure.

Robust linear mixed-effects modeling
A comparison of the effects of MDCF-2 versus RUSF was repeated using a robust linear mixed-effects modeling approach that controls for heavy-tailed distributions (i.e., outliers) attributable to both betweenand within-subject variation (15). Equation (2) was used to estimate model coefficients using the rlmer function in the R package robustlmm v2.3 (16). The differential rate of growth as a function of treatment arm (MDCF-2 versus RUSF) reported in Table S5 is β1 in Equation (2), and represents how much more WLZ, WAZ, LAZ, or MUAC improved in children receiving MDCF-2 compared to those receiving RUSF per week. P-values were computed using the pt function in R, which calculates a p-value by drawing from the Student's t-distribution given a t-statistic (from rlmer) and an estimated degrees-offreedom (approximated by the Welch-Satterthwaite equation).

Analysis of illness and co-morbidities during supplementation
For each intervention, the change in prevalence of cough, rhinorrhea, fever or diarrhea was quantified using a generalized mixed-effects linear model with a 'logit' link function that predicted a given comorbidity from weeks in the intervention, controlling for a random intercept for each participant. The model took the form: The within-treatment log-odds ratio reported in Table S6 is β 1 in Equation (3) and represents how much more/less likely it would be to have a co-morbidity each week during the intervention period.
A comparison of the effects of MDCF-2 and RUSF on the prevalence of cough, rhinorrhea, fever and diarrhea was performed using a generalized mixed effects linear model with a 'logit' link function predicting a particular co-morbidity from the interaction between weeks in the intervention and treatment, ONLINE SUPPLEMENT 14 controlling for the main effects of treatment, weeks in the intervention, and a random intercept for each participant. The model took the form: The differential prevalence of co-morbidity as a function of treatment arm (MDCF-2 versus RUSF) reported in Table S6 is β 1 in Equation (4) and represents how much more likely a given co-morbidity is to be reported each week in the MDCF-2 arm compared to the RUSF arm.

Analysis of dietary habits during supplementation
Data from a 66-question food-frequency questionnaire (FFQ) was collected at enrollment as well as one, two, four, five, nine and 12 weeks after starting the intervention. Each question prompted either a binary response (yes/no) or a count-based (integer) response. Caregivers of participants were asked to recall the foods that their children had consumed over the previous 24 hours. Food items were organized into seven groups according to WHO infant and young child feeding guidelines: (i) grains, roots and tubers; (ii) legumes and nuts; (iii) dairy products; (iv) flesh foods; (v) eggs; (vi) vitamin A rich fruits and vegetables; and (vii) other fruits and vegetables.
A comparison of the effects of MDCF-2 and RUSF on responses to FFQ items was performed using a generalized mixed-effects linear model. The underlying regression model followed either a binomial distribution for questions prompting binary answers or a Poisson distribution for questions prompting count-based answers. Each model predicted the response to a question as a function of the interaction between weeks in the study and treatment, controlling for the main effects of treatment, weeks in the intervention, as well as a random intercept for each participant. The model took the form: The differential change in prevalence (for binary answers) or number of times a food group was consumed (for count-based answers) in Table S4A is β1 in Equation (5) and represents how much more likely a food type was reported to be consumed in the FFQ each week in the MDCF-2 arm compared to the RUSF arm. Estimates from models that did not converge are reported as 'NA'. The overall difference in FFQ response between the two treatment arms is denoted by β2 in Equation (5) and is reported in Table   S4B.
Additionally, a mixed-effects linear model was created for each FFQ response to determine whether any items on the questionnaire were related to ponderal growth. Each model predicted the FFQ response as a function of WLZ, controlling for the fixed effects of treatment, weeks in the intervention, the interaction of weeks in the intervention and treatment, as well as a random intercept for each child.
The model took the form: The relationship between FFQ response and WLZ reported in Table S4C is β1 in Equation (6) and represents the effect of a one-unit increase in FFQ response on WLZ. Because WLZ and FFQ responses were collected at different times, timepoints were converted to ranks (integer values between 1 through 7, reflecting the number of times WLZ was measured and FFQs were administered during the intervention) prior to building each model. Estimates from models that did not converge are reported as 'NA'.

Identification of 'WLZ-associated' proteins
Pearson correlations between changes in protein abundances and ponderal growth rates were used to nominate 'WLZ-associated' proteins. Because WLZ was measured every 15 days (a total of seven measurements throughout the course of the intervention) while plasma protein abundances were only quantified at three time points (baseline, one and three months after the start of intervention), we

Differential abundance analysis of plasma proteins
Differential abundance analyses between timepoints, intervention, or the interaction between timepoint and intervention were performed using limma (18). The 'duplicateCorrelation' function, which corrects for correlations within a blocked design, was used to account for the repeated measurements taken from each participant, resulting in the equivalent of a mixed effects linear model with a random intercept for each child. Statistical significance was defined as a limma q<0.1.
Enrichment for GO 'biological processes' or the set of 70 proteins whose changes in abundances were significantly correlated with changes in WLZ was performed by rank-ordering proteins by their limma test-statistic (for within-treatment comparisons) or log-fold-difference (for between-treatment comparisons), then employing the fgsea package in R to calculate enrichment using 10,000 permutations.
The term 'significant' was reserved solely for statistical tests that resulted in p<0.05 (e.g., enrichment for WLZ-associated proteins) or q-value<0.1 for analyses requiring correction for multiple hypotheses (e.g., enrichment for GO terms).

Analysis of gut microbiota reconfiguration
Mixed-effects linear models (R packages lme4 v1.1.23 and lmerTest v3.1.1) were used to relate the abundances of ASVs (variance-stabilizing transformed counts) in each trial participant to the same participant's anthropometric characteristics using model formulas of the form: ANOVA was used to determine the significance of relationships between model terms and WLZ. WLZassociated ASVs were identified as those exhibiting false-discovery-rate adjusted p-values  0.05.
Differences in abundance were calculated for each ASV in each trial participant between the beginning and end of the intervention and between the end of intervention and the one-month follow-up timepoint.
These ASV responses were averaged within and compared between the (i) MDCF-2 and RUSF trial arms and (ii) participants with upper-quartile and lower-quartile −WLZ responses. The enrichment of WLZassociated ASVs in these comparisons was calculated using the fgsea package in R (10,000 permutations).
Mixed effects linear models were also used to relate the abundances of WLZ-associated ASVs to time under treatment in the current study and to time in a set of healthy, age-matched, longitudinally sampled children from a previous study in Mirpur (7). Reference samples (n = 309, 29 individuals) from the healthy cohort were selected based on inclusion within the upper and lower bounds of age in the current study (11-23 months) and WLZ above -1 at all timepoints. Model formulas took the following form: ASV abundance ~ β 1 (month of age or month in study) + (1|PID) Model coefficients for age were compared between each treatment arm and the healthy cohort using ztests, with false-discovery-rate adjusted p-values  0.05 indicating a significant result.

ONLINE SUPPLEMENT
18 The durability of ASV responses was determined by comparing the beginning to end of treatment response of each taxon to the end of treatment to 1-month follow-up response in each trial participant.
The effects of starting microbiota configuration and growth were assessed in the 59 children who received MDCF-2 by normalizing baseline ASV abundances with DESeq2 and regressing them against -WLZ.

Determining associations between food frequency questionnaire responses and ASV abundances
Mixed-effects linear models were used to determine associations between FFQ responses and the abundances of the 23 WLZ correlated ASVs. The model took the form: ASV abundance ~ β 1 (FFQ response) + β 2 (weeks since baseline) β 3 (FFQ response: weeks since baseline) + (1|PID) ASV abundance was parameterized by the variance-stabilizing transformed counts computed by DESeq2.
1 represents the effect of a one-unit change in FFQ response on ASV relative abundance, while 3 represents the rate of change of FFQ response on ASV abundance. The number of weeks since baseline was matched to the closest week in which both FFQ response and ASV data were available.  coefficients listed in Table S16A and Table S16B represent 1 and 3 in Equation (9), respectively. Q-values listed in Table S16 represent false-discovery rate-adjusted p-values; q-values less than 0.1 are bolded and considered statistically significant. Models that did not converge are listed as "NA" in Table S16 and were due to nearly homogenous responses to the FFQ item.

Negative binomial singular value decomposition (NB-SVD) analysis
We previously described cross-correlation singular value decomposition (CC-SVD), an analytical technique that can be used to reveal associations between disparate feature types measured in the same individual (8). However, because bacterial abundances measured in fecal samples from this study followed a negative binomial distribution, the statistical assumptions of CC-SVD were violated. Thus, we developed negative binomial SVD (NB-SVD), a statistical method that can be used to identify associations between disparate feature types measured from the same individual when one feature type ONLINE SUPPLEMENT 19 follows a negative binomial distribution. NB-SVD analysis begins with two abundance matrices-one for the abundances of ASVs, the other for the abundances of proteins. Each element of the ASV abundance matrix A MxN contains Ai,j-the abundance of ASV j in fecal sample i-while each element of the protein abundance matrix P MxP contains element Pi,k, which is the abundance of protein k in plasma sample i.
Each row i represents abundances quantified in matched plasma and fecal samples taken from the same individual at the same timepoint (baseline, one month, or three months after starting intervention). All 118 participants who had available fecal and plasma samples at baseline, one month, and three months were included as rows in A MxN and P MxP . Additionally, A MxN was filtered to remove any ASV that was present in less than 5% of samples.
Next, a cross-association matrix between proteins and ASVs was created. For a given plasma protein k, negative binomial regression with Empirical Bayes shrinkage was used to predict the expected counts of each ASV from the abundance of protein k (11). This procedure was implemented using the R Because C NxP contains the association (i.e. the negative binomial regression test-statistic calculated by DESeq2) between the abundances of ASVs and proteins, an SV represents a cross-association profile between these two feature types. Therefore, ASVs or proteins with the largest magnitude projections will have a cross-association profile most similar to that of the SV they most strongly project on. The most positively projecting ASVs will be strongly associated with the most positively projecting proteins and negatively associated with the most negatively projecting proteins. Similarly, the most negatively projecting ASVs will be strongly associated with the most negatively projecting proteins and negatively associated with the most positively projecting proteins. Rank-ordering features by their projections and choosing the top most positively and negatively projecting features (20 in each direction for ASVs, 50 in each direction for proteins) provided a rational way for identifying coordinated groups of bacteria and proteins whose abundances are tightly coupled.
Because SVD identifies uncorrelated components, each SV represents a unique cross-association profile distinct from that of other SVs. To determine the number of SVs that contain cross-association information above noise, a random-matrix approximation was employed (19). Briefly, C NxP was shuffled along each column to produce a randomized association matrix without any information about the relationship between ASVs and proteins. SVD was performed on the randomized matrix, and the percent variance explained by SV1 was used as the noise threshold; any SV calculated from the SVD of C NxP that explained less variation than SV1 of the shuffled matrix was deemed noise (Figure S8C). Using this method, the first 10 SVs were retained for downstream enrichment analyses.
To identify whether any of the first 10 bacterial SVs were enriched for WLZ-associated taxa, GSEA was performed on the rank-ordered ASV projections along each SV, using the list of 'WLZassociated' taxa (described above) as the reference set. The same procedure was performed for protein projections to determine whether any protein SVs were enriched for WLZ-associated proteins.

DATA DEPOSITION
Bacterial V4-16S rDNA sequences in raw format (prior to post-processing and data analysis) have been deposited at the European Nucleotide Archive under study accession PRJEB38494. Code pertaining to statistical analysis is available from Gitlab. Fecal and plasma biospecimens used for the analyses described were provided to Washington University under a Materials Transfer Agreement with icddr,b.

Effects of MDCF-2 and RUSF on mid-upper arm circumference (MUAC)
Mid-upper arm circumference (MUAC) is another measure of ponderal growth status that complements WLZ (20,21). MUAC improved in children in both groups during the intervention ( Table 2). Over the 4month period between initiation of the intervention and the end of the 1-month follow-up, children who received MDCF-2 had faster improvements in MUAC (-MUAC) compared to those receiving RUSF ( Table 2).

Effects of MDCF-2 and RUSF on co-morbidities
The effects of treatment on four symptoms [cough, rhinorrhea, fever and diarrhea] were quantified (Supplementary Methods). The prevalence of cough and runny-nose reported each week was reduced by MDCF-2 compared to RUSF supplementation (Figure S2A,B; Table S6). MCDF-2 administration was also associated with a greater reduction in fever ( Figure S2C, Table S6). In contrast, RUSF was associated with a greater reduction in diarrhea ( Figure S2D, Table S6).

Changes in dietary habits during MDCF-2 and RUSF supplementation
Over the course of the study, there was no statistically significant difference in the mean number of food groups represented in the diets of children in the MDCF-2 and RUSF arms (4.1 versus 3.88, p=0.30).
More than half of the participants in both groups received an acceptable diet: at the end of the 3-month and 61% for minimum acceptable diet (MAD); the corresponding values in the RUSF group were 58%, ONLINE SUPPLEMENT 22 97% and 51%. No significant differences in MAD were observed between the two groups before or after the interventions.
Food frequency questionnaires were administered to mothers of study participants throughout the trial in order to monitor changes in dietary habits in addition to MDCF-2 or RUSF supplementation. Of the food frequency questionnaire items that had quantitative responses (i.e., mother asked to describe the number of times a food was consumed in the past 24 hours), six were significantly altered by treatment type over the course of the intervention; egg, sweet potato, and dairy product consumption were significantly increased, while consumption of commercially available complementary foods (e.g. cereal), legumes and the frequency of breast-feeding were significantly decreased by MDCF-2 compared to RUSF ( Figure S3A-F, Table S4A). Children in the MDCF-2 trial consumed significantly more dairy products and reported a higher minimum dietary diversity compared to children in the RUSF arm ( Table   S4B). None of the items that were significantly different or that changed significantly between the two arms were associated with ponderal growth (generalized linear mixed-effects model; p>0.05; Table S4C).

Associations between food frequency questionnaire responses and ASV abundances
Linear mixed-effects models were used to determine whether changes in dietary habits, as judged by FFQ responses, were associated with the 23 WLZ-associated bacterial taxa. An affirmative answer to item FFQ101 (Are you breastfeeding your child?) was significantly positively associated with Bifidobacterium sp. (q<0.001) and significantly negatively associated with Faecalibacterium prausnitzii abundances (q=0.04) ( Table S16A). The abundance of Ruminococcus gnavus was significantly positively associated with responses to FFQ132 (How would you describe the participant's appetite?), which measured subjective appetite quality with higher scores indicating better appetite. Finally, abundances of F.
prausnitzii showed a significant association with the interaction between FFQ118a (How many times did the participant consume foods made with beans, lentils, peas, corn, ground nuts, or any other legumes in the past week?) and time, indicating that consumption of legumes was associated with a faster accumulation of F. prausnitzii (Table S16B).

Relating features of the plasma proteome to members of the gut microbiota
We determined whether and how features of the plasma proteome co-vary with members of the gut microbiota, especially WLZ-associated taxa. We had previously described cross-correlation singular value decomposition (CC-SVD), a method for relating disparate features in the same individual (8).
However, the distribution of ASV abundances measured in fecal samples from children in this study followed a negative binomial distribution, invalidating the statistical assumptions of CC-SVD. Therefore, we generalized CC-SVD to account for this distributional difference and developed negative binomial SVD (NB-SVD). We performed NB-SVD analysis by first creating an association matrix where each row represents a bacterial taxon, each column represents a plasma protein, and each element of the matrix represents the test-statistic describing how strongly plasma protein k predicts the abundance of taxon j under an Empirical Bayes negative binomial regression model -a 'correlation' equivalent for count-based data ( Figure S8A). SVD was then performed on this association matrix to identify groups of plasma proteins that were 'correlated' with similar sets of bacterial taxa (ASVs). Each singular vector (SV) represents a unique 'association profile' between proteins and ASVs that is distinct from other SVs. NB-SVD revealed that of the ten SVs that carried cross-association information above noise ( Figure S8C), SV8 was the only one that was significantly enriched for WLZ-associated taxa (GSEA p=0.002). Therefore, we focused on the plasma protein-ASV cross-association profile represented by SV8 (see Table S14 for projections of each taxon and protein along the first 10 SVs). The association strength of the top 20 positively projecting ASVs and the top 50 most positively projecting proteins are shown in Figure S9B (see Figure S10 for negatively projecting ASVs and proteins).
The top 20 taxa with positive projections on SV8 included several that were identified as having a statistically significant association with -WLZ (e.g., Bifidobacterium adolescentis, Prevotella copri, an Olsenella sp., and two Blautia sp.; Figure S9B). Remarkably, SV8 + proteins (i.e., those that are positively associated with SV8 + taxa) were significantly enriched for mediators of cartilage development and bone growth; they include SFRP4, COMP, THBS4, ROBO2, and IGF1 (discussed in Results), as well as collagen type VI α-3 chain (COL6A3), a key regulator of skeletal muscle development and bone density (22,23) (Figure S9A,B, Table S14B, Table S15). Additionally, the SV8 + proteins were significantly enriched for members of the set of 70 WLZ-associated proteins (GSEA p<0.001). In contrast, SV8proteins (i.e., those that are negatively associated with SV8 + taxa) were significantly enriched for mediators of acute phase response, interleukin-6 (IL-6) activation, fatty acid oxidation, and bone resorption ( Figure S9A,B, Figure S10, Table S15). The top 20 SV8taxa included several Bacteroides sp., Campylobacter sp., and the Bifidobacterium sp. that was significantly negatively associated with β-WLZ. These bacteria were negatively associated with SV8 + proteins involved in bone growth and positively associated with SV8proteins related to inflammation, fatty acid oxidation, and bone resorption ( Figure S9A, Figure S10, see Table S15 for the proteins contributing most to each pathway). The results provided by NB-SVD analysis reveal that the abundances of proteins involved in bone growth and inflammation are coupled to the representation of WLZ-associated taxa, providing further evidence of potential mechanisms by which components of the gut microbiota can operate to regulate growth.

MDCF-2 responsiveness and durability of response
To further characterize mechanisms underlying the ponderal growth response to MDCF-2, we divided the cohort of children given MDCF-2 into upper and lower quartiles based on their ponderal growth rates (β-WLZ) over the course of the 3-month intervention (n=15 children/group). Those in the upper-quartile were significantly more wasted at baseline (p=0.008, two-sided Student's t-test), but after the first month of intervention showed complete catch-up growth to the lower-quartile responders (p=0.82; Figure S6A).
By the end of the 3-month intervention, these children had significantly higher WLZ scores than those in the lower-quartile (p<0.0001 two-sided Student's t-test), suggesting that the differences in growth rates were not simply due to regression toward the mean (Figure S6A).

ONLINE SUPPLEMENT 25
Comparison of the plasma proteomes of the two groups revealed that at baseline, those in the upper-quartile had higher levels of proteins involved in anti-viral immune activation including interferon α-1 (IFNA1), interferon λ-2 (IFNL2), IL-1β, IL-6, and CXCL9, and to a lesser degree, protein mediators of antimicrobial humoral immune responses ( Figure S6B, Table S10A, Table S11A). Conversely, they had lower baseline levels of proteins belonging the set of 70 'WLZ-positively associated' proteins (e.g., COMP, COL6A3, THBS3, SLITRK3, SLITRK5, and LEP) as well as other proteins assigned to the GO terms 'bone growth' and 'axonogenesis', compared to children in the lower-quartile ( Figure S6B, Table   S10A, Table S11A). These results indicate that children with a pro-inflammatory, growth mediatordepleted state at baseline exhibit the most rapid increases in WLZ during MDCF-2 treatment.
After one month of supplementation, proteins positively associated with WLZ increased, while those involved in anti-viral defense and antimicrobial immune activation decreased more in children in the upper-compared to lower-quartile ( Figure S6B, Table S10B, Table S11B). Interestingly, proteins associated with amino acid catabolism and fatty acid oxidation increased significantly more after one month of MDCF-2 intervention in those in the upper-quartile compared to the lower-quartile, but showed no differences at the end of treatment ( Figure S6B, Table S10B, Table S11B Table S10C, Table S11C). The inhibitory IGF binding protein IGFBP-2, and growth factor differentiation factor 15 (GDF15), which is associated with anorexia and lipolytic biomarkers in children with severe acute malnutrition (9), were significantly decreased in the upper quartile of MDCF-2 responders ( Table S10C). Levels of proteins related to axonogenesis and nervous system development were also significantly increased more in the upper-quartile responders (GSEA, q<0.1; they include cellular retinoic acid-binding protein 2 (CRABP2), a facilitator of the conversion of dietary carotenoids to Vitamin A (24), SLITRK5, NTRK2, and the axon guidance receptor UNC5B (Table S10C, Table S11C).
Proteins involved in antimicrobial humoral immune response exhibited significantly greater decreases after 3 months of MDCF-2 supplementation in children with upper-compared to the lowerquartile of β-WLZ responses ( Figure S6B, Table S10C, Table S11C (Table S10C).
An analysis of changes in the microbiota after three months of MDCF-2 treatment revealed that children in both the upper-and lower-quartiles of β-WLZ response had significant increases in the abundances of the set of 21 positively WLZ-associated taxa (pupper<0.001, plower=0.01, GSEA; Figure   S13A, Table S20). A direct comparison between the two groups of children revealed that those in the upper-quartile of β-WLZ response exhibited larger changes in the set of 23 taxa whose abundances were significantly associated with WLZ (p=0.04, GSEA).  indicate the statistical significance of the association between FFQ response and WLZ. See Table S4 for unadjusted P-values which indicate the interaction between treatment and weeks since starting the intervention on FFQ response.       Table S1. Nutritional composition of MDCF-2 and RUSF.      Table 2 were computed using a standard linear mixed-effects model (LMM) and are compared against a variation of the LMM approach that is robust to outliers (Robust LMM).
(B) Intent-to-treat analysis. † Values represent the mean ± SD. ‡ Linear model predicting anthropometry at the start of the intervention as a function of treatment group, controlling for baseline age, gender and any illness 7 days prior to starting the intervention. β indicates the mean difference in anthropometry between participants who were assigned to the MDCF-2 and RUSF arms at the start of the intervention. CI, confidence interval. § Mixed effects linear model predicting anthropometry as a function of weeks since starting nutritional supplementation, controlling for the main effects of baseline age, gender, any illness 7 days prior to starting the intervention, and a random intercept for each participant. β indicates the growth rate in unit (anthropometric measure) per week.
¤ Mixed effects linear model predicting anthropometry as a function of the interaction between treatment group and weeks since starting nutritional supplementation, controlling for the main effects of baseline age, gender, any illness 7 days prior to starting the intervention, weeks in the intervention, treatment ONLINE SUPPLEMENT 47 group, and a random intercept for each participant. β indicates the interaction between treatment and growth rate in unit/week (a positive value indicates a faster growth rate in children receiving MDCF-2). Table S6. Effects of supplementation on co-morbidity and illness CI: confidence interval. † Generalized mixed effects linear model predicting prevalence of co-morbidity (binomial presence or absence) as a function of weeks since starting nutritional supplementation, controlling for a random intercept for each participant. β indicates the log-odds of reporting a co-morbidity per week; positive indicates increasing likelihood while negative indicates decreasing likelihood during intervention. ‡ Generalized mixed effects linear model predicting prevalence of co-morbidity (binomial presence or absence) as a function of the interaction between treatment group and weeks since starting nutritional supplementation, controlling for a random intercept for each participant. β indicates the increased/decreased weekly likelihood of reporting a co-morbidity in the MDCF-2 arm over the RUSF arm; positive indicates a higher likelihood while negative indicates a lower likelihood of reporting a comorbidity in children receiving MDCF-2 compared to those receiving RUSF. Table S7. Correlations between changes in protein abundances and ponderal growth. Table S8. Gene set enrichment analysis of WLZ-associated proteins.