Antibiotics and the developing intestinal microbiome, metabolome and inflammatory environment: a randomized trial of preterm infants

Antibiotic use in neonates can have detrimental effects on the developing gut microbiome, increasing the risk of morbidity. A majority of preterm neonates receive antibiotics after birth without clear evidence to guide this practice. Here microbiome, metabolomic, and immune marker results from the Routine Early Antibiotic use in SymptOmatic preterm Neonates (REASON) study are presented. The REASON study is the first trial to randomize symptomatic preterm neonates to receive or not receive antibiotics in the first 48 hours after birth. Using 16S rRNA sequencing of stool samples collected longitudinally for 91 neonates, the effect of such antibiotic use on microbiome diversity is assessed. The results illustrate that type of nutrition shapes the early infant gut microbiome. By integrating data for the gut microbiome, stool metabolites, stool immune markers, and inferred metabolic pathways, an association was discovered between Veillonella and the neurotransmitter gamma-aminobutyric acid (GABA). These results suggest early antibiotic use may impact the gut-brain axis with the potential for consequences in early life development, a finding that needs to be validated in a larger cohort.


7
Bray-Curtis and Jaccard distance indices were used to assess community structure, taking into 126 account quantitative ASV abundance and qualitative ASV presence/absence information, 127 respectively. Principle coordinates analysis (PCoA) did not reveal any immediately obvious 128 clustering differences between groups for either metric (Fig. 2), which may suggest little or no 129 persistent effect of antibiotic use beyond 48 hours after birth. Beta dispersion was not

greater spread in variance) between group A infants and infants in 136
other groups, which could be explained by group A infants often receiving antibiotics beyond 48 137 hours after birth. However, when the non-parametric permutational analysis of variance 138 (PERMANOVA) test was applied to each timepoint across groups, there were no significant 139 differences in Bray-Curtis or Jaccard distances among all groups at any given corrected GA 140 timepoint (Fig. 2). 141

Feeding patterns drive changes in gut diversity and bacterial load 142
For preterm infants, diet generally consists of mother's breast milk (MBM), pasteurized 143 donor breast milk (DBM), formula, or some combination of these sources. Some infants also 144 experienced periods of no enteral feeding (NPO: nil per os). To investigate effects of feeding 145 while still considering effects from antibiotic use, feeding types were compared within each 8 respective analysis group. In addition, for purposes of comparing feeding types at each corrected 147 GA timepoint, feeding types with only a single sample at each timepoint (n=1) were removed. 148 This reduced the total number of stool samples from 522 to 461. The number of samples in each 149 group at each timepoint, and also by feeding type, is summarized in Supplementary Table S3. 150 Using the calculated alpha diversity metrics described previously, feeding type was significantly 151 different in bacterial richness only at corrected GA week 32 in group A infants 152 p= 0.0069), where samples collected during feeding with all or partial mother's milk tended to 153 have higher bacterial richness (Fig. 3A). Furthermore, Shannon diversity was significantly lower 154 in infants not fed orally at corrected GA week 36 in group A infants (p=0.042) (Fig. 3B). The 155 log10-transformed number of 16S rRNA copies were not significant at any timepoint in any 156 group, although notably feeding with MBM alone typically had higher 16S rRNA copies 157 compared with formula in all groups except for group A, which might be explained by continued 158 antibiotic use beyond 48 hours (Fig. 3C). 159 Beta diversity between feeding types within groups was likewise only significant at few 160 specific timepoints. For group A infants, Bray-Curtis distances between formula and MBM 161 feeding were significantly different at corrected GA week 34 using PERMANOVA (R2=0.093, 162  Applying LME modelling by feeding type according to analysis group allows the 168 observation of feeding effect over time, focusing again on the 12 weeks of corrected GA after 169 removal of feeding type singletons at each timepoint (Fig. 4). Perhaps not surprisingly, among all 170 groups, periods of NPO led to a lower trend in Shannon diversity over time (p=0.003) (Fig. 4G). 171 While bacterial richness appeared to trend lower, the trend was not significant (p=0.341) (Fig.  172 4A). For group A infants, which received antibiotics in 48 hours after birth and often beyond, 173 MBM was associated with a slight increase in richness (p=0.009) (Fig. 4B), and periods of NPO 174 led to a lower trend in Shannon diversity (p=0.031) (Fig. 4H). in Group B infants, who never 175 received antibiotics and tended to be older, larger and healthier, all feeding types including 176 formula (p=0.004), MBM (p<0.001) and MBM+formula (p=0.018) led to increasing trends in 177 Shannon diversity (Fig. 4I)

Gut microbial community development is highly variable and unique to each infant 190
Although the infants in this study were analyzed among 5 groups, each infant's stay in 191 the NICU is highly personalized, by type, frequency, number or length of antibiotic use, type and The more subtle effects of antibiotic use and feeding patterns become visually apparent 206 by taking this individualized approach. To illustrate, infant 5 was exclusively fed mother's milk 207 from day 14 through day 71 post birth. At day 31, antibiotics vancomycin and piperacillin were 208 administered (Supplementary Figure S1). During treatment, Veillonella was entirely removed 209 from the stool microbiome, falling from 3.23E+06 cells per gram (73%) from the pre-treatment 210 time point to an undetectable level while antibiotics were used. At day 47, 9 days after antibiotic 211 treatment ceased, Veillonella again dominated the stool at 3.83E+06 cells per gram (80%), 212 almost a complete replacement of levels before treatment. Also, the proportions of the other 2 213 genera found in the stool, Escherichia and an unclassified Enterobacteriaceae spp., were nearly 214 identical after treatment and continuation of mother's milk as before treatment (pre-treatment: 215 Escherichia -21%, Enterobacteriaceae spp. -4.7%; post-treatment: Escherichia -15%, 216 Enterobacteriaceae spp. -3.3%). Thus, either that mother's milk effectively restored the stool 217 microbiome to its pre-treatment state or this effect occurred due to removal of antibiotic selective 218 pressure, or both. A similar effect can be seen in infant 12 between 25 and 50 DOL where the 219 microbiome is restored post-antibiotics. In this case, the restoration is observed with MBM, 220 DBM, and formula. In some cases, antibiotic use had no effect on the microbiome composition 221 (e.g. infant 42, 84), suggesting the presence of resistance mechanisms in the dominant gut 222 microbes (in these 2 cases, members of Enterobacteriaceae). In fact, Enterobacteriaceae presence 223 followed administration of ampicillin and gentamicin, the 2 most commonly prescribed 224 antibiotics immediately after birth. This occurred in 24 of the 91 infants. Other times antibiotic 225 use appears to dramatically and irreversibly change microbiome composition (e.g. infant 25). 226

Bacterial genera correlate with stool metabolites and inferred metabolic pathways 227
In addition to 16S rRNA profiling, 90 stool samples from 10 infants were analyzed for 228 metabolomic profiling ( Table 1). Four of the 5 groups were included in these samples for 229 comparison (group B samples not included). Peak height responses were recorded for 454 230 identifiable metabolites. To determine if gut bacteria were associated with relative concentrations 231 of metabolites in stool, the top 10 most abundant bacterial genera associated with identified 232 metabolites were determined. Repeated measures correlation values were plotted using a 233 heatmap, which indicated numerous significant, positive and negative, associations between 234 bacteria and metabolites (Fig. 6A). Interestingly, Veillonella were positively associated with the 235 neurotransmitter 4-aminobutanoate (GABA) (R = 0.27, p = 0.013) and Veillonella counts were 236 significantly different between groups A and C2 (p =0.0475), C1 and C2 (p=0.029), and C2 and 237 C2Bailed (p=0.042) using the Wilcoxon paired test (Fig. 6C). Also, Veillonella counts were not 238 significantly different between samples of infants that received antibiotics, i.e. A and C1 239 (p=0.57) or C1 and C2Bailed (p=0.17). GABA peak height responses followed similar trends as 240 Veillonella counts, that is, responses were significantly different between groups that received 241 and did not receive antibiotics (A vs. C2, C1 vs. C2, C2 vs. C2Bailed) but not between groups 242 that both received antibiotics (A vs. C1, C1 vs. C2Bailed) (Fig. 6B). 243 Furthermore, using PICRUSt2, functional pathway abundances were inferred based on 244 the rarefied 16S rRNA data29. The Veillonella counts of predicted pathways were strongly 245 correlated with biosynthesis of the GABA precursor L-glutamate (R = 0.88, p = 3.02E-27) 246 (Supplementary Figure S2). Thus, it may be that Veillonella could be at least partially 247 responsible for GABA neurotransmitter production and that this function is negatively impacted 248 by antibiotic use early in life. Alternatively, Veillonella may instead be involved in biosynthesis 249 and export of L-glutamate in the gut, which is then converted to GABA by the host glutamate 250 decarboxylase. However, PICRUSt2 results are based on inferred pathways from reference 251 13 genomes closely related to the 16S rRNA data used here and are, at best, predictions in the 252 absence of functional data specific to this cohort. 253 A negative correlation between Bifidobacterium counts and glycocholic acid was 254 observed (R = -0.39, p = 0.0098), which was also impacted by antibiotic use between groups. In 255 addition, bifidobacteria were negatively associated with other conjugated bile acids including 256 taurocholic (R = -0.22, p = 0.045) and glycocholic acids (R = -0.21, p = 0.048), but positively 257 associated with deconjugated cholic (R = 0.25, p = 0.027) (Fig. 6A). Thus, gut microbiota 258 affected by antibiotic use may be responsible for modification of neuroactive metabolites (i.e. 259 deconjugated bile salts) in addition to production of neurotransmitters. 260

Immune markers in stool correlate with bacterial abundance 261
Antibiotic use was examined for its correlation with inflammatory marker levels in stool. 262 These levels were also correlated with gut bacterial abundances. Twelve immune markers were 263 measured in 110 stool samples across 18 of the first enrolled infants. A summary of immune 264 marker data samples including infants per group and number of samples per infant is given in 265 Table 1. Ten bacterial genera had at least one significant correlation with an immune marker (p < 266 0.05) (Fig. 7A). Significant correlations between the bacterial genera and stool immune markers 267 were classified as either inflammatory or anti-inflammatory based on the known function of the 268 marker (Fig. 7B). Interestingly, Enterococcus counts were negatively correlated with levels of 269 TNF-alpha and macrophage inflammatory protein 1-alpha (MIP1). Citrobacter were positively 270 correlated with MIP1 and IL6 (R = 0.21, p = 3.74E-05), and were significantly higher in group 271 C1 compared to group C2 (p=7.7E-07) and group C2 compared to C2Bailed (p=0.00022) by the 272 correlated with levels of epidermal growth factor (EGF), which was the strongest correlation 274 within the dataset. Escherichia/Shigella counts were highest among group A samples, but not 275 significantly higher compared to other groups (Wilcoxon, p>0.05). 276

Discussion 277
There is an urgent need for evidence supporting or refuting the widespread practice of 278 routine antibiotic use after birth in symptomatic preterm neonates. The REASON study 279 represents a significant step as it is the first randomized controlled trial to test the feasibility of 280 randomizing symptomatic preterm infants to antibiotics versus no antibiotics, evaluating the 281 effect of antibiotic treatment on the developing gut microbiome, metabolome, and inflammatory 282 environment. Our results expand upon previous reports that early routine antibiotic use leads to 283 alterations in the early life gut microbiome, even after discontinuation of antibiotics17,30,31. The 284 results presented here suggest that antibiotic use 48 hours after birth did not tend to have a 285 lasting effect on the development of gut microbiome diversity over time, and that the gut 286 microbiota diversity was recoverable. However, use of antibiotics extending beyond 48 hours 287 after birth often did have significant impacts on the microbiome over time, as evidenced in group 288 A infants compared to the other enrollment groups. The power to detect significant associations 289 in this study was hampered however, mainly because many of the infants randomized to not 290 receive antibiotics were changed to antibiotic administration. Furthermore, there were few 291 infants who were enrolled in group B (an important antibiotic-free control group) and those 292 enrolled in group B had few samples due to short stays in the NICU. A larger multi-center 293 randomized study is needed to validate and expand upon the extended effect of antibiotics on the 294 developing gut microbiome. 295 Our results support the notion that feeding types likely also have a significant influence 296 on gut microbiome richness and diversity, though in this case only at specific timepoints32-34. 297 Exclusive or partial feeding with mother's own milk appeared to have higher bacterial load 298 compared to formula and NPO, though not significantly. This observation is backed by previous 299 evidence that breast milk harbors maternal-originating bacteria, as well as nutritional 300 components (prebiotics) that support bacterial proliferation in the intestinal tract35,36. 301 Interestingly, formula-fed infants had comparable levels of richness and diversity as mother's 302 milk. This supports the idea that mother's milk drives early colonization of a limited set of 303 dominant microbes through nutrient and antimicrobial-mediated selection37-39. Feeding trends 304 over time were able to be assessed for the main feeding types such as MBM, DBM and formula 305 however again the ability to detect meaningful results for rarer feeding types (particularly 306 combinations of sources) and group B infants was hampered by small sample size and will 307 require a larger cohort. 308 Integrating detailed and personalized records of clinical and laboratory data led us to 309 identify overlooked patterns in the data. One such peculiar pattern was that stool samples taken 310 during administration of the anti-fungal fluconazole had lowered copies of 16S rRNA, 311 suggesting lower bacterial load. A previous study reported that fluconazole, though not 312 inherently bactericidal, increased the bactericidal activity of neutrophils40. Immune marker data 313 were collected for 18 of the first enrolled infants, and one or more of those markers, such as 314 calprotectin which is secreted by neutrophils, may help explain this pattern41. However, only 3 of 315 the 18 infants that had immune marker data received fluconazole. Therefore more data are 316 needed to test this hypothesis. Interestingly, counts of Enterococcus were negatively correlated 317 with levels of pro-inflammatory markers such as TNF and MIP1, an odd finding considering 318 Enterococci have been associated with risk for infection in preterm neonates42,43. On the other 319 hand, Citrobacter counts were associated with increased levels of the macrophage chemokine 320 MIP1 and counts were significantly higher in infants randomized to receive or were bailed to 321 which is reduced in preterm infants, is critical for early brain development, and possesses 339 immunomodulatory properties51,52. Antibiotic use was negatively affected the abundance of 340 Veillonella and that Veillonella were positively correlated with GABA concentrations in the gut. 341 Furthermore, Veillonella correlated strongly with the L-glutamine biosynthesis pathway, the 342 precursor to GABA production. Aside from production of neurotransmitters, negative 343 correlations were identified between Bifidobacterium abundance and concentrations of 344 conjugated bile acids, particularly glyco-and taurocholic acid. Conjugated bile acids were also 345 significantly different based on antibiotic use. Bifidobacteria, which were more abundant in 346 infants that did not receive antibiotics, are known to deconjugate bile acids to primary forms 347 including cholic acid53,54. Cholic acid can passively diffuse into the brain where it blocks 348 signaling in the GABAA receptor55. Bifidobacteria may therefore be essential in regulating 349 GABA signaling in the developing brain. These are significant findings, for they suggest routine 350 antibiotic use could be disrupting processes involved in the gut-brain axis and 351 immunomodulatory pathways critical for neonatal and future childhood development. randomized studies with greater infant enrollment will be crucial in our understanding of the 361 effects current neonatal practice has on health which will allow for the reevaluation of practices. 362 Such trials will need to expand on the findings from this pilot study from a multi-omic standpoint 363 to identify direct links between antibiotic-induced dysbiosis and health outcomes. 364

Experimental design, enrollment, and clinical sample and data collection 366
The REASON study was conducted from January 2017 -January 2019 at the University 367 of Florida and was approved by the institutional review board (IRB201501045). This study is 368

Absolute bacterial abundance by copy number correction 412
Absolute bacterial abundance was calculated on a per gram of stool basis by correcting 413 the relative sequencing abundance by the variable number of copies of the 16S rRNA gene in 414 each observed organism. This correction was done using the "Estimate" tool provided as part of 415 the rrnDB copy number database62. Briefly, after rarefying each sample to an even sequencing 416 depth, the ASV sequences were submitted through the rrnDB online portal where they were 417 classified down to the genus level using the RDP classifier version 2.12 and copy number 418 adjusted using rrnDB copy number data version 5.662,63. The copy number adjusted relative 419 abundance for each observed taxon was multiplied by the total number of 16S rRNA copies 420 obtained by qPCR, resulting in the absolute abundance of each taxon per gram of stool. 421

Fecal inflammatory markers 422
Inflammatory markers were analyzed using a combination of multiplex technologies 423 using the Bio-Rad Bio-Plex platform (Bio-Rad, California, USA). The markers evaluated include 424 common markers of intestinal inflammation including calprotectin and S100A12, in addition to 425 other markers such as IL-6, TNF, IL-10 and other cytokines and chemokines that may play a role 426 in inflammatory or anti-inflammatory processes. The data were analyzed using direct 427 comparisons of all infant groups using ANOVA and subsequent individual comparisons. Fecal 428 calprotectin and S100A12 levels were measured using the fCal ELISA kit from BUHLMANN 429 Laboratories AG (Schonenbuch, Switzerland) and the Inflamark S100A12 kit from Cisbio

Metabolomics 439
The infant stool samples were suspended in 400 µl 5 mM ammonium acetate. Mzmine was used to identify features, deisotope, align features and perform gap filling to fill in 460 any features that may have been missed in the first alignment algorithm66. Features were 461 matched with SECIM internal compound database to identify metabolites. All adducts and 462 complexes were identified and removed from the data set prior to statistical analysis. 463

Statistical Analysis 464
The ASV and taxonomy tables resulting from DADA2 were manipulated using the 465 phyloseq R package v1.30.067. Inferred metabolic pathway abundances were determined from 466 the rarefied 16S rRNA data using PICRUSt229. Alpha diversity measures, including the observed 467 number of ASVs and the Shannon diversity index were calculated using the microbiome R 468 package v1.8.0 (https://bioconductor.org/packages/devel/bioc/html/microbiome.html). Box plots 469 (including statistical testing where applicable) were generated using the ggpubr R package v0.2.4 470 (https://github.com/kassambara/ggpubr), which serves as a wrapper for ggplot268. The linear 471 mixed-effects modeling and associated plots were done using the longitudinal plugin "q2-472 longitudinal" offered in Qiime2 v2019.426-28. The biomformat R package (https://biom-473 format.org) was used to convert data in phyloseq format to BIOM format for import into 474 Qiime269. Bray-Curtis and Jaccard distance dissimilarities were calculated using the vegan R 475 package v2.5.6 (https://github.com/vegandevs/vegan) and PCoA plots were made using ggplot2 476 v3.3.068. Individual infant charts were also generated using ggplot2. Non-parametric statistical 477 tests including the Wilcoxon and Kruskal-Wallis tests were used for pairwise and overall 478 comparisons of 3 or more factors, respectively70,71. The permutational analysis of variance 479 (PERMANOVA) test was used in the vegan package to compare overall microbiome 480 dissimilarities between antibiotic use, feeding type, and enrollment groups. P-values were 481 adjusted for false discovery rate (FDR) via the Benjamin-Hochberg method72. Repeated 482 measures correlation values (for non-independent repeated samples for multiple subjects) were 483 calculated using the rmcorr R package73. 484

Data availability 485
The demultiplexed 16S rRNA sequencing data generated in this study is deposited in the 486 NCBI Sequence Read Archive (SRA) under BioProject PRJNA515272. 487

Trial Registration 488
This project is registered at clinicaltrials.gov under the name "Antibiotic 'Dysbiosis' in 489 Preterm Infants" with trial number NCT02784821.