Two commercial, freestall dairies (Farm 1, Farm 2) in northern Colorado and representing a combined population of 2350 lactating cattle participated in this pilot study. Both dairies consisted of predominantly Holstein Friesian cows fed total mixed rations (TMR) formulated by the same nutritionist and with nutrient contents similar to one another. Forage was predominantly corn silage and alfalfa hay with additional carbohydrates provided in the form of hominy, brewer’s grain, distiller’s grain and wheat middlings. Each TMR was supplemented with vitamin and trace minerals designed for high-producing cows and Diamond V original line XPC supplement (Cedar Rapids, IA). Early lactation cows were targeted for the study, as they are a cohort at higher risk for O157 shedding, and this enabled our detection of both shedding and nonshedding individuals (Mechie et al. 1997; Venegas-Vargas et al. 2016).

Our sampling design was employed to assess the change in microbial communities between all O157-positive and -negative samples, and between shedding events and shedding patterns of individual cows. On sample day 1, all cows within the first 21 days postpartum on each dairy (n = 74) were sampled by obtaining >10 g faeces via rectal palpation. Samples were kept on ice prior to laboratory characterization. On-farm record systems were used to gather animal life-history features: lactation number (parity), days in milk (DIM), disease during current lactation and disease treatments during current lactation (Dairy Comp 305™, Valley Agricultural Software, Tulare, CA; DHI-Plus, DHI Computing Service Inc., Provo, UT). Diseases recorded included retained placenta, mastitis, metritis, fever of unknown origin, pneumonia, enteric disease, dystocia (including severity), ketosis and lameness. Recorded treatments included penicillin, oxytetracycline, ceftiofur, flunixin meglumine, drench (oral electrolytes) and propylene glycol.

Laboratory enrichment and latex agglutination procedures described below were used to identify bacterial isolates containing the rfb (O157) gene in day 1 faecal samples. Based on those preliminary results, 10 cows that shed O157 on day 1, and 10 cows that did not, were selected per farm and tested in the same fashion for five consecutive days (n = 40 cows, n = 200 samples).

O157 isolation was performed via selective enrichment and detection ‘gold standard’ procedures with slight modification, as described previously (Stenkamp-Strahm et al. 2017a,b). Briefly, samples were mixed 1 : 10 in buffered peptone water (BPW) for both enrichment and initial direct plating. One hundred microlitre was spread plated on sorbitol MacConkey agar with BCIG (Oxoid Diagnostic Reagents, Basingstoke, Hampshire, UK) containing 1·25 mg potassium tellurite and 0·025 mg cefixime (CT-SMAC-BCIG; HiMedia Laboratories, Mumbai, India). These direct plates were incubated at 37°C for 24 h (March and Ratnam 1986). As pathogenic O157 has been known to adapt a sorbitol fermenting phenotype within 24 h, ‘suspect’ O157 colonies seen on plates throughout experiments were deemed as those with straw, grey, pink–grey or too small/difficult to characterize colony coloration (Schmidt et al. 1999; Ayaz et al. 2014).

Direct plates containing ≥100 suspect colonies after incubation were chosen for latex agglutination. Three to 15 colonies per plate were tested for O157 by agglutination using an E. coli O157 latex kit, following manufacturer’s instructions (Oxoid Diagnostic Reagents). Positive colonies were enriched in BPW for 6 h and stored at −80°C in 10% sterile glycerol. For PCR experiments, 10 μl of thawed isolates were centrifuged at 5000 g for 5 min and re-suspended in 30 μl molecular grade water. A volume of 5 μl re-suspended culture template was placed into Qiagen Multiplex PCR Plus Kit reactions, according to the manufacturer’s instructions (Qiagen, Venlo, the Netherlands). In brief, each 25-μl PCR reaction consisted of 12·5 μl master mix, 2·5 μl primer mix containing 0·2 μmol l⁻¹ each primer, 5 μl molecular grade water and 5 μl culture template. The thermal cycling conditions consisted of an initial incubation at 95°C for 5 min to activate the polymerase, followed by 40 cycles of amplification with denaturation at 95°C for 30 s, annealing at 57°C for 1 min and 30 s and extension at 72°C for 30 s, ending with a final extension at 68°C for 10 min. Thermocycling was performed using a Bio-Rad S1000 Thermal Cycler (Bio-Rad, Sydney, Australia). PCR products were analysed by agarose gel electrophoresis using a 2% agarose gel (Lonza Group Ltd., Basel, Switzerland).

The faecal dilution remaining after direct plating was enriched for 6 h at 37°C. Enriched samples not confirmed as O157 positive through direct plating were subjected to immunomagnetic separation (IMS) using Dynabeads anti-E. coli O157 and a BeadRetriever System (Life Technologies, Oslo, Norway). IMS samples were subsequently plated onto CT-SMAC-BCIG and incubated for 24 h at 37°C. Suspect colonies were confirmed by latex agglutination and PCR targeting O157 rfb (O-antigen) gene (Wang et al. 2002). All rfb-positive preliminary isolates were subsequently PCR tested for stx1, stx2 (shiga toxin genes) and eaeA (a variant of the eae intimin gene) using the same PCR protocol outlined above (primers; Paton and Paton 1998). O157 were not enumerated after IMS detection and PCR confirmation.

Samples were deemed to be enterohaemorrhagic E. coli (EHEC) O157 positive when isolates contained rfb, eaeA and any stx genes. Samples were deemed to be atypical enteropathogenic E. coli (aEPEC) O157 positive when isolates contained rfb and eaeA genes. For subsequent statistical analysis, all pathogenic strains (EHEC and aEPEC) were considered together and are hereafter referred to as ‘O157’.

DNA library preparation and 16S rRNA sequencing were performed at the Argonne National Laboratory and followed Earth Microbiome Project suggested protocols (Gilbert et al. 2014; www.earthmicrobiome.org, accessed August 2015). Briefly, genomic DNA was extracted from faecal samples stored at −80°C using the PowerSoil DNA Isolation Kit (MoBio/Qiagen, Carlsbad, CA). To support pooling of all collected samples during a paired-end 2 × 150-base pair Illumina sequencing run, the amplification primer set contained nine extra bases in the adapter region of the forward primer and a 12-base Golay barcode sequence in the reverse amplification primer (Caporaso et al. 2012). For amplification of the V4 hypervariable region of the 16S rRNA gene, 515F (5′-GTGYCAGCMGCCGCGGTAA-3′) and 806R (5′-GGACTACNVGGGTWTCTAAT-3′) primers with the defined barcodes and Illumina flow cell adapter sequences were used for amplification. Each 25-μl PCR reaction contained 12 μl of certified DNA-free water (MoBio/Qiagen), 10 μl of 5-Prime HotMasterMix (1×, Quanta Biosciences, Beverly, MA), 1 μl of forward primer (5 μmol l⁻¹ concentration, 200 pmol l⁻¹ final), 1 μl Golay Barcode Tagged Reverse Primer (5 μmol l⁻¹ concentration, 200 pmol l⁻¹ final) and 1 μl of genomic DNA. The PCR conditions were as follows: 94°C for 3 min, 35 cycles of 94°C for 45 s, 50°C for 60 s, 72°C for 90 s and a final single extension of 10 min at 72°C. Following PCR, amplicons were quantified using a Quant-iT™ PicoGreen^® dsDNA Assay Kit (Invitrogen, Waltham, MA). Based on different quantification values, volumes of each sample were pooled to achieve equal representation. Pools were cleaned using the UltraClean^® PCR Clean-Up Kit (MoBio/Qiagen) and quantified de novo. Pool molarity was determined, and the pool was then diluted to 2 nmol l⁻¹, before denaturing with NaOH. The sample was then repeat diluted to a final concentration of 2 pmol l⁻¹. A 30% PhiX spike was added, prior to loading on an Illumina HiSeq sequencer. Samples were sequenced using a 300-cycle V2 reagent cartridge (Illumina, San Diego, CA).

Sequence read quality was analysed via FastQC. Raw fastq files were demultiplexed with a maximum barcode error of 0 using default methods in QIIME (Caporaso et al. 2010b). Reads not assigned via barcode were removed. QIIME was then used to preprocess reads to OTUs. During this approach, reads were clustered using USEARCH, and chimeras were removed using the UCHIME algorithm (ver. 9.0, Edgar 2010). Reads were aligned to the Greengenes core alignment with 97% sequence identity using PyNAST (DeSantis et al. 2006; Caporaso et al. 2010a). Taxonomy was assigned using the Ribosomal Database Project 2.2 classifier and the Greengenes 13.8 taxonomy reference (Wang et al. 2007; McDonald et al. 2012). A maximum likelihood approximated tree was built using FastTree 2.1.9 (Price et al. 2010).

During this preprocessing, an open reference approach was utilized with the packages listed above (Rideout et al. 2014). The open reference algorithm allowed for an initial clustering and closed reference sequence alignment, followed by de novo clustering of reads that failed to align. During the second clustering, representative sequences were created using the centroid of those that failed to align, and these were used during a closed reference picking process. OTUs that were only represented by a single read were discarded.

Nonmetric multi-dimensional scaling with a Bray–Curtis distance was used to ordinate OTUs from O157-positive and -negative samples in space. An unweighted unifrac distance was used to perform a principal coordinate analysis of OTUs from O157-positive and -negative samples. The relative abundance of taxa within each sample was measured using the R package Phyloseq (McMurdie and Holmes 2013).

O157 categories were used to classify the presence of O157 at either the sample or the cow level. These categories are listed in Table 1 and include for the sample level: pathotype, day prior vs day of and day of vs day after. ‘Pathotype’ classified samples as containing aEPEC/EHEC, or no pathogenic O157. For ‘day prior vs day of’ and ‘day of vs day after’ samples were paired to compare either the day prior to shedding or the day after shedding to the day of the shedding event in that individual. At the cow level, categories included pattern and ever vs never. ‘Pattern’ was a category that defined cows as multi-day (≥2 days shed), intermittent (1 day shed) or never shedders (0 days shed). ‘Ever vs never’ classified cows as having shed or not shed pathogenic O157 at least once during the study period. For ‘pattern’ and ‘ever vs never’, each cow was classified the same across days.

Using these samples and cow defined categories, two different statistical approaches were used to analyse data from microbial count tables: epidemiologic modelling of microbial alpha diversity and measurement of taxa differential abundance. These approaches are described, in turn, below. As taxa within microbial communities are defined at the sample level, for cow-level categories (pattern, ever vs never) only the epidemiologic modelling of diversity measures was used. For matched samples (day of vs day after, day prior vs day of), only measuring differential abundance of taxa was used.

Relative abundance of OTUs aggregated to different taxa levels were quantified using Phyloseq. Alpha-diversity metrics were quantified using Phyloseq and Vegan, including observed richness (R), Shannon’s index (S) and Pielou’s measure of species evenness ((S)/Log(R)) (McMurdie and Holmes 2013; Oksanen et al. 2017). Previous studies have shown that rarefying 16S rRNA count data prior to analysis is necessary to control for sequencing depth, but can negatively influence results (McMurdie and Holmes 2014). For analyses, data were normalized using cumulative sum scaling (CSS) in metagenomeSeq to correct for this bias (Paulson et al. 2013).

Random effects, logistic and multinomial modelling were completed to measure the association between O157 categories and diversity measures. In brief, for the category ‘pathotype’, random effects regression was used with sample-level diversity measures, and a random effect for cow. For the category ‘ever vs never’, a logistic regression model was used with diversity measures averaged across days for each individual cow. For the ‘pattern’ analysis, multinomial models were used with diversity measures averaged across days for each individual cow. Separate animal-level covariates that modify the risk of shedding in dairy cows have been reported previously (Menrath et al. 2010; Venegas-Vargas et al. 2016; Stenkamp-Strahm et al. 2017a). As these variables may cause, confound or mediate the association between alpha diversity and O157 shedding status, a directed acyclic graph was used to evaluate the role of these factors in the association between O157 categories and diversity (Fig. S1). The distribution of the variables parity, days in milk, disease status, farm and treatment were measured across O157 categories. These variables were assessed individually for associations with diversity, and those that met the screening criteria (P ≤ 0·2) were included in the O157 models of diversity. Odds ratios and confidence intervals were calculated from model coefficients and standard errors using the interquartile range (IQR) of diversity values.

Using metagenomeSeq, taxa were aggregated to the family, genus and species levels (Paulson et al. 2013). Aggregated tables were filtered to reflect taxa present in at least 25–30% of samples, and differential abundance at each taxonomic level was measured in turn. The ideal methodology used to test for differential abundance between nonpaired 16S rRNA count tables is not agreed upon in the literature (McMurdie and Holmes 2014; Thorsen et al. 2016; Weiss et al. 2017); therefore, two modelling approaches were used to assess changes in abundance for the ‘pathotype’ metric. In metagenomesSeq, zero-inflated Gaussian (ZIG) regression was used to evaluate changes between CSS normalized count tables (Paulson et al. 2013). In DESeq2, negative binomial regression was used to evaluate the same changes in count tables internally normalized by calculating geometric means and median count ratios (Love et al. 2014). For matched samples (‘day prior vs day of’ and ‘day of vs day after’), aggregated count tables were normalized with CSS and compared using a Wilcoxon rank-sum test on table differences. After adjusting for multiple testing using the Benjamini and Hochberg (1995) correction, a P-value cutoff of <0·1 was used to detect differences. This P-value was selected due to the preliminary nature of the project, and relatively low power.

All procedures were approved by Colorado State University’s Institutional Animal Care and Use Committee. All statistical analyses were performed in R ver. 3.3.2 or later (R Core Team 2016).

The study population consisted of 40 cows (20 per farm) that were sampled for five consecutive days. Of these individuals, 14 (35%) shed EHEC, 4 (10%) shed aEPEC and 31 (78%) shed rfb isolates without virulence genes. Samples taken on day 1 of the study were preliminarily assessed for rfb (38% shed rfb+ strains, 62% did not), prior to continued sampling of O157-positive and -negative cows on days 2–5. PCR characterization of isolates revealed many fewer cows harbouring O157 with virulence genes (EHEC and aEPEC) than those identified as rfb positive on day 1. Of the 200 study samples, 20 (10%) contained EHEC isolates, 4 (2%) contained aEPEC isolates and 67 (34%) contained rfb isolates that did not have virulence genes.

During 16S library preparation of faecal sample DNA, four samples either did not amplify well during PCR or did not sequence adequately, and were omitted from further analysis. Of these samples, three were from Farm 1 and one was from Farm 2. None of the omitted samples were identified as having aEPEC or EHEC. After sequencing and prior to OTU analysis in QIIME, all Illumina DNA sequence reads were seen to have an average quality PHRED score ≥30 via FASTQC. The data set revealed a total count of 12 225 598 reads, and 159 354 unique OTUs. Mean sequence read count per sample was 39 059.

Bacterial communities and member taxa were strikingly similar among cow samples, and did not cluster in ordination space based on O157 presence or absence (Fig. 1). Measures related to animal life-history characteristics and other shedding categories also did not show clustering of samples in space. Relative abundances of taxa at the family and phylum levels further revealed that communities with and without O157 were similar (Fig. 1a,b). Differences, where noted, were only slight. O157-positive and -negative bacterial communities contained 62 and 56% Firmicutes, 35 and 31% Bacteroidetes and 2·2 and 2·4% Tenericutes respectively. Actinobacteria was slightly higher in non-O157 samples (2·5 vs 1·9% in O157-positive samples) as were Spirochaetes (1·6 vs 1·0% in O157-positive samples). Proteobacterial percentages were 4·2% in non-O157 samples and 4·6% in O157 containing samples.

When looking at community composition at the family level, several taxa were seen to be slightly higher in abundance in non-O157 samples compared to O157 containing samples: Paraprevotellaceae (5·4 vs 4·7% in O157 positive), Bacteroidaceae (4·7 vs 3·6% in O157 positive), Bifidobacteraceae (2·4 vs 1·7% in O157 positive) and Spirochaetaceae (1·5 vs 1·0% in O157 positive). Meanwhile, Lachnospiraceae (6·0% in non-O157 vs 7·1% in O157 positive), Erysipelotrichaceae (0·9% in non-O157 vs 1·4% in O157 positive) and Peptostreptococcaceae (2·2% in non-O157 vs 2·8% in O157 positive) were slightly higher in O157-positive samples.

Five different variables described previously to influence a dairy cow’s risk of shedding pathogenic O157 (farm, DIM, parity, disease status and treatment status) may confound or mediate the association between microbial diversity and O157. Cows in the current study experienced ketosis (n = 5), metritis (n = 4), mastitis (n = 2), retained placenta (n = 2), lameness (n = 2), enteric disease (n = 2), fever of unknown origin (n = 1) and pneumonia (n = 1). The variable ‘disease’ was collapsed into a binary yes/no category due to sparseness of most disease types (no disease: n = 21; disease: n = 19). Treatments given to cows in the current study included penicillin (n = 3), oxytetracycline (n = 5), ceftiofur (n = 7), flunixin meglumine (n = 2), drench (n = 3) and propylene glycol (n = 4). Due to sparseness of treatment types, the variable ‘treatment’ was collapsed into a categorical variable with three levels. These levels represented cows that were never treated (n = 23), cows that were ever treated with any antibiotic (penicillin, oxytetracycline, ceftiofur; n = 11) and cows that were not treated with an antibiotic but were treated with a nonantibiotic agent (flunixin meglumine, propylene glycol, drench; n = 6). Enrolled cattle were parity 1 (n = 19), parity 2 (n = 7), parity 3 (n = 8), parity 4 (n = 3), parity 5 (n = 1) and parity 6 (n = 2). The variable parity was also collapsed into three categories due to sparseness of cells and biological reasoning (parity 1, n = 19; parity 2, n = 7; parity ≥3, n = 14).

All of the variables known to influence shedding of O157 were grouped by O157 categories, described in Table 2. Of these variables, farm, parity and treatment differed by O157 ‘pathotype’ status (Table 2; P ≤ 0·08). Treatment and parity had variable distributions across ‘pattern’ categories (P = 0·03 and P = 0·33 respectively) and treatment varied between categories of ‘ever vs never’ (P = 0·13).

Richness, evenness and Shannon’s diversity were computed for all samples using normalized 16S rRNA read counts. The IQR of these values by O157 categories are presented in Fig. 2. In multinomial models, cows that were classified as multi-day shedders were seen to have lower average richness (Fig. 3; OR = 0·51; 95% CI: 0·41–0·64) compared to cows that never shed during the study period. Odds ratios and confidence intervals were nonsignificant for other diversity models comparing O157 categories.

Variables that may confound or mediate the association between O157 categories and diversity measures (Fig. S1, Table 2) were assessed for individual associations with richness, Shannon’s and evenness values (Table S1). Variables with P < 0·2 model results were considered further in influencing the regression outcomes between O157 category and diversity seen in Fig. 3. These variables included parity, farm and treatment, and were the same variables seen to be different when distributed by O157 shedding metrics (Table 2).

These three variables were included in models of O157 and alpha diversity. Each variable was first assessed by itself with an O157 metric, and then in turn with other variables. For most models, the nonsignificant associations seen in the original crude values were not altered with adjustment (Table S2). Cows with a multi-day shedding pattern still had lower average richness compared to never shedding cows, regardless of adjusting with any or all variables. However, when the variable parity was added to pattern models with any combination of farm, treatment or no other confounders, cows with an intermittent shedding pattern were seen to have significantly higher average richness compared to never shedding cows (Table S2; parity added alone: OR = 1·23; 95% CI: 1·07–1·40). Due to the discrepant nature of these conclusions, raw values of richness for cows of all parity levels were evaluated. A single intermittently shedding, parity 1 cow had much higher richness values than nearly all other cows in the study, including on the day she shed O157. The average richness value for this cow on days 1–5 was 7644, compared to average values that fell in the range of 3000–5000 for other cows. Given the relatively small sample size in this pilot project, it was hypothesized that this individual was driving the increase in average richness seen between intermittently and never shedding cows, when controlling for parity. A sensitivity analysis was performed removing this outlier (Fig. 4). When re-running models of richness and O157 pattern adjusting for confounders, intermittently shedding cows were seen to have significantly lower average richness (OR = 0·568; 95% CI: 0·49–0·66) compared to never shedding cows. Similar to results when the single high-richness cow was included, multi-day shedding cows still had significantly lower average richness (OR = 0·415; 95% CI: 0·33–0·52) compared to never shedding cows.

To examine taxa that may be driving associations seen during diversity modelling, differential abundance testing was performed using sample-level O157 metrics. Both ZIG and negative binomial regression were used to evaluate differences in taxa defined at the species, genus and family levels. When aggregating taxa to these different taxonomic levels and filtering to those seen in 25% of samples, there were 195 unique species, 590 unique genus and 290 unique family-level taxa for analysis. Comparing samples with O157 to those without O157 and controlling for ‘cow’, one genus (Moryella; P = 0·04) was seen to be higher in abundance and one species (Bacillus coagulans; P = 0·09) was seen to be lower in abundance using ZIG models (Fig. 5a). Negative binomial models did not show similar conclusions; there were no family-, genus- or species-level taxa seen to have differential abundance between O157-positive and -negative samples (Fig. 5b).

We compared day prior vs shedding day (n = 12 pairs), as well as shedding day vs day after shedding (n = 10 pairs), by running Wilcoxon rank-sum tests on the differences in family, genus and species taxa abundance. After correcting for multiple testing, no specific taxa were observed to differ in days prior or after O157 shedding, when compared to the O157 shedding day.