Exploring associations between the teat apex metagenome and Staphylococcus aureus intramammary infections in primiparous cows under organic directives

ABSTRACT

The primary objective of this study was to identify associations between the prepartum teat apex microbiome and the presence of Staphylococcus aureus intramammary infections (IMI) in primiparous cows during the first 5 weeks after calving. We performed a case-control study using shotgun metagenomics of the teat apex and culture-based milk data collected longitudinally from 710 primiparous cows on five organic dairy farms. Cases had higher odds of having S. aureus metagenomic DNA on the teat apex prior to parturition compared to controls (OR = 38.9, 95% CI: 14.84–102.21). Differential abundance analysis confirmed this association, with cases having a 23.8 higher log fold change (LFC) in the abundance of S. aureus in their samples compared to controls. Of the most prevalent microorganisms in controls, those associated with a lower risk of post-calving S. aureus IMI included Microbacterium phage Min 1 (OR = 0.37, 95% CI: 0.25–0.53), Corynebacterium efficiens (OR = 0.53, 95% CI: 0.30–0.94), Kocuria polaris (OR = 0.54, 95% CI: 0.35–0.82), Micrococcus terreus (OR = 0.64, 95% CI: 0.44–0.93), and Dietzia alimentaria (OR = 0.45, 95% CI: 0.26–0.75). Genes encoding for Microcin B17 AMPs were the most prevalent on the teat apex of cases and controls (99.7% in both groups). The predicted abundance of genes encoding for Microcin B17 was also higher in cases compared to controls (LFC 0.26).

IMPORTANCE

Intramammary infections (IMI) caused by Staphylococcus aureus remain an important problem for the dairy industry. The microbiome on the external skin of the teat apex may play a role in mitigating S. aureus IMI risk, in particular the production of antimicrobial peptides (AMPs) by commensal microbes. However, current studies of the teat apex microbiome utilize a 16S approach, which precludes the detection of genomic features such as genes that encode for AMPs. Therefore, further research using a shotgun metagenomic approach is needed to understand what role prepartum teat apex microbiome dynamics play in IMI risk.

INTRODUCTION

Bovine mastitis is an inflammatory disease of the udder caused by microorganisms that overcome the physical and immunological barriers of the mammary gland (1). Despite more than a century’s worth of research into the etiology of this disease, mastitis remains an issue for dairy cows and producers (2). One major pathogen of concern is Staphylococcus aureus, a contagious gram-positive bacterium that results in persistent infections of the mammary gland and that is notoriously difficult to treat—even with the use of antibiotics (3). This makes bovine mastitis caused by S. aureus an important health and welfare issue for dairy cows, and a major financial burden for dairy farmers; therefore, further research is needed to prevent and treat these infections (2, 3).

Perhaps one of the most effective approaches to preventing S. aureus intramammary infections (IMIs) is to maintain a low prevalence of the pathogen within the herd by implementing strict biosecurity protocols (4). To achieve this objective, common practices include culling infected animals, maintaining high udder hygiene, separating infected from non-infected animals, and avoiding the purchase of infected animals (4–6). Preventive and treatment practices such as the use of antibiotics and internal teat sealants also provide an effective means for treating existing infections and preventing new ones during the dry period, i.e., when the mammary gland is not producing milk (7, 8). However, both antibiotics and internal teat sealants cannot be used in dairy cows from organic-certified milk production systems (9). Indeed, restrictions on antibiotic use have been proposed as a contributing factor to the higher prevalence of S. aureus infections observed in some organic farms compared to conventional farms (9, 10). This is especially problematic for nulliparous dairy cows because they often calve with S. aureus IMIs (11). More broadly, the lack of effective measures to treat mastitis on organic farms has led to the use of alternative therapies such as feeding whey products, multivitamins, and probiotics, but their impact on mastitis remains unclear (9, 12). Even within conventional dairy farms (i.e., those not organically certified), control of S. aureus remains a challenge in some circumstances and is considered a major contributor to antimicrobial use (3). Therefore, for both conventional and organic dairy production, research into new tools and therapies for S. aureus IMI is critical.

One possible tool that has recently garnered significant attention is the teat apex (or teat end) microbiome (13), with several studies reporting potential associations with mastitis risk (14–16). Using culture-independent methods, one study showed that mammary quarters with a history of mastitis had lower microbial diversity compared to teat canals without a history of mastitis (16). Higher microbial richness has also been observed on the teat apices of cows free of an IMI compared to those with an IMI (17). Culture-dependent methods have been used to identify associations between specific microbes on the teat apex and mastitis risk and somatic cell count numbers (18). The presence of non-aureus staphylococci (NAS) such as Staphylococcus chromogenes on the teat apex during prepartum has been associated with lower somatic cell counts shortly after calving, which is used as an indicator for the presence of subclinical mastitis (18). Staphylococcus chromogenes isolated from the teat apex has also been shown to partially or fully inhibit S. aureus in vitro (19). The reason for this apparent inhibition is thought to be the production of antimicrobial peptides (AMPs) capable of targeting and attenuating growth of other microorganisms (20). Indeed, NAS isolated from teat apices (including S. chromogenes, Staphylococcus equorum, and Staphylococcus saprophyticus) have been shown to produce AMPs, resulting in growth inhibition of multiple S. aureus strains in vitro (15). Due to this growing body of literature, the teat apex microbiome and AMPs have been recognized as important areas of further research with regard to mastitis risk (3, 13, 15); however, few high-throughput sequencing studies of the teat apex microbiome exist, precluding the identification of other microorganisms or AMPs that may be involved in the complex etiology of mastitis.

The primary objective of this study was to investigate differences in the prepartum teat apex microbiome of cows that were culture positive for a postpartum S. aureus IMI compared to cows that were not. Secondary objectives were to explore the functional capacity of the teat apex microbiome to encode genes responsible for AMP production and to identify AMP-encoded genes that may be associated with postpartum S. aureus IMIs.

MATERIALS AND METHODS

Study design

In this case-control study, we collected skin swab samples from the teat apex and quarter milk samples from 710 organically reared dairy cows from 5 farms in Colorado, Minnesota, New Mexico, and Texas. Eligibility for farm and cow enrollment has been described previously (21). This study was approved by the University of Minnesota Institutional Animal Care and Use Committee (#1807-36109A), the Colorado State University Animal Care and Use Committee (#1402), and the Texas Tech University Animal Care and Use Committee (#18068-10). Collection of skin swab samples began 8 weeks prior to parturition and ended 4–5 weeks after parturition. Weekly sample collection of quarter milk samples began immediately after parturition and ended 4–5 weeks after parturition.

Animal and herd characteristics

Animal and herd characteristics for each of the five enrolled farms in this study are described in Table 1. The predominant animal breed for four of the five farms was Holstein, and a mix of Holstein crosses for one of the five farms. The number of milking cows for enrolled farms ranged from 100 to 3,000. For four of the five farms, Dairy Herd Improvement testing was confirmed by farm personnel; quarter milk culture was performed on quarters with clinical mastitis; and the use of fly control strategies was practiced. Pre- and post-milking teat disinfectants were used on all farms.

TABLE 1.

Herd and management information of enrolled farms^a

	Farm A	Farm B	Farm C	Farm D	Farm E
Average number of milking cows	1,700	3,000	1,500	275	100
Housing system lactation	N/A	Dry lot	Dry lot	N/A	N/A
Bedding material lactation	Sand	N/A	N/A	Straw	N/A
Housing dry period	N/A	Pasture	Pasture	N/A	N/A
Pre- and post-dipping	Y	Y	Y	Y	Y
Clinical mastitis records	Y	Y	Y	Y	N/A
Milk culture on fresh cows	N	Y	Y	N	N/A
Milk culture on clinical mastitis cases	Always	Always	Always	Always	N/A
Use of fly control strategies	Y	Y	Y	Y	N/A
Predominant breed	Holstein	Holstein	Holstein	Cross	Holstein

^a Cells labeled with “N/A” indicate information that was not collected or available.

Sample collection

Skin swab samples from the teat apex were collected in the milking parlor for both prepartum and postpartum animals. Teat disinfectants were not applied by farm personnel prior to sampling. University personnel collected skin swab samples from each animal by adhering to the following protocol: when working with a new animal, a new pair of gloves was put on; if manure contaminated any teats, then the teats were cleaned with a paper towel; a 4 in² gauze square pre-moistened with 1× phosphate buffer saline was removed from a whirl-pak bag; the gauze square was scrubbed against each of the cow’s teat ends; placed back inside the original whirl-pak bag, tightly sealed and labeled with the cow identifier, date of sample collection, and farm name; and then moved into a Ziplock bag that was stored in a cooler with ice. Samples collected in Minnesota were driven back to the University of Minnesota and stored in a freezer at a temperature of −80°C within 8 h of sample collection. Samples collected outside of Minnesota were driven back to the participating university and placed in a freezer at a temperature of −80°C within 8 h of sampling. Ziplock bags containing samples from outside of Minnesota were then placed inside cryoboxes, filled with dry ice, and shipped overnight to the University of Minnesota for long-term storage at a temperature of −80°C.

Quarter milk samples from postpartum cows were collected in the milking parlor. University personnel collected quarter milk samples from each animal by adhering to the following protocol: when working with a new animal, a new pair of gloves was put on; for each quarter, milk was forestripped and the ends of each quarter disinfected with a gauze square soaked in 70% EtOH; a flip top tube was then placed under the quarter and approximately 10 mL of milk dispensed into it; the top of the tube was closed and labeled with the quarter that was sampled (e.g., front-left, front-right, rear-left, or rear-right); and then placed on a metal rack that was moved into a cooler with ice once sampling was completed. Coolers with milk samples collected from Minnesota farms were driven back to the University of Minnesota and stored in a freezer at a temperature of −20°C within 8 h. Coolers with milk samples collected from outside of Minnesota were driven back to the participating university and stored in a freezer at a temperature of −20°C within 8 h. Milk samples collected from outside of Minnesota were then placed inside cryoboxes, filled with ice packs, and shipped overnight to the University of Minnesota for long-term storage at a temperature of −20°C.

Bacterial culture

Quarter milk samples from each animal and sample collection time point were aseptically pooled into composite samples and submitted to the Laboratory for Udder Health at the University of Minnesota for bacterial culture and taxonomic identification (21). Pooled milk samples were plated onto Columbia CNA agar with 5% sheep blood and MacConkey agar using 100 µL of milk (22). Pooled milk samples were considered positive for bacterial growth when the number of colony-forming units was greater than or equal to one (10 CFU/mL) (22). Pooled milk samples with the growth of more than three unique microorganisms were classified as contaminated (21). Taxonomic identification of non-contaminated bacterial cultures was carried out using matrix-assisted laser desorption ionization-time-of-flight mass spectrometry (MALDI-TOF MS) (23). The threshold for species classification was based on a minimum confidence score of 2.0 as described previously (22).

Participants

The eligibility criteria for being considered a case were based on the presence of S. aureus in any of the postpartum milk samples based on bacterial culture and MALDI-TOF MS. Colonies from milk samples that could only be taxonomically assigned to the genus level (i.e., Staphylococcus) were not eligible to be cases, nor were cows with missing calving dates. Eligibility criteria for being considered a control was based on the absence of S. aureus in any of the postpartum milk samples based on bacterial culture and MALDI-TOF MS results. However, this does not imply that control cows were free of any intramammary infection; indeed, control cows often were not. While it may be tempting to only consider those cows free of an intramammary infection as possible controls, doing so could lead to bias because it excludes animals from the source population that were also at risk of acquiring a S. aureus intramammary infection (24). Cases and controls were individually matched by farm and days in milk (±3 DIM). If a case animal had multiple postpartum milk samples testing positive for S. aureus, only the first positive sample was used for matching. If more than one control could be matched to a case animal, then one was selected at random using a random number generator. Cows with a milk sample classified as “contaminated” were not eligible to be considered a case in subsequent weeks. Similarly, cows were not eligible to be matched to cases if a pooled milk sample had been classified as “contaminated” in any previous postpartum milk sample. These decisions were made because contaminated milk samples were not submitted for taxonomic identification using MALDI-TOF; therefore, the presence or absence of S. aureus was unknown. Skin swab samples from matched pairs collected up to and prior to the time of matching were subjected to DNA extraction, library preparation, and sequencing.

Sample preparation

Prior to DNA extraction, batches of 12 whirl-pak bags containing gauze squares were removed from −80°C storage and thawed at room temperature. During thawing, a biosafety cabinet was cleaned with 70% EtOH and a glass-bead sterilizer preheated to 250°C. Once gauze squares were thawed, whirl-pak bags were moved inside the biosafety cabinet. Each whirl-pak bag was processed in the following way: the bag was opened and the gauze square was cut into thirds using metal scissors and sterilized with a glass-bead sterilizer; the dirtiest third of the gauze square was visually identified and placed inside a PowerBead Pro Tube using metal clamps sterilized with a glass-bead sterilizer (Qiagen, Cat No. 9002864, Hilden, Germany); the whirl-pak bag was then closed. Once each PowerBead Pro Tube was filled, 800 µL of CD1 lysis buffer was dispensed into each (Qiagen, Cat No. 9002864, Hilden, Germany). PowerBead Pro Tubes were then placed inside a bead-beater for mechanical cell disruption using the following parameters: 20 s shake at 2,200 revolutions per minute (rpm) and then 30 s idle (BioSpec Products, Cat. No. 1001, Bartlesville, OK). This process was repeated three times. PowerBead Pro tubes were then subjected to centrifugation for 2 min at 16,000 RCF. PowerBead Pro tubes were then moved inside a biosafety cabinet that was sterilized with 70% EtOH. For each PowerBead Pro tube, 600 µL of the supernatant was dispensed into separate rotor adapters (Qiagen, Cat No. 9002864, Hilden, Germany). Once complete, rotor adapters were placed inside a QIAcube Connect instrument for DNA extraction (Qiagen, Cat No. 9002864, Hilden, Germany).

DNA extraction, library preparation, and sequencing

DNA was extracted from gauze samples collected from the teat apex using the PowerSoil Pro Kit (Qiagen, Cat No. 47016, Hilden, Germany) on the QIAcube Connect instrument (Qiagen, Cat No. 9002864, Hilden, Germany). Extracted DNA was enzymatically sheared and dual-indexed using the Illumina Nextera XT DNA Library Preparation Kit (Illumina Inc., San Diego, CA). Libraries were created using a quarter (12.5 µL) of the recommended reaction volume (50 µL) of the transposase enzyme and without a AMPure XP Bead Cleanup step, which was shown to generate data with comparable quality and quantity as compared to a full reaction with or without a bead cleanup (File S1; Fig. S1 to S5). DNA concentration was measured using a Nanodrop Spectrophotometer (ThermoFisher Scientific). Prior to full-depth sequencing, three separate library pools (batches) underwent MiSeq Nano (Illumina Inc., San Diego, CA; 2 × 150 base pairs) quality control sequencing to ensure an equal balance of barcodes across pools (N = 340, 340, and 169 samples in pools 1, 2, and 3, respectively). Once barcode balance was confirmed, each pool was spread across four lanes of an Illumina NovaSeq 6000 S4 flow cell (Illumina Inc., San Diego, CA; 2 × 150 bp). Three positive control samples were included in each pool (ZymoBIOMICS, Cat No. D6310). Batch effects by lane were minimized by spreading libraries across each of the four lanes of the S4 flow cell. Batch effects by sequencer were minimized by randomizing samples onto separate sequencing runs. In other words, samples from cases were not sequenced on one sequencing run and controls sequenced on another.

Bioinformatics

The Nextflow pipeline framework was used for all bioinformatic processing of sequence data (25).

Quality control

Paired-end sequence reads were inspected for data quality and adapter contamination using FastQC (https://github.com/s-andrews/FastQC; 0.11.7) and MultiQC version 1.8 (26). Sequence reads were subjected to quality trimming using TrimGalore v0.6.6 (27): one-base pair was trimmed from the 3′ ends of each sequence read, and then subsequences with a quality score less than 25 or at least a four base-pair overlap with the Nextera adapter sequence were trimmed (CTGTCTCTTATA). After trimming, sequence reads were removed if they contained any ambiguous base pairs, had a sequence length less than 50 base pairs, or were missing their mate pair. Only sequence reads that survived this QC process were used in downstream analyses.

Background DNA removal

Sequence reads were aligned to a custom reference library to identify bovine sequence reads. The custom library was built from the Bos taurus (RefSeq assembly accession: GCF_002263795.2) reference genome using Kraken2’s “download-taxonomy,” “add-to-library,” and “build” commands v2.1.2 (28). Sequence reads that were not classified as bovine were used as input for taxonomic assignment.

Taxonomic assignment

Sequence reads were aligned to the CHOCOPhlan marker-gene database using Metaphlan with the “add_viruses” parameter v 3.1.0 (29). The queries included alignments to archaea, bacteria, eukaryotes, and DNA viruses. We noted that some of the taxonomic assignments inferred by Metaphlan were likely incorrect for some virus classifications; therefore, we chose to describe these viruses at a broader taxonomic rank in the text and within graphs when appropriate (e.g., at the order level instead of the species level).

Identification of antibacterial peptides

The Data Repository of Antimicrobial Peptides (DRAMP version 3.0) was selected as the reference database for AMP characterization (30). Non-host, paired-end sequence reads were merged using the “read_merger” program implemented in Kraken (31). Merged sequence reads were aligned to the DRAMP protein database using PALADIN v1.5.0 (32). The “index” and “align” commands in PALADIN were used in series to index the DRAMP database and then align genomic sequences to the DRAMP database. Sequence alignment map (SAM) files output from PALADIN were converted to sorted binary alignment map (BAM) files using SAMtools v1.9 (33). Using the BAM file, a count matrix was generated by counting the number of sequence reads in each sample that aligned to each reference sequence in the DRAMP database using seqkit v2.4.0 (34) (bit 4 unset). Only primary alignments were considered for downstream analysis. Reads with multiple hits to the reference database (indicated by the XA tag in the SAM file) were retained. Alignments to peptides labeled as “synthetic” in the “protein existence” column of the DRAMP reference database were discarded, as were those with a eukaryotic origin.

Linkage analysis

Metagenomic assembly of non-host sequence reads was performed with metaSPAdes v3.15.5 (35). Taxonomic assignment of contigs was performed using Kraken2 v2.1.2 (28). A custom DIAMOND database was built from the DRAMP reference database (36), filtered to only include the subset of AMPs that were identified using the alignment-based approach described in the previous section (see Identification of Antibacterial Peptides section). Annotated contig sequences were then aligned to the subsetted DRAMP database using the DIAMOND aligner v0.9.13 (36). Alignments with a percent identity less than 80% were discarded. When multiple genes encoding AMPs were detected in the same genomic position within a single contig, the AMP with the highest percent identity alignment was retained. If the percent identity was equivalent among such alignments, then the AMP with the lowest e-value alignment was retained. If the e-values for such alignments were also the same, we removed the relevant alignment results from further analysis.

Analysis of positive controls

Reference genomes for the positive mock community were downloaded from https://s3.amazonaws.com/zymo-files/BioPool/ZymoBIOMICS.STD.refseq.v2.zip (ZymoBIOMICS, Cat No. D6310). Sequence reads from 10 positive control samples were aligned to the Zymo reference database using BWA-MEM v0.7.17 (37). The resulting SAM file was converted to a sorted BAM file using SAMtools v1.9 (33). Counts of reads aligning to each reference sequence—and associated plasmids—were tallied from each BAM file using the seqkit program with the “bam” option and “count” flag v2.4.0 (34). Only counts of primary alignments were considered. Counts from each sample were converted to relative abundances and visualized using stacked bar charts in ggplot2 v3.4.2 (38).

Sequencing statistics

Linear mixed effect models from the “lme4” package (v1.1.34) were used to determine if the raw sequencing depth (primary outcome) of teat apex samples differed between cases and controls after adjusting for potential confounding variables, such as sequencing batch, weeks to infection, days to infection, and farm (39). Confounding factors remained in the model if their effect was significantly associated with the primary outcome. This was determined by comparing a full model—with each of the above variables—to a null model with one of the variables removed using a likelihood ratio test. The same procedure was used to determine if the host sequencing depth (secondary outcome) differed between cases and controls. The “emmeans” package (v1.8.7) was used to calculate adjusted raw or host sequencing depths between cases and controls when significantly associated with the outcome (40). Univariate linear regression from the “stats” package (v4.3.1) and ordinal logistic regression from the “MASS” package (v7.3.60) were used to determine if days in milk, days to infection, or weeks to infection differed between cases and controls (41). A permutational multivariate analysis of variance was used to determine if sequencing batch was significantly associated with teat apex microbiome beta-diversity using vegan v2.6.4 (42, 43). The distance metric used for beta-diversity was the Aitchison distance (44).

Microbiome statistics

Logistic regression was used to determine the association between the presence of a S. aureus IMI (as determined from the milk sample cultures) and the odds of observing each microorganism in the metagenomic data generated from the teat apex samples. The term species will be used throughout the text to refer to the microbial features (i.e., microorganisms) inferred by Metaphlan (29). Logistic models were fit for each species using the glmer function in the “lme4” package v1.1.34 (39). For each species, a logistic model with a “logit” link function and “binomial” family was fit using notation from (45):

$${\displaystyle \mathrm{l}\mathrm{n}\left[{\displaystyle \frac{p}{1-p}}\right]={\beta }_0+{\beta }_1{\mathrm{I}\mathrm{M}\mathrm{I}}_i\ast {\beta }_2{\mathrm{W}\mathrm{e}\mathrm{e}\mathrm{k}\mathrm{s}\mathrm{T}\mathrm{o}\mathrm{I}\mathrm{n}\mathrm{f}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}_i+{\beta }_3\left({\mathrm{B}\mathrm{a}\mathrm{t}\mathrm{c}\mathrm{h}}_i\right)+{\mathrm{F}\mathrm{a}\mathrm{r}\mathrm{m}/\mathrm{C}\mathrm{o}\mathrm{w}}_k}$$

where P represents the probability of observing the species, ln(P/(1 − P)) represents the logit transformation, IMI is a two-level factor representing the presence or absence of a postpartum S. aureus IMI, WeeksToInfection is a seven-level factor representing the number of weeks prior to IMI at which the sample was taken, Batch is a three-level factor representing sequencing batch, Farm is a five-level factor representing the farm the sample was collected from, and Cow is a multi-level factor representing the cow that was sampled. A nested random effect was input into the model to account for cows nested within each farm. The “emmeans” package was used to calculate adjusted risks and odds ratios for each species on the response scale (40). Adjusted risks were defined as the proportion of samples with the presence of the species after accounting for other covariates in the model. Confidence intervals were used to determine if the odds of harboring a given species were significantly higher in cases or controls; therefore, P-values were not subjected to adjustment for multiple comparisons.

Negative binomial regression models were used to investigate the association between the presence of a S. aureus IMI and the abundance of each species in cases and controls. Species present in fewer than 5% of samples were discarded. The negative binomial framework was chosen due to the overdispersion observed in our data set (File S2; Fig. S6). Indeed, these are common characteristics of high-throughput sequencing experiments (46–49) and numerous software packages exist to account for these very features (46, 50–52). However, few provide convenient and robust methods for handling data sets produced from observational studies, e.g., testing for interactions, specifying multiple (or nested) random effects, and dealing with repeated measurements; therefore, we chose to use the models implemented in the “glmmTMB” package v1.1.8 (53) because they offer these features and integrate well with downstream statistical packages such as emmeans (40). Negative binomial regression models were fit for each species in the following way:

$${\displaystyle E\left[{\mathrm{A}\mathrm{b}\mathrm{u}\mathrm{n}\mathrm{d}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}}_i\right]={\beta }_0+{\beta }_1\left({\mathrm{I}\mathrm{M}\mathrm{I}}_i\right)\ast {\beta }_2\left({\mathrm{W}\mathrm{e}\mathrm{e}\mathrm{k}\mathrm{s}\mathrm{T}\mathrm{o}\mathrm{I}\mathrm{n}\mathrm{f}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}_i\right)+{\beta }_3\left({\mathrm{B}\mathrm{a}\mathrm{t}\mathrm{c}\mathrm{h}}_i\right)+\mathrm{o}\mathrm{f}\mathrm{f}\mathrm{s}\mathrm{e}\mathrm{t}\left({\mathrm{D}\mathrm{e}\mathrm{p}\mathrm{t}\mathrm{h}}_i\right)+{\mathrm{F}\mathrm{a}\mathrm{r}\mathrm{m}/\mathrm{C}\mathrm{o}\mathrm{w}}_k}$$

An offset term for sequencing depth was included in the model to account for differences in sequencing depth per sample, as in (47). Our response variable for each statistical model (i.e., abundance) was not normalized a priori simply because the negative binomial model expects counts and not fractional or negative integers, which are commonly produced by many normalization methods and transformations [cumulative sum scaling (50), base 10 logarithm or centered-log ratio (44)]. The significance of the interaction term was tested using a likelihood ratio test. If the term did not significantly improve model fit, then the term was removed, and the additive model ran instead. Significance values were corrected for multiple comparisons using Benjamin Hochberg correction (54). Fits from each model were used as input to the emmeans function in the “emmeans” package to calculate adjusted species abundances. The adjusted abundances from cases and controls were used to calculate a log fold change for each species. Intraclass correlation coefficients were computed for variance components—when possible—using the icc function in the “performance” package v0.10.4 (55).

Linear mixed effect models were used to investigate the association between the presence of an S. aureus IMI and antimicrobial peptide (AMP) gene richness and inverse Simpson diversity. AMP richness and diversity were measured using the estimate_richness function in “phyloseq” v1.44.0 (56). For each alpha diversity outcome (richness or diversity), the following linear model was fit:

$${\displaystyle E\left[{\mathrm{A}\mathrm{l}\mathrm{p}\mathrm{h}\mathrm{a}}_i\right)]={\beta }_0+{\beta }_1\left({\mathrm{I}\mathrm{M}\mathrm{I}}_i\right)\ast {\beta }_2\left({\mathrm{W}\mathrm{e}\mathrm{e}\mathrm{k}\mathrm{s}\mathrm{T}\mathrm{o}\mathrm{I}\mathrm{n}\mathrm{f}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}_i\right)+{\beta }_3\left({\mathrm{B}\mathrm{a}\mathrm{t}\mathrm{c}\mathrm{h}}_i\right)+{\mathrm{F}\mathrm{a}\mathrm{r}\mathrm{m}/\mathrm{C}\mathrm{o}\mathrm{w}}_k}$$

The “emmeans” package was used to estimate adjusted richness and diversity values between cases and controls for each alpha diversity measure on the response scale (40). AMP-encoded gene abundances were compared between cases and controls using the same method and formula described for the comparison of species abundances.

RESULTS

Enrollment and sample collection information

The number of animals enrolled in this study and the number of samples collected from participating farms are described in Table 2. A total of 710 animals were enrolled from five farms, with a range of between 28 and 288 animals enrolled on each farm. The total number of teat apex samples collected from the 710 animals was 4,827. The median number of teat apex samples collected from each animal was 8. Samples were collected in 2019 and parts of 2020.

TABLE 2.

Animal enrollment and sample collection information

	Farm A	Farm B	Farm C	Farm D	Farm E	Overall
Animals (N)	288	159	164	71	28	710
Teat apex samples (N)	1,437	1,220	1,287	678	205	4,827
Median number of teat apex samples per animal	5	9	9	10	7
Median number of milk samples per animal	4	5	5	4	4
Year(s) of sample collection	2019	2019	2019–2020	2019	2019–2020
Month(s) of sample Collection	July–Nov	Feb–Aug	Aug–Jan	Feb–June	Aug–Jan

Case-control matching statistics

The number of animals selected for this case-control study and their sample characteristics are described in Table 3. One of the five farms—Farm E—did not have any cases of S. aureus IMIs; therefore, no samples from this farm were selected for this case-control study. Historically, this farm had undertaken a comprehensive plan to eliminate S. aureus from its herd. While the farm had not been actively testing for S. aureus for several years when it was enrolled in this study, it seems that their biosecurity protocols for incoming replacement cows had been successful in maintaining a S. aureus-free herd. Based on our eligibility criteria, the total number of cows included in this case-control study was 162 (cases = 81, controls = 81). The number of controls with a non-aureus IMI was 50 (61.7%). The number of cases and controls with an IMI caused by more than one microorganism, i.e., a co-infection, was 37 (45.6% of 81 cows) and 17 (20.9% of 81 cows), respectively. The distribution of the number of microorganisms isolated from cases and controls is shown in File S2; Fig. S7. Previous analysis of the milk samples for this population reported that 67 cows with a S. aureus IMI had a persistent infection, defined as two or more milk samples with an S. aureus IMI (22). A total of 839 teat apex samples from these 162 cows were available for metagenomic sequencing (418 from cases and 421 from controls). The median number of samples sequenced from cases was 5 (min = 3; max = 9) and controls was 5 (min = 3; max = 8). The median number of days after calving for when a case became a case was 6 (min = 0, max = 32). The median number of days after calving for when a control was matched to a case was 6 (min = 0, max = 29).

TABLE 3.

Demographics of cases and controls

	Case	Control	Total
No. of animals	81	81	162
No. of samples	418	421	839
No. of animals by farm
Farm A	20	20
Farm B	31	31
Farm C	9	9
Farm D	21	21
Farm E	0	0
No. of samples per animal (min, max, median)	3,9,5	3,8,5
Median DIM at time of matching	6	6
Number of any IMIa	81	50
Number of co-infectionsa	37	17
Number of NASa	19	33
Number of Corynebacteriuma	3	0

^a Denominator is the number of case and control animals, i.e., 81.

Sequencing statistics

A total of 26.5B (billion) sequence reads was generated from 839 teat apex samples sequenced on three separate runs of an Illumina NovaSeq using an S4 flow cell (Table S2) (Mean = 31.6M reads per sample; SD = 6.0M reads). After QC, 25.7B sequence reads remained (Mean = 30.6M reads per sample; SD = 5.8M reads), and after removal of B. taurus reads, 18.0B sequence reads remained (Mean = 21.5M reads per sample; SD = 9.9M reads).

Sequencing batch, case or control group, farm, days to infection, and weeks to infection were not significantly associated with raw sequencing depth (Fig. 1A through E, ANOVA P > 0.05). Sequencing batch was not significantly associated with host sequencing depth (Fig. 1F; i.e., B. taurus DNA), but case or control group, farm, and weeks to infection were (ANOVA P < 0.05). On average, cases had 4.25 million more B. taurus DNA reads compared to controls after adjusting for weeks to infection and farm (95% CI: 2.62M–5.87M; P < 0.0001; Fig. 1G through J).

Fig 1.

Sequencing and quality control plots of teat apex samples. (A) Raw sequencing depth as a function of sequencing batch; (B) raw sequencing depth as a function of case or control group and sequencing batch; (C) raw sequencing depth as a function of case or control group; (D) raw sequencing depth as a function of case or control group and days to infection; (E) raw sequencing depth as a function of case or control group and weeks to infection; (F) host sequencing depth as a function of sequencing batch; (G) host sequencing depth as a function of case or control group and sequencing batch; (H) host sequencing depth as a function of case or control group; (I) host sequencing depth as a function of case or control group and days to infection; (J) host sequencing depth as a function of case or control group and weeks to infection; (K) count of the number of teat apex samples by case or control group and days in milk; (L) count of the number of teat apex samples by case or control group and days to infection; (M) count of the number of teat apex samples by case or control group and weeks to infection; and (N) PCA plot of teat apex microbiome colored by sequencing batch.

Days in milk, days to infection, and weeks to infection in cases did not differ between cases and controls (Fig. 1K through M). Clustering of the teat apex microbiome by sequencing batch was not visually apparent using PCA (Fig. 1N); indeed, the amount of variation in the teat apex microbiome that could be explained by sequencing batch was <1% (R² = 0.0036, F = 1.53, P = 0.0009).

Positive control statistics

Ten positive control samples (i.e., mock communities) were sequenced on three separate runs of the Illumina NovaSeq instrument (Illumina Inc., San Diego, CA). The composition of the mock community included eight bacteria: Listeria monocytogenes, Pseudomonas aeruginosa, Bacillus subtilis, Escherichia coli, Salmonella enterica, Limosilactobacillus fermentum (formerly Lactobacillus fermentum), Enterococcus faecalis, and S. aureus and two eukaryotes: Cryptococcus neoformans and Saccharmomyces cerevisiae (Zymo Research Cat. No. D6310). All 10 members of the mock community were identified from 100% of positive control samples (File S2; Fig. S8); however, the relative abundance of these mock community members in our data differed from the theoretical composition. The most abundant member of the mock community (L. monocytogenes) was under-represented with a mean relative abundance of 60.9% (theoretical: 89.1%). All other members of the mock community tended to be overrepresented compared to their theoretical compositions (File S2; Table S1).

Taxonomic classification statistics

A list of all microorganisms identified from shotgun metagenomic sequencing of the teat apex microbiome is listed in Table S3. Non-host sequencing reads were taxonomically assigned to 1,395 archaeal, bacterial, eukaryotic, and viral species. The number of distinct species classified to each kingdom was 18 for archaeal species (1.3% of all 1,395 identified species), 987 for bacterial species (70.8%), 10 for eukaryotic species (0.7%), and 380 for viral species (27.2%).

The top 60 most prevalent microorganisms and their abundances from the metagenomic data (crude prevalence from 839 samples) are shown in Fig. 2. Prevalent bacterial species included NAS: S. chromogenes (n = 472 samples, 56.2%), S. haemolyticus (n = 470, 56.0%), S. auricularis (n = 412, 49.1%), and S. devriesei (n = 388, 46.2%); Corynebacterium species: C. efficiens (n = 732, 87.2%), C. xerosis (n = 733 87.3%), and C. marinum (n = 732, 87.2%); Bifidobacterium species: B. pseudolongum (n = 827, 98.5%) and B. merycicum (n = 548, 65.3%); Jeotgalicoccus species (n = 679, 80.9%); Dietzia alimentaria (n = 660, 78.6%); and Kocuria polaris (n = 610, 72.7%). The most prevalent archaeal species were methane-producing archaea: Methanosarcina mazei (n = 276, 32.8%), Methanobrevibacter thaueri (n = 261, 31.1%), and Methanobrevibacter millerae (n = 125, 14.8%). The most prevalent viral species were two herpes-associated viruses, which we describe at the order-level taxonomy: Herpesvirales I (n = 836, 99.6%) and Herpesvirales II (n = 346, 41.2%); bacteriophages: Microbacterium phage Min 1 (n = 579, 69.0%), Staphylococcus virus EW (n = 268, 31.9%), and Enterobacteria phage P4 (n = 267, 31.8%) and a bovine-associated polyomavirus (n = 518, 61.7%). The most prevalent eukaryotic species was the fungus Aspergillus fumigatus (n = 13, 1.5%).

Fig 2.

Taxonomic composition, prevalence, and abundance of the most frequently sequenced microorganisms from the teat apex microbiome. The signature heatmap displays the median normalized abundance (clr, center-log ratio) of each microorganism for each week relative to infection in cases. The bar graph shows the prevalence of microorganisms in teat apex samples (denominator 839 samples). The dot plot shows the relative abundance of each microorganism within each sample. The red triangles represent the mean relative abundance of each microorganism across all samples.

The most relatively abundant microorganisms were bacteria from the NAS, Corynebacterium, Bifidobacterium, Dietzia, and Jeotgalicoccus genera, accounting for much of the metagenomic DNA from the teat apex (Fig. 2). The three most relatively abundant species from the metagenomic data were B. pseudolongum (18.3%), C. marinum (10.0%), and D. alimentaria (9.8%). Among the NAS group, S. chromogenes (2.5%), S. haemolyticus (1.8%), and S. devreisei (1.8%) were the most relatively abundant. Other relatively abundant Corynebacterium species included C. marinum (10.0%), C. efficiens (3.5%), C. xerosis (2.7%), and C. pollutisoli (2.6%). The Jeotgalicoccus sp. (2.6%), Sarcina sp. (3.0%); Kocuria polaris (2.2%) and K. rosea (1.3%); and Ruminococcaceae P7 species (2.2%) were abundant as well.

Presence of Staphylococcus aureus and skin-associated viruses on the teat apex prior to parturition were the strongest risk factors for Staphylococcus aureus IMIs after parturition

The presence of S. aureus and skin-associated viruses in the metagenomic teat apex data prepartum was significant risk factors for postpartum S. aureus IMIs (Fig. 3A). Samples from cases had significantly higher odds of harboring S. aureus DNA on the teat apex metagenome compared to controls (OR = 38.95, 95% CI: 14.8–102.2). The adjusted risk of S. aureus DNA in samples from cases was 33.6% (95% CI: 22.5%–46.8%) and controls was 1.2% (95% CI: 0.4%–3.3%). The odds of harboring S. aureus DNA varied more between cows [intraclass correlation coefficient (ICC): 0.39] than it did between farms (ICC: 0.01). In some cows, the presence of S. aureus DNA on the teat apex could be observed ≥7 weeks prior to the detection of the postpartum S. aureus IMI (Fig. 3B), almost 62 days away from detection of the infection (Fig. 3C). A majority of S. aureus DNA found on the teat apex was observed prior to parturition, with few samples containing any detected S. aureus DNA after parturition (Fig. 3D). Cases also had higher odds of containing DNA from a herpes-associated virus—Herpesvirales II virus (species1326)—(OR = 2.09, 95% CI: 1.46–2.98) compared to controls. The adjusted risk of Herpesvirales virus II DNA in cases was 43.57% (95% CI: 28.74%–59.65%) and controls was 26.97% (95% CI: 16.07%–41.59%). The variation in the odds of harboring Herpesvirales virus II DNA was similar between cows (ICC: 0.08) and farms (ICC: 0.08). Samples from cases also had a higher odds of containing DNA from an unspeciated Psychrobacter bacterium (OR = 1.83, 95% CI: 1.82–1.83), but the prevalence of this bacterium was low (n = 82 total samples) and only found on a single farm.

Fig 3.

(A) Forest plot describing the relationship between having an S. aureus intramammary infection and the odds of each species being present on the teat apex. Adjusted risks (Adj. Risk) represent the probability of the species being present on the teat apex, stratified by cases and controls. Odds ratios (OR) represent the odds of the species being present in cases relative to controls. The dashed vertical line in the forest plot indicates the null association. Points represent ORs and horizontal bars represent 95% confidence intervals. The number of samples (“frequency”, y-axis) with S. aureus in the teat apex metagenomic data (No = orange, Yes = blue) as a function of weeks to infection (B), days to infection (C), and days relative to calving (D).

A total of 31 bacteria and 2 bacteriophages were identified as potential microbial protectives (i.e., species associated with lower odds of S. aureus IMI, Fig. 3A). The most prevalent microorganisms in controls associated with protection were Microbacterium phage Min 1 (OR = 0.37, 95% CI: 0.25–0.53), Corynebacterium efficiens (OR = 0.53, 95% CI: 0.30–0.94), Kocuria polaris (OR = 0.54, 95% CI: 0.35–0.82), Micrococcus terreus (OR = 0.64, 95% CI: 0.44–0.93), Dietzia alimentaria (OR = 0.45, 95% CI: 0.26–0.75), Corynebacterium maris (OR = 0.44, 95% CI: 0.25–0.74), Salinicoccus kekensis (OR = 0.55, 95% CI: 0.32–0.92), Turicibacter sanguinis (OR = 0.59, 95% CI: 0.41–0.83), and Corynebacterium camporealensis (OR = 0.53, 95% CI: 0.32–0.85).

The odds of NAS such as S. chromogenes (OR = 0.97, 95% CI: 0.56–1.69), S. devriesei (OR = 0.78, 95% CI: 0.44–1.38), or S. haemolyticus (OR = 0.81, 95% CI: 0.47–1.38) being present on the teat apex were not different between cases and controls. The adjusted risk of S. chromogenes DNA in samples from cases was 67.1% (95% CI: 46.2–82.9) and controls was 67.6% (95% CI: 46.8–83.2). The adjusted risk of S. devriesei DNA in cases was 46.6% (95% CI: 22.0–73.0) and controls was 52.7% (95% CI: 26.5–77.6). The adjusted risk of S. haemolyticus DNA in cases was 66.4% (95% CI: 54.2%–76.8%) and controls was 70.8% (95% CI: 59.0–80.3).

Abundance of Staphylococcus aureus on the teat apex was higher in cases compared to controls

The abundance of sequence reads originating from S. aureus was significantly higher in cases compared to controls [log fold change (LogFC) = 23.8, BH-adjusted P < 0.0001, Fig. 4A]. Across all farms, the abundance of S. aureus was consistently higher in cases compared to controls (Fig. 4B). There was more variation in S. aureus abundance between cows (ICC: 0.66) than between farms (ICC: 0.006). The abundance of sequence reads originating from Psychrobacter immobilis (LogFC = 10.23, BH-adjusted P = 0.002) and Bacteroides pyogenes (LogFC = 7.34, BH-adjusted P = 0.04) was significantly higher in samples from controls compared to cases (Fig. 4A). Unlike S. aureus, these two bacterial species were primarily found on a single farm (Fig. 4B), and this was reflected in their variance components for the farm-specific random effects for P. immobilis (ICC: 0.96) and B. pyogenes (ICC: 0.56). The abundance of sequence reads originating from NAS—S. chromogenes, S. devriesei, S. haemolyticus, and S. auricularis—were not significantly different between cases and controls (Fig. 4A and B).

Fig 4.

(A) Bar plot displaying log fold change in abundance (x-axis) of each microorganism (y-axis) between cases and controls. Bars colored red indicate microorganisms with a significantly higher abundance in samples from cases compared to controls (i.e., risk factor); bars colored green indicate microorganisms with a significantly higher abundance in samples from controls compared to cases (i.e., protective); bars colored gray indicate microorganisms without a significant effect size (i.e., not sig). (B) Heatmap displaying normalized microbial abundances between cases and controls, stratified by farm. Abundances represent centered log ratios (i.e., clr).

Antimicrobial peptides

We identified 171 unique putative AMP genes across all 839 metagenomic data sets generated from teat apex samples. After removing AMP genes with a synthetic (N = 16) and eukaryotic (N = 81) origin, 74 AMP genes remained for analysis (Table S4). Two samples—one in the case and one in the control group—did not have any bacterial AMP genes. Many of the putative bacterial AMP genes that we identified from the teat apex were rare, with 64 of 74 (86.4%) being found in fewer than 10 samples (Fig. 5). The mean number of bacterial AMPs found in cases was 1.45 (95% CI: 1.33–1.75) and controls was 1.63 (95% CI: 1.42–1.84). The difference in the mean number of AMPs was not statistically different between cases and controls (contrast: case-control; estimate = −0.09, P = 0.13). The mean AMP gene inverse Simpson diversity in cases was 1.09 (95% CI: 1.05–1.14) and controls was 1.08 (95% CI: 1.03–1.13), with a non-statistically significant mean difference of 0.13 (contrast: cases-control; P = 0.36). The most prevalent and abundant predicted bacterial AMP genes were Class I (Microcin B17) and IIa microcins (Colicin-V). Microcin B17 was found in 837 of 839 teat apex samples (99.7%), and Microcin-V was found in 43 of 839 teat apex samples (5.1%). The adjusted mean abundance of Microcin B17 in cases was 27.9 (95% CI: 23.5–33.2) and controls was 33.6 (95% CI: 28.3–39.9), with a significant mean difference (P = 0.0014). As the remaining putative AMP genes identified from the teat apex were rare, we chose not to perform a statistical analysis on them; therefore, crude statistics are reported. Two homology-inferred (DRAMP03172 and DRAMP03170) and three protein-level (DRAMP02795, DRAMP02794, DRAMP02793) AMP genes belonging to the staphylococcal hemolytic protein family were found in 108 of 839 samples (12.8%): antibacterial protein 1 homolog in 42 samples (5.0%), antibacterial protein 3 homolog in 62 samples (7.3%), antibacterial protein 1 in one sample (0.1%), antibacterial protein 2 in 28 samples (3.3%), and antibacterial 3 protein in 29 samples (3.4%). The number of samples with any of these AMP genes was 42 in cases (5.0%) and 66 in controls (7.8%). A gene encoding a class I bacteriocin named Sonorensin, typically found in Bacillus sonorensis MT93 strains, was found in 31 samples (3.6%). Class II bacteriocins typically found in S. aureus were found in 22 samples (2.6%): Aureocin A70 (AurA, 9 samples), Aureocin A70 (AurB, 5 samples), Aureocin A70 (AurC, 3 samples), Aureocin A70 (AurD, 5 samples).

Fig 5.

Antimicrobial peptides detected in the teat apex shotgun metagenomic data. Genes encoding for AMPs were inferred by aligning translated sequence reads to the DRAMP protein database using PALADIN (32). AMP-encoded genes present in three or more samples are shown.

Genomic linkage of antimicrobial peptides

We de novo assembled metagenomic sequences from 839 samples into contigs, classified the contigs into taxonomic units, and then aligned the contigs to the 74 bacterial AMP genes we identified previously. After discarding low-quality alignments and those mapping equally well to the same genomic loci, 365 samples (43.5%) containing 911 alignments from 652 contigs remained for analysis. The most frequently identified AMP genes from the 911 alignments belonged to the staphylococcal hemolytic protein family (57, 58): antibacterial protein 1 (DRAMP03170, 21.8% of the 911 alignments), antibacterial protein 2 (DRAMP02794, 20.4%), and antibacterial protein 3 (DRAMP03172, 21.7%). Indeed, AMP genes from the staphylococcal hemolytic protein family were most often localized to contigs taxonomically classified to S. haemolyticus, i.e., 379 contigs contained a staphylococcal hemolytic protein family AMP gene, and 298 of these were classified as S. haemolyticus. Plasmid-encoded class II bacteriocins typically isolated from S. aureus were also found (59): Aureocin A70 A (1.7% of 911 alignments), Aureocin A70 B (1.7%), Aureocin A70 C (1.8%), and Aureocin A70 D (3.0%), and these were predominantly localized to contigs classified as plasmids (i.e., 39 contigs contained an Aureocin AMP and 34 of these were colocalized with plasmids). Genes encoding for S. aureus bacteriocins called PSMα1 and PSMα2 (2.3% and 1.8% of 911 alignments, respectively) were also identified and strictly localized to contigs taxonomically assigned to S. aureus (60). Genes encoding bacteriocins called Serracin P (2.8% of 911 alignments) and Carnobacteriocin B2 (3.7%) were also identified.

DISCUSSION

We found that samples from cows culturing positive for an S. aureus IMI after parturition had significantly higher odds of S. aureus metagenomic DNA being detected from the teat apex compared to cows that cultured negative (OR = 38.95, 95% CI: 14.8–102.2). These findings are consistent with culture-based results reporting that nulliparous cows with teats culturing positive for S. aureus have 3.34-times higher odds of testing positive for an S. aureus IMI after parturition compared to heifers with teats that were culture negative for S. aureus (61). The same study also showed that the presence of S. aureus cultured from lacteal secretions collected prior to parturition was associated with 19.5-higher odds of having an S. aureus IMI after parturition (61). A recent cross-sectional study of 287 multiparous cows found that when the teat skin from a quarter was culture positive for S. aureus, the odds of the same quarter being culture positive for an S. aureus IMI was 7.8-times higher compared to quarters with teats that were not culture positive for S. aureus (62). A smaller cross-sectional study of 57 multiparous cows also found that teat skin samples that were culture positive for S. aureus had 4.5-times higher risk of S. aureus IMI in the same quarter compared to teat skin samples that were not culture positive for S. aureus (63). However, the cross-sectional nature of the latter two study designs and the time periods at which they occurred (i.e., during lactation) precluded establishing temporality between S. aureus colonization of the teat skin and S. aureus IMIs.

A strength of our study is that we attempted to address the issue of temporality by sampling the teat-skin up to 8 weeks prior to parturition and on weekly intervals thereafter. However, our ability to definitively establish temporal relationships between the teat apex microbiome and S. aureus IMIs was complicated by the fact that many of the observed post-calving S. aureus IMIs occurred during the first week post-calving. In addition, the high number of case samples with the presence of S. aureus DNA on the teat apex compared to controls prior to parturition suggests that IMIs could have been acquired prior to calving (Fig. 3D), as was observed by reference (61). If this is true, then the route of exposure may have been due to “open quarters,” which (64) observed in 60% of teat canals as far back as 60 days prior to parturition in a study of 84 Holstein heifers. Indeed, prepartum IMIs with both minor and major pathogens such as S. aureus were common among open quarters, as was the proportion of postpartum IMIs (77%) caused by pathogens observed during the prepartum period (64).

Previous culture-based studies of the teat apex have also reported protective effects of specific bacterial groups against major mastitis pathogens such as S. aureus (65, 66). One popular bacterial group with the purported effect has been NAS. Consistent with previous studies of the teat apex, NAS DNA was highly prevalent in our study, with S. chromogenes, S. devriesei, S. haemolyticus, and S. auricularis all attaining at least 45% prevalence across our sample set (67–69). However, in contrast to observations reported in previous field and in vitro studies, we were not able to provide strong support for the hypothesis that the presence of NAS DNA on the teat apex confers protection against S. aureus colonization of the mammary gland (15, 66). In fact, there was no evidence of an association between the presence of DNA from S. chromogenes (OR = 0.97, 95% CI: 0.56–1.69), S. devriesei (OR = 0.78, 95% CI: 0.44–1.38), or S. haemolyticus (OR = 0.81, 95% CI: 0.47–1.38) on the teat apex and S. aureus IMI risk. Nor was there a difference in DNA abundance of these bacteria between cases and controls. In many cases, the inhibition of S. aureus by S. chromogenes in previous studies was observed within in vitro models (66), which is not likely to be reflective of the teat apex environment. Indeed, recent concerns about the utility and reproducibility of in vitro experiments using culture models that are not reflective of the complex microbial environment under study have been raised within the life science community (70). A meta-analysis investigating the association between the presence of minor pathogens (NAS and C. bovis) on major pathogen (S. aureus, S. agalactiae, S. uberis, S. dysgalactiae, E. coli) IMI risk found that, on average, there was no association in observational studies, but there was a significantly lower risk, on average, observed in experimental studies (71). As the authors point out, heterogeneity between both observational and experimental studies does exist (71), which may explain the null findings observed in our study. Many of the protective associations conferred by NAS have also been explored within milk, whereas our study focused on the skin, which may also explain the null associations observed for most NAS species investigated in our study. Finally, our analysis utilized metagenomic data, which can contain DNA from both live and dead organisms; whereas previous studies reporting on the protective effect of NAS utilized culture-based assays wherein only viable NAS bacteria were included.

Our study also identified several microorganisms on the teat apex that were associated with lower odds of acquiring an S. aureus IMI. The microorganisms associated with the largest protective effects included Corynebacterium camporealensis, Corynebacterium maris, Kocuria polaris, Kocuria rosea, Dietzia alimantera, Cutibacterium acnes, Psychrobacter immobilis, and Microbacterium phage min 1. Many of these genera and species have been reported in the relatively few existing studies that utilize high-throughput sequencing to interrogate the teat apex microbiome of dairy cows (72–74). However, little is known about the specific effects these microorganisms have on udder health, as most of these microorganisms have not been widely studied using traditional culture-based methods. The exception to this is C. acnes [formerly Propionibacterium acnes and Corynebacterium parvum (75)], which had a short history of being evaluated as an intramammary immunostimulant during the dry period to prevent new IMIs after parturition (76) and as a treatment for chronic S. aureus IMIs (75). The main takeaway from both studies was that C. acnes cultured from milk samples had little to no effect on udder health outcomes (75, 76). The positive effect observed in our study may be explained by the fact that C. acnes was found on the teat apex and, therefore, may act differently in this environmental niche compared to inside the mammary gland itself.

Bacteria belonging to the Corynebacterium genus are generally considered to be minor IMI pathogens, but their presence within the mammary gland has also been associated with lower odds of S. aureus IMI in experimental studies, though findings are mixed (71). Information about the specific species within the Corynebacterium genus that we identified as being protective (i.e., C. camporealensis and C. maris) is scarce, as these bacteria have not been reported in most culture-based studies involving udder health. This is likely because many bacteria are not easily culturable, and the MALDI-TOF MS instrumentation used to assign taxonomic labels is limited to the taxa that are well-represented in reference databases. The microorganism with the strongest protective effect was Microbacterium phage Min 1 (OR = 0.37, 95% CI: 0.25–0.53), a bacteriophage that has been found on plasmids within bacterial cells of the Microbacterium genus (77). Indeed, eight species of bacteria belonging to the Microbacterium genus were observed on the teat apex (Table S3). Bacteria belonging to the Microbacterium genus have also been reported as common laboratory contaminants (78), but this seems to be an unlikely source in our study as we would expect an equal probability of these bacteriophage to be found in both cases and controls since our samples were processed using the same lot of laboratory reagents (adjusted risk in cases = 63.0% in cases, adjusted risk in controls 82.0%). However, given the specificity of bacteriophage, it seems unlikely that there would be a direct relationship between the presence of Microbacterium phage and S. aureus, which was an assumption of the statistical model that investigated this relationship. In practice, this assumption may not be appropriate and could have been explored using alternative methods that consider interactions with other microorganisms, such as Bayesian or co-occurrence networks (79).

An unexpected finding of our metagenomic study was the difference in abundance of reads that aligned to the B. taurus genome between cases and controls (Fig. 1G through J). Because we sequenced DNA, it is impossible to determine what types of host cell populations were sequenced, but somatic cells such as keratinocytes (i.e., epidermal cells) are likely as we sampled skin (80). Indeed, keratinocytes are the most common cell type found on skin, and they play a fundamental role in maintaining the integrity of the skin in the presence of microorganisms (81–83). One possible explanation for the presence of somatic cells on the teat skin may be attributed to teat-end lesions. While we did not quantify teat-end lesions in our study population, we found that the teat skin of case animals had significantly higher odds of harboring metagenomic DNA from a herpes virus compared to controls (OR = 2.09, 95% CI: 1.46–2.98). This is an interesting finding because herpes-associated viruses are thought to cause skin lesions that can act as a reservoir for bacterial colonization of the teat skin, thus leading to an increased risk of acquiring an IMI (84). The presence of herpes virus DNA on the teat skin was also significantly associated with a higher host DNA read count (estimate = 8.7M host DNA reads, 95% CI: 7.5 M–9.9M), indicating that the association between herpes virus and host DNA read counts may be mediated by the presence of teat skin lesions.

Another type of somatic cell that may have been sampled from the teat skin are neutrophils, which are common within the mammary gland of dairy cows (85), especially during an intramammary infection (3). In fact, the concentration of neutrophils and other somatic cells in bovine milk (termed “somatic cell count”) is a widely used indicator for udder health and mastitis screening (86). Although we did not sample milk, neutrophils can also be found in the skin when responding to S. aureus infections in humans (87), which may explain the elevated abundance of B. taurus DNA in skin swab samples from cases in our study (Fig. 1I and J). A limitation of our study is that we did not measure cow or quarter-level SCC from postpartum milk samples, which may have yielded additional information as to why cases had significantly higher amounts of B. taurus DNA compared to controls. However, increased somatic cells in the milk would only provide a partial explanation for this finding, as B. taurus DNA was consistently higher in samples from cases compared to controls throughout the entire sampling period, i.e., prior to lactation. Additionally, the immune status of cows during the pre-partum period is known to be highly dynamic (88), and this could have also contributed to differences both in IMI susceptibility and in host DNA levels at the teat apex. It may be tempting to attribute this finding to a technical artifact of our study design, but we found no evidence for this, as the elevated levels of B. taurus DNA in cases versus controls was found consistently across all sequencing batches (Fig. 1G) and there was no difference in raw sequencing depth between samples collected from cases and controls (Fig. 1B and C). Although we observed a statistically significant batch effect, the effect size was small and unlikely to be a major contributing factor. Taken together, this finding appears to reflect a real biological phenomenon that deserves further attention.

In addition to exploring associations between the teat apex metagenome and the presence of a postpartum S. aureus IMI, we also investigated the functional capacity of the teat apex microbiome to produce AMPs that could also be associated with the presence of a S. aureus IMI. Our metagenomic analysis identified 74 unique putative AMP-encoded genes with a presumed bacterial origin based on annotation information provided in the DRAMP reference database (30). The most prevalent of these AMP-encoded genes were Class I and II microcins, thought to be primarily produced by bacteria belonging to the Enterobacteriaceae family (89). Although we could not verify the specific bacterial origin of these AMP-encoded genes using our colocalization approach, our finding that Microcin abundance was greater in controls compared to cases (P = 0.0014) is interesting, though their antibacterial effect is thought to be restricted to closely related bacteria, i.e., those within the Enterobacteriaceae family (90), of which S. aureus is not a member. We also identified genes encoding for AMPs belonging to the staphylococcal hemolytic protein family (57, 58). Indeed, the only putative AMPs identified within contigs assigned to S. haemolyticus (298 contigs) were AMPs within the staphylococcal hemolytic protein family. Intriguingly, the genes encoding for these AMPs were more prevalent in controls versus cases (crude prevalence of 7.8% versus 5.0%), suggesting a potential protective effect against S. aureus IMIs. S. haemolyticus belongs to the NAS group of bacteria that are thought to play a protective role in S. aureus IMI risk through the production of these types of molecules (15). It is important to note that our data suggest that bacteria from the teat apex have the functional capacity to produce these peptides, but as we identified them using an in silico and DNA-based approach, we must be cautious about interpreting these results—they are simply hypotheses. Currently, little is known about the role of these naturally occurring peptides on IMI risk, but further research on their effect using proteomic assays could be explored, especially considering the current need for alternative antimicrobial therapies within agricultural settings (91).

Although we identified numerous microorganisms that were significantly associated with the presence of post-calving S. aureus IMIs, we also found that these results need to be interpreted within the context of their variance components, i.e., cows and farms. As an example, our model investigating the association between S. aureus DNA on the teat apex and postpartum S. aureus IMI risk found that there was more variation between cows (ICC: 0.32) than between farms (ICC: 0.10). However, this was not the case for a similar model investigating the association between Psychrobacter immobilis DNA and S. aureus IMI risk, in which the variation between cows (ICC: 0.04) was less than that between farms (ICC: 0.77). Indeed, DNA sequences from the latter microbe were primarily found within a single farm (high farm ICC), while the former was found consistently across each farm (low farm ICC). This is visually apparent when comparing the abundances of these two specific microorganisms between cases and controls across each farm (Fig. 4B). Environmental factors such as bedding material, housing conditions, and seasonal effects are all known to affect the concentration and types of microorganisms that can be found on the teat skin (92–94). While our study was not sufficiently powered at the farm level to account for these farm-level factors, they likely played a major role in the variation of the teat skin microbiome that we observed across farms. More broadly, this implies that the teat apex microbiome is not likely to be homogenous in animals between farms, and that future studies of the teat apex microbiome should consider accounting for this biological effect when interpreting their results, as results with large effect sizes or significant associations cannot simply be generalized across all herds. This certainly poses an interesting dilemma for future studies of the teat apex microbiome, as the investigator will need to think carefully about the number of herds to enroll and if enrolling multiple herds in similar locations or with similar management practices will mitigate the differences observed across farms.

Conclusions

The presence of Staphylococcus aureus DNA on the prepartum teat apex was the strongest risk factor for S. aureus IMIs after parturition and exhibited the largest differential abundance between cases and controls. Further research into the role of bacteriophage and other bacteria associated with protection against S. aureus IMIs could be explored to further understand their putative impact on udder health. Based on our results, the genera Psychrobacter, Corynebacterium, Dietzia, Kocuria, and Cutibacterium and phages associated with Microbacterium may be promising candidates for this future research. The continued exploration of AMPs (particularly Microcin B17 and AMPs produced by S. haemolyticus) could yield non-antibiotic alternatives for improved udder health.

DATA AVAILABILITY

The raw sequence data generated in this study has been submitted to the Sequence Read Archive on NCBI (Accession: PRJNA984925).

SUPPLEMENTAL MATERIAL

The following material is available online at https://doi.org/10.1128/aem.02234-23.

Supplemental file 1. aem.02234-23-s0001.docx.

Extended materials and methods.

Supplemental file 2. aem.02234-23-s0002.docx.

Extended results.

Supplemental Table S2. aem.02234-23-s0003.csv.

Sample meta-data and sequencing statistics.

Supplemental Table S3. aem.02234-23-s0004.csv.

Taxonomy table.

Supplemental Table S4. aem.02234-23-s0005.csv.

Antimicrobial Peptide sequence meta-data.

ASM does not own the copyrights to Supplemental Material that may be linked to, or accessed through, an article. The authors have granted ASM a non-exclusive, world-wide license to publish the Supplemental Material files. Please contact the corresponding author directly for reuse.