Skip to main content
NCBI home page
Journal of Animal Science logoLink to Journal of Animal Science
. 2018 May 18;96(6):2113–2124. doi: 10.1093/jas/sky096

Poor feed efficiency in sheep is associated with several structural abnormalities in the community metabolic network of their ruminal microbes1

Rocky D Patil 1, Melinda J Ellison 2, Sara M Wolff 1, Courtney Shearer 3, Anna M Wright 4, Rebecca R Cockrum 5, Kathy J Austin 6, William R Lamberson 1, Kristi M Cammack 7, Gavin C Conant 1,8,9,10,✉
PMCID: PMC6095279  PMID: 29788417

Abstract

Ruminant animals have a symbiotic relationship with the microorganisms in their rumens. In this relationship, rumen microbes efficiently degrade complex plant-derived compounds into smaller digestible compounds, a process that is very likely associated with host animal feed efficiency. The resulting simpler metabolites can then be absorbed by the host and converted into other compounds by host enzymes. We used a microbial community metabolic network inferred from shotgun metagenomics data to assess how this metabolic system differs between animals that are able to turn ingested feedstuffs into body mass with high efficiency and those that are not. We conducted shotgun sequencing of microbial DNA from the rumen contents of 16 sheep that differed in their residual feed intake (RFI), a measure of feed efficiency. Metagenomic reads from each sheep were mapped onto a database-derived microbial metabolic network, which was linked to the sheep metabolic network by interface metabolites (metabolites transferred from microbes to host). No single enzyme was identified as being significantly different in abundance between the low and high RFI animals (P > 0.05, Wilcoxon test). However, when we analyzed the metabolic network as a whole, we found several differences between efficient and inefficient animals. Microbes from low RFI (efficient) animals use a suite of enzymes closer in network space to the host’s reactions than those of the high RFI (inefficient) animals. Similarly, low RFI animals have microbial metabolic networks that, on average, contain reactions using shorter carbon chains than do those of high RFI animals, potentially allowing the host animals to extract metabolites more efficiently. Finally, the efficient animals possess community networks with greater Shannon diversity among their enzymes than do inefficient ones. Thus, our system approach to the ruminal microbiome identified differences attributable to feed efficiency in the structure of the microbes’ community metabolic network that were undetected at the level of individual microbial taxa or reactions.

Keywords: feed efficiency, metabolic network, metagenomics, rumen, vertebrate microbiome

INTRODUCTION

Ruminant animals have a symbiotic relationship with their ruminal microbiota, where microbial enzymes degrade complex polysaccharides indigestible to most animals (Stevens and Hume, 1998). Several studies have found a link between the presence of some bacterial lineages and differences in feed efficiency between individual ruminants (Guan et al., 2008; Hernandez-Sanabria et al., 2010; Rius et al., 2012; McCann et al., 2014). Greater feed efficiency would reduce animal agriculture’s environmental impact while at the same time improving its economic efficiency (Herrero et al., 2013). Hence, modifying the ruminal microbiome represents a potentially powerful mechanism for improving producer outcomes. Unfortunately, as with many other features of microbial ecosystems, making a link between the structure of the microbiome and feed efficiency is difficult (Hernandez-Sanabria et al., 2012). In part, this difficulty is due to the fact that the techniques used to date have been taxonomically focused, even though many microbial ecosystems can display considerable taxonomic variation while preserving their underlying functions (Moore et al., 2009; Doolittle and Booth, 2017).

Using a metabolic network approach (Lacroix et al., 2008), we compared the enzymes found in the ruminal microbiomes of feed-efficient and feed-inefficient ewes. Strikingly, no single metabolic reaction was significantly different in its abundance between the 2 groups. However, by analyzing their community metabolic networks, we were able to show that, as we hypothesized, the low residual feed intake (RFI; efficient) animals used reactions that were located in closer proximity to the host network than were those of the high RFI animals, presumably allowing for the more efficient extraction of metabolites in the former. Furthermore, low RFI animals had metabolic networks with greater mean Shannon entropy, an increased network complexity that might enable them to exploit a greater proportion of the metabolites they consume.

MATERIALS AND METHODS

Feeding Trial and Rumen Sample Collection

All animal procedures were approved by the University of Wyoming Animal Care and Use Committee. Sixteen Targhee ewe lambs showing extremes of RFI (selected from a surveyed group of 59 animals with initial BW = 45.8 ± 2.5 kg) were selected for metagenomic sequencing (Ellison et al., 2017). All lambs were fed a forage-based diet primarily consisting of alfalfa and wheat middlings, and their individual feed intakes were measured with the GrowSafe System for each day of the 70-d trial (Airdrie, Alberta, Canada). The average daily gain (ADG) was computed using the difference between initial and final BW divided by the 70-d period. The metabolic midweight (MMWT) was calculated from the mid-BW, which was the BW measured on the middle day of trial. For this dataset, RFI was calculated as the difference between actual feed intake and the expected feed intake. The expected feed intake for the animal’s production was determined by regressing actual feed intake on ADG and MMWT (Cammack et al., 2005). The sheep were then ranked on their computed RFI values. We selected the 15% of animals showing the most extreme low and high values of RFI for our metagenomics analyses. Thus, the 8 sheep having the lowest RFI (the most feed efficient) were chosen as our low RFI group, whereas the 8 animals with the largest RFI were our high RFI group.

Rumen fluid from these 16 animals (≥30.0 mL) was suctioned through a Tygon tube (length: 1 m, diameter: 1.5 cm) using a dosing syringe (400 mL) placed through the mouth down into the lamb’s esophagus and running into the rumen. Rumen samples were collected after the feeding trial before ewes were returned to a private producer’s breeding flock and immediately stored on dry ice until the samples were placed in long-term storage at −80 °C.

DNA Extraction and Illumina Library Preparation

Extraction of particulate matter from the fluid was achieved by centrifugation (Ellison et al., 2014). One milliliter of lysis buffer, silicon (0.1 g of 0.5 mm), and zirconia (0.3 g of 0.1 mm) beads were combined with thawed rumen samples to isolate DNA. The lysis buffer consisted of 500 mM NaCl, 50 mM Tris–CL, 50 mM EDTA, and 4% SDS. The mixture was homogenized with a Mini-Beadbeater-8 set at maximum speed for centrifugation of 16,000 × g for 3 min followed immediately by incubation at 70 °C for 15 min with gentle mixing every 5 min and centrifuged for 5 min at 4 °C. After homogenization, the supernatant was transferred into a new tube followed by the addition of lysis buffer (300 µL). This procedure was repeated for a second time, and the supernatants combined. The QIAamp DNA Stool Mini Kit (Qiagen, Santa Clarita, CA) was used in the removal of proteins, RNA, and the precipitation of nucleic acids. Using a NanoDrop spectrophotometer (NanoDrop, Wilmington, DE), we found that the quality of the DNA from the rumen samples was within acceptable ranges, with an A260/280 absorbance ratio ≥ 1.8.

Genomic libraries were constructed for the 16 rumen samples based on the recommendations described in the Illumina’s TruSeq DNA sample prep kit (Ellison et al., 2014). Diagenode BioRuptor was used to generate fragments of appropriately 350 bp by the shearing of genomic DNA. The fragmented DNA was made blunt-ended through an end repair reaction that removed 3′ and 5′ overhangs. At the 3′ end of the blunt DNA molecule, a single adenosine was added with the ligation of Illumina adapters. The Agilent BioAnalyzer High Sensitivity DNA assay determined the fragment size, whereas the Qubit assay quantified each of the 16 purified libraries. Next, these libraries were diluted, multiplexed into 4 samples per lane, and sequenced using an Illumina HiSeq 2500, producing 100 bp paired-end shotgun reads.

Sequence Quality Filtering and Operational Taxonomic Unit Analysis

We quality filtered the resulting reads by removing all bases in each read after the first 3 successive bases having a phred quality score of less than 15 (Ewing et al., 1998). Any read pair with 1 member less than 85 bases in length (after the above filtering) or having an average phred quality score of less than 25 was completely removed from the analysis (Table 1). We compared the filtered reads to 16S rDNA sequences from the Ribosomal Database Project (RDP, release 10.24) using Bowtie: read pairs both having sequence matches with greater than 97% identity to 1 or more database sequences were retained (Cole et al., 2009). Operational taxonomic units (OTU) were defined using the sequences in the RDP itself: sequences in the database were connected by edges if their sequence identities were greater than 97% and OTU defined as connected components in this network (Ellison et al., 2014). We considered any of the above read pair hits that matched to sequences from one and only one such database-generated OTU to be an instance of that OTU.

Table 1.

Animal read statistics

Class Animal ID Total reads after quality filtering1 Number of reads matching MetaCyc2 Number of reads matching OTU3
Low RFI4 1 119,833,042 310,921 (0.26%) 1,192
2 115,699,275 307,898 (0.27%) 1,252
3 17,712,966 253,800 (1.43%) 849
4 16,336,193 183,859 (1.13%) 645
9 136,044,846 439,771 (0.32%) 1,562
10 129,561,726 323,016 (0.25%) 1,033
11 126,647,671 242,727 (0.19%) 828
12 83,677,521 220,034 (0.26%) 707
High RFI 5 119,789,299 320,006 (0.27%) 1,050
6 126,030,561 316,015 (0.25%) 1,133
7 131,776,487 375,399 (0.28%) 1,713
8 113,135,369 336,808 (0.30%) 1,149
13 202,547,203 706,605 (0.35%) 2,528
14 100,366,944 197,717 (0.20%) 686
15 101,143,440 528,514 (0.52%) 889
16 109,260,821 313,093 (0.29%) 1,101
Open in a new tab

1Total reads sequenced after quality filtering.

2Number and percentage of reads that were mapped to nodes from MetaCyc.

3Number of specific OTUs identified in the sequences. OTU = operational taxonomic unit.

4RFI = residual feed intake.

MetaCyc

MetaCyc is a comprehensive enzyme and metabolic pathway database that contains more than 2,491 metabolic pathways from 2,816 organisms, primarily bacteria and archeans but also some vertebrates (Caspi et al., 2014). The database describes enzymes, metabolites, and reactions derived from the literature. It thus includes a set of enzyme sequences, each of which is associated with 1 or more biochemical reactions that it is believed to catalyze. In our previous work, we described how the contents of this database were used to infer a metabolic network, where nodes represent metabolic reactions and edges connect nodes that share a metabolite. Reactions with identical metabolites were merged into a single reaction (Wolff et al., 2017). The result was a set of 6,140 reaction nodes that could be potentially observed in a community metabolic network. Isolated reactions were omitted from this network.

In this study, 2 metabolic networks were constructed: one for the ovine host and one for the ruminal microbes. The MetaCyc database does not include an ovine metabolic network: instead, we inferred one from those of humans and cattle. To do so, we used orthology inference to determine which of the enzymes for the 2,863 annotated reactions of the human metabolic network and the 1,404 reactions of the bovine network we could identify orthologs for in the ovine genome (Jiang et al., 2014). There were 15,562 human–sheep and 15,479 bovine–sheep ortholog pairs found with our synteny-based orthology program (Bekaert and Conant, 2011, 2014), which uses annotations from Ensembl release 75 (Flicek et al., 2014). The ovine metabolic network was constructed from these ortholog pairs under the assumption that the presence of an orthologous enzyme in sheep indicates the presence of the corresponding reaction. Because MetaCyc has no vertebrate reaction allowing the assimilation of butyrate, a known nutrient for ruminants (Masson and Phillipson, 1951), we added a pseudoreaction converting butyrate to butyryl-CoA to the sheep metabolic network. In total, we inferred 1,989 nodes in this ovine metabolic network: 1,834 from the human network, 154 from the bovine one, and the butyrate to butyryl-CoA reaction just described.

To prevent the presence of ubiquitous metabolites (often called “currency metabolites”) from creating a nearly fully connected network, we excluded the most common metabolites when constructing network edges, using 3 thresholds to do so: namely, metabolites that occurred in a total of more than 25, more than 50, or more than 100 reactions. We will hereafter refer to the resulting networks as N25, N50, and N100 (Pérez-Bercoff et al., 2011). For N25, N50, and N100, there were 261, 206, and 174 currency metabolites removed, respectively.

Matching Translated Metagenomic Reads to the Microbial Network

The Illumina data from the 16 samples were translated to AA sequences in all 6 possible open reading frames (ORF). Only paired AA sequences each having a translated ORF longer than 29 residues were retained and searched against the MetaCyc protein database. The sequence search against MetaCyc was made using a search tool based on the SeqAn library (Döring et al., 2008). We required two 7 residue identical word matches to seed a Smith–Waterman local alignment between a database sequence and the translated read (Smith and Waterman, 1981). A match between the read pair and database sequences was confirmed if both reads produced alignments having greater than 80% AA identity over 80% of the metagenomic ORF. We assigned such a read pair to 1 or more metabolic network nodes so long as all of the enzyme sequences that the pair matched to belonged to the same overlapping set of nodes. Translated read pairs were not mapped if they matched 2 or more reactions with nonoverlapping annotated enzymes: see (Wolff et al., 2017) for further details.

Inference of the Interface Between the Host and Microbial Metabolic Networks

To model the exchange of metabolites between microbes and the host, we defined sets of interface metabolites that were likely to be absorbed from the rumen by the host animal. Because the details of this transfer are not fully known, we used 3 different metabolite sets. The first set (VFA) is composed of the most common VFA in the rumen: acetate, butyrate, and propionate (Bergman, 1990). Other VFA are unfortunately not annotated in MetaCyc. The second interface set (VFA + AA) was defined by adding the 20 universal AA to the VFA set (for a total of 23 metabolites). The third interface set was composed of all the metabolites in the VFA + AA set with the addition of a subset of 234 metabolites, which are known to be absorbed in the human gut and which were taken from the global reconstruction of the human metabolic network (Duarte et al., 2007). These compounds were matched to MetaCyc, and inorganic compounds, the DNA/RNA nucleotides, carbon dioxide, and NAD+ and NADP+ were removed. The remaining 204 metabolites were added to the VFA and AA to yield the 227 metabolites in the “ALL” set. The complete list of these metabolites is available as supplementary material from our previous article (Taxis et al., 2015). The host and microbial networks were linked by adding edges between pairs of nodes, one from each network, that each use the same interface metabolite (Fig. 1).

Figure 1.

Figure 1.

Open in a new tab

The interaction between the host (purple, left) and microbial metabolic networks (right) for efficient and inefficient sheep. Each node (circle) represents an enzyme-catalyzed reaction: edges (lines) connect pairs of reactions that share metabolites. The colored lines in the center are interface metabolites that are potentially transferred from the microbes to the host animal and hence connect the 2 metabolic networks (Materials and Methods). For the illustrated network, these metabolites consisted of the 20 universal AA and 3 VFA that are the primary nutrients of the host (VFA + AA: Materials and Methods). The 10 interface metabolites with microbial reactions with the largest number of mapped reads are shown with individual colors (upper left). The upper half of each microbial node is colored based on the relative abundance reads mapping to that node for the low RFI animals, whereas the lower half of the node is similarly colored for the high RFI animals: see the scale diagram at the lower left. There were no metabolic reactions that showed significantly different abundance between the 2 RFI groups (P > 0.05, Wilcoxon test).

Metabolic Network Layers

The creation of this single metabolic network allows for the calculation of distances between reactions in the host and microbial metabolic networks. Dijkstra’s algorithm was used to determine the shortest path between every pair of nodes in the merged metabolic network (Dijkstra, 1959). A node’s distance to the other subnetwork then defined as its distance to the closest node in that subnetwork. Hence, nodes using interface metabolites are at the minimal distance of 0, with other nodes being successively more distant. The result of this distance definition is to decompose the merged metabolic network into the layers shown in Fig. 1.

Node Read Density Analysis

We used a 2-sample Wilcoxon test to assess if there were significant differences in the normalized read counts between the high and low RFI animals for the nodes in the microbial network. Only nodes with at least one mapped read pair in each of the 2 RFI groups were analyzed. The P-values from the Wilcoxon test were subjected to a 5% false-discovery rate correction (Benjamini and Hochberg, 1995).

Network Structure Analysis

The mean network layers for mapped reads were calculated separately for the low and high RFI status animals for all possible combinations of the different networks (N25, N50, and N100) and the interface metabolites (3 × 3 = 9 comparisons). To determine whether there were significant differences between the mean layers of the 2 RFI classes, we used a randomization approach. First, the mapped reads for the low and high RFI groups were pooled together and then the same number of reads as were originally mapped to the high RFI animals were randomly selected and assigned to a new “pseudo-high” set, with the remaining reads assigned to the “pseudo-low” set. We then compared the difference in mean layer number for the 2 classes for each of 1,000 of these randomized datasets to the difference in mean layer observed in the actual data. If the real difference was larger than all but 5% of the randomized datasets, we considered this evidence that the RFI classes differed in mean layer.

The association of RFI status and metabolic network structure was also assessed using several network statistics: “carbon sum,” betweenness centrality, node degree, and clustering coefficient. We define carbon sum to be the sum of the number of carbon atoms present in a reaction’s suite of products and reactants. Betweenness centrality is defined as the number of shortest paths between all pairs of nodes in a network that pass through the specified node (Freeman et al., 1991). The clustering coefficient measures the proportion of a node’s neighbors that are also each other’s neighbors (Watts and Strogatz, 1998), and node degree is simply the total edge count for a node (Watts and Strogatz, 1998). For each node and RFI class, the 4 statistics were weighted by the normalized number of mapped reads for that class and node, and the mean of this weighted value computed over all nodes. The weighted differences in these statistics between the 2 RFI classes were then compared to weighted differences from the randomized networks just described. If 5% or fewer of the randomized networks displayed a difference in the statistic in question as large as or larger than that observed in the actual data, we concluded that the RFI classes differed for that statistic (Table 2).

Table 2.

Network structure

Statistic1 Network2 Low3 High3 Real difference (low–high)4 Mean random difference5 P a
Carbon sum N25 33.07 33.15 0.08 0.03 0.021
N50 33.07 33.15 0.08 0.03 0.022
N100 33.07 33.15 0.08 0.03 0.020
Betweenness centrality N25 8,938.85 8,970.73 31.88 19.22 0.196
N50 18,090.94 18,093.78 2.84 34.03 0.954
N100 18,136.57 18,115.52 21.05 31.32 0.591
Degree N25 3.52 3.53 0.01 0.004 0.187
N50 9.17 9.19 0.02 0.01 0.119
N100 12.08 12.14 0.06 0.01 <0.001
Clustering coefficient N25 0.88 0.88 <0.001 <0.001 0.075
N50 0.86 0.86 <0.001 <0.001 0.622
N100 0.85 0.85 <0.001 <0.001 0.919
Open in a new tab

a P-value for the test of the null hypothesis that there is no difference in the mean statistic. P was computed by randomly assigning reads to RFI groups and comparing the difference in the 2 mean values for 1,000 such randomizations to the actual data (see Materials and Methods). P-values < 0.05 were judged significant and are shown in bold. RFI = residual feed intake.

1Network statistic compared between the low and high RFI animals. Carbon sum: total number of carbon atoms in a reaction. Betweenness centrality: number of shortest paths crossing a node. Degree: number of edges of a node. Clustering coefficient: number of fully connected triangles a node participates in relative to its total number of edges.

2Network defined by currency metabolite removal (Materials and Methods).

3Mean value for selected network statistic mapped for low or high RFI animals.

4Difference between the mean statistic for low and high RFI animals.

5Mean random difference between low and high RFI groups when reads are randomly assigned.

Principal Component Analysis

Principal component analysis (PCA) was performed with the R princomp() (Team, 2008) function to explore the patterns of variation using the normalized read or OTU counts from the 16 animals.

Shannon Entropy of OTU and Node Distributions

The mean Shannon entropies (Shannon, 1948) of the node and OTU distributions were calculated for the high and low RFI animals. To assess the differences between the RFI classes in these entropy measures, the differences between classes were again compared with the differences seen for networks where reads were randomly assigned to animals without reference to their RFI status.

RESULTS

Metagenomic Sequencing and Community Metabolic Network Interference

Using an Illumina HiSeq 2500, we shotgun sequenced microbial DNA from rumen fluids sampled from 8 animals judged to be efficient in their feed usage (low RFI) and 8 animals judged to be inefficient (high RFI). After quality filtering, we obtained approximately 1.75 billion 100 bp paired-end reads (Table 1).

From the MetaCyc database (Caspi et al., 2014), we inferred pair of reaction-centric metabolic networks that model the flow of metabolites between the microbes and the host. Interface metabolites, defined as compounds produced by the microbes and potentially absorbed by the host as nutrients, define the interface between these 2 networks (Materials and Methods). We also employed 3 different stringencies for removing currency metabolites such as ATP: from high to low stringency, these networks are defined as N25, N50, and N100, respectively. Starting from MetaCyc’s database of enzyme sequences, we used sequence similarity searches against the metagenomic read data to infer which enzymes (and hence metabolic reactions) were present in each rumen sample. For each such sample, metagenomic reads were mapped to the microbial metabolic network, which was connected to the inferred sheep metabolic network by 1 of 3 predefined sets of interface metabolites, denoted as VFA, VFA + AA, and ALL (Materials and Methods). Network N50 with the inference set VFA + AA is illustrated in Fig. 1.

Nodes Do Not Differ by RFI in Their Proportion of Mapped Reads

We explored whether differences in RFI were associated with differences in enzyme abundances between the 2 efficiency groups. In Fig. 1, the upper half of each microbial node (right-hand network) is colored by relative read abundance in the low RFI animals and the lower half by abundance in high RFI animals. Visually, there are very few nodes with any apparent difference in relative enzyme-coding gene abundance between the 2 groups. To formally test whether such differences were present, we applied a Wilcoxon test. There were no metabolic reactions with differing read abundance between the low and high RFI animals (P > 0.05). Differences in the microbiomes of low and high RFI animals are hence not apparent at the level of single reactions.

Low RFI Animals Employ Reactions Nearer to the Host Network and Involving Shorter Carbon Chains

We next tested 4 network-level statistics for differences between the 2 RFI groups. The average number of carbon atoms involved in a reaction was significantly greater in the high RFI animals than in the low RFI animals (Table 2; P < 0.05: see also Supplementary Fig. 1). On the other hand, the 2 diet classes did not differ on average in their betweenness centrality or clustering coefficients (Table 2; P > 0.05), with a weak difference in node degree seen only for network N100.

However, although little difference was seen between the low and high RFI animals for individual nodes and for most network statistics, the gross structure of the microbial network does differ between the groups. In particular, the reads from the microbes found in the low RFI animals map to reactions that are significantly closer to the host metabolic network than do the reads from the high RFI (inefficient) animals (Table 3 and Supplementary Fig. 1). This result is in accordance with our hypothesis that low RFI animals have community metabolic networks that are able to interface more easily with that of the host animal because they use reactions that are nearer to it in network space.

Table 3.

RFI group and network position

Group1 Currency cutoff2 Mean layer: low RFI3 Mean layer: high RFI3 Real difference: mean layers4 Maximum random difference5 P a
VFA N25 1.976 1.974 0.003 0.009 0.304
N50 1.869 1.882 0.013 0.008 <0.001
N100 1.772 1.787 0.016 0.006 <0.001
VFA + AA N25 0.798 0.805 0.007 0.004 <0.001
N50 0.863 0.872 0.009 0.004 <0.001
N100 0.899 0.910 0.011 0.004 <0.001
ALL N25 0.504 0.507 0.003 0.003 <0.001
N50 0.456 0.457 0.001 0.002 0.313
N100 0.500 0.502 0.002 0.002 <0.001
Open in a new tab

a P-value for the test of the null hypothesis that there is no difference in the mean layer of the 2 RFI groups. Reads were randomly assigned to RFI status and mean layers were computed 1,000 times and the difference in these random networks was compared with the real network (Materials and Methods). P-values < 0.05 were judged significant and are shown in bold.

1Interface metabolite set.

2Network (see Table 2).

3Mean layer number for reads mapped from low or high RFI animals. RFI = residual feed intake.

4Difference between the mean layer for low and high RFI.

5Maximum random difference between the mean layer for low and high RFI.

Principal Component Analysis

We used PCA to visualize the differences between animals in the pattern of OTU and nodes mapped in their microbiomes. For the taxonomic analysis, it generally appears that low RFI animals cluster together, whereas the high RFI animals separate into 2 groups (Fig. 2A). However, there was no clear difference between the low and high RFI animals when the normalized read count per metabolic network node was used as the underlying data for the PCA (Fig. 2B).

Figure 2.

Figure 2.

Open in a new tab

First 2 principal components (x- and y-axes, respectively) from PCAs of the OTUs (A) and normalized node abundances (B). The low RFI animals are shown in green, whereas the high RFI animals are in blue.

Shannon Entropy

We compared the Shannon entropies (Shannon, 1948) of the distributions of reads onto OTU and onto reactions for the low and for the high RFI animals. For the metabolic network nodes, the efficient animals showed significantly greater mean entropy that the inefficient ones (P < 0.001; Fig. 3B). For the taxonomic data, no significant difference in mean entropy between the 2 classes was seen (P = 0.073; Fig. 3B). For both the taxonomic and metabolic network measures, however, the mean entropy of the samples was smaller than that seen when reads were randomized within an RFI group (P < 0.001; Fig. 3A and C, see also Supplementary Fig. 2).

Figure 3.

Figure 3.

Open in a new tab

Shannon entropies of the distribution of microbial reads onto OTUs (x-axis) and onto metabolic network nodes (y-axis) for the low (green) and high (blue) RFI animals. (A) Distribution of randomized node-wise entropies for low (green) and high (blue) RFI animals (Materials and Methods). (B) Shannon entropies for OTUs and metabolic network nodes for each sampled animal. The mean Shannon entropies for the 2 classes are marked by the open squares. The mean node-wise entropy for the low RFI animals is significantly greater than that for the high RFI animals (P < 0.001). There is not a significant difference between the mean taxonomic entropy of the 2 classes (P = 0.073). In all cases, the mean entropy for the 2 classes is smaller than the mean entropies seen when reads were randomized among individuals within a class [histograms in (A) and (C); P < 0.001, Materials and Methods]. (C) As for (A) but with taxonomic values.

DISCUSSION

Understanding the degree to which feed efficiency differences are causally associated with the observed microbial differences between efficient and inefficient animals (Guan et al., 2008; Hernandez-Sanabria et al., 2010; Rius et al., 2012; McCann et al., 2014) is obviously of considerable importance. Unfortunately, making such links can be difficult (Hernandez-Sanabria et al., 2012). For instance, it appears efficient cattle produce less methane than do inefficient ones (Hegarty et al., 2007). As methane production diverts the flow of carbon atoms toward methane and away from other reduced organic compounds that ruminants absorb for nutrition, a simple hypothesis would be that inefficient animals simply possess more methanogens (Pinares-Patiño et al., 2013). However, the actual situation considerably more complex and not mechanistically understood (Zhou et al., 2009). Shi et al. (2014) found that although sheep differing in methane production showed no statistical differences in the number genes from methane-production pathways present between animals with low and high methane production, the expression of those genes was higher in individuals with high methane production. We suggest that the complexity of these results suggest the need for systems-level studies of the ruminal ecosystem.

A second reason such approaches are needed is the limited association between taxonomic and functional variation in the microbiome (Moore et al., 2009; Doolittle and Booth, 2017). Indeed, it is possible that this disconnect between taxonomy and function may contribute to the difficulties seen when trying to use the taxonomy of the ruminal microbiome to predict feed efficiency (Hernandez-Sanabria et al., 2010; Rius et al., 2012). Thus, in our own work, we have shown that functional approaches employing metabolic networks uncover features of microbial ecosystems that are not apparent from taxonomic analyses (Taxis et al., 2015). We have also used this metabolic network approach to show that a diet shift in a group of sheep initially fed a forage-like diet induced complex changes in the community metabolic networks of their rumens. In particular, when switched to a grain-based diet, the set of microbial reactions present shifted to favor those that are closer in network space to the reactions of the host sheep (Wolff et al., 2017). Likewise, we found differences in almost every aspect of the networks induced by the 2 diets, from the abundances of reads mapped to individual nodes to the overall network structure. These results induced us to conduct the work just described, exploring whether network analyses of the microbiome could uncover new insights as to the linkage between the microbiome and feed efficiency.

Given the importance of feed efficiency, it would be very desirable to find equally dramatic differences between high and low RFI animals. Unfortunately, such is not the case: when we considered standard comparisons of node abundances and network statistics, we generally found no significant difference between the RFI classes. Indeed, for no single metabolic network node was there a statistically significant difference between the RFI classes after multiple test corrections. Likewise, our PCA did not suggest that RFI status was a major differentiating factor between the animals (Fig. 2B).

However, when we used a set of analyses that were more explicitly based on hypotheses regarding why inefficient and efficient animals might differ in their ruminal microbiomes, the picture was strikingly different. Ruminants obtain nutritional energy from a relatively small number of compounds they absorb from their symbiotic microbes. When we structured the metabolic network to account for this fact, we saw that reads from the low RFI animals mapped nearer (on average) to the host metabolic network that did those from high RFI animals. When coupled to the fact that low RFI animals also on average used reactions with shorter carbon chains than did high RFI animals, a picture emerges of RFI being partly a function of how well the microbial ecosystem is structured to allow the host animal to extract metabolites from it. This proposal is further supported by the analyses of the diversity (Shannon entropy) of the metabolic networks found in the 2 groups, where we found that low RFI ewes had a greater mean node entropy than did high RFI ones. This observation is actually the converse of the pattern observed in obese humans, whereby the individuals who gain weight more easily have reduced taxonomic diversity (Turnbaugh et al., 2009). Several studies have also shown that a decrease in microbial diversity is linked to digestive diseases such as Crohn’s disease (Sha et al., 2013) and irritable bowel syndrome (Carroll et al., 2012) in humans. The increased application of antibiotics in animal production has resulted in decreased resilience in the gut microbiota due to reduced biodiversity (Mosca et al., 2016). However, microbiomes with reduced diversity are associated with poor health more generally, suggesting low diversity may indeed be the pathological case (Cotillard et al., 2013). These observations are contradicted to a degree, in the specific case of cattle fed grain-based diets, by the observation that low taxonomic diversity is associated with higher feed efficiency (Shabat et al., 2016; Li and Guan, 2017). However, we suspect that this example is the exception rather than the rule, both because the diet in question is much less complex than the evolved bovine diet and because it focuses on taxonomy alone, whereas our results in Fig. 3 speak to functional diversity. We suggest that studies of the microbiome would benefit from critically considering the ecological research suggesting that complex, diverse ecosystems are more efficient and stable (Cardinale et al., 2002; Cotillard et al., 2013).

It is important to note that nothing in our data speak to causality (itself a challenging concept in complex systems like these; Wagner, 1999) with respect to feed efficiency. Host genetics could very well drive the microbiome differences we have detected, as could environmental differences or developmental trajectories (Li and Guan, 2017). Likewise, the structure of the microbiome will interact with host physiology in complex ways (such as changes in ruminal pH), and the causes and effects in these relationships may not be yet understood. The importance of studying the microbiome lies in improving our mechanistic understanding of the processes of ruminal fermentation and feed efficiency, regardless of the ultimate origins of these differences.

A few other caveats are also in order. Our analyses of the microbial metabolic networks were limited to known and identifiable enzymes, and it is possible that the story would grow in complexity with a more complete view of microbial metabolism. Likewise, the linkage between ecosystem diversity and efficiency is still debated (Fridley, 2001). However, we note that several factors increase our confidence in our results. In all 3 cases, the significant differences between the 2 RFI classes were in the direction we predicted a priori: namely, efficient animals should use shorter carbon chain metabolites that are near to the host network and should show increased diversity in their networks. Likewise, the mean entropies of the actual samples are lower than that seen when reads were randomized within RFI class, as would be expected if there are further rules structuring these microbial communities such as patterns of required taxa co-occurrence (Carr and Borenstein, 2014; Wolff et al., 2017).

Synthesizing from our observations, we propose that it may be unhelpful to view inefficiency in feed usage as a unitary trait. Figure 3 and Supplementary Figs 1 and 2 suggest that inefficient animals show a good deal of animal-to-animal variation. From a system perspective, this fact may not be surprising: constructing an ecosystem that allows efficient growth is likely subject to errors in many dimensions: perhaps some inefficient animals have ecosystems that are both nondiverse and far from the host network (e.g., animals 15 and 16 in Supplementary Fig. 2), whereas others are well positioned but use longer carbon chains that would be optimal (animal 7 in Supplementary Fig. 1). To gain a better understanding of the gut microbiome, we argue that a system perspective is invaluable (Consortium, 2012; Franzosa et al., 2014), both as a means of generating hypotheses and as a starting point for developing interventions. Such perspectives will also allow for the development of predictive models providing mechanistic insights into the sources of inefficiency in feed use (Karlsson et al., 2011).

SUPPLEMENTARY DATA

Supplementary data are available at Journal of Animal Science online.

Supplementary Data
Click here for additional data file. (96.6KB, pdf)

Footnotes

1

This project was supported by USDA National Research Initiative (NRI) grant no. 2011-68006-30185 (M.J.E., W.R.L., K.M.C., G.C.C.) and National Science Foundation grant no. NSF-DBI-1358997 (C.S., A.M.W., G.C.C.).

LITERATURE CITED

  1. Bekaert M., and Conant G. C.. 2011. Copy number alterations among mammalian enzymes cluster in the metabolic network. Mol. Biol. Evol. 28:1111–1121. doi:10.1093/molbev/msq296 [DOI] [PubMed] [Google Scholar]
  2. Bekaert M., and Conant G. C.. 2014. Gene duplication and phenotypic changes in the evolution of mammalian metabolic networks. PLoS ONE 9:e87115. doi:10.1371/journal.pone.0087115 [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Benjamini Y., and Hochberg Y.. 1995. Controlling the false discovery rate: A practical and powerful approach to multiple testing. J. R. Stat. Soc Ser B (Methodol.) 57:289–300. [Google Scholar]
  4. Bergman E. N. 1990. Energy contributions of volatile fatty acids from the gastrointestinal tract in various species. Physiol. Rev. 70:567–590. doi:10.1152/physrev.1990.70.2.567 [DOI] [PubMed] [Google Scholar]
  5. Cammack K. M., Leymaster K. A., Jenkins T. G., and Nielsen M. K.. 2005. Estimates of genetic parameters for feed intake, feeding behavior, and daily gain in composite ram lambs. J. Anim. Sci. 83:777–785. doi:10.2527/2005.834777x [DOI] [PubMed] [Google Scholar]
  6. Cardinale B. J., M. A. Palmer, and Collins S. L.. 2002. Species diversity enhances ecosystem functioning through interspecific facilitation. Nature 415:426–429. doi:10.1038/415426a [DOI] [PubMed] [Google Scholar]
  7. Carr R. and Borenstein E.. 2014. Comparative analysis of functional metagenomic annotation and the mappability of short reads. PLoS ONE 9:e105776. doi:10.1371/journal.pone.0105776 [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Carroll I. M., Ringel-Kulka T., Siddle J. P., and Ringel Y.. 2012. Alterations in composition and diversity of the intestinal microbiota in patients with diarrhea-predominant irritable bowel syndrome. Neurogastroenterol. Motil. 24:521–530, e248. doi:10.1111/j.1365-2982.2012.01891.x [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Caspi R., T. Altman R. Billington K. Dreher H. Foerster C. A. Fulcher T. A. Holland I. M. Keseler A. Kothari A. Kubo, et al. 2014. The metacyc database of metabolic pathways and enzymes and the biocyc collection of pathway/genome databases. Nucleic Acids Res. 42:D459–D471. doi:10.1093/nar/gkt1103 [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Cole J. R., Q. Wang E. Cardenas J. Fish B. Chai R. J. Farris A. S. Kulam-Syed-Mohideen D. M. McGarrell T. Marsh G. M. Garrity, et al. 2009. The ribosomal database project: Improved alignments and new tools for rRNA analysis. Nucleic Acids Res. 37:D141–D145. doi:10.1093/nar/gkn879 [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Consortium H. M. P. 2012. Structure, function and diversity of the healthy human microbiome. Nature 486:207–214. doi:10.1038/nature11234 [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Cotillard A., S. P. Kennedy L. C. Kong E. Prifti N. Pons E. Le Chatelier M. Almeida B. Quinquis F. Levenez N. Galleron, et al. ; ANR MicroObes Consortium. 2013. Dietary intervention impact on gut microbial gene richness. Nature 500:585–588. doi:10.1038/nature12480 [DOI] [PubMed] [Google Scholar]
  13. Dijkstra E. W. 1959. A note on two problems in connexion with graphs. Numer. Math. 1:269–271. doi:10.1007/BF01386390 [Google Scholar]
  14. Doolittle W., and Booth A.. 2017. It’s the song, not the singer: An exploration of holobiosis and evolutionary theory. Biol. Philos. 32:5–24. doi:10.1007/s10539-016-9542-2 [Google Scholar]
  15. Döring A., D. Weese T. Rausch, and Reinert K.. 2008. Seqan an efficient, generic C++ library for sequence analysis. BMC Bioinformatics 9:11. doi:10.1186/1471-2105-9-11 [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Duarte N. C., S. A. Becker N. Jamshidi I. Thiele M. L. Mo T. D. Vo R. Srivas, and Palsson B. Ø.. 2007. Global reconstruction of the human metabolic network based on genomic and bibliomic data. Proc. Natl. Acad. Sci. U.S.A. 104:1777–1782. doi:10.1073/pnas.0610772104 [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Ellison M. J., Conant G. C., Cockrum R. R., Austin K. J., Truong H., Becchi M., Lamberson W. R., and Cammack K. M.. 2014. Diet alters both the structure and taxonomy of the ovine gut microbial ecosystem. DNA Res. 21:115–125. doi:10.1093/dnares/dst044 [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Ellison M., Conant G., Lamberson W., Cockrum R., Austin K., Rule D., and Cammack K.. 2017. Diet and feed efficiency status affect rumen microbial profiles of sheep. Small Rumin. Res. 156:12–19. doi:10.1016/ j.smallrumres.2017.08.009 [Google Scholar]
  19. Ewing B., L. Hillier M. C. Wendl, and Green P.. 1998. Base-calling of automated sequencer traces using phred. I. Accuracy assessment. Genome. Res. 8:175–185. doi:10.1101/gr.8.3.175 [DOI] [PubMed] [Google Scholar]
  20. Flicek P., Amode M. R., Barrell D., Beal K., Billis K., Brent S., Carvalho-Silva D., Clapham P., Coates G., Fitzgerald S., et al. 2014. Ensembl 2014. Nucleic Acids Res. 42:D749–D755. doi:10.1093/nar/gkt1196 [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Franzosa E. A., Morgan X. C., Segata N., Waldron L., Reyes J., Earl A. M., Giannoukos G., Boylan M. R., Ciulla D., Gevers D., et al. 2014. Relating the metatranscriptome and metagenome of the human gut. Proc. Natl. Acad. Sci. U.S.A. 111:E2329–E2338. doi:10.1073/pnas.1319284111 [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Freeman L. C., Borgatti S. P., and White D. R.. 1991. Centrality in valued graphs: A measure of betweenness based on network flow. Soc. Netw. 13:141–154. doi:10.1016/0378-8733(91)90017-N [Google Scholar]
  23. Fridley J. D. 2001. The influence of species diversity on ecosystem productivity: How, where, and why?Oikos 93:514–526. doi:10.1034/j.1600-0706.2001.930318.x [Google Scholar]
  24. Guan L. L., Nkrumah J. D., Basarab J. A., and Moore S. S.. 2008. Linkage of microbial ecology to phenotype: Correlation of rumen microbial ecology to cattle’s feed efficiency. FEMS Microbiol. Lett. 288:85–91. doi:10.1111/j.1574-6968.2008.01343.x [DOI] [PubMed] [Google Scholar]
  25. Hegarty R. S., J. P. Goopy R. M. Herd, and McCorkell B.. 2007. Cattle selected for lower residual feed intake have reduced daily methane production. J. Anim. Sci. 85:1479–1486. doi:10.2527/jas.2006-236 [DOI] [PubMed] [Google Scholar]
  26. Hernandez-Sanabria E., L. A. Goonewardene Z. Wang O. N. Durunna S. S. Moore, and Guan L. L.. 2012. Impact of feed efficiency and diet on adaptive variations in the bacterial community in the rumen fluid of cattle. Appl. Environ. Microbiol. 78:1203–1214. doi:10.1128/AEM.05114-11 [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Hernandez-Sanabria E., L. L. Guan L. A. Goonewardene M. Li D. F. Mujibi P. Stothard S. S. Moore, and Leon-Quintero M. C.. 2010. Correlation of particular bacterial PCR-denaturing gradient gel electrophoresis patterns with bovine ruminal fermentation parameters and feed efficiency traits. Appl. Environ. Microbiol. 76:6338–6350. doi:10.1128/AEM.01052-10 [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Herrero M., P. Havlík H. Valin A. Notenbaert M. C. Rufino P. K. Thornton M. Blümmel F. Weiss D. Grace, and Obersteiner M.. 2013. Biomass use, production, feed efficiencies, and greenhouse gas emissions from global livestock systems. Proc. Natl. Acad. Sci. U.S.A. 110:20888–20893. doi:10.1073/pnas.1308149110 [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Jiang Y., Xie M., Chen W., Talbot R., Maddox J. F., Faraut T., Wu C., Muzny D. M., Li Y., Zhang W., et al. 2014. The sheep genome illuminates biology of the rumen and lipid metabolism. Science 344:1168–1173. doi:10.1126/science.1252806 [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Karlsson F. H., I. Nookaew D. Petranovic, and Nielsen J.. 2011. Prospects for systems biology and modeling of the gut microbiome. Trends Biotechnol. 29:251–258. doi:10.1016/j.tibtech.2011.01.009 [DOI] [PubMed] [Google Scholar]
  31. Lacroix V., L. Cottret P. Thébault, and Sagot M. F.. 2008. An introduction to metabolic networks and their structural analysis. IEEE/ACM Trans. Comput. Biol. Bioinform. 5:594–617. doi:10.1109/TCBB.2008.79 [DOI] [PubMed] [Google Scholar]
  32. Li F., and Guan L. L.. 2017. Metatranscriptomic profiling reveals linkages between the active rumen microbiome and feed efficiency in beef cattle. Appl. Environ. Microbiol. 83:e00061-17. doi:10.1128/AEM.00061-17 [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Masson M. J., and Phillipson A. T.. 1951. The absorption of acetate, propionate and butyrate from the rumen of sheep. J. Physiol. 113:189–206. [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. McCann J. C., L. M. Wiley T. D. Forbes F. M. Rouquette Jr., and Tedeschi L. O.. 2014. Relationship between the rumen microbiome and residual feed intake-efficiency of Brahman bulls stocked on Bermudagrass pastures. PLoS ONE 9:e91864. doi:10.1371/journal.pone.0091864 [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Moore S. S., F. D. Mujibi, and Sherman E. L.. 2009. Molecular basis for residual feed intake in beef cattle. J. Anim. Sci. 87(Suppl 14):E41–E47. doi:10.2527/jas.2008-1418 [DOI] [PubMed] [Google Scholar]
  36. Mosca A., M. Leclerc, and Hugot J. P.. 2016. Gut microbiota diversity and human diseases: Should we reintroduce key predators in our ecosystem?Front. Microbiol. 7:455. doi:10.3389/fmicb.2016.00455 [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Pérez-Bercoff Å., A. McLysaght, and Conant G. C.. 2011. Patterns of indirect protein interactions suggest a spatial organization to metabolism. Mol. Biosyst. 7:3056–3064. doi:10.1039/c1mb05168g [DOI] [PubMed] [Google Scholar]
  38. Pinares-Patiño C. S., Hickey S. M., Young E. A., Dodds K. G., MacLean S., Molano G., Sandoval E., Kjestrup H., Harland R., Hunt C., et al. 2013. Heritability estimates of methane emissions from sheep. Animal 7(Suppl 2):316–321. doi:10.1017/S1751731113000864 [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Rius A. G., S. Kittelmann K. A. Macdonald G. C. Waghorn P. H. Janssen, and Sikkema E.. 2012. Nitrogen metabolism and rumen microbial enumeration in lactating cows with divergent residual feed intake fed high-digestibility pasture. J. Dairy Sci. 95:5024–5034. doi:10.3168/jds.2012-5392 [DOI] [PubMed] [Google Scholar]
  40. Sha S., B. Xu X. Wang Y. Zhang H. Wang X. Kong H. Zhu, and Wu K.. 2013. The biodiversity and composition of the dominant fecal microbiota in patients with inflammatory bowel disease. Diagn. Microbiol. Infect. Dis. 75:245–251. doi:10.1016/ j.diagmicrobio.2012.11.022 [DOI] [PubMed] [Google Scholar]
  41. Shabat S. K., G. Sasson A. Doron-Faigenboim T. Durman S. Yaacoby M. E. Berg Miller B. A. White N. Shterzer, and Mizrahi I.. 2016. Specific microbiome-dependent mechanisms underlie the energy harvest efficiency of ruminants. ISME J. 10:2958–2972. doi:10.1038/ismej.2016.62 [DOI] [PMC free article] [PubMed] [Google Scholar]
  42. Shannon C. 1948. A mathematical theory of communication. AT&T Tech. J. 27:379–423 and 623–656. doi:10.1002/j.1538-7305.1948.tb01338.x [Google Scholar]
  43. Shi W., C. D. Moon S. C. Leahy D. Kang J. Froula S. Kittelmann C. Fan S. Deutsch D. Gagic H. Seedorf, et al. 2014. Methane yield phenotypes linked to differential gene expression in the sheep rumen microbiome. Genome Res. 24:1517–1525. doi:10.1101/gr.168245.113 [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Smith T. F. and Waterman M. S.. 1981. Identification of common molecular subsequences. J. Mol. Biol. 147:195–197. doi:10.1016/0022-2836(81)90087-5 [DOI] [PubMed] [Google Scholar]
  45. Stevens C. E. and Hume I. D.. 1998. Contributions of microbes in vertebrate gastrointestinal tract to production and conservation of nutrients. Physiol. Rev. 78:393–427. doi:10.1152/physrev.1998.78.2.393 [DOI] [PubMed] [Google Scholar]
  46. Taxis T. M., S. Wolff S. J. Gregg N. O. Minton C. Zhang J. Dai R. D. Schnabel J. F. Taylor M. S. Kerley J. C. Pires, et al. 2015. The players may change but the game remains: Network analyses of ruminal microbiomes suggest taxonomic differences mask functional similarity. Nucleic Acids Res. 43:9600–9612. doi:10.1093/nar/gkv973 [DOI] [PMC free article] [PubMed] [Google Scholar]
  47. Team R. D. C. 2008. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. [Google Scholar]
  48. Turnbaugh P. J., M. Hamady T. Yatsunenko B. L. Cantarel A. Duncan R. E. Ley M. L. Sogin W. J. Jones B. A. Roe J. P. Affourtit, et al. 2009. A core gut microbiome in obese and lean twins. Nature 457:480–484. doi:10.1038/nature07540 [DOI] [PMC free article] [PubMed] [Google Scholar]
  49. Wagner A. 1999. Causality in complex systems. Biol. Philos. 14:83–101. doi:10.1023/A:1006580900476 [Google Scholar]
  50. Watts D. J., and Strogatz S. H.. 1998. Collective dynamics of ‘small-world’ networks. Nature 393:440–442. doi:10.1038/30918 [DOI] [PubMed] [Google Scholar]
  51. Wolff S. M., Ellison M. J., Hao Y., Cockrum R. R., Austin K. J., Baraboo M., Burch K., Lee H. J., Maurer T., Patil R., et al. 2017. Diet shifts provoke complex and variable changes in the metabolic networks of the ruminal microbiome. Microbiome 5:60. doi:10.1186/s40168-017-0274-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
  52. Zhou M., Hernandez-Sanabria E., and Guan L. L.. 2009. Assessment of the microbial ecology of ruminal methanogens in cattle with different feed efficiencies. Appl. Environ. Microbiol. 75:6524–6533. doi:10.1128/AEM.02815-08 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary Data
Click here for additional data file. (96.6KB, pdf)

Articles from Journal of Animal Science are provided here courtesy of Oxford University Press

RESOURCES