Protozoal populations drive system-wide variation in the rumen microbiome

Abstract

While rapid progress has been made to characterize the bacterial and archaeal populations of the rumen microbiome, insight into how they interact with keystone protozoal species remains elusive. Here, we reveal two distinct system-wide rumen community types (RCT-A and RCT-B) that are not strongly associated with host phenotype nor genotype but instead linked to protozoal community patterns. We leveraged a series of multi-omic datasets to show that the dominant Epidinium spp. in animals with RCT-B employ a plethora of fiber-degrading enzymes that present enriched Prevotella spp. a favorable carbon landscape to forage upon. Conversely, animals with RCT-A, dominated by genera Isotricha and Entodinium, harbor a more even distribution of fiber, protein, and amino acid metabolizers, reflected by higher detection of metabolites from both protozoal and bacterial activity. Our results indicate that microbiome variation across key protozoal and bacterial populations is interlinked, which should act as an important consideration for future development of microbiome-based technologies.

Introduction

The herbivore rumen is a highly specialized organ that has co-evolved in symbiosis with a complex microbiome, made up of thousands of microbial populations whose interactions collectively convert plant material into energy-yielding metabolites for the host. The rumen microbiome acts as an interface between the nutrient potential of the feed and the metabolism of the host animal, and includes members from all domains of life: Archaea, Bacteria, and Eukarya (ciliate protozoa and fungi)^1,2. From ingested plant material, cellulose, pectin, xylans, xyloglucans, and other polysaccharides are degraded by microbially encoded carbohydrate-active enzymes (CAZymes) down to their component monosaccharide units, which are subsequently fermented into several intermediates. Most importantly, pyruvate is converted to volatile fatty acids (VFAs) such as acetate, propionate, and butyrate³. Along this fermentation pathway, hydrogen (H₂) is produced, which predominantly flows into methanogenesis but can also be incorporated into VFAs through alternative hydrogen sinks such as the reduction of fumarate⁴. The rumen epithelial wall is able to transport most of the VFAs directly into the blood, whereas more complex metabolites take a longer path, being assimilated by the posterior gastrointestinal tract⁵.

Rumen microbiome structure and function is shaped by many dynamic host-associated variables, such as diet, age, health status, animal husbandry, behavior, and breed. Efforts to monitor and predict overall rumen microbiome function for the purpose of improved animal production have mainly focused on recovering isolates and genomes of the various microbial populations. However, the superior amenability of bacteria and archaea to current molecular microbiology techniques has created significant domain-specific information bias, with recovery of greater than 50,000 bacterial and archaeal genomes compared to ~50 for eukaryotic species^1,2. The ciliate protozoa, specifically the class Litostomatea, subclass Trichostomatia, have a relatively large biomass in the rumen (up to 50%¹), and are ubiquitous among ruminants. Although single-celled, they have complex organelles and physiological features, such as mouthlike adoral openings that lead to a tongue-like extrusible peristome, which ingests feed particles into an esophagus-like structure⁶. This, combined with their outside being covered with undulating cilia for propulsion, makes many of them voracious predators⁶. To add to their versatility some ciliate species are able to degrade plant fibers² while others can act as a microhabitat for archaea and bacteria⁷, especially Methanobrevibacter spp., which form metabolic mutualisms by recycling the H₂ produced by the ciliates as a main metabolic end product^8,9. Coexistence or exclusion patterns of protozoal species in ruminants have also been observed for over half a century, initially via microscopy-based studies suggesting that certain protozoal species are mutually exclusive, both in sheep and cattle¹⁰ Early microscopy-based studies were later supported by molecular phylogenetic marker genes studies^9,10,11, which supported the classification of protozoal populations into two community types: type A, defined by the presence of Polyplastron multivesiculatum, often accompanied by Diploplastron affine, and type B, characterized by Epidinium and Eudiplodinium spp., together or alone¹⁰.

Providing in vivo context to the wider ecological impacts of rumen protozoal populations has proven immensely challenging but is necessary to advance microbiome-based solutions to animal productivity and sustainability, for example in the context of methane mitigation and improving feed efficiency. Rapid advancement of biotechnological tools has improved the availability of data for resident rumen microbiota, yet information on how species interact within these multidomain ecosystems is still limited. In this study, we applied long-read metagenomics, existing single-cell amplified eukaryotic genomes, and genome-centric multi-omics of both host and its microbiome to improve resolution of inter-domain relationships and the influence they exert at a system level. Two breeds of cattle from a highly controlled experiment were phenotyped for key performance traits, while rumen contents, epithelial, and liver samples were analyzed across all molecular layers—genes, transcripts, proteins, and metabolites (Fig. 1). Taxonomic analysis identified two clear rumen microbiome structure types across the entire animal cohort that were not strongly correlated to breed, any of the recorded animal performance metrics, or methane emissions. Deeper analyses across the microbial domains confirmed two distinct protozoal population types that we hypothesize to drive system-wide microbiome differences, ultimately affecting the interlinked metabolisms that channel the flow of nutrients across the feed-microbiome-host axis.

Fig. 1: Experimental design of the animal trial and microbiome analyses.

a Animal trial setup. A total of 80 steers across two breeds (Aberdeen-Angus cross and Luing) were enrolled of which 71 completed the 3-6 month study period that culminated in their slaughter. Key performance traits such as average daily gain (ADG), dry matter intake (DMI), feed conversion ratio (FCR) and methane yield (g/kg DMI) were measured for all animals and rumen samples periodically collected across the duration of the trial. The rumen prokaryote community structure was determined using 16S rRNA gene analysis for all 71 animals at slaughter. b Multi-omic sampling and analysis design for a subset of 24 animals, selected on the widest recorded level of natural methane yield variation. At slaughter, three sample locations were collected: Rumen digesta, rumen wall tissue, and liver. Samples were characterized on several molecular layers: Genomes, transcripts, proteins, untargeted metabolomics. For a further subset of 6 animals, metagenomes, metatranscriptomes and metabolomes were analyzed across 5 temporal sampling points collected during the experimental period using nasogastric tubing. Temporal samples were collected prior to the adaptation phase, before and after the performance test as well as immediately after leaving the respiration chambers. c Key performance traits and other animal production metrics that were determined for all enrolled animals. IQR: interquartile range. Significant P values (Welch two-sample t-test; two-sided, without multiple comparison correction) are given in bold. Created in BioRender. Aho, V. (https://BioRender.com/pqc6vo3).

Results

One controlled animal experiment reveals two distinct rumen microbiome structure types

As part of an effort to improve the depth of understanding within the rumen microbiome, we analyzed samples from a controlled feedlot trial of adult beef cattle fed a total mixed ration of forage and concentrate (ratio: 51:49). From an initial 80 animals representing two breeds that commenced the trial, 36 Aberdeen-Angus cross (AAX) and 35 Luing animals completed the experimental period with all planned measurements, including key performance traits (KPTs) such as dry matter intake (DMI), average daily gain (ADG), feed conversion ratio (FCR) and methane yield (g/kg DMI). For microbiome analysis, rumen samples were taken for all 71 animals at five timepoints across the experimental period and subjected to 16S rRNA gene amplicon sequencing, with an additional final time point sampled at slaughter (Fig. 1a). A subset of 24 animals (12 AAX, 12 Luing) that represented natural levels of methane yield but additionally included those with highest and lowest measurements, were sampled across both the host and its microbiome at slaughter. The datasets generated from these 24 animals included long-read metagenomics for metagenome-assembled genome (MAG) reconstruction as well as RNA, protein and metabolite analysis of rumen digesta, rumen epithelia and liver tissue (Fig. 1b). As expected, the recorded KPTs showed breed-dependent differences in animal metrics, such as a higher liveweight and dry matter intake (DMI) in AAX animals, and a trend for higher methane emissions (g/kg DMI) in Luing animals (Fig. 1c).

For microbiome characterization, long-read metagenomic sequencing of rumen samples from the 24-animal subset produced a total of 700 high- and medium-quality dereplicated metagenome-assembled genomes (MAGs, 656 classified as bacterial, 44 as archaeal; Supplementary Data 1a). These sample-specific MAGs, together with previously published fungal genomes (n = 9)¹² and protozoal single amplified genomes (SAGs) (n = 53)^2,13, formed the reference database for metatranscriptomic and metaproteomic analyses. Rumen metatranscriptomics identified 1,669,849 expressed genes (1,299,827 from bacteria, 80,325 from archaea, 252,768 from protozoa, and 9529 from fungi), whereas metaproteomics identified 35,655 protein groups (16,823 from bacteria, 380 from archaea, 18,000 from protozoa, 137 from fungi, and 315 from the cattle host) (Supplementary Data 1b). To further assist our interpretations of host and microbial metabolic activity we generated untargeted metabolomic data from the three different sample types available (numbers of identified metabolites: rumen: 496; rumen epithelium: 517; liver: 859; Supplementary Data 1b). Finally, we performed VFA measurements from rumen fluid, as well as Microarray Polymer Profiling (MAPP) of rumen digesta, determining the composition and relative abundance of glycans available to the rumen microbiome.

Community structure analysis of the 71 animals using the 16S rRNA gene sequence data did not reveal clear associations for any alpha or beta diversity metrics with breed, methane yield, or any of the other measured animal KPT (Supplementary Fig. 1, and Supplementary Data 2). However, beta diversity plots illustrated two groups of animals whose prokaryote structure distinctly clustered together, which could also be captured using probabilistic modeling (Dirichlet Multinomial Mixtures¹⁴) (Fig. 2a). Surprisingly, these two clusters, hereafter referred to as Rumen Community Type-A and -B (RCT-A and RCT-B), did not correspond to any measured animal KPT nor to any technical grouping that arose from the experimental workflow (Supplementary Fig. 2a). Furthermore, these community types were stable across time: the animals consistently stayed in the same cluster over the six timepoints sampled during the experiment (Supplementary Fig. 2b). Using genome-scale metabolic models to predict the metabolic reaction abundances of archaea and bacteria, we found little variation in functional potential between RCT-A and RCT-B, supporting the importance of including all microbial domains (e.g., eukaryotes) as well as functional omics beyond genes and genomes in our analyses (Supplementary Fig. 3).

Fig. 2: Rumen microbiome analyses showing two distinct groups, labeled Rumen Community Type (RCT) -A and -B.

a, b Principal Coordinates Analysis of robust Aitchison distances of (a) 16S rRNA gene sequence data (n = 71), showing two optimal clusters from Dirichlet Multinomial Mixtures; b metagenomic data (n = 24). c–f Principal Component Analysis of microbiome-derived c metatranscriptomic, d metaproteomic, and e metabolomic data as well as f host-omic data, showing the first Principal Component (PC) with p < 0.1 (two-sided t-test, no multiple comparison correction) between clusters for each data type. Box hinges: 1st and 3rd quartiles, whiskers: hinge to highest and lowest values within 1.5*IQR of hinge. g Taxonomic classifications of features (transcripts or proteins) with the strongest contributions (i.e., loadings) to PC1 in digesta metatranscriptomics and metaproteomics, grouped by association with RCT-A (positive) or RCT-B (negative), showing top 100 features for each omic and RCT; “other” summarizes taxa with only one feature in the top 100. Data for this figure are available in Supplementary Data 3.

Curiously, RCT-A and -B were detectable across several omic layers in the 24-animal subset that we analyzed in more detail. Principal coordinates analysis (PCoA) of MAG abundances reflected the same pattern that was detected in the 16S rRNA gene sequence data (Fig. 2b). In Principal Component Analyses (PCA) of digesta and rumen wall epithelium metatranscriptomics as well as digesta metaproteomics, the first principal components (PCs) clearly differentiated RCT-A and -B, thus mirroring the 16S rRNA gene sequence analysis results of the entire animal cohort (Fig. 2c, d). The congruence between the molecular layers affirmed that elements of metabolism are affected by this distinct compositional difference in microbial communities of RCT-A and RCT-B. Untargeted metabolomics of digesta and the rumen wall epithelium also reflected this pattern on PCs 4 and 3, respectively (Fig. 2e). Finally, host proteomics and transcriptomics from wall and liver data showed a trend towards the two community types, although not always statistically significant (p < 0.05 only for PC15 from liver proteomics; 0.1 > p > 0.05 for other host data; Fig. 2f).

Protozoal patterns associate with rumen community types

To further investigate the defined RCT-A or -B types, we extended our analyses to include domains of life present within the rumen samples of this study, incorporating the archaeal and bacterial MAGs as well as the single amplified genomes (SAGs) for protozoal populations. The taxonomic classifications of the transcripts and proteins with the strongest contributions (i.e., loadings) to the significant principal component patterns that were observed (Fig. 2g, and Supplementary Fig. 4, Supplementary Data 3) clearly indicated that the RCT-A and -B clustering extended to the abundance profiles of detected protozoal species. Based on these and the differential abundance comparisons of metatranscriptomic and metaproteomic data, animals that exhibited the RCT-A pattern were enriched with families Entodiniinae and Isotrichidae, including high abundances of Entodinium bursa, Entodinium caudatum, Entodinium longinucleatum, and Isotricha intestinalis, as well as Ostracodinium gracile and Polyplastron multivesiculatum (Fig. 3, Supplementary Data 4). Conversely, animals with the RCT-B pattern were enriched with subfamilies Diplodiniinae and Ophryoscolecinae² and typically included the species Diplodinium dentatum, Epidinum cattanei, Epidinium caudatum, and Ophryoscolex caudatus (Fig. 3, and Supplementary Data 4).

Fig. 3: Differential abundances of taxa across omic approaches.

a, b Summaries of differential abundance comparisons between Rumen Community Types (RCTs). Numbers indicate taxa with p < 0.05 (out of total n of taxa). metaG: metagenomic relative abundances (n = 24, Wilcoxon rank-sum tests), metaT: metatranscriptome count data (n = 24, DESeq2), metaP: metaproteome label-free quantification (LFQ) intensities (n = 22, Wilcoxon rank-sum tests). All tests were two-sided and multiple comparison corrected with Benjamini-Hochberg false discovery rate. a Metagenome-assembled genomes (MAGs) of bacteria and archaea, showing genera with differentially abundant MAGs supported by at least two omic analyses in the same direction. b Single-amplified genomes (SAGs) of protozoa, showing species with at least two differentially abundant SAGs regardless of direction. c, d Abundances of protozoal species from (b). Diamonds: medians, whiskers: IQR. c Metatranscriptomic data (variance stabilizing transformed (VST) counts; n = 24), d Metaproteomic data (sums of LFQ intensities; n = 22). Data for this figure are available in Supplementary Data 4.

The protozoal abundance patterns observed in our RNA and protein data bore remarkable similarities to these community types first reported via microscopy and later validated via phylogenetic marker gene analysis^9,10,11. A noticeable difference between the classical community types and our results was the coexistence of P. multivesiculatum and Epidinium spp. in our RCT-B animals. Such coexistence has been suggested to constitute an AB community type^9,15,16, and postulated to represent a transitional stage from type B to A⁹. To explore the interrelationships of these protozoa, we examined the metatranscriptomes and metaproteomes of rumen samples collected from six animals over a period of three to six months. In line with the temporal stability of the bacterial and archaeal community structure (Supplementary Fig. 2b), our data indicated a constant, low but detectable presence of P. multivesiculatum together with Epidinium spp. in RCT-B animals (Supplementary Fig. 5). This brings to question whether these two protozoal genera are indeed mutually exclusive.

Protozoal community types affect bacterial and archaeal structure and function

We sought to better identify the microbial populations driving the observed system-wide variation, as well as its function implications, by further examining the metatranscriptome and metaproteome data. Differential expression analysis and the features with the strongest loadings in our abovementioned PCA analysis both highlighted that specific bacterial, archaeal and protozoal populations were indeed more prevalent in either RCT-A or -B (Figs. 2g, and 3). Collectively, for animals categorized as RCT-A, the metaproteomes from their rumen were largely dominated by Isotricha spp, Entodinium spp, and the clostridial lineage Acutalibacteraceae (RUG762) while transcriptomes for various Methanobrevibacter spp., Sodaliphilus spp., Faecousia spp. and Lachnospiraceae (UBA1066) were also prevalent. In contrast, both the metatranscriptome and metaproteome for rumen samples from RCT-B animals showed far higher detection of Epidinium spp., while metatranscriptomics also indicated an enrichment in Prevotella spp. (Fig. 3).

To link biology to these observed structural patterns, we explored the annotated functions of the differentially detected populations more deeply with specific attention to the key functional stages of rumen digestion, namely fiber hydrolysis, fermentation of organic material and production of energy-yielding volatile fatty acids (Fig. 4, and Supplementary Data 5). By far the most detectable fibrolytic population observed in RCT-B animals was Epidinium spp. which contained a plethora of carbohydrate-active enzymes (CAZymes) predicted to act upon cellulose, arabinoxylans, beta-mannans and arabinogalactan protein glycans commonly found in grasses and grains (Fig. 4b). Epidinia are the most reputable among the rumen ciliates to actively attach and degrade plant material, as visually confirmed across a series of prior studies¹⁷. In a scenario where epidinia are more proliferant and engaging in plant material deconstruction, it is reasonable to expect their activity and size will impact the glycan landscape that is available for neighboring microbial populations. Indeed, our MAPP analysis of rumen digesta particles was suggestive of differences in various beta-glucan, xylans, xyloglucans, and arabinogalactan proteins between the epidinia-dominated RCT-B animals and the RCT-A animals (Fig. 5a). In concert, many fiber-degrading bacterial lineages, such as Sodaliphilus and Prevotella spp., were additionally detected at higher levels in metatranscriptomic data arising from RCT-B animals (Fig. 3a), supporting our hypothesis that system-wide effects are likely defined by protozoal activity. While metaproteomic detection of central butyrate-producing enzymes was observed in both RCT’s, we suspect the elevated activity of epidinia species was influential towards higher butyrate levels in RCT-B animals (Fig. 4b, c), which is supported by prior meta-analysis of protozoa that showed defaunation will substantially decrease ruminal butyrate levels¹⁸.

Fig. 4: Functional differences between rumen community types.

a, b Summaries of KEGG Orthologs (KOs) in pathways of interest (based on²⁰) and CAZymes in (a) bacterial and (b) protozoal genera of interest (based on Fig. 2g). CAZymes and KOs were chosen by selecting the 10 pathways or substrates (right-side labels) including the largest amount of features with p < 0.1 between RCTs, then trimming the list to CAZymes and KOs (y-axis labels) detected as differentially abundant at least 5 times across each row. metaT: metatranscriptomics (n = 24; DESeq2), metaP: metaproteomics (n = 22; Wilcoxon rank-sum tests). “CAZy: multiple substrates” contains CAZymes that may act upon various linkages and/or substrates. c. Metabolomic comparisons: volatile fatty acids (VFAs; n = 21) and metabolites from untargeted metabolomics (annotation level 1 or 2a; n = 24) with p < 0.1 (Wilcoxon rank-sum tests). Diamonds: medians, whiskers: IQR. ** p < 0.01, *: p < 0.05, ·: 0.1 > p > 0.05. All tests (a–c) were two-sided and multiple comparison corrected with Benjamini-Hochberg false discovery rate. Data for this figure are available in Supplementary Data 5.

Fig. 5: Metabolic predictions from major populations strongly featured in RCT-A and -B animals that are predicted to influence rumen function.

a Epidinium cattanei, the protozoal species most strongly associated with RCT-B, predicted to engage a broad range of CAZymes to degrade plant material. Given the size and activity of E. cattanei their fiber-degrading metabolism is predicted to impact rumen structure and function of RCT-B animals, which was supported via Microarray Polymer Profiling (MAPP) of plant fibers. MAPP inset: n = 24; diamonds: medians, whiskers: IQR; ** p < 0.01, *: p < 0.05 (two-sided Wilcoxon rank-sum tests without multiple comparison correction); data for this plot are available in Supplementary Data 5e. b RUG762 (Acutalibacteraceae), strongly associated with RCT-A, highlighting sugar fermentation, amino acid (red boxes) metabolism and a partial Wood-Ljungdahl Pathway, supported by associated energy conservation machinery such as the electron-bifurcating hydrogenase (HndABCD, [FeFe] group A), ferredoxin:NAD-oxidoreductase (Rnf) complex, and components of a FoF1 ATP synthase. Bold text: differentially abundant metabolites from Fig. 4c. Created in BioRender. Aho, V. (https://BioRender.com/w9cdcqz).

In the absence of elevated epidinia metabolism within RCT-A animals, both PCA and differential abundance analyses indicated the primary responsibilities for carbohydrate digestion was shared more broadly across the protozoal species Entodinium spp. and P. multivesiculatum as well as bacteria affiliated to family Acutalibacteraceae or genera Prevotella, Sodaliphilus, Faecousia and Merdiplasma (Fig. 4a, b). The Isotricha species that dominated RCT-A animals were, as expected¹⁸, not primarily degraders of plant material, though we suspect their influence still impacted heavily upon bacterial populations. For example, populations affiliated to RUG762 (Acutalibacteraceae), had some of the strongest loadings for RCT-A animals within the metaproteomic PCA analysis (Fig. 2g) and similar to Isotricha species were consistently enriched in RCT-A (Fig. 3a). Functional annotation of RUG762 MAGs suggested that these populations were engaged largely in protein and amino acid metabolism, which was supported by metaproteomic enrichment of predicted RUG762 enzymes (Fig. 5b) and metabolites for aspartate, glutamine and branched chain amino acid metabolism in RCT-A animals (Figs. 4c, 5b). Fermentation end products were predicted to be acetate and possibly propionate and branched-chain volatile fatty acids, which were also detected at higher proportions in RCT-A animals (Fig. 4c). The protein and amino acids for ruminal metabolism could plausibly arise from the grain fraction of the animal’s diet (355 g/kg dry matter in the concentrate component). However, Isotricha spp. have been shown to excrete cellular nitrogen in the form of amino acids, principally alanine, proline, glutamic acid, and aspartic acid^17,19. If such excretion of amino acids indeed occurs in RCT-A animals dominated by Isotricha spp. our observations of elevated RUG762 metabolism are plausibly connected, though we acknowledge this hypothesis must be tested in future experiments that examine cellular proximity and nutrient transfer between these populations.

It was interesting to note that for RCT-A animals a grouping of Methanobrevibacter-affiliated populations were detected at significantly higher abundance and/or with PCA loadings clearly associating them with RCT-A, despite there not being significant differences in measured methane yield across the two groups of animals (Supplementary Fig. 2a). The holotrich Isotricha species have been repeatedly shown¹⁸ to associate with different methanogenic populations than entodiniomorphids (e.g., epidinia), and our data also followed this trend with Methanobrevibacter populations in epidinia-dominated RCT-B animals seemingly of distinct strains compared to RCT-A (Fig. 3a, and Supplementary Fig. 6). Functional examination of bacterial populations enriched in RCT-A animals (Fig. 3a) also identified several taxa, including Faecousia and Merdiplasma species²⁰, that were predicted to encode multimeric electron-bifurcating [FeFe] group A hydrogenases (HndABCD) as well as selected features putatively associated with the Wood-Ljungdahl pathway (WLP) (Fig. 4a, and Supplementary Fig. 7). The WLP potentially facilitates reductive acetogenesis and can act as an alternative hydrogen sink to methanogenesis. While reductive acetogens indeed co-exist with methanogens, under normal rumen conditions they are believed to be outcompeted energetically and thus are often observed in low abundance. However, in a methane-inhibited rumen with elevated hydrogen partial pressure, Faecousia and Merdiplasma species have been observed to flourish²¹. Indeed, aforementioned RUG762 populations were also suspected to encode a partial WLP as well as the associated energy conservation machinery such as the bidirectional electron-bifurcating hydrogenase (HndABCD), ferredoxin:NAD-oxidoreductase (Rnf) complex, and FoF1 ATP synthase (Fig. 5b). However, the absence of the acetyl-CoA synthase/carbon monoxide dehydrogenase complex that is required for the complete reduction of CO₂ to acetate, leads us to speculate that RUG762 populations are instead producing methionine via a cobalamin-dependent 5-methyltetrahydrofolate–homocysteine methyltransferase. In the context of higher methanogenic features in RCT-A animals, the non-differential methane yield levels observed across the RCT-A and -B animal are fascinating and are likely arising from as-yet undefined hydrogen flow that influences differing methanogen strains and other hydrogenotrophs in the rumen.

Implications for the host animal

Despite the distinct systems-wide microbiome shifts that were reconstructed for RCT-A and -B animals, we were surprised to observe only limited effects of these microbial community differences on host biology. Animal performance measurements (Supplementary Fig. 2a), microbial and host metabolomic data, and host expression data in gut epithelial and liver tissues showed only minor changes to a small number of features (Fig. 2). The clearest difference was the relative composition of several amino acids and VFAs, with propionate and branched chain volatile fatty acids higher in RCT-A animals, while butyrate levels were higher in RCT-B (Fig. 4c). Since VFAs are the major energy source for the host animal and are taken up directly through the rumen wall epithelium³, we further applied a series of network analyses using rumen and epithelial proteomic data to ascertain if underlying expression patterns were indeed evident between metabolically linked microbial and host pathways. From rumen metaproteomes, weighted gene correlation network analysis (WGCNA)²² identified a wide variety of co-expression modules (ME) that contained mixtures of protozoal, bacterial and archaeal proteins; many of these modules were, unsurprisingly, strongly correlated with the RCT variable (Supplementary Fig. 8, Supplementary Data 6). In the epithelial proteomics data, WGCNA identified only two co-expression modules, comprised largely of host proteins, that were correlated with the RCT groupings, none of which were enriched with proteins functionally inferred in VFA metabolism (Supplementary Fig. 9, Supplementary Data 6). Of note, interlinked patterns of rumen digesta (ME9 and ME13) and epithelial (ME1) modules were enriched in proteins annotated in cysteine and methionine metabolism and RUG762 populations suggesting possible metabolic interplay of amino acids, though this needs future testing for validation. The lack of striking host effects arising from microbiome differences between RCT-A and -B animals highlights the extraordinary plasticity and functional redundancy of the rumen microbiome.

Discussion

Rumen protozoa are large and complex compared to their bacterial and archaeal neighbors and their presence and distribution within the livestock rumen has been heavily documented for well over 130 years¹⁷. Despite their long-standing history their impact across the total rumen ecosystem remains poorly understood at a molecular level due to technical restrictions that have impeded their study, and which have only recently been overcome with omics methodologies. Herein we were excited to link the molecular patterns and functional interpretations in our data to community types first postulated over 60 years ago via light microscopy¹⁰. When first describing protozoal community types in 1962 J. Margaret Eadie concluded that inter-relationships of the type described may play an important role in determining the components of a particular rumen microfauna¹⁰. We show that for the animals in this study, the system-wide rumen microbiome structure indeed extended beyond the protozoal components originally proposed in community types A and B to encompass bacterial and archaeal populations.

Advancing the original Eadie hypothesis, our multi-layered omic datasets offered plausible interpretations on how two independent modes of metabolic interactions are interlinked across the rumen microbiome of RCT-A and -B animals. Of particular note was the seemingly direct influence certain protozoal species (e.g., Epidinium spp.) play at higher trophic levels such as fiber hydrolysis, which likely impacts fiber structural configuration and availability for bacterial fibrolytic populations. On the other hand, protozoal metabolism of Isotricha spp. was predicted to indirectly affect how nutrients enter the food chain via excretion of metabolites such as amino acids and hydrogen, which impacted the structure and function of intermediate fermenters. While this study goes some way into explaining the microbiome-wide effects that particular protozoa can exert, major questions regarding the origin of their structural configuration still remain. Our experimental design using endogenous microbiome data analyzed at slaughter does not enable definitive explanations whether protozoa or prokaryotic population structure act as the primary selective pressure for RCT structure, nor has it examined diurnal variations or the impact of predation, which is well known to shape bacterial populations^23,24,25,26. We also speculate the original protozoal seeding took place via animal-animal contact likely during early life transition that started with mother-calf contact and gradually extended to other animals across the greater herd. Unfortunately, behavioral data prior to animal enrollment and pen groupings used in this animal trial were not recorded, though it was clear that grouping of RCT-A and -B animals together in randomized pens had no immediate nor long term influence upon microbiome structure.

In this study we show that the acceleration in genome recovery of protozoal populations and their supplementation into rumen microbiome databases has massively impacted our ability to estimate the transdomain microbial trophic cascades that convert complex plant material into energy-yielding nutrients for the host animal. Moving forward, several outstanding knowledge gaps need to be prioritized so that greater microbiome resolution can be routinely gained. Laboratory-based experiments that validate both proximity and metabolic interactions between protozoa, bacteria and archaea will lead to improved interpretations of how protozoa modulate rumen biology and formulate tools to potentially intervene where desired. Furthermore, characterizing diurnal variations as well as protozoal predatory relationships will require a more tailored sampling and microbiome monitoring regime and/or in vitro experiments that control for the concentration of key protozoal species. Finally, more extensive surveys of increased animal numbers, varying diets, breeds and management practices will need to be analyzed at a depth comparative to the present study to ascertain the wider implications of protozoal-bacterial-archaeal interactions, and how that knowledge can be applied to improve microbiome modulation strategies that make meaningful impact.

Methods

Ethics statement

The animal experiment was conducted at the Beef and Sheep Research Center of Scotland’s Rural College (6 miles south of Edinburgh, UK). The experiment was approved by the Animal Experiment Committee of SRUC and was conducted in accordance with the requirements of the UK Animals (Scientific Procedures) Act 1986.

Experimental design and measurement of key performance traits

An initial group of 80 steers (castrated males) representing two breeds of Bos taurus beef cattle (Aberdeen-Angus cross (AAX, n = 40), and Luing (n = 40)) was selected for the experiment. Of these, 71 (AAX: n = 36; Luing: n = 35) successfully completed the designed sampling scheme. All animals were provided a typical basal diet consisting of whole crop barley (300 g/kg DM), grass silage (200 g/kg DM), barley (355 g/kg DM), maize dark grains (120 g/kg DM), molasses (15 g/kg DM) and minerals (10 g/kg DM). Feed was offered to animals on an ad libitum basis. For half of the animals, the experimental design originally involved supplementation with Asparagopsis taxiformis red algae vegetative tissue (thallus) at 0.3% of the organic mass (OM). A. taxiformis is a feed additive which has been shown to reduce methane emissions in past studies^27,28,29,30. However, due to adverse effects observed in animals during the planned three-week seaweed adaptation phase, supplementation was terminated after just 15 days. All animals were given a further 5 weeks to adapt to basal feed before performance testing was carried out. Temporal 16S rRNA gene amplicon analysis of samples collected before and after the seaweed supplementation indicated no long-lasting microbiome effects (Supplementary Fig. 2c). Due to this delay, the heaviest 32 animals, balanced for breed, underwent a shorter performance test period of 4 weeks instead of the normal 8 weeks. During performance testing, daily feed intake was recorded using electronic feeders (HOKO, Insentec, Marknesse, The Netherlands). Twice weekly, duplicated samples of each diet component were collected to determine dry matter content and to calculated dry matter intake (DMI). Body weight of each animal was measured weekly to estimate average daily gain (ADG) using a linear regression model including time on test. Feed conversion ratio (FCR) was calculated for each animal as average daily DMI divided by ADG.

At the end of the experimental period, the animals’ methane emissions were measured in respiration chambers. One week prior to entering the respiration chambers, the cattle were single-housed in training pens, identical in size and shape to the pens inside the chamber, to adapt to individual housing. The cattle were allocated to six respiration chambers based on the criterion of minimization of the variation in body weight. They remained in the respiration chambers for 3 days, which included one day for adaptation and a 48 h measurement period for methane emissions.

Of the 71 animals that completed the trial, 24 were selected for multi-omic analysis, including equal numbers of the two breeds, and representing the full range of methane emissions. For a further subset of six animals (out of 24), samples were also analyzed for a time series collected during the experimental period using orogastric tubing, as described below.

Rumen content and tissue sample collection

On live animals, longitudinal rumen fluid samples (50 ml) were collected using a stomach tube (16 × 2700 mm Equivet Stomach Tube; Jørgen Kruuse A/S, Langeskov, Denmark) nasally and aspirating manually. Samples were collected prior to the adaptation phase to seaweed, before and after the performance test as well as immediately after leaving the respiration chambers. Additionally, rumen fluid samples (50 ml) were obtained after the animals were slaughtered in a commercial abattoir, immediately after the rumen was opened to be drained. Immediately after sampling, the rumen digesta was filtered through four layers of muslin and a 5 ml sample of the filtered liquid was transferred into a 30 ml universal tube and stored in a freezer at −80 °C.

Rumen cell wall samples were collected from the central region of the ventral sac before the rumen had been washed. The ruminal tissue was dipped into a 125 ml beaker containing a PBS solution to remove the ruminal digesta. The tissue was sliced using a sterile scalpel and transferred to a 30 ml universal tube containing 5 ml RNALater. Additionally, liver samples were taken by the meat inspector, with a section cut out using a sterile scalpel and then stored in a 30 ml universal tube with 5 ml RNALater. All tissue samples were stored in a freezer at -80 °C before being analyzed. Further details regarding the sampling and experimental procedures carried out at SRUC can be found in previously published studies^31,32 which followed a similar protocol.

16S rRNA gene amplicon sequence data

Rumen digesta sample (n = 71 post-slaughter, n = 118 longitudinal) DNA extraction, PCR amplification and sequencing of 16S rRNA gene amplicons was performed at DNASense ApS (Aalborg, Denmark).

Sample DNA extraction

Rumen digesta DNA was extracted using the FastDNA Spin kit for Soil (MP Biomedicals, USA) with the following exceptions to the standard protocol: 500 μL of sample, 480 μL Sodium Phosphate Buffer and 120 μL MT Buffer were added to a Lysing Matrix E tube. Bead beating was performed at 6 m/s for 4x40s. Gel electrophoresis using Tapestation 2200 and Genomic DNA screentape (Agilent, USA) was used to validate product size and purity of a subset of DNA extracts. DNA concentration was measured using Qubit dsDNA HS/BR Assay kit (Thermo Fisher Scientific, USA).

Sequencing library preparation

Amplicon libraries for the 16S rRNA gene variable region 4 (abV4-C) were prepared using a custom protocol based on an Illumina protocol³³. Up to 10 ng of extracted DNA was used for PCR amplification. Each reaction (25 μL) contained (12.5 μL) PCRBIO Ultra mix and 400 nM of each forward and reverse tailed primer mix. The PCR program was as follows: initial denaturation at 95 °C for 2 min, 30 cycles of amplification (95 °C for 15 s, 55 °C for 15 s, 72 °C for 50 s) and a final elongation at 72 °C for 5 min. Duplicate reactions were performed for each sample and the duplicates pooled afterwards. The primers targeting the abV4-C region were the following, designed according to³³: [515FB] GTGYCAGCMGCCGCGGTAA and [806RB] GGACTACNVGGGTWTCTAAT³⁴, with tails that enable attachment of Illumina Nextera adapters necessary for sequencing in a subsequent round of PCR. The amplicon libraries were purified using the standard protocol for CleanNGS SPRI beads (CleanNA, NL) with a bead to sample ratio of 4:5. DNA was eluted in 25 μL of nuclease free water (Qiagen, Germany). DNA concentration was measured using Qubit dsDNA HS Assay kit (Thermo Fisher Scientific, USA). Gel electrophoresis using Tapestation 2200 and D1000/High sensitivity D1000 screentape (Agilent, USA) was used to validate product size and purity of a subset of libraries.

Sequencing libraries were prepared from purified amplicon libraries using a second PCR. Each reaction (25 μL) contained PCRBIO HiFi buffer (1x), PCRBIO HiFi Polymerase (1 U/reaction) (PCRBiosystems, UK), adapter mix (400 nM of each forward and reverse) and up to 10 ng of amplicon library template. PCR was done with the following program: initial denaturation at 95 °C for 2 min, 8 cycles of amplification (95 °C for 20 s, 55 °C for 30 s, 72 °C for 60 s) and a final elongation at 72 °C for 5 min. The resulting libraries were purified following the same protocol as above for the first PCR.

DNA sequencing

The purified sequencing libraries were pooled in equimolar concentrations and diluted to 2 nM. The samples were paired-end sequenced (2 × 300 bp) on a MiSeq (Illumina, USA) using a MiSeq Reagent kit v3 (Illumina, USA) following the standard guidelines for preparing and loading samples on the MiSeq. > 10 % PhiX control library was spiked in to overcome low complexity issues often observed with amplicon samples.

Sequence data analysis

Quality trimming and amplicon sequence variant (ASV) inference for the 16S rRNA gene amplicon sequence data was performed with dada2³⁵ following the recommended Big Data Paired-end workflow³⁶ using default parameters, except for the following choices for the filterAndTrim step: truncLen = 240 for forward, 200 for reverse reads; trimLeft = 20 for forward, 30 for reverse reads; maxEE = 2, and truncQ = 6. The reference database for taxonomic classification was the dada2 formatted version of release 214 of the Genome Taxonomy Database (GTDB)³⁷.

Metagenomics

DNA extraction and sequencing as well as initial metagenomic sequence data analysis for rumen digesta samples was performed at DNASense ApS (Aalborg, Denmark).

DNA extraction

DNA intended for sequencing on the Illumina platform (n = 24 post-slaughter, n = 29 longitudinal) was extracted during the workflow for 16S rRNA gene amplicon data, as described above. DNA intended for ONT sequencing (n = 24 post-slaughter) was extracted with the DNeasy PowerSoil Kit (Qiagen, Germany) and further cleaned with the DNeasy PowerClean Pro Cleanup Kit (Qiagen, Germany). A custom SPRI (Solid Phase Reversible Immobilization) short fragment removal step was implemented to remove fragments shorter than ~1500–2000 bp. DNA concentration and purity was assessed using the Qubit dsDNA HS Assay kit (Thermo Fisher Scientific, USA) and the NanoDrop One device (Thermo Fisher Scientific, USA). DNA size distribution was evaluated using the Genomic DNA ScreenTape on the Agilent Tapestation 2200 (Agilent, USA).

Illumina sequencing

Extracted DNA was fragmented to approximately 550 bp using a Covaris M220 with microTUBE AFA Fiber screw tubes and the settings: Duty Factor 10 %, Peak/Displayed Power 75 W, cycles/burst 200, duration 40 s and temperature 20 °C. The fragmented DNA was used for metagenome preparation using the NEB Next Ultra II DNA library preparation kit. The DNA library was paired-end sequenced (2  × 150 bp) on a NovaSeq S4 system (Illumina, USA).

Oxford Nanopore sequencing

SQK-LSK114 sequencing libraries were prepared according to manufacturer recommendations with a minor custom modification to allow for native barcoding using kits EXP-NBD104 and EXP-NBD114 (Oxford Nanopore Technologies, Oxford, UK). Briefly; before initiating the SQK-LSK114 protocol, native barcodes were ligated onto end-prepped sample DNA (100–200 fmol) using NEB Blunt/TA ligase mastermix (New England Biolabs, USA). ~10–20 fmol barcoded DNA library were loaded onto primed PromethION FLO-PRO114M (R10.4.1) flow cells and sequenced on the PromethION P2 Solo device running MinKNOW Release 22.07.3 (MinKNOW Core 5.3.0-rc3-p2solo).

Data preprocessing

Raw Illumina reads were filtered for PhiX using Usearch11³⁸ and trimmed for adapters using cutadapt³⁹ (v. 3.5). Forward and reverse read files were concatenated using a custom python script. Raw Oxford Nanopore fast5 files were basecalled and demultiplexed in Guppy v. 6.1.15 using the dna_r10.4.1_E8.2_400bps_sup algorithm. Adapters were removed in Porechop v. 0.2.4 using default settings. NanoStat v.1.4.0⁴⁰ was used to assess quality parameters of the basecalled data. The adapter-trimmed data was then filtered in Filtlong v. 0.2.1 with –min_length set to 1500 bp and –min_mean_q set to 96 (q-score of 14).

Metagenome de novo assembly and binning

Metagenomes were assembled and binned using two independent pipelines in parallel. The resulting MAGs were lastly dereplicated in a single pool to produce the final 700 MAGs.

The first pipeline performed draft de novo co-assembly for metagenomes in six groups of samples/animals (combinations of control and treatment, corresponding to the seaweed supplementation, and a three-category methane variable representing low, medium and high emission levels) using Flye⁴¹ (v. 2.9.1-b1780) by setting the metagenome parameter (--meta). Draft metagenomes were first polished with Medaka⁴² (v. 1.7.1) using quality-filtered Oxford Nanopore R10.4.1 data, following further polishing with minimap2⁴³ (v. 2.24-r1122) and racon⁴⁴ (v. 1.5.0) using Illumina data covering the relevant metagenome sample trajectory. Each metagenome assembly was subjected to independent and automated genome binning using Metabat2⁴⁴ (v. 2.15) and Vamb⁴⁵ (v. 4.1.1). MAGs from each metagenome were subsequently dereplicated using dRep⁴⁶ v. 3.2.2 setting minimum MAG length to 250000 bp (-l). All dereplicated MAGs from each metagenome assembly were finally pooled and dereplicated again (cross-dereplicated) with dRep.

The second pipeline accepted samples containing paired short-read and nanopore metagenomes. These were processed using a hybrid assembly approach, followed by MAG recovery through the Aviary⁴⁷ (v. 0.5.7) pipeline (https://github.com/rhysnewell/aviary) using the recover workflow with default settings. The resulting assemblies were manually inspected using Bandage to identify and verify closed genomes.

The bins from both the first and second parallel pipelines were pooled, showcasing a total of 4,469 redundant recovered MAGs. Completeness and contamination rates were assessed with CheckM2⁴⁸ (v. 1.0.1) using the lineage wf command. Only MAGs with > 70% completeness and < 10% contamination were retained for further analysis. To address potential multi-mapping issues during meta-omic relative abundance calculations, the genomes were dereplicated using a custom script. Pairwise Average Nucleotide Identity (ANI) values were calculated for all MAGs using Skani⁴⁹. Genomes with > 97% ANI and > 50% alignment were clustered using complete linkage clustering. The highest-quality MAG within each cluster was selected as the representative genome. The quality score was calculated using the following metric: completeness - 5*contamination - 5*num_contigs/100 - 5*num_ambiguous_bases/100000, as described by Parks et al. (2020)⁵⁰. This clustering process was iteratively repeated until no further clustering of representative MAGs was possible. This resulted in a nonredundant set of 700 MAGs.

Genome-scale metabolic reconstruction and analysis

We built a genome-scale metabolic model (GEM) from each MAG using the automated metabolic reconstruction tool CarveMe⁵¹, gap filling for anaerobic growth on a complete medium (import allowed for all possible nutrients but not oxygen) to capture the metabolic environment of the rumen. We used the GEMs to convert relative archaeal and bacterial abundances from 16S rRNA sequencing and metagenomics data to metabolic reaction abundances. Annotating abundance data with the GTDB taxonomy⁵², we mapped the GEMs to the data by matching taxa on the genus level for the ASVs and the species level for the MAGs. We computed the frequency of each metabolic reaction for each ASV and MAG by taking the average of reaction presence (0 = reaction absent, 1 = reaction present) across all GEMs mapped on the genus level for the ASVs and by directly using reaction presence in GEMs for the MAGs. Then, we computed the abundance of each reaction in each sample by multiplying the reaction frequencies by the ASV or MAG abundances. We performed PCA separately for the ASV and MAG reaction abundances, standardizing features by removing the mean and scaling to unit variance.

Rumen microbial genome database for metatranscriptomics and metaproteomics

For metatranscriptomic and metaproteomic data analyses, we built databases consisting of six parts representing different sources and taxonomic domains:

700 MAGs assembled from our digesta samples, representing both archaea and bacteria

• A. Bos taurus host genome ARS-UCD1.3⁵³ GCF_002263795.2 (NCBI Bioproject PRJNA391427).

• B. Entodinium caudatum genome¹³ (NCBI Bioproject PRJNA380643).

• C. 52 protozoal SAGs² (NCBI Bioproject PRJNA777442).

• D. 9 fungal genomes from phylum Neocallimastigomycota¹².

• E. 14 bacterial genomes of genus Campylobacter⁵⁴.

This total rumen microbial genome database consisted of 4.2 M proteins with an average length of 426.8 amino acids totaling 1.8 G amino acid letters.

Annotation of genomes and characterization of proteins

The different parts of the rumen microbial genome database (A-F) were annotated using several tools. For the 700 recovered MAGs (rumen microbial genome database part A) and 14 Campylobacter spp. genomes (F), Prokka⁵⁵ (v. 1.14.6) was used for annotation and to translate the coding sequences. CheckM2⁴⁸ (v. 1.0.2) was used for assessment of completeness and contamination. The remaining database parts (B–E) were downloaded as amino acid sequences. Translated genes of the complete rumen genome database (A–F) were characterized functionally using eggnog-mapper⁵⁶ (v. 2.1.12), resulting in the identification of PFAM⁵⁷, CAZy⁵⁸, and KEGG⁵⁹ orthologs. The eggnog-mapper results were predominantly used for interpretation of the (meta)proteomic analysis. Pathway enrichment analysis was calculated using the KEGG orthologs and KEGG pathway database⁶⁰ (downloaded on 2023-08-28) via clusterProfiler⁶¹ (v. 4.10.0). Taxonomic identification of MAGs were done with GTDB-tk⁵² (v. 2.4.0) using database r214. The genomic characterization tools mentioned above were run via CompareM2⁶² (v. 2.11.1). For screening of metabolic capacities, DRAM⁶³ (v. 1.4) was used on the translated amino acid sequences of the complete rumen microbial genome database with the following parameters: DRAM.py annotate_genes --use_uniref --threads 64. The DRAM results were predominantly used for interpretation of the (meta)transcriptomic analysis. CoverM⁶⁴ (v. 0.6.1) was used to calculate read coverage and estimate relative abundances of bacterial and archaeal MAGs (A).

Meta- and host transcriptomics

RNA extraction and sequencing for rumen digesta (n = 24 post-slaughter, n = 29 longitudinal), wall (n = 17) and liver samples (n = 24), as well as bioinformatic analyses for rumen wall and liver sequence data, were performed at DNASense ApS (Aalborg, Denmark).

RNA extraction

RNA extraction for rumen digesta, rumen wall and liver samples was performed with the standard protocol for RNeasy PowerMicrobiome Kit (Qiagen, Germany) with minor modifications: custom reagent volumes were used, PM4 buffer was replaced with 70 % ethanol in initial extraction mix, and bead beating was performed at 6 m/s for 4x40s. Gel electrophoresis using Tapestation 2200 and RNA screentape (Agilent, USA) was used to validate product integrity and purity of RNA extracts. RNA concentrations were measured using Qubit RNA HS/BR Assay kit (Thermo Fisher Scientific, USA). The extracted RNA was treated with the TURBO DNAfree (Thermo Fisher Scientific, USA) to ensure removal of all DNA in the samples. Afterwards the RNA was quality controlled using RNA screentape (Agilent, USA) and Qubit RNA HS/BR Assay kit (Thermo Fisher Scientific, USA).

Sequencing library preparation

RNA extracts were rRNA depleted using the Ribo-Zero Plus rRNA Depletion Kit (Illumina, USA), and residual DNA from RNA extraction was removed using the DNase MAX kit (MoBio Laboratories Inc.). The samples were purified using the standard protocol for CleanPCR SPRI beads (CleanNA, NL) and further prepared for sequencing using the NEBNext Ultra II Directional RNA library preparation kit (New England Biolabs). Library concentrations were measured using Qubit HS DNA assay and library DNA size estimated using TapeStation with D1000 ScreenTape. The samples were pooled in equimolar concentrations and sequenced (2 × 150 bp, PE) on a Novaseq platform (Illumina, USA). All kits were used as per the manufacturer’s instructions with minor modifications.

Host transcriptome mapping

Forward and reverse sequencing cDNA reads were quality-filtered and trimmed for Illumina adapters using Cutadapt v. 3.7³⁹ used in paired-end mode. For liver and rumen wall data, the reads were subsequently mapped against the Bos Taurus Genome Reference ARS-UCD1.3 (Genbank assembly accession GCA_002263795.3). The genome and its associated gene transfer format file (GTF) were downloaded and indexed using STAR⁶⁵ (v. 2.7.10a), setting a sjdbOverhang of 149 bp. Adapter-trimmed sample reads were mapped against the indexed genome of ARS-UCD1.3 using STAR (v. 2.7.10a) in paired-end mode, with the option -outReadsUnmapped Fastx enabled. Alignments were ported to coordinate-sorted BAM files, and FeatureCounts (v. 2.0.1) of the SubRead package⁶⁶ was used to quantify CDS mappings as counts. Where nothing else is stated, the default settings were used for all bioinformatic tools.

Rumen wall metatranscriptome mapping

For rumen wall samples, the forward and reverse cDNA reads that did not map against the Bos taurus genome were bioinformatically depleted for rRNA using Ribodetector v. 0.2.7⁶⁷ and then mapped against the rumen MAGs. Prior to mapping, the concatenated MAGs were indexed using STAR⁶⁵ (v. 2.7.10a). The rRNA-depleted and quality filtered DNA reads were mapped against the MAGs with STAR, setting alignIntronMax to 1. All alignments were ported to coordinate-sorted BAM files.

Rumen content metatranscriptomics

Rumen content data were mapped against the Bos taurus genome (Genome Reference ARS-UCD1.3) using minimap2 v 2.2. All non-paired mapped reads were retrieved using samtools v 1.17⁶⁸ with the following parameters samtools fastq -f 12 -F 256 -c 7 -1 read1.fq.gz -2 read2.fq.gz. rRNA reads present in the samples were bioinformatically removed using SortMeRNA v 4.3.6⁶⁹ with the following SILVA databases: silva-bac-16s-id90, silva-arc-16s-id95, silva-bac-23s-id98, silva-arc-23s-id98, silva-euk-18s-id95 and silva-euk-28s-id98, and the parameters –out2–paired_out –fastx–thread 64. These reads were used to quantify the expression of coding sequences (CDS) encoded in the rumen microbial genome database using Kallisto⁷⁰ (v. 0.50.0). The resulting ‘raw-counts’ tables were gathered into a single table using the Bioconductor tximport⁷¹ (v. 1.26.1) library in R 4.2.2.

Meta- and host proteomics

Proteomic and metaproteomic measurements and all bioinformatic analyses for rumen digesta (n = 24 post-slaughter, n = 29 longitudinal), wall (n = 24) and liver (n = 24) samples were performed at the Norwegian University of Life Sciences (NMBU; Ås, Norway).

Protein extraction and digestion

Protein extraction was performed following a previously published protocol¹. Briefly, for rumen samples we used 300 μL of fluid for downstream processing; for liver samples we used ~300 μL of finely chopped/liquified liver (with sterile scalpel); finally for rumen wall samples we used sterile tweezers and scalpel to carefully remove the wall papillae from the remainder of the tissue and finely chop them into liquified mass (~300 μL). Each sample was combined with 150 μL lysis buffer (30 mM DTT, 150 mM Tris-HCl (pH = 8), 0.3% Triton X-100, 12% SDS) and 4 mm glass beads (≤160 μm), then vortexed and rested on ice for 30 mins. Sample lysis was performed with a FastPrep-24 Classic Grinder (MP Biomedical, Ohio, USA) for 3 × 60 s at 4.0 m/s⁷², followed by centrifugation at 16,000 × g for 15 min at 4 °C. Lysate was removed and its absorbance measured at A750 on BioTek Synergy H4 Hybrid Microplate Reader (Thermo Fisher Scientific Inc., Massachusetts, USA). 40–50 μg of protein was prepared in SDS-buffer, heated in a water bath for 5 min at 99 °C, and analyzed by SDS-PAGE with Any-kD Mini-PROTEAN TGX Stain-Free gels (Bio-Rad, California, USA) in a 2 minute run for sample clean-up, before staining with Coomassie Blue R-250. Visible bands were excised and divided into 1 mm² pieces before reduction, alkylation and trypsin digestion. Peptides were concentrated and eluted using C18 ZipTips (Merck Millipore, Darmstadt, Germany) following manufacturer’s instructions.

Mass spectrometry

Peptide samples were analyzed by coupling a nano UPLC (nanoElute, Bruker) to a trapped ion mobility spectrometry/quadrupole time of flight mass spectrometer (timsTOF Pro, Bruker). Peptides were separated with a PepSep Reprosil C18 reverse-phase (1.5 µm, 100 Å) 25 cm × 75 μm analytical column coupled to a ZDV Sprayer (Bruker Daltonics, Bremen, Germany). Column temperature was kept at 50 °C using the integrated oven. Equilibration of the column was performed before the samples were loaded (equilibration pressure 800 bar). The flow rate was set to 300 nl/min and the samples separated using a solvent gradient from 5 % to 25 % solvent B over 70 minutes, and to 37 % over 9 minutes. The solvent composition was then increased to 95 % solvent B over 10 min and maintained at that level for an additional 10 min. In total, a run time of 99 min was used for the separation of the peptides. Solvent A consisted of 0.1 % (v/v) formic acid in milliQ water, while solvent B consisted of 0.1 % (v/v) formic acid in acetonitrile.

The timsTOF Pro was run in positive ion data dependent acquisition PASEF mode with the control software Compass Hystar version 5.1.8.1 and timsControl version 1.1.19 68. The acquisition mass range was set to 100 – 1700 m/z. The TIMS settings were: 1/K0 Start 0.85 V⋅s/cm2 and 1/K0 End 1.4 V⋅s/cm2, Ramp time 100 ms, Ramp rate 9.42 Hz, and Duty cycle 100 %. Capillary Voltage was set at 1400 V, Dry Gas at 3.0 l/min, and Dry Temp at 180 °C. The MS/MS settings were the following: number of PASEF ramps 10, total cycle time 0.53 sec, charge range 0–5, Scheduling Target Intensity 20000, Intensity Threshold 2500, active exclusion release after 0.4 min, and CID collision energy ranging from 27–45 eV.

Data analysis

The raw spectra were analyzed using mspipeline1⁷³ (v. 2.0.0) based on FragPipe⁷⁴ (v. 19.1). Using Philosopher⁷⁵ (v. 4.8.1), MSFragger⁷⁶ (v. 3.7) and IonQuant (v. 1.8.10). Spectra were analyzed slicing the rumen microbial genome database into 16 parts using msfragger.misc.slice-db=16. Mass calibration was disabled with msfragger.calibrate_mass=0. The maximum length of peptides to be generated during in-silico digestion was 35 with msfragger.digest_max_length=35. Allowed number of missed cleavages 1 and 2 was set to 1 with msfragger.allowed_missed_cleavage_{1,2} = 1. Otherwise, default settings were used. The processing was performed on an AMD x86-64 “Threadripper Pro” 5995WX 64 cores, 8 memory channels, 512GiB DDR4 3200 MHz ECC (8 × 64 GiB) and 4 2TB SSDs in raid0.

Proteomic intensities were log2-transformed prior to any statistical analysis. Genes in the proteomic database were annotated using eggnog e-mapper (v. 2.1.12) using CompareM2 (v. 2.11.1). Missing values were imputed using missRanger⁷⁷ (v. 2.6.0). Two post-slaughter rumen digesta samples were excluded from all downstream analyses due to low number of identified proteins.

Untargeted metabolomics

Untargeted metabolomic analyses for rumen digesta, rumen wall, and liver samples were carried out by MS-Omics Aps (Vedbæk, Denmark). Compound identification was performed at four levels: Level 1: identification by retention times (compared against in-house authentic standards), accurate mass, and MS/MS spectra; Level 2a: identification by retention times (compared against in-house authentic standards), and accurate mass; Level 2b: identification by accurate mass, and MS/MS spectra; Level 3: identification by accurate mass alone. A deviation of 3 ppm was accepted for accurate mass identification.

Sample extraction

Rumen digesta samples were vortexed and an aliquot (100 µl) transferred to a spin filter (0.22 µm). The aliquot was diluted with water (100 µl) and filtered by centrifugation (4656 × g, 2 × 5 min, 4 °C). Filtered extracts were diluted 10 times in mobile phase eluent A and fortified with stable isotope labeled standards before analysis.

Rumen wall and liver samples were mixed with ceramic beads and precooled methanol/water (1:2) fortified with stable isotope labeled standards. The samples were then placed in a pre-cooled (–20 °C) bead beater and homogenized (4 × 30 s, 30 Hz) followed by ultrasonication (5 min). After centrifugation (18,000 × g, 5 min, 4 °C), the supernatant of each tube was collected. The sample pellets were re-extracted as described above. The two extract supernatants were pooled and passed through a phosphor removal cartridge (Phree, Phenomenex). A precise aliquot of the extract was evaporated to dryness under a gentle stream of nitrogen, before reconstitution with 10% Eluent B in Eluent A.

LC-MS method

All biological samples were analyzed together with three types of technical control samples: procedural blanks (water), blanks (solvent), and pooled quality control samples (QC) generated by pooling together equal volumes from all corresponding samples. Biological samples were analyzed in single technical injection per matrix. Numbers of samples per type were as follows: rumen digesta: 53 biological samples (n = 24 post-slaughter, n = 29 longitudinal), 4 blanks, 1 procedural blank, 8 QC samples, 8 QC injections used for iterative MS2 acquisition, 2 system suitability tests, and 3 diluted QCs used for correlation dilution testing; rumen wall: 24 samples, 4 blanks (solvent), 1 procedural blank (water), 7 QC samples, 10 QC injections used for iterative MS2 acquisition, 2 system suitability tests, 3 diluted QCs used for correlation dilution testing; liver: 24 samples, 4 blanks (solvent), 1 procedural blank (water), 7 QC samples, 10 QC injections used for iterative MS2 acquisition, 2 system suitability tests, 3 diluted QCs used for correlation dilution testing. All samples, QCs and injections can be found in the MetaboLights repository (accession: MTBLS12283).

Metabolite separation was carried out on a Thermo Scientific Vanquish UHPLC System (VH-P10-A, VH-A10-A, VH-C10-A) running XCalibur Version 4.4.16.14, using an ACQUITY UPLC HSS T3 C18 column (2.1 × 150 mm, 1.8 μm particle size; Waters Corporation). The column was maintained at a temperature of 30 °C, and the flow rate was set to 300 μL/min. The mobile phase consisted of 10 mM ammonium formate with 0.1% formic acid in water (pH 3.1) as solvent A, and 10 mM ammonium formate with 0.1% formic acid in methanol as solvent B. Sample injection was performed in partial loop mode with an injection volume of 5 μL. The chromatographic separation was achieved using the following gradient elution profile: the initial composition was held at 0% B for 2.0 minutes, followed by a linear increase to 35% B at 12.0 minutes. The gradient was then ramped to 90% B by 13.0 minutes and held at 90% B until 14.0 minutes. The column was re-equilibrated to initial conditions (0% B) by 15.0 minutes.

Mass spectrometry analyses were run on a QExactive-HF Orbitrap instrument (Thermo Scientific) with XCalibur version 4.4.16.14. The ionization source was a heated ESI (HESI II) with alternating positive and negative polarity and the following parameters: spray voltage negative 2.6 kV/positive 3.5 kV, sheet gas (50 N2), ion source temperature 350 °C, ion transfer tube temperature 290 °C. Fragmentation was performed using higher-energy collisional dissociation (HCD) at stepped 20, 40, 60 normalized collision energy. The instrument resolution was 60,000 at a mass accuracy of under 1 ppm, with adaptive scan rates (data dependent Top N acquisition; changes the scan rate during the run). Samples used for quantification and peak integration were acquired in FullScan MS1 mode from 65 – 975 m/z. Pooled QC injections used for feature annotation were recorded in iterative exclusion/inclusion ddTopN MSMS2 mode.

Volatile fatty acid quantification

Rumen digesta samples (n = 21 post-slaughter) were thawed on ice and centrifuged when still cold. 450 µL of each sample was transferred to a new tube and 50 µL of a 50% formic acid solution was added to reach a 5 % concentration of formic acid. Samples were then centrifuged again and 400 µL of the supernatant was transferred to a GC-vial, with 1000 µL of an internal standard solution added. Volatile fatty acids were separated using gas chromatography (Trace 1300 GC with autosampler, Thermo Scientific) with a Stabilwax-DA column (30 m, 0.52 mm ID, 0.25 µm, Restek).

Microarray polymer profiling

Microarray polymer profiling (MAPP) entails the printing of extracted glycans as high-density microarrays which are then probed with monoclonal antibodies with specificities for different glycan epitopes. The output from MAPP provides insight into the relative abundance of epitopes across the sample set.

Alcohol insoluble residues (AIR) were prepared from each rumen digesta sample (n = 24) as follows: samples were homogenized to a fine powder using a tissue lyser (Qiagen). Approximately five volumes of 70% ethanol were added, the samples vortexed for 10 minutes then centrifuged at 2700 × g for 10 minutes and the supernatants discarded, This step was repeated. Approximately five volumes of 1:1 methanol and chloroform were added to the pellet and the samples were again vortexed and centrifuged as previously. Finally, approximately five volumes of acetone were added and the same vortexing and centrifugation steps performed. The resulting pellets were AIR.

To extract glycans, 300 μL of 50 mM diamino-cyclo-hexane-tetra acetic acid, pH 7.5, were added to 10 mg AIR. After agitation in a tissue lyser (27 s-1 for 2 minutes and 10 s-1 2 h), samples were centrifuged at 2700 × g for 10 minutes. The supernatant was removed, 300 μL 4 M NaOH with 1% v/v NaBH4 added to the pellet and the agitation and centrifugation steps repeated. The resultant NaOH extraction supernatants were diluted sequentially (1/2,1/5,1/5,1/5) in microarray printing buffer (55.2% glycerol, 44% water and 0.8% Triton X-100), and the four dilutions were printed in quadruplet onto nitrocellulose membranes using a non-contact microarray robot (Arrayjet, Roslin). Thus, every replicate was represented by a 16-spot subarray (four concentrations and four printing replicates). Arrays were probed with monoclonal antibodies, scanned, uploaded into microarray analysis software (Array Pro Analyzer 6.3, Media Cybernetics) and mean spot signals from each sub array calculated.

Statistics and data visualization

Unless otherwise specified, statistical analyses and visualizations were performed in the R statistical programming language⁷⁸ (v. 4.3.2). The knitr⁷⁹ package (v. 1.45) was used for reporting, renv⁸⁰ (v. 1.0.7) for package management, ggplot2⁸¹ (v. 3.5.1) for visualizations, cowplot⁸² (v. 1.1.3) for composing multipanel figure layouts, and ComplexHeatmap⁸³ (v. 2.15.4) for heatmaps.

16S rRNA gene ASV data was managed with phyloseq⁸⁴ (v. 1.46.0), which was also used to calculate alpha diversity indices. Rumen community types (RCTs) were defined using the ASVs data and Dirichlet Multinomial Mixtures¹⁴ clustering implemented with mia⁸⁵ (v. 1.10.0), selecting the optimal number of clusters based on the Laplace method. Only ASVs that were present in at least half of all slaughter timepoint samples (n = 35) were used for this analysis.

All beta diversity comparisons for ASV counts and MAG relative abundances were performed using vegan⁸⁶ (v. 2.6-6), with robust Aitchison distances statistically compared with PERMANOVA (adonis2; 9999 permutations), and visualized with PCoA (package ape⁸⁷, v. 5.8). Principal Component Analysis (PCA) for all other omic data types was run with the “prcomp” function. For (meta)transcriptomic data (variance stabilizing transformed (VST) counts) and (meta)proteomic data (log2 transformed LFQ intensities with imputed missing values), the 1000 features with the highest variance were selected for PCA. For untargeted metabolomic data, where the number of features was orders of magnitude lower, all features were used, except for rumen digesta, where features with annotation level 3 were excluded.

Where statistical testing was done between two categorical variables, Fisher’s exact tests were used. Continuous variables were compared either with two-sided t-tests (KPTs and other animal-related metrics, Principal Component scores), or two-sided Wilcoxon rank-sum tests (alpha diversity indices, metagenomic relative abundances, proteomic LFQ intensities, MAPP intensities, normalized intensities of metabolites from untargeted measurements, and molar percentages of volatile fatty acids) with multiple comparison correction using the “fdr” option of the “p.adjust” function. Differential abundance comparisons for count data (ASVs and meta- and host transcriptomics) were run with DESeq2⁸⁸ (v. 1.42.1), with default parameters for transcriptomic data, and the “sfType” parameter changed to “poscounts” for ASV data.

Network analysis (WGCNA)

Correlation-network based analysis was applied on the proteomic and metaproteomic samples to group co-expressed proteins into clusters. Weighted gene co-expression network analysis²² (WGCNA) (v. 1.73) was applied on data that included imputed missing values to construct clusters independently in the digesta, rumen wall epithelium, and liver samples. These clusters were then correlated via their eigengenes across samples to obtain host-microbiome boundary-links.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.