Microbiome composition
The sequencing of microbiome rRNA amplicons from ruminal and fecal samples of the experimental group B yielded a total of 10,573,763 paired-end reads (4,628,604 paired-end reads for bacteria, 4,443,390 for archaea and 1,501,769 for protozoa), reaching 20,241,296 paired-end reads with the addition of sequencing data from Group A. After quality control, and singleton exclusion, a total of 4,519 bacterial ASVs (2,680 ruminal ASVs and 1,839 fecal ASVs), 1,023 archaeal ASVs (421 ruminal ASVs and 602 fecal ASVs) and 151 ruminal protozoan ASVs across treatments. Rarefaction curves based on the alpha-diversity metrics of Shannon-Wiener (diversity) reached a plateau, which indicated that additional sequences would not likely result in additional features.
Comparison of samples from different treatment groups using alpha-diversity metrics (Observed ASVs and Shannon-Wiener indexes) under the Kruskal-Wallis testing method revealed that rumen bacterial diversity was significantly more abundant and richer in animals fed the conventional diet (Group A) than those fed the byproducts diet (Group B) (P = 0.006 and P = 0.04, respectively). Similarly, the ruminal archaea diversity was also richer (P = 0.0004), but not more abundant. There was no significant difference when contrasting alpha diversity metrics of fecal samples. Comparisons of the beta-diversity metric Unweighted Unifrac using the PERMANOVA approach, revealed that samples of archaea and bacteria tended to form two significant clusters, which represented the treatment groups (adjusted P< 0.01) (Supplementary Figures 1-3), a tendency most pronounced in fecal populations.
Phenotypic description
Methane emission was calculated for each experimental group as the average value of all visits to the GrowSafe feedlot during the finishing period. Animals from the experimental group A presented a mean methane emission of 179,1 g/day and a standard error of 26,18 g/day, while animals from the experimental group B presented a mean methane emission of 161 g/day and a standard variation of 26,05 g/day. Mixed Models showed that the difference in CH4 emission between experimental groups was significant (P<0.0001).
Taxonomic composition of the experimental group B
Herein, we will present only results concerning the taxonomic composition of dietary treatment group B. Please refer to [28] for an extensive exploration of the taxonomic diversity of the group A.
The phylum Bacteroidetes was the most relatively abundant bacterial phylum identified in the rumen microbiome (38.18% ± 3.86%), followed by Firmicutes (35.72%), Proteobacteria (8.96%), Sphirochaetes (5.40%) and Fibrobacteres (4.6%). Differently, the Phylum Firmicutes was the most abundant in the fecal microbiome (52.59%), followed by Bacteroidetes (30.87%), Proteobacteria (13.3%) and Tenericutes (1.31%). At the genus level, Prevotella was the most abundant genus in the rumen microbiome (19.87%), followed by Treponema (6.28%), Ruminobacter (5.78%), Fibrobacter (5.56%) and Christensenellaceae R-7 (5.56%). Conversely, the genus Ruminococcaceae UCG-005 was the most relatively abundant in the fecal microbiome (13.63%), followed by Succinivibrio (12.75%), Bacteroides (9.71%), Prevotella (6.69%) and Rikenellaceae RC9 (5.07%) (Figure 1).
Regarding the archaea domain, Euryarchaeota was the only phylum identified in both microbiomes. At the species level, these microbiomes were populated by Methanobrevibacter gottschalkii (59.16% and 74.89% for rumen and feces, respectively), Methanobrevibacter ruminantium (31.98% and 17.11%) and Methanosphaera sp. ISO3-F5 (7.27% and 7.5% (Figure 2A). As for protozoa, Ciliophora was the only phylum identified in rumen, and was populated by 3 genera, Bozasella/Triplumaria (70.73%), Entodinium (28.82%) and Ostracodinium (0.44%) (Figure 2B).
Differential abundant ASVs in dietary treatment groups
We applied the analysis of composition of microbiomes (ANCOM) to investigate the influence of dietary treatments in the microbiome composition at the ASV level. Seventeen ruminal ASVs of bacterial origin were differentially more abundant (DA) in the group A, from which the most prominent were classified as Bacteroidales F082 group (ASV 20 and 23, CLR: 1.51), Christensenellaceae (ASV 112, CLR: 1.3), Pedosphaeraceae families (ASV 145, CLR: 1.09) and the genus Succiniclasticum (ASV 170, CLR: 1.04). Ten ASVs were DA in the group B, of which the most abundant were classified as Succiniclasticum (ASV 97, CLR: 0.48), Acetitomaculum (ASV 116, CLR:1.07), Lachnospiraceae family (ASV 247, CLR: 0.98), Fibrobacter (ASV 96, CLR: 0.98) and Succinivibrio genus (ASV 118, CLR: 0.94) (Supplementary Figure 4). Also, three fecal ASVs were DA in our experimental groups; one was classified as a member of the family Rikenellaceae (ASV 361, CLR: 0.59) and was more abundant in the group A, while an ASV was classified as a member of the family Prevotellaceae (ASV 332, CLR: 0.51) and another as the genus Oscillibacter (ASV 526, CLR: 0.51) were both more abundant in the group B (Supplementary Figure 5).
Eight archaeal ASVs were DA among treatment groups in the rumen microbiome. Four ASVs classified as M. gottschalkii (ASVs 1, 2, 13 and 11, CLR > 1), one as M. ruminantium (ASV 23, CLR: 1.13) and one ASV belonging to the Methanomassiliiicoccaceae family (ASV 36, CLR: 0.78) were all more abundant in the group A, while one classified as M. ruminantium (ASV 4, CLR: 1.79) and other as Methanosphaera group ISO3-F5 (ASV 33, CLR: 0.33) were more abundant in the group B (Supplementary Figure 6). Seven archaeal ASVs were DA in the fecal microbiome. From these, the ASVs classified as M. gottschalkii (ASVs 2, 13 and 11, CLR > 1.5) and M. smithii (ASV 28, CLR: 1.19) were more abundant in the group A, while M. ruminantium (ASV 4, CLR: 2) and Methanosphaera group ISO3-F5 (ASVs 5 and 33, CLR > 0.8) were more abundant in the group B (Supplementary Figure 7). No DA ASVs of protozoa origin were observed.
Discrimination between dietary treatment groups with Random Forest classification models
Random forest (RF) classification models were trained using CLR transformed relative abundances of each dataset, to test if the microbiome populations at ASV level could be used to discriminate the treatment group. Random forest has been shown to be the most accurate Machine Learning (ML) model for microbiome data analysis [36]. This method has the ability to discriminate groups, while considering interrelationships in high dimensional data [37]. The trained models resulted in high cross-validation scores for the bacteria test sets (r2=0.89 for rumen, r2=0.84 for feces), for archaea (r2=0.86 for rumen and r2=0.82 for feces) but not for protozoa (r2=0.57).
The feature importance function was used to select ASVs that contributed the most to the model's accuracy and to optimize the models. In short, the number of predictors were reduced to those with a contribution value >=0.01 to retrain the models, this resulted in 16 of 1683 as predictors for bacteria, 27 of 118 for archaea, and 30 of 52 for protozoa in rumen, while 22 of 1077 ASVs were predictors for bacteria and 33 of 88 for archaea in feces, respectively. This feature reduction resulted in an increased cross-validation for bacteria (r2=0.91 for rumen and r2=0.94 for feces), archaea (r2=0.91 for rumen and r2=0.86 for feces) and for protozoa (r2=0.71) with high recall and precision scores (Supplementary Table 1). Predictors used are available in the Supplementary Table 2.
Association between bacterial and archaeal ASVs found in rumen and feces and CH4 emission.
Previous analysis showed a significant difference in the mean CH4 emission of experimental groups, with group A (estimated mean = 179.11) emitting more methane than group B (estimated mean = 160.97). In order to investigate the proportion of variation of CH4 emissions explained by the microbiome composition of these animals, Linear mixed models were used with experimental groups information as fixed effects, weight and slaughter groups as co-variables, daily mean CH4 emission (g/day) as the dependent variable and individual log-transformed ASVs abundances as independent variables. This analysis identified significant associations between bacteria and archaea and CH4 emission in both environments. Within the rumen microbiome, the ASV 40, a Pseudobutyrivibrio (β=16.5contribution = 9.7%) and the ASV 44, a Bacteroidales (β=-2.6, contribution = -1.3%), were associated with CH4 emission phenotype (Figure 3).
Furthermore, we identified two bacterial ASVs in the fecal microbiome that were positively associated with CH4 emission: ASV 0, a Succinivibrio (β=10.2, contribution =6%) and the ASV 36, a Parabacteroides (β=2.9, contribution = 1.7%). Also, there were four bacterial ASV in this biome that were negatively associated with CH4 emission, ASV 35, a Ruminococcaceae UCG-005 (β=-12.5, contribution = -7.3%), ASV 39, a Phascolarctobacterium (β=-3.6, contribution = -1.8%), ASV 43, a Bacteroides (β=-2.9, contribution = -1.7%) and ASV 51, an Akkermansia (β=-2.5 contribution = -1.5%) (Figure 3). In addition, a single archaea ASV classified as M. gottschalkii was identified as positively associated with CH4 emission, the ARQ ASV 1 (β=4.21, contribution = 2.4%). There was no significant associations between CH4 emission and fecal archaea ASVs and protozoa ASVs.