An artificial intelligence approach of feature engineering and ensemble methods depicts the rumen microbiome contribution to feed efficiency in dairy cows.

Genetic selection has remarkably helped U.S. dairy farms to decrease their carbon footprint by more than doubling milk production per cow over time. Despite the environmental and economic benefits of improved feed and milk production efficiency, there is a critical need to explore phenotypical variance for feed utilization to advance the long-term sustainability of dairy farms. Feed is a major expense in dairy operations, and their enteric fermentation is a major source of greenhouse gases in agriculture. The challenges to expanding the phenotypic database, especially for feed efficiency predictions, and the lack of understanding of its drivers limit its utilization. Herein, we leveraged an artificial intelligence approach with feature engineering and ensemble methods to explore the predictive power of the rumen microbiome for feed and milk production efficiency traits, as rumen microbes play a central role in physiological responses in dairy cows. The novel ensemble method allowed to further identify key microbes linked to the efficiency measures. We used a population of 454 genotyped Holstein cows in the U.S. and Canada with individually measured feed and milk production efficiency phenotypes. The study underscored that the rumen microbiome is a major driver of residual feed intake (RFI), the most robust feed efficiency measure evaluated in the study, accounting for 36% of its variation. Further analyses showed that several alpha-diversity metrics were lower in more feed-efficient cows. For RFI, [Ruminococcus] gauvreauii group was the only genus positively associated with an improved feed efficiency status while seven other taxa were associated with inefficiency. The study also highlights that the rumen microbiome is pivotal for the unexplained variance in milk fat and protein production efficiency. Estimation of the carbon footprint of these cows shows that selection for better RFI could reduce up to 5 kg of diet consumed per cow daily, potentially reducing up to 37.5% of CH4. These findings shed light that the integration of artificial intelligence approaches, microbiology, and ruminant nutrition can be a path to further advance our understanding of the rumen microbiome on nutrient requirements and lactation performance of dairy cows to support the long-term sustainability of the dairy community.

Host and rumen microbiome contributions to feed efficiency traits in Holstein cows.

It is now widely accepted that dairy cow performance is influenced by both the host genome and rumen microbiome composition. The contributions of the genome and the microbiome to the phenotypes of interest are quantified by heritability (h2) and microbiability (m2), respectively. However, if the genome and microbiome are included in the model, then the h2 reflects only the contribution of the direct genetic effects quantified as direct heritability (hd2), and the holobiont effect reflects the joint action of the genome and the microbiome, quantified as the holobiability (ho2). The objectives of this study were to estimate h2, hd2,m2, and ho2 for dry matter intake, milk energy, and residual feed intake; and to evaluate the predictive ability of different models, including genome, microbiome, and their interaction. Data consisted of feed efficiency records, SNP genotype data, and 16S rRNA rumen microbial abundances from 448 mid-lactation Holstein cows from 2 research farms. Three kernel models were fit to each trait: one with only the genomic effect (model G), one with the genomic and microbiome effects (model GM), and one with the genomic, microbiome, and interaction effects (model GMO). The model GMO, or holobiont model, showed the best goodness-of-fit. The hd2 estimates were always 10% to 15% lower than h2 estimates for all traits, suggesting a mediated genetic effect through the rumen microbiome, and m2 estimates were moderate for all traits, and up to 26% for milk energy. The ho2 was greater than the sum of hd2 and m2, suggesting that the genome-by-microbiome interaction had a sizable effect on feed efficiency. Kernel models fitting the rumen microbiome (i.e., models GM and GMO) showed larger predictive correlations and smaller prediction bias than the model G. These findings reveal a moderate contribution of the rumen microbiome to feed efficiency traits in lactating Holstein cows and strongly suggest that the rumen microbiome mediates part of the host genetic effect.

Characterization of rumen microbiome and metabolome from oro-esophageal tubing and rumen cannula in Holstein dairy cows.

Less invasive rumen sampling methods, such as oro-esophageal tubing, became widely popular for exploring the rumen microbiome and metabolome. However, it remains unclear if such methods represent well the rumen contents from the rumen cannula technique. Herein, we characterized the microbiome and metabolome in the rumen content collected by an oro-esophageal tube and by rumen cannula in ten multiparous lactating Holstein cows. The 16S rRNA gene was amplified and sequenced using the Illumina MiSeq platform. Untargeted metabolome was characterized using gas chromatography of a time-of-flight mass spectrometer. Bacteroidetes, Firmicutes, and Proteobacteria were the top three most abundant phyla representing ~ 90% of all samples. Although the pH of oro-esophageal samples was greater than rumen cannula, we found no difference in alpha and beta-diversity among their microbiomes. The overall metabolome of oro-esophageal samples was slightly different from rumen cannula samples yet more closely related to the rumen cannula content as a whole, including its fluid and particulate fractions. Enrichment pathway analysis revealed a few differences between sampling methods, such as when evaluating unsaturated fatty acid pathways in the rumen. The results of the current study suggest that oro-esophageal sampling can be a proxy to screen the 16S rRNA rumen microbiome compared to the rumen cannula technique. The variation introduced by the 16S rRNA methodology may be mitigated by oro-esophageal sampling and the possibility of increasing experimental units for a more consistent representation of the overall microbial population. Studies should consider an under or over-representation of metabolites and specific metabolic pathways depending on the sampling method.

Technical note: Validation of a system for monitoring individual behavior of Holstein cows.

The objectives of the 2 studies reported herein were to validate the accuracy of an automated monitoring device (AMD) to detect side lying, resting, activity, rumination, eating, walking, and panting in nonlactating and lactating dairy cows. Additionally, we aimed to determine whether the total time per cow-state recorded by the AMD within a 30-min interval corresponds to the total time per cow-state recorded simultaneously by visual observation. Study personnel (n = 2) observed pregnant nonlactating Holstein cows (n = 10) for 30 min in the morning and 30 min in the afternoon for 6 consecutive days and recorded continuously each cow-state. In study 2, study personnel (n = 2) observed lactating Holstein cows (n = 10) for 30 min in the morning and 30 min in the afternoon for 6 consecutive days. In both studies, cow-state was recorded every second, and within 1 min, the most prevalent cow-state was considered to be the behavior presented by the cow during that interval. Using the observer as the gold standard, test characteristics were calculated for the minute-by-minute interval analyses. For the 30-min interval analyses, the concordance correlation coefficient (pc) and the coefficient of determination (R2) between the total minutes for each cow-state recorded by the observer and the AMD were calculated. In study 1, for the minute-by-minute interval analyses, test characteristics were high for rumination (≥90.1%) and eating (≥73.8%), moderate for resting (≥62.9%), but negligible for medium activity (≥17%). For the 30-min interval analyses, the correlations between the total time of visual observations compared with the total time recorded by AMD for rumination (R2 = 0.97, pc = 0.98) and eating (R2 = 0.91, pc = 0.94) were very high, for resting (R2 = 0.77, pc = 0.79) was high, and for medium activity (R2 = 0.41, pc = 0.41) was low. In study 2, for the minute-by-minute interval analyses, test characteristics were high for rumination (≥79.4%), eating (≥74.2%), and resting (≥73.0%), but they were low for panting (≥31.3%) and negligible for medium activity (≥22.2%). For the 30-min interval analyses, the correlations were similar to study 1 (rumination: R2 = 0.85, pc = 0.91; eating: R2 = 0.95, pc = 0.97; resting: R2 = 0.84, pc = 0.90; medium activity: R2 = 0.44, pc = 0.57; and panting: R2 = 0.21, pc = 0.42). In summary, the AMD used in this study provided accurate data regarding resting, rumination, and eating of pregnant nonlactating and lactating Holstein cows.