Abstract
Knowledge of diet composition and intake levels in beef cattle is valuable for post hoc feed traceability and for more accurate modelling of the diet impact on methane emissions and performance traits. However, a direct measure of this information can be costly and labour-intensive and is not always feasible. In this study, rumen metagenomic data combined with machine learning algorithms were used to predict diet type, nutritional composition, and intake levels. An external validation to assess the generalizability of the models was also performed. Rumen samples were collected from 142 animals belonging to two breeds, Luing (n = 70) and Charolais crossbred (n = 72), with 425.6 ± 43.5 d old and 461.9 ± 70.2 kg body weight. The animals participated in a 56-d feeding trial and were assigned to diets differing in forage-to-concentrate ratio, with 72 animals receiving a concentrate-based diet and 70 receiving a forage-based diet. Liquid ruminal contents were collected immediately postmortem and subsequently subjected to metagenomic sequencing. Based on these sequences, the relative abundance of microbial genes (MGs), microbial genera (MTs), and phyla were determined. The log-ratio between the abundances of Verrucomicrobia and Chlorobi discriminated diet type with an average classification accuracy of 0.86 ± 0.05, while using the log-ratio transformed abundances of 4769 MTs and MGs as predictors reached 0.90 ± 0.05. All this microbiome information was used in a random forest model to predict continuous values for nutritional diet components starch, crude protein, neutral and acid detergent fibre, and metabolizable and gross energy with external validation prediction accuracy values between 0.77 and 0.83. Microbiome features important for prediction of diet components such as fibre and starch included Mitsuokella, Selenomonas, and MGs involved in flagellar assembly and aminoacyl-tRNA biosynthesis. Microbiome data were more informative for predicting the feed composition than the amount of feed consumed, which reached a prediction accuracy of 0.27 ± 0.12 for dry matter intake (DMI). However, microbiome data can still be used as a screening tool to classify DMI into low, medium, or high with a classification accuracy of 0.74. Incorporating dietary information into linear phenotypic and genetic models to predict methane production (MP) and DMI reduced root mean square error (RMSE) by 26.9% and 9.6%, respectively, in the phenotypic model. In the genetic model, only MP showed a reduction in RMSE, with a 31% improvement. These findings highlight rumen microbiome data as a valuable tool for the post hoc prediction of feed composition in beef cattle.