This study aimed at predicting feed conversion ratio (FCR) of young rabbits from abundances of amplicon sequence variants (ASVs) to improve this trait by selecting animals with the most favorable microbiota and identifying the most relevant microorganisms involved in feed efficiency. Data come from two rabbit populations coming from paternal INRA 1001 line (the G10, selected for 10 generations for decreased residual feed intake and the G0 control produced from frozen embryos of the common ancestor line). There were 296 and 292 FCR data from G10 and G0 individuals, respectively. Phenotypic data were pre-corrected for the systematic effects of group, batch, litter size and sex and the random litter effect. Sequence quality control and chimera removal were performed with the DADA2 pipeline. Samples with less than 5,000 final sequence counts and doubleton ASV were removed. The ASV counts of the final table (including 918 ASVs) were centered log-ratio transformed and corrected for batch effects with a surrogate variable analysis. Nested resampling for hyper-parameter tuning and prediction validation was implemented leading to 25 pairs of training/test sets. Bayesian regression models (Bayesian Lasso, Bayesian Ridge Regression and Reproducing Kernel Hilbert Spaces) and machine learning algorithms (Support vector machine and Elastic net) were fitted to all ASVs leading to an almost null prediction accuracy in all cases. Then, ASVs were ranked for their prediction importance using the permutation accuracy importance score in a Random Forest algorithm based on conditional inference and, different subsets of increasing size (50, 100, 150, 200, 300, 400, 500, All) of the most important ASVs and surrogate variables were used as predictors in the machine learning algorithms. The best performance and the most stable results were obtained with machine learning using the 100 most important ASVs being most of them assigned to order Clostridiales. The medians of the Spearman correlation (interquartile range) were 0.33 (0.09) and 0.32 (0.06) for SVM and ENET, respectively.