Rumen metagenomics of Tan sheep with divergent growth rate: functional gene and MAG profiles do not distinguish the groups

Abstract Background Feed conversion ratio (FCR) is a key economic trait in lamb production, and the rumen microbiome governs the energy the host obtains from feed. Because genus-level biomarkers are inconsistent across studies, we examined whether efficient and inefficient Tan sheep differ at the level of metabolic genes and metagenome-assembled genomes (MAGs). Sixteen Tan sheep, drawn from a finishing cohort of 100 animals, were ranked by average daily gain (ADG) and split into a high-growth/low-FCR group (n = 10) and a low-growth/high-FCR group (n = 6); individual feed intake was not recorded, so ADG, not measured FCR, defined the groups. Rumen contents were deep-shotgun sequenced, assembled and binned. Of 3,283 retained KEGG orthologues (KOs), a candidate required both a Wilcoxon between-group difference (P < 0.05) and a Spearman ADG correlation (P < 0.05) with concordant direction, because no adjusted P reached 0.05 (smallest Benjamini-Hochberg-adjusted P = 0.41). Because the two layers are not independent - the groups were formed by ranking the same ADG values - we calibrated the size of the resulting candidate set against a label-permutation null distribution, and we evaluated a Lasso-plus-logistic-regression classifier under nested leave-one-out cross-validation, in which feature selection was repeated inside every fold and the permutation test re-ran the entire pipeline. Results Global community structure did not differ (ANOSIM taxonomy R = -0.139, P = 0.95; function R = -0.064, P = 0.682). The two-layer screen yielded 36 direction-concordant KOs (54 between-group and 94 ADG-correlated; union 112), enriched for sugar-transport (ABC/PTS) and energy genes, whereas canonical VFA genes showed no detectable difference (buk, pct, acs; all P > 0.7; absence of evidence rather than evidence of absence). However, 36 double-evidence KOs did not exceed the label-permutation null expectation (null mean 67, 95th percentile 195; P = 0.59), and the same was true for each single layer (between-group: 54 observed versus null mean 131, P = 0.83; ADG-correlated: 94 observed versus null mean 167, P = 0.67). A classifier built on the 36 KOs and evaluated with the feature set fixed in advance reached AUC = 1.000 and 87.5% accuracy, but under nested cross-validation the same pipeline gave AUC = 0.767 and 56.2% accuracy (9/16), which did not differ significantly from the permutation null (null mean AUC = 0.525, 95th percentile 0.883; P = 0.190). Initial body weight was confounded with group (Spearman rho with ADG = 0.67, P = 0.005). Only three MAGs were double-evidence, all inefficient-enriched and negatively correlated with ADG, representing unnamed novel genera (two Firmicutes-affiliated and one Bacteroidota-affiliated). Conclusions In this cohort of 16 Tan sheep, growth-efficiency divergence was not accompanied by any statistically detectable remodelling of rumen functional genes: the carbohydrate-uptake and energy-metabolism candidate panel did not exceed permutation-based null expectations, no KO approached FDR significance, and no classifier generalised under nested validation. The 591 MAGs recovered here provide a breed-specific genome resource for Tan sheep, and the contrast between the non-nested (AUC = 1.000) and nested (AUC = 0.767) estimates quantifies how much apparent signal feature selection performed outside cross-validation can manufacture at this sample size.