Impact of reducing metagenomic sequencing depth on phenotypic prediction accuracy of feed intake and average daily gain in beef cattle

Metagenomic information can aid in both genomic and phenotypic predictions of economically relevant traits. Financial restraints often result in a trade-off between the number of samples sequenced and the depth of sequencing. Therefore, it is critical to understand how changes in sequencing depth impact phenotypic prediction accuracy to make optimal use of resources. This study utilized host genomic and rumen metagenomic information of 717 beef cattle to make phenotypic predictions for average daily dry matter intake (ADDMI) and average daily gain (ADG). Metagenomic samples were sequenced at an average depth of 20 million reads (20M set) and were downsampled to 50% (10M set), 25% (5M set), and 10% of the reads (2M set). Rumen microbial open reading frames (ORF) were predicted from each set of reads and used to define a random metagenomic effect in a mixed model framework. Variance components were estimated for each model using all available data, i.e., no masking of phenotypes. Cross-validation schemes were utilized to determine prediction accuracy. Models which incorporated host genomic and metagenomic information explained more variation and generally had greater prediction accuracies than models with either effect alone. Models using the 2M or 5M set resulted in smaller microbiability estimates and lower prediction accuracy for both ADDMI and ADG compared to models using the 10M or 20M sets, though these differences were often not large when measures of uncertainty were considered. For ADDMI, there were only slight differences in microbiability and prediction accuracy between different downsampled sets in most scenarios. For ADG, the 20M set had roughly equivalent microbiability estimates as the other sets but also had a notably greater prediction accuracy, dependent on cross-validation scheme. Spearman correlations of metagenomic effect solutions, termed the estimated metagenomic value (EMV), between all sets for all models were always >0.90. However, the correlations between the EMV for models with the 5M, 10M, and 20M sets were always higher than those with the EMV from the 2M set. The 10M and 20M EMV always had correlations >0.98. Thus, dependent on trait and reference population composition, metagenomic predictions from data sequenced at a depth of 2-10 million reads per sample may yield results approximately equivalent to those from data sequenced at 20 million reads per sample in terms of variance explained and phenotypic prediction accuracy.