Comprehensive evaluation of metabolite imputation methods and the effect of imputation accuracy on the performance of metabolite-trait association using large-scale metabolite datasets in cattle.

Abstract

As metabolomics data are typically acquired across multiple batches over months or even years, the integrated metabolite data sets often face complex missing-value challenges arising from fluctuations in instrument performance, variation in detection windows, and heterogeneity among analytical platforms. Accurate imputation of missing values is therefore a crucial strategy to addressing these challenges and enhancing the performance of downstream metabolomics analyses. In this study, we utilized large-scale authentic plasma and serum metabolite data sets (n = ∼3,000 metabolites) in Holstein cattle (n = ∼3,000 individuals), with each data set consisting of 2 batches, to evaluate the performance of 5 imputation methods (MIRTH, PIXANT, EM, KNN, and RF) to impute missing metabolites within batches and non-overlapping metabolites between batches. We demonstrated that RF yielded the best imputation performance in most cases, with accuracies reaching approximately 0.6 ∼0.7 for within-batch imputation and approximately 0.3 ∼0.4 for cross-batch imputation. Subsequently, we assessed the effect of imputation accuracy on the performance of metabolite-trait association analysis for 5 milk production traits using RF-imputed metabolite data. Compared with significantly associated metabolites identified using the original metabolite data, analysis using the imputed metabolite data resulted in substantial reductions in statistical power as well as increases in false-positive associations, particularly when imputed data from different batches were merged. Further efforts are needed to improve the accuracy for cross-batch imputation by developing new imputation methods and increasing the number of overlapping metabolites across batches.