Biological impact of missing-value imputation on downstream analyses of gene expression profiles

Motivation: Microarray experiments frequently produce multiple missing values (MVs) due to flaws such as dust, scratches, insufficient resolution or hybridization errors on the chips. Unfortunately, many downstream algorithms require a complete data matrix. The motivation of this work is to determine the impact of MV imputation on downstream analysis, and whether ranking of imputation methods by imputation accuracy correlates well with the biological impact of the imputation. Methods: Using eight datasets for differential expression (DE) and classification analysis and eight datasets for gene clustering, we demonstrate the biological impact of missing-value imputation on statistical downstream analyses, including three commonly employed DE methods, four classifiers and three gene-clustering methods. Correlation between the rankings of imputation methods based on three root-mean squared error (RMSE) measures and the rankings based on the downstream analysis methods was used to investigate which RMSE measure was most consistent with the biological impact measures, and which downstream analysis methods were the most sensitive to the choice of imputation procedure. Results: DE was the most sensitive to the choice of imputation procedure, while classification was the least sensitive and clustering was intermediate between the two. The logged RMSE (LRMSE) measure had the highest correlation with the imputation rankings based on the DE results, indicating that the LRMSE is the best representative surrogate among the three RMSE-based measures. Bayesian principal component analysis and least squares adaptive appeared to be the best performing methods in the empirical downstream evaluation.