The influence of missing value imputation on detection of differentially expressed genes from microarray data

Area of Science:

• Bioinformatics
• Genomics
• Statistical Analysis

Background:

• Missing values in microarray data analysis pose significant challenges.
• Previous imputation method comparisons focused on value accuracy, not analytical impact.
• Existing research primarily addresses missing at random data, neglecting missing not at random entries.

Purpose of the Study:

• To investigate the influence of imputation methods on detecting differentially expressed genes in cDNA microarray data.
• To introduce and evaluate a novel imputation method, LinImp, against KNNimpute.
• To assess the impact of different missing data types (random vs. not at random) on gene detection.

Main Methods:

• Applied ANOVA for microarrays and Significance Analysis of Microarrays (SAM).
• Evaluated gene detection loss due to imputation using a novel measure.
• Proposed and implemented LinImp, a linear model-based imputation method.
• Compared LinImp with KNNimpute for various missing data percentages and types.

Main Results:

• The type of missingness significantly affects gene detection; 5% missing not at random is comparable to 10-30% missing at random.
• LinImp outperforms KNNimpute, resulting in fewer lost differentially expressed genes (twice as many lost with KNNimpute for 10% missing at random).
• The proposed measure of imputation influence provides insights beyond traditional root mean squared error.

Conclusions:

• Imputation strategy critically influences the identification of differentially expressed genes.
• LinImp offers improved performance over KNNimpute for microarray data imputation.
• Considering the type of missingness is crucial for accurate gene expression analysis.