Iterative bicluster-based least square framework for estimation of missing values in microarray gene expression data

DNA microarray experiment inevitably generates gene expression data with missing values. An important and necessary pre-processing step is thus to impute these missing values. Existing imputation methods exploit gene correlation among all experimental conditions for estimating the missing values. However, related genes coexpress in subsets of experimental conditions only. In this paper, we propose to use biclusters, which contain similar genes under subset of conditions for characterizing the gene similarity and then estimating the missing values. To further improve the accuracy in missing value estimation, an iterative framework is developed with a stopping criterion on minimizing uncertainty. Extensive experiments have been conducted on artificial datasets, real microarray datasets as well as one non-microarray dataset. Our proposed biclusters-based approach is able to reduce errors in missing value estimation.

► Estimate missing values by ignoring unrelated genes and conditions. ► Iteratively select similar genes and conditions to improve accuracy in estimation. ► Estimation accuracy is significanlty improved, especially in bicluster region. ► Our algorithm is guaranteed to converge.