A cluster-directed framework for neighbour based imputation of missing value in microarray data

DNA microarray has been the most widely used functional genomics approach in bioinformatics. However, microarray data suffer from frequent missing values due to various experimental and data handling reasons. Leaving this unsolved may degrade the reliability of any consequent downstream analysis. As such, missing value imputation has been recognised as an important pre-processing step, which can yield the quality of data and its interpretation. Several techniques found in the literature have successfully exploited the characteristics and relations among a set of genes closest to the one under examination. However, the selection of so-called nearest neighbours is based simply on proximity between gene pairs, without taking the structural or grouping information into account. In response, this paper proposes a novel cluster-directed framework (CFNI: Cluster-directed Framework for Neighbour-based Imputation), in which data clustering is uniquely used to guide the identification of nearest neighbours. This allows a more accurate imputed value to be derived. Not only it performs better than several benchmark methods on published microarray data sets; it is also generalised such that any neighbour-based imputation technique can be coupled with the proposed model. This has been successfully demonstrated with both single pass and iterative models.