Missing Value Estimation Algorithms on Cluster and Representativeness Preservation of Gene Expression Microarray Data

TL;DR

This paper investigates how different missing value imputation methods affect clustering outcomes in gene expression microarray data, emphasizing the importance of preserving original data clusters for accurate biological interpretation.

Contribution

It highlights the significant impact of imputation methods on clustering results and advocates for developing techniques that maintain original data structures rather than solely focusing on value accuracy.

Findings

Extensive differences in clustering outcomes due to imputation methods.

Imputation quality significantly influences data interpretation.

Traditional methods may distort original data clusters.

Abstract

Missing values are largely inevitable in gene expression microarray studies. Data sets often have significant omissions due to individuals dropping out of experiments, errors in data collection, image corruptions, and so on. Missing data could potentially undermine the validity of research results - leading to inaccurate predictive models and misleading conclusions. Imputation - a relatively flexible, general purpose approach towards dealing with missing data - is now available in massive numbers, making it possible to handle missing data. While these estimation methods are becoming increasingly more effective in resolving the discrepancies between true and estimated values, its effect on clustering outcomes is largely disregarded. This study seeks to reveal the vast differences in agglomerative hierarchal clustering outcomes estimation methods can construct in comparison to the precision exhibited (presented through the cophenetic correlation coefficient) in comparison to their high efficiency and effectivity in value preservation of true and imputed values (presented through the root-mean-squared error). We argue against the traditional approach towards the development of imputation methods and instead, advocate towards methods that reproduce a data set's original, natural cluster. By using a number of advanced imputation methods, we reveal extensive differences between original and reconstructed clusters that could significantly transform the interpretations of the data as a whole.