SPRING: A METHOD FOR IDENTIFYING DIFFERENTIALLY EXPRESSED GENES IN MICROARRAY DATA

Analysis of 'omics' data is a central issue of system biology. As one of the most widely used 'omics' data, gene expression profiles from microarray experiments are applied to many frontier studies. The first and important step to analyze microarray data is to identify differentially expressed genes (DEGs) under two experimental conditions. Thereby, several DEG-identifying algorithms have been proposed. However, both traditional algorithms, such as Fold-Change, T-test and Significance Analysis of Microarrays (SAM), and modern ones, such as Rank Product, Outlier Robust t-statistic and Outlier Sums, are statistics-based approaches with the same core idea, which considers DEGs as the differences between two series of numbers. We present a novel view based on the hypothesis that DEGs are the differences between two input modes rather than the differences between two digital series, and then propose a novel non-statistical algorithm based on this idea, named Spring (SPG), which uses a Self-Organization Map (SOM) neural network to detect the input modals of DEGs under two sets of conditions. Firstly, the input matrix for SOM is constructed by reconstruction of the gene expression matrix, amplification of the difference of DEG and use of pairs of units divided from reconstructed gene expression matrix; and then, the strategy to improve the accuracy and stability is proposed by the Mass Spring Model, Minimum Spanning Tree Clustering and fuzzy clustering matrix. Compared with T-test and SAM, our algorithm obtains more DEGs in higher accuracy from both simulation and Homo sapiens datasets. Especially, we describe the details to transform SPG to a meta-analysis algorithm at the end.