Post-processing of Graph Regularized Non-negative Matrix Factorization Algorithm for Cancer Gene Clustering

With the rapid development of DNA microarray technology, the application of informatics research methods in oncology is becoming more and more popular. Because the gene expression data extraction experiment has the characteristics of a large number of genes and complex and changeable experimental conditions, clustering technology has been introduced into the field of molecular biology to assist research and analyze gene expression data. Currently, many clustering algorithms are applied in gene expression analysis. Nevertheless, owing to the high dimensionality of gene expression data, many algorithms face the problem of low computational efficiency. The method based on the graph theory thinks of the sample of gene expression data as the point in high-dimensional space. Its low sample performance determines that the constructed matrix is small in size. Therefore, it has lower computational complexity. Therefore, the graph regularized non-negative matrices factorization (GNMF) proposed by Cai et al. has been widely used for gene clustering. However, the deficiency of the GNMF algorithm is unstable with data variation for gene clustering. In this paper, we propose post-processing of Graph Regularized Nonnegative Matrix Factorization Algorithm for Gene Clustering, called pGNMF. In the pGNMF method, we first normalize the solution of GNMF, thereby reducing the sensitivity of the traditional GNMF method to the prior selection of genes or initial conditions, and effectively improving the robustness of the algorithm. Experimental results show that the proposed algorithms outperform existing GNMF algorithms for gene clustering.