Selections of data preprocessing methods and similarity metrics for gene cluster analysis

Clustering is one of the major exploratory techniques for gene expression data analysis. Only with suitable similarity metrics and when datasets are properly preprocessed, can results of high quality be obtained in cluster analysis. In this study, gene expression datasets with external evaluation criteria were preprocessed as normalization by line, normalization by column or logarithm transformation by base-2, and were subsequently clustered by hierarchical clustering, k-means clustering and self-organizing maps (SOMs) with Pearson correlation coefficient or Euclidean distance as similarity metric. Finally, the quality of clusters was evaluated by adjusted Rand index. The results illustrate that k-means clustering and SOMs have distinct advantages over hierarchical clustering in gene clustering, and SOMs are a bit better than k-means when randomly initialized. It also shows that hierarchical clustering prefers Pearson correlation coefficient as similarity metric and dataset normalized by line. Meanwhile, k-means clustering and SOMs can produce better clusters with Euclidean distance and logarithm transformed datasets. These results will afford valuable reference to the implementation of gene expression cluster analysis.

Selections of data preprocessing methods and similarity metrics for gene cluster analysis

Clustering is one of the major exploratory techniques for gene expression data analysis. Only with suitable similarity metrics and when datasets are properly preprocessed, can results of high quality be obtained in cluster analysis. In this study, gene expression datasets with external evaluation criteria were preprocessed as normalization by line, normalization by column or logarithm transformation by base-2, and were subsequently clustered by hierarchical clustering, k-means clustering and self-organizing maps (SOMs) with Pearson correlation coefficient or Euclidean distance as similarity metric. Finally, the quality of clusters was evaluated by adjusted Rand index. The results illustrate that k-means clustering and SOMs have distinct advantages over hierarchical clustering in gene clustering, and SOMs are a bit better than k-means when randomly initialized. It also shows that hierarchical clustering prefers Pearson correlation coefficient as similarity metric and dataset normalized by line. Meanwhile, k-means clustering and SOMs can produce better clusters with Euclidean distance and logarithm transformed datasets. These results will afford valuable reference to the implementation of gene expression cluster analysis.