A data structure and function classification based method to evaluate clustering models for gene expression data

摘要

Objective:To establish a systematic framework for selecting the best clustering algorithm and provide an evaluation method for clustering analyses of gene expression data. Methods: Based on data structure (internal information) and function classification (external information), the evaluation of gene expression data analyses were carried out by using 2 approaches. Firstly, to assess the predictive power of clusteringalgorithms, Entropy was introduced to measure the consistency between the clustering results from different algorithms and the known and validated functional classifications. Secondly, a modified method of figure of merit (adjust-FOM) was used as internal assessment method. In this method, one clustering algorithm was used to analyze all data but one experimental condition, the remaining condition was used to assess the predictive power of the resulting clusters. This method was applied on 3 gene expression data sets (2 from the Lyer's Serum Data Sets, and 1 from the Ferea's Saccharomyces Cerevisiae Data Set). Results: A method based on entropy and figure of merit (FOM) was proposed to explore the results of the 3 data sets obtained by 6 different algorithms, SOM and Fuzzy clustering methods were confirmed to possess the highest ability to cluster. Conclusion: A method based on entropy is firstly brought forward to evaluate clustering analyses.Different results are attained in evaluating same data set due to different function classification. According to the curves of adjust_FOM and Entropy_FOM, SOM and Fuzzy clustering methods show the highest ability to cluster on the 3 data sets.