Autor: |
Bozanis, Panayiotis, Houstis, Elias N., Xiong, Fei, Huang, Heng, Ford, James, Makedon, Fillia S., Pearlman, Justin D. |
Zdroj: |
Advances in Informatics; 2005, p437-447, 11p |
Abstrakt: |
Microarray gene expression data contains informative features that reflect the critical processes controlling prominent biological functions. Feature selection algorithms have been used in previous biomedical research to find the "marker" genes whose expression value change corresponds to the most eminent difference between specimen classes. One problem encountered in such analysis is the imbalance between very large numbers of genes versus relatively fewer specimen samples. A common concern, therefore, is "overfitting" the data and deriving a set of marker genes with low stability over the entire set of possible specimens. To address this problem, we propose a new test environment in which synthetic data is perturbed to simulate possible variations in gene expression values. The goal is for the generated data to have appropriate properties that match natural data, and that are appropriate for use in testing the sensitivity of feature selection algorithms and validating the robustness of selected marker genes. In this paper, we evaluate a statistically-based resampling approach and a Principal Components Analysis (PCA)-based linear noise distribution approach. Our results show that both methods generate reasonable synthetic data and that the signal/noise rate (with variation weights at 5%, 10%, 20% and 30%) measurably impacts the classification accuracy and the marker genes selected. Based on these results, we identify the most appropriate marker gene selection and classification techniques for each type and level of noise we modeled. Keywords: Gene Selection, Stability Measurement, Sample Classification, Microarray Data [ABSTRACT FROM AUTHOR] |
Databáze: |
Supplemental Index |
Externí odkaz: |
|