Statistical Analysis of Microarray Data Clustering using NMF, Spectral Clustering, Kmeans, and GMM
IEEE/ACM Transactions on Computational Biology and Bioinformatics
|September 21, 2020
Summary
Statistical analysis reveals Nonnegative Matrix Factorization (NMF) and k-means clustering perform similarly on gene expression data, outperforming spectral clustering. Gaussian Mixture Models (GMM) show superior performance, suggesting data follows Gaussian distributions.
Area of Science:
- Bioinformatics
- Computational Biology
- Machine Learning
Background:
- Clustering algorithms like Nonnegative Matrix Factorization (NMF), spectral clustering, k-means, and Gaussian Mixture Models (GMM) are widely used for analyzing microarray gene expression data.
- Limited statistical analysis exists to rigorously compare the performance differences between these unsupervised learning methods in this domain.
Purpose of the Study:
- To conduct a statistical analysis comparing the performance of NMF, spectral clustering, GMM, and k-means algorithms.
- To evaluate these algorithms on eleven publicly available microarray gene expression datasets.
- To determine the significance of performance differences across various cluster numbers (2-10).
Main Methods:
- Implementation and evaluation of ten NMF variations, six spectral clustering algorithms, four GMM approaches, and standard k-means.
- Clustering of eleven diverse microarray gene expression datasets.
- Statistical analysis to assess performance differences and multidimensional scaling for visual inspection of data structures.
Main Results:
- Nonnegative Matrix Factorization (NMF) and k-means demonstrated statistically similar performance, outperforming spectral clustering.
- Visual inspection via multidimensional scaling plots suggested a lack of manifold structures in the datasets.
- Gaussian Mixture Models (GMM) exhibited superior performance, indicating that the gene expression datasets likely follow Gaussian distributions, evidenced by elliptical cluster boundaries.
Conclusions:
- NMF and k-means are robust choices for gene expression clustering, while spectral clustering may be less suitable due to the absence of manifold structures.
- The superior performance of GMM suggests that gene expression data often adheres to Gaussian distributions.
- Further research could explore tailored algorithms for non-Gaussian or complex data structures in transcriptomics.
