ECCB 2002 Poster sorted by: Author | Number

Next | Previous poster (in order of the view you have selected)

Title: Time series microarray data analysis using ISA grouping biologically related functions
Hei Jin, Kim,
Intelligent Multimedia lab. C.S.E. of POSTECH (POhand University of Science and TECHnology

We applied Independent Subspace Analysis (ISA) to gene expression data, deriving independent gene expression patterns under some gene interaction. ISA algorithm is made by Aapo Hyvärinen(2001). Both of ICA and ISA take independence to extract important feature. In contrast to ordinary ICA, however, ISA resembling Gabor functions is assumed that some components can be divided into n-tuples, such that components inside a given n-tuple may be dependent each other, instead different tuples are independent. Hence it can minimize information loss by keeping the dependency and extracts invariant features. In the microarray data analysis case, ISA includes two features which make it useful: the ability to assign genes to multiple co-expression pattern groups and the capability to cluster key genes which determine each critical point of cell cycle. ISA leads to have simultaneously activated gene a cluster and to group similar patterns. The pattern group highlights biological functions to reduce noise and also to maintain the useful interaction functions. The distribution of every gene corresponding to a pattern cluster can be applied not only to predict the cell cycle pattern but also to give valuable information to explore unknown gene's function. To cluster genes according to their functions based on gene expression data, various algorithms such as PCA, ICA, Bayesian decomposition etc. have been applied. Those algorithms ignore the dependency between genes or between biological functions. We compared our results to previously applied algorithms, PCA and ICA. Using PCA, the dominant G1's (37.5% of the whole genes ) make a well-constructed cluster but hardly get other significant clusters ICA case, difficult to understand the component features because of the dependencies in biological functions. A gene belonging to one of patterns of ISA was selected by the local energies, which maximize the likelihood ISA find independent subspaces allowing the dependencies of the patterns in the subspace ISA is based on similar principal of Garbor filter and therefore ISA has feature invariance. Components in the same subspace look slightly different but the functions were revealed similar to the other.
In short, ISA is an unsupervised learning method for gene expression data analysis, based on independent subspace analysis (ISA) which aims at finding independent feature subspace of multivariate data in dependent sub-components. In a feature subspace, patterns of each component are shown the properties: slight time invariance or the reverse sign invariance. As shown by Liebermeister, each component correspond to one of biological functions. Our results make a set of function( a feature space) assumed that biologically related each other. Further information about ISAcycle is freely available at
[1] Wolfram Liebermeister (2002), Linear modes of gene expression determined by independent component analysis, Bioinformatics vol.18 51-60.
[2] T.D.Moloshok et al.(2002), Application of Bayesian Decomposition for analyzing microarray data, Bioinformatics vol.18 566-575.
[3] Aapo Hyvärinen et al.(1999), Emergence of phase and shift invariant features by decomposition of natural images into independent feature subspace, Neural Computation August 1999.
[4] Aapo Hyvärinen et al.,(2001) A two-layer sparse coding model learns simple and complex cell receptive fields and topography from natural images, Vision Research 41 2413-2423.
[5] Paul T. Spellman(1998) et al., Comprehensive Identification of Cell Cycle-regulated Genes of the Yeast Saccharomyces cerevisiae by Microarray Hybridization, Molecular Biology of the Cell, vol.9 3273-3297.
[6] Cardoso, J.F.(1998) Multidimensional independent component analysis. In Proc. ICASSP'98, Seattle.
[7] Kohonen, T.(1996) Emergence of invariant-feature detectors in the adaptive-subspace self-organizing map. Biological Cybernetics, 75:281-291.