CLUSFAVOR 5.0: hierarchical cluster and principal-component analysis of microarray-based transcriptional profiles
© Peterson, licensee BioMed Central Ltd 2002
Published: 24 June 2002
CLUSFAVOR (CLUSter and Factor Analysis with Varimax Orthogonal Rotation) 5.0 is a Windows-based computer program for hierarchical cluster and principal-component analysis of microarray-based transcriptional profiles. CLUSFAVOR 5.0 standardizes input data; sorts data according to gene-specific coefficient of variation, standard deviation, average and total expression, and Shannon entropy; performs hierarchical cluster analysis using nearest-neighbor, unweighted pair-group method using arithmetic averages (UPGMA), or furthest-neighbor joining methods, and Euclidean, correlation, or jack-knife distances; and performs principal-component analysis.
DNA microarrays are useful for identifying genes that are co-expressed in different phenotypes, experiments, or both. Genomic regions containing cis-regulatory motifs can be identified by exon mapping cDNAs of co-expressed genes using BLAST [1,2,3,4]. Because cis-regulatory motifs act as binding sites for transcription factors that control expression, co-expressed genes sharing the same regulatory motifs are likely to be networked and under the same regulatory control [5,6,7,8].
To identify co-expressed genes using DNA microarray data, one typically uses classification and data-reduction methods such as cluster analysis and principal-component analysis (PCA). The CLUSFAVOR 5.0 computer program was developed for hierarchical cluster analysis (HCA) and PCA of DNA microarray expression data. CLUSFAVOR 5.0 was developed under the Windows operating system using Microsoft Visual Basic , and can therefore be installed and run on any of the 32-bit versions of Windows (95, 98, NT, 2000, or XP). CLUSFAVOR can standardize expression data, sort, and perform HCA and PCA of arrays and genes. The program accommodates missing data, can calculate replicate averages, and determine replicate outliers and drop them from the analysis. CLUSFAVOR 5.0 has primarily been used for DNA microarray data, but can be used for numerical taxonomy to identify natural groupings of variables for non-genetic data. This report reviews user specifications, input file formats, numerical methods, and output formats for the CLUSFAVOR 5.0 computer program.
The CLUSFAVOR 5.0 program can open either a tab-delimited text file for a new run or a pre-formatted file containing output parameters and data from a previous run. The option to view results of previous runs quickly eliminates the long wait time needed when obtaining results of large processing jobs. When opening tab-delimited text files for a new run, CLUSFAVOR 5.0 will recognize any text file whose filename ends with .txt. File formats are described in the User's Guide, available for download (see Downloading files section).
Standardization of input data
Results of multivariate statistical methods such as HCA and PCA depend strongly on the scale (the range) of data used. A common method for standardizing input data during HCA involves treating the variables (or records) being clustered as the 'objects' and the records (or variables) as the 'attributes'. When HCA is being performed on the arrays as objects, standardized expression is calculated by subtracting the gene-specific average expression and dividing by the gene-specific standard deviation, as genes are the attributes. When cluster-analyzing the genes as objects, standardization is based on the array-specific average and standard deviation of expression. In this fashion, zero means are obtained for the attributes rather than the objects being clustered. The distance functions used for HCA in CLUSFAVOR 5.0 can be based on either Euclidean distance or correlation. When Euclidean distance is specified as the distance function and input data are standardized, a single round of standardization is performed which removes additive and multiplicative size displacements among expression profiles so that residual differences are detected. However, when correlation is specified as the distance function, the double round of standardization is more effective at removing size displacements between expression profiles. Standardization is also used by default for all PCA runs on arrays and genes. Summary statistics for array- and gene-specific average and standard deviation in expression can be saved in a tab-delimited output file.
When working with DNA microarray data, biologists often want to know which genes have the greatest or least expression or standard deviation across the arrays. When sorting is specified at run-time, CLUSFAVOR 5.0 calculates the gene-specific coefficient of variation, average and total expression, standard deviation, and performs an ascending quicksort of each parameter being considered. Results are saved in JPG image files using a color gradient of expression specified by the user, and are also written in tabular form to tab-delimited text files. All sorting results (JPG images and text) are linked with HTML (hypertext markup language) files for viewing.
Hierarchical cluster analysis (HCA)
Hierarchical cluster analysis (HCA) is an exploratory multivariate statistical method for identifying 'natural' groupings of objects considered in an analysis. The least distance, D(r,s), between two objects r and s (arrays r and s) consisting of n r , and n s elements is first identified. The distance function, D(r,s), can be based on either the Euclidean distance (0 ≤ D(r,s) < + ∞) or 1-correlation (0 ≤ D(r,s) ≤ 2). Objects r and s are joined to form a new 'node' u with n u = n r + n s elements. Next, distances between the newly formed node u (comprised of objects r and s) and all other nodes v (or objects) are calculated as
where D(r,v) is the distance between nodes r and v joined previously, D(s,v) is the distance between nodes s and v joined previously. Single linkage (nearest neighbor), unweighted pair-group method using arithmetic averages (UPGMA), or complete linkage (furthest neighbor) is specified by the user, and n u , n r , and n s are the number of objects in the nodes. After all the new distances between node u and other nodes have been calculated, a search for the smallest distance is conducted, followed by calculation of new distances. This is done repeatedly until all clusters have joined. The CLUSFAVOR 5.0 algorithm first clusters the arrays, and then the genes.
Principal-component analysis (PCA)
Principal-component analysis (PCA) is useful for reproducing the total variance among a large number of variables using a much smaller number of unobservable variables or dimensions called latent factors. CLUSFAVOR 5.0 uses the principal-component solution to the factor model for extracting factors (components). This is accomplished by use of the principal-axis theorem, which says that for a gene-by-gene (n × n) correlation matrix R, there exists a rotation matrix E and diagonal matrix Λ such that ERE' = Λ. The principal form of R is given as
where columns of E and E' are the eigenvectors and diagonal entries of Λ are the eigenvalues. In CLUSFAVOR 5.0, only components whose eigenvalues exceed unity, λ j > 1, are extracted from Λ and sorted such that λ1 ≥ λ2 ≥ ... ≥ λ m ≥ 1. The 'loading' or correlation between genes and extracted components is represented by a matrix in the form
Viewing results in HTML
Cluster image displays (such as that in Figure 1) for each group of 100 genes are saved separately in JPG format and are linked to an HTML file for viewing with a web browser. This enables the user to view all cluster images for thousands of genes and also to export results quickly to either public or password-protected directories for web publication or collaborative data analysis review. PCA does not depend on the number of genes in a run, and always generates JPG image files (such as Figures 4 and 5) that are linked to an HTML file for browser viewing. During HTML viewing of PCA output, a user can click on a command button in the HTML file to retrieve cDNA sequences (in FASTA format from the National Center for Biotechnology Information (NCBI)) for each group of genes identified. The cDNA sequences can then be used to search for upstream regions using BLAST; these can be used for cis-regulatory motif searching [2,3,4,5,6,7,8].
Benchmarking CLUSFAVOR with the statistical analysis package SPSS
Simulated gene expression data for 60 arrays and 120 genes were generated for benchmarking. Data in arrays (columns) 1-10 were based on 120 pseudo-random uniform variates, U(0,1). Arrays 11-20 were filled with 120 standard normal variates, N(0,1). Arrays 21-30 had elements distributed N(3,1), arrays 31-40 N(-3,1), 41-50 with N(0,3), and 51-60 with 120 variates distributed N(0,0.3). Various additive and multiplicative translations were applied in the rows so that expression also varied over the genes. Expression values in rows 1-40 (genes 1-40) were not transformed. However, in rows 41-60 a constant of 3 was added to all array elements, in rows 61-80 a constant of 3 was subtracted from all array values, in rows 81-100 a constant of 3 was multiplied with all array entries, and in rows 101-120 a constant of 0.3 was multiplied with all array elements.
CLUSFAVOR 5.0 can be downloaded from . Users must first install Version 2.0 in order to obtain the base set-up, and then can download the executable file (clusfavor.exe) for Version 5.0, and save this file into the directory where Version 2.0 was installed. CLUSFAVOR 5.0 is copyright protected against commercial gain, and has a 90-day non-exclusive license that can be extended free of charge for non-profit institutions.
Algorithm development for CLUSFAVOR 5.0 was supported by NCI grant CA-78199-04.
- BLAST. [http://0-www.ncbi.nlm.nih.gov.brum.beds.ac.uk/BLAST/]
- Suzuki Y, Ishihara D, Sasaki M, Nakagawa H, Hata H, Tsunoda T, Watanabe M, Komatsu T, Ota T, Isogai T, et al: Statistical analysis of the 5' untranslated region of human mRNA using "Oligo-Capped" cDNA libraries. Genomics. 2000, 64: 286-297. 10.1006/geno.2000.6076.PubMedView ArticleGoogle Scholar
- Arnone MI, Davidson EH: The hardwiring of development: organization and function of genomic regulatory systems. Development. 1997, 124: 1851-1864.PubMedGoogle Scholar
- Manson McGuire A, Church GM: Predicting regulons and their cis-regulatory motifs by comparative genomics. Nucleic Acids Res. 2000, 28: 4523-4530. 10.1093/nar/28.22.4523.PubMedView ArticleGoogle Scholar
- Wuensche A: Genomic regulation modeled as a network with basins of attraction. Pac Symp Biocomput. 1998, 89-102.Google Scholar
- D'haeseleer P, Liang S, Somogyi R: Genetic network inference: from co-expression clustering to reverse engineering. Bioinformatics. 2000, 16: 707-726. 10.1093/bioinformatics/16.8.707.PubMedView ArticleGoogle Scholar
- Tavazoie S, Hughes JD, Campbell MJ, Cho RJ, Church GM: Systematic determination of genetic network architecture. Nat Genet. 1999, 22: 281-285. 10.1038/10343.PubMedView ArticleGoogle Scholar
- Bussemaker HJ, Li H, Siggia ED: Regulatory element detection using correlation with expression. Nat Genet. 2001, 27: 167-171. 10.1038/84792.PubMedView ArticleGoogle Scholar
- Microsoft Visual Basic.NET. [http://msdn.microsoft.com/vstudio/]
- Ross DT, Scherf U, Eisen MB, Perou CM, Rees C, Spellman P, Iyer V, Jeffrey SS, Van de Rijn M, Waltham M, et al: Systematic variation in gene expression patterns in human cancer cell lines. Nat Genet. 2000, 24: 227-235. 10.1038/73432.PubMedView ArticleGoogle Scholar
- Heyer LJ, Kruglyak S, Yooseph S: Exploring expression data: identification and analysis of coexpressed genes. Genome Res. 1999, 9: 1106-1115. 10.1101/gr.9.11.1106.PubMedPubMed CentralView ArticleGoogle Scholar
- Kaiser HF: The varimax criterion for analytic rotation in factor analysis. Psychometrika. 1958, 23: 187-200.View ArticleGoogle Scholar
- SPSS. [http://www.spss.com]
- CLUSFAVOR. [http://mbcr.bcm.tmc.edu/genepi/]
- Eisen MB, Spellman PT, Brown PO, Botstein D: Cluster analysis and display of genome-wide expression patterns. Proc Natl Acad Sci USA. 1998, 95: 14863-14868. 10.1073/pnas.95.25.14863.PubMedPubMed CentralView ArticleGoogle Scholar
- SAS. [http://www.sas.com]
- Statistica. [http://www.statsoft.com]
- Insightful. [http://www.insightful.com]
- The R project for statistical computing. [http://www.r-project.org]