>Hmm, your answer left me thinking about how to measure distances. Why >doesnt a distace function just calculate the distance between the values >that are there and leave out the NA:s? I have filtered away with the >B-test the spots that are supposedly not to be differentially expressed >and have only a subset of the total number of spots. Three slides of my 18 >have many NA:s. Should I exclude them therefor because the distance is to >affected? > >/ Marcus > > >At 08:21 2003-10-30 +0100, you wrote: > >>>Subject: RE: [BioC] Heatmap function >>>Date: Wed, 29 Oct 2003 13:55:43 -0500 >>>X-MS-Has-Attach: >>>X-MS-TNEF-Correlator: >>>Thread-Topic: [BioC] Heatmap function >>>Thread-Index: AcOeQFWAqheKzfUtT1WkYP+4GKRURgACytsA >>>From: "Furge, Kyle" <kyle.furge@vai.org> >>>To: "Marcus" <marcusb@biotech.kth.se> >>>Cc: <bioconductor@stat.math.ethz.ch> >>>X-MIME-Autoconverted: from quoted-printable to 8bit by >>>kiev.biotech.kth.se id h9TJK8Pr030736 >>> >>>This question has come up often at our institute... so here goes for an >>>brief, unformal, explanation. >>> >>>Clustering depends on some type of distance metric to determine how the >>>samples are related >>> >>>If you have a set of vectors : >>>x <- c(1,NA,2,NA) >>>y <- c(NA,2,NA,1) >>> >>>The pattern of missing values make computing any type of distance >>>between the vectors incomprehensible. The dist functions return NA for >>>these types of comparisons. >>> >>>Clustering functions don't like NA because their job is just to organize >>>the data. A distance of NA is not understandable. >>> >>>What some programs do (like Eisen's cluster) is "threshold" the NA >>>values to some arbitrary "large" distance. >>> >>>The following dist function computes distances and then replaces any NA >>>values with an arbitrarily large distance (10% greater then the largest >>>actually distance). This function may be helpful for input into hclust >>>because NA values are replaced >>> >>>na.dist <- function(x,...) { >>> t.dist <- dist(x,...) >>> t.dist <- as.matrix(t.dist) >>> t.limit <- 1.1*max(t.dist,na.rm=T) >>> t.dist[is.na(t.dist)] <- t.limit >>> t.dist <- as.dist(t.dist) >>> return(t.dist) >>>} >>> >>>I typed this from memory, so it may contain typo's, but you see the idea. >>> >>>I hope this helps, >>> >>>-kyle ********************************************************************** ********************* Marcus Gry Bj?rklund Royal Institute of Technology AlbaNova University Center Stockholm Center for Physics, Astronomy and Biotechnology Department of Molecular Biotechnology 106 91 Stockholm, Sweden Phone (office): +46 8 553 783 39 Fax: + 46 8 553 784 81 Visiting adress: Roslagstullsbacken 21, Floor 3 Delivery adress: Roslagsv?gen 30B