thanks for the reply, shuagay. actually I am not building a classifier. what I am doing is try to calculate the likelyhood ratio under different conditions and use this ration to construct a network. in this process, the basic unit we are dealing with is gene-gene pair. the condition can be,for example, gene a and gene b co-expressed( with a coexpression value) in expression datasets. The probability we want to calculate is Pr(this pair of gene has a co-expression value r given that this pair of genes co-existed in a disease database record)/Pr(co-exp value r given that this pair of genes didn't co-existed in that record). In order to get this ratio, we have to use a set of training data(disease database record, expression data, annotation data etc..) and a model(possibly a linear model) to estimate it. it would be very appreciated for any suggestion.
Hi all I am using bayesian model to construct a diease related network(gene gene interacts in disease related processes). do you have any idea on how to define a set of "negative disease genes" that can be used as the negative training data?
•
written
by
ewre
1 answer
If gene expression is to be used, then what first comes in mind for "negative disease genes" are the ones that are not differentially expressed in diseased condition vs control (e.g. housekeeping genes). Your question, however, is a little confusing. I believe that in expression-based classifier, you select some set of genes that reflect your condition and then use expression data in diseased samples and controls to train the network. So the negative training set is the expression in control samples, not a set of genes. What exactly are the features and instances used in your classifier?
Ok now I see. I suggest you use some gene-disease annotation based on Disease Ontology (DO, http://disease-ontology.org/), e.g. from here http://doa.nubic.northwestern.edu/pages/search.php. Then you can make permutations and select gene pairs that share common DO as positive set and don't share any common DO terms as negative set.
I have tried DO, it's cool. the only problem is that is the information it provided reliable. the postive training data i used by now comes from two part, one is DisGenNet(http://ibi.imim.es/web/DisGeNET/v01;jsessionid=1w86jcj81w6sb1im5qu54f8ct), the other one is malacard(http://www.malacards.org/). the negative one is randomly sampled from the whole genome genes pairs reminded. I think it is better to do a filtration with DO on the randomly selected negative gene pairs provided that DO is reliable. besides, I have one more question: are there any papers which have and estimation on the ratio of "number of guilty gene pairs"/"number of non-guilty gene pairs" in human genome?
Log in to answer this question.
-
Combine k-fold cv and lasso regression
written by flynndwight •
-
Expression of genes in KEGG pathway
written by nhaus •
-
Machine learning using microarray data
written by Gene_MMP8
-
Drug targets for hub genes
written by Gene_MMP8
-
How to find the hub genes in the gene co-expression network constructed by WGCNA
written by wangdp123
-
Visualize regulatory network using cytoscape
written by kim •
-
Classifier Optimization In Weka!
written by kandoigaurav
-
Gene-Disease Associations From Mass Spec Data
written by syntonic.c •
-
How To Perform A Pathway Or Network Level Analysis Of Rare Variants?
written by ewre
-
How Do You Explain The Correlation Over A Gene Set And Pathways And Other Ideas To Deal With The Ge…
written by J.F.Jiang
Selection of negative set depends on the kind of information you are using to build your classifier. Are you using mutations or gene expression or something else?
gene expression data will be used.