How to select and output the informative genes in the resulting model based on SVM?

Question

[a511512345]

Hello everyone, I am learning SVM, according to http://compdiag.molgen.mpg.de/ngfn/docs/2005/sep/exercises-classif.pdf Since I am completely a novice, when I follow the book's instructions to learn to encounter the following questions: 1. How to select and output the informative genes in the resulting model? 2.How many genes do you need to still get a reasonable CV error? 3. How to use the model to do ROC curve analysis? Very much looking forward to your reply and script thank you very much!

[Kevin Blighe]

Please validate (by up-voting and/or accepting) previous answers to your questions, or at least acknowledge them by replying.

Here are un-responded answers to your questions:

• Which are the promoters?
• how to predict lncRNA based on gene symbol?
• how to use "marray"and "lumi"packages
• what kind of data can be directly WGCNA?

[a511512345]

Dear Kevin Blighe, Thanks for your help and reminders. In fact, every time before I want to reply your help, but I do not know what is the reason, not every successful reply to you. Maybe for me in China, and not well linked to biostars. I feel so guilty I can not let you know in time that you have helped me to learn a lot, I hope you will continue to help me.

[Kevin Blighe]

No problem. Have you read the entire manual?

They perform the following:

1. select the top 1000 genes based on high variance
2. apply a support vector machine with linear kernel (for more, read here )
3. cross-validate the SVM 10 times (for info on cross validation see A: Multinomial elastic net implementation on microarray dataset , error at lognet a )
4. choose the best predictors based on low error after cross validation
5. further refine the predictors using a simple t-test

[a511512345]

Sincerely thank you You always help me in detail About SVM, I have read the manual thoroughly. It is frustrating that I still do not fully understand, do not know which of the important genes to choose. Probably due to my extremely poor programming ability as a doctor. Thank you again!

[Kevin Blighe]

The tutorial uses multiple indicators in order to choose important genes, but the most relevant to the tutorial is a low error margin after cross-validation of the SVM linear kernel.

The cross-validation repeats the SVM multiple times (10 times), and then compares the results each time. This allows for the production of error margins. If a gene has very similar values in each cross validation, it's error will be low; if a gene has very different values in each cross-validation, it's error will be high.

谢谢

[a511512345]

ok, I bite again to study hard! Hope to learn! thank you In addition, your Chinese is very good, ha ha ha ha Thank you