Open Notebook: SVM

Showing posts with label SVM. Show all posts

Wednesday, February 18, 2015

weka, lifespan prediction

https://weka.waikato.ac.nz/dataminingwithweka/preview

Friday, January 30, 2015

todo, SVM project. find out the support vectors. verify the predictions in ken rls database.

SVM project. find out the support vectors. verify the predictions in ken rls database.

This project could take a long time

Sunday, January 11, 2015

done, pool single gene mutant data from rls.csv

I studied ken's old codes, my old codes, and wrote "_export_rls20150111.R".

On 20150111, I found out that I did the work in "_explore.20140705.R".

Friday, January 2, 2015

compare He14 with White2010 predictions

In He14Plos Genetics, there are 137 LL genes based on SGD annotations.

C white 2010 used Kaeberlein 564 data set for prediction. In this dataset, I have 113 known long-lived mutants. Comparison with He14Plos, I found 56 of the White10 LL genes are non-LL and only 32 have LL in the He14 data sets.

To do: I need use the ken rls.db to find out the RLS measures.

He 14, ibuprofen on RLS

He C, Tsuchiyama SK, Nguyen QT, Plyusnina EN, Terrill SR, et al. (2014) Enhanced Longevity by Ibuprofen, Conserved in Multiple Species, Occurs in Yeast through Inhibition of Tryptophan Import. PLoS Genet 10(12): e1004860. doi:10.1371/journal.pgen.1004860

He14 partitioned the yeast genes into LL long-lived and NLL (non LL). This partition can be used to verify my svm project. There were 137 LL genes and 3840 non-LL genes.

0.2mM ibuprofen was used for yeast cells (a bit too high?)

Used Giaever's drug screen to identify the tryptophas pathway in ibuprofen effect.

Did miscropscopic screen on bud sizes and growth. This is an useful data resource.

tb = read.xlsx("journal.pgen.1004860.s010.xlsx", 1)
tb = tb[,1:6]; tb[,1] = as.character(tb[,1])
names(tb)= c("ORF","RLS","BirthSize","MeanSize","G1DNAcontent", "Fitness")

#Birthsize is from

Hoose SA, Rawlings JA, Kelly MM, Leitch MC, Ababneh QO, et al. (2012) A

systematic analysis of cell cycle regulators in yeast reveals that most factors act

independently of cell size to control initiation of division. PLoS Genet 8:

e1002590.

Truong SK, McCormick RF, Polymenis M (2013) Genetic Determinants of Cell

Size at Birth and Their Impact on Cell Cycle Progression in Saccharomyces

cerevisiae. G3 (Bethesda) 3: 1525–1530.

# Fitness is from Giaever2002Nature,
# http://genomics.lbl.gov/YeastFitnessData/websitefiles/cel_index.html

http://www.plosgenetics.org/article/info%3Adoi%2F10.1371%2Fjournal.pgen.1004860
http://news.sciencemag.org/biology/2014/12/ibuprofen-boosts-some-organisms-life-spans

Wednesday, December 10, 2014

Shedden et al, 2008, Nature medicine, gene expression based survival prediction in lung adenocarcinoma

Shedden et al, 2008, Nature medicine, gene expression based survival prediction in lung adenocarcinoma
http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2667337/

health disparity, expression dataset, supposed to have ethnic group information.

several classifier algorithms were used. cross-validation were applied, ROC curves used.

good reference for my lifespan prediction study.

Method A (Gene clusters and ridge regression)

binary tree-structured vecgor quantization -> binary cluster tree using expression profiles

Method B (Stratified Cox model on univariately selected genes)

Method C (clustering of samples combined with minimum gene selection).

Method D. (clustering of samples combined with minimum gene selection).

Sunday, November 2, 2014

SVM, reading notes

See http://hongqinlab.blogspot.com/2014/11/elements-of-statistical-learning-video.html

SVM kernel trick

trial and error to separate data in high dimenstional space

cross validation

predict True Negative?

Mathews correlation coefficient (MCC) (for binary classification)

In general the equation for a hyperplane has the form

SVM maximize soft margin.

Data should be standardized for SVM analysis, because SVM treats every columns the same.

On researchgate, someone argues: Perform different normalization such as Z-Score or Min-Max before using PCA. Z-Score normalization before using PCA might be beneficial.

For principal component (PCA) and svm,
http://www.softcomputing.net/isda2010_2.pdf
On researchGate: Principal components are linear combinations of original variables x1, x2, etc. So when you do SVM on PCA decomposition you work with these combinations instead of original variables.

Support vector classifer in the enlarged spaced solves separation problem in the lower-dimensional space.

Question: Kernel is used to computer inner products of vectors. Why are there different types of kernels for computing the same thing (inner products)?

SVM for more than 2 classes: