发信人: feir (菲儿), 信区: Biology
标 题: Re: Also microarray questions
发信站: Unknown Space - 未名空间 (Wed Nov 24 12:03:38 2004) WWW-POST
It's indeed hard to argue. Irizarry wrote a paper to compare the 3 algorithms
in 2003 on Nucleic Acids Res. We've done some simulation studies and found
that his conclusions are pretty objective. While RMA is better with the
lower-expressed genes, it underestimates fold changes in a great deal, which
often times is very annoying to explain to the clinicians. But it's
interesting that in Affymetrix's to-be-released new algorithm (not called
MAS6, but some new name), they inherit the quantile normalization idea of RMA
which shows that it's not that bad.
【 在 leohawk (leohawk) 的大作中提到: 】
: Based on my limited experience:
:
: MAS 5 is the best, Dchip is ok, RMA is the worst in doing probe level
: analysis.
:
: I am talking about affymetrix platform... The reason I think RMA is no so
good
: probably arise from its presumtion about the data...
:
: 别扔砖, 个人意见.
:
:
: 【 在 feir (菲儿) 的大作中提到: 】
: : First of all, if you could find a biostatistician in your department or
your
: : school you can let him/her do the analysis for you. All that you are
asking
: : is standard microarray data analysis.
: :
: : You should use the p-values instead of fold changes in defining
: differentially
: : expressed genes. What puppeteer mentioned (0.05/N) is called Bonferroni
: : adjustment for multiple comparison, which is the most conservative way
: : possible. Other common and less-conservative way is FDR (false-discovery
: : rate) adjustment. Based on my experience, if you have any gene that's
: : significant using Bonferroni you'll have dozens of significant genes using
: : FDR.
: :
: : Another important thing is what algorithm you used to pre-process your
data.
:
: : If you used Affymetrix's MAS 5 you probably want to get rid of all the
genes
: : whose average expression are too low (say 16 or 32) because MAS 5 is not
: great
: : at processing the low-expressed genes so they are likely to be an artifact
: : when they are significant. RMA is a little better on that issue but RMA
: tends
: : to underestimate the fold changes.
: :
: : One more thing you can try is to log-transform your data before you do the
T
: : test comparison, which is almost a standard in microarray analysis. You
may
: : get more signficant genes by doing that because your data would be less
: : skewed.
: :
: : After you get the gene list you can put the probe-set names into EASE
(very
: : simple, just copy-paste), choose the correct chip name in EASE and run it.
: : EASE is a freeware in recgonizing gene themes. Other things you could do
: : include making some pretty clustering pictures, building a prediction
model,
: : matching your list with other lists people have generated before, etc.
: :
: :
: : 【 在 dentsu (xixi) 的大作中提到: 】
: : : I had done microarray and gotten the data in excel sheet. My design is
: : simple:
--
※ 来源:.Unknown Space - 未名空间 mitbbs.com.[FROM: 140.163.]
|