the correlation of RNA-seq data.

[kentnf]

I generated an correlation value table for all biological replicates.
Most of them have good correlation 0.9;
2-3 set around 0.8;
6-7 sets are < 0.8.

How about the data? How to improve it? Can I use all of them for differential expression analysis?
Thank you!

[mrawlins]

The nature of taking samples and the techniques makes it possible for some biological (or even technical) replicates to seem dissimilar for any given metric (p-values, correlations, etc.)

I would use any or all of them for DE analysis. If you account for the variance between replicates in your DE analysis you will reduce the number of false positives. That gets into the question of how you intend to determine differential expression. Some methods will allow you to use that replicate variance, while others won't. Either way, it shouldn't hurt you to include them all.

[frankyue50]

For me, most replicates have correlation > 0.95. Less than 0.8 is too low. You can try qq normalization to see if it improves.

I generated an correlation value table for all biological replicates.
Most of them have good correlation 0.9;
2-3 set around 0.8;
6-7 sets are < 0.8.

How about the data? How to improve it? Can I use all of them for differential expression analysis?
Thank you!

[vyellapa]

Could any of you post the method/code for how you're finding the correlation values.
I did a simple cor(FPKM_1,FPKM_2) on the FPKM values from cufflinks on biological replicates and got a value of .2.

It would be really useful to know how you guys are doing it.

[kopi-o]

A simple cor() (in R, I assume) can get very skewed by large outlier values such as those that can appear in Cufflinks output. You can try a Spearman correlation cor(FPKM_1, FPKM_2, method="spearman"), or take logs of the FPKM values, or do an arcsine transformation (http://bridgecrest.blogspot.fi/2011/...technical.html)

[vyellapa]

A simple cor() (in R, I assume) can get very skewed by large outlier values such as those that can appear in Cufflinks output. You can try a Spearman correlation cor(FPKM_1, FPKM_2, method="spearman"), or take logs of the FPKM values, or do an arcsine transformation (http://bridgecrest.blogspot.fi/2011/...technical.html)

I tried the spearman method which increased my correaltion value to ~.4. I did remove the outliers, ie. all values with exponential values such as 8.1667e-05, 6.162e+05, etc before I ran the correlation test. I also tried the cor of logs(after adding 1's to elimintae 0's) ie. cor(log10(FPKM_1), log10(FPKM_2)) but it did not show any better correlation value. The arcsine method is something I have not tried yet but seems interesting.

A scatter plot and correlation values using count data seemed good though.

http://seqanswers.com/forums/showthread.php?p=79112&posted=1#post79112

[kopi-o]

If you are only interested in the gene-level FPKMs, you could always divide the counts by gene lengths and the number of mapped reads (divided by a million) which would then lead to well-correlated FPKMs (since you are only scaling the values linearly). If you are interested in isoform-level FPKMs you obviously need something like Cufflinks, but you could try alternatives like RSEM, MISO (etc) to check how they perform.