Hi,
I have access to RNASeq for two tissues (A/B) for (Dis) and Control (Ctl) groups. The data is pooled and the three pools contain data from 3, 5 and 4 patients per Dis/Ctl, respectively. Pools 1 and 2 contain data for all 4 groups (A_Dis,B_Dis,A_Ctl,B_Ctl) and pool 3 contains data for B_Dis and B_Ctl only. So, all in all I have 10 distinct datasets each consisting of 3 to 5 lanes of Illumina sequencing data.
I want to use this data to detect possible different levels of RNA editing (actually A-to-I editing, this means I'm looking for A to G mutations in my data) between my Dis and Ctl groups. I use a small script that returns the number of reads and how many of them are edited for each putative RNA editing site from darned.
For my initial analysis I used edgeR to detect differential editing. But since edgeR needs raw reads I had to somehow "normalise" my data.
For example one row (within a single pool) I want to analyse looks like this (total reads/edited reads):
A_Dis: 100/30 B_Dis: 120/40 A_Ctl: 90/40 B_Ctl: 80/50
So of course I can't just compare the edited reads (30, 40, 40 and 50) because the total numbers of reads are different in each group. Therefore I normalise them according to the lowest number of reads in each row:
(30/100)*80=24, (40/120)*80=32, (40/90)*80=35.6; (50/80)*80=50
Then I use these "normalised" reads (24, 32, 35.6 and 50) as input for edgeR and performed the analysis like in manual. So far so good. But now I want to use the data from all pools.
I basically tried two setups:
These two setups had (except from the top element) very different results. I also tried to simply scale up my normalised values by multiplying them by e.g. 10 and the results were totally different again.
So I think that my results are in fact more the results from my "normalisation" procedure than from the experiment.
Do you have any ideas to get rid of the normalisation? Do you think that I should treat differential RNA editing like DE at all or is there an other model that might fit better?
Regards,
Thomas
I have access to RNASeq for two tissues (A/B) for (Dis) and Control (Ctl) groups. The data is pooled and the three pools contain data from 3, 5 and 4 patients per Dis/Ctl, respectively. Pools 1 and 2 contain data for all 4 groups (A_Dis,B_Dis,A_Ctl,B_Ctl) and pool 3 contains data for B_Dis and B_Ctl only. So, all in all I have 10 distinct datasets each consisting of 3 to 5 lanes of Illumina sequencing data.
I want to use this data to detect possible different levels of RNA editing (actually A-to-I editing, this means I'm looking for A to G mutations in my data) between my Dis and Ctl groups. I use a small script that returns the number of reads and how many of them are edited for each putative RNA editing site from darned.
For my initial analysis I used edgeR to detect differential editing. But since edgeR needs raw reads I had to somehow "normalise" my data.
For example one row (within a single pool) I want to analyse looks like this (total reads/edited reads):
A_Dis: 100/30 B_Dis: 120/40 A_Ctl: 90/40 B_Ctl: 80/50
So of course I can't just compare the edited reads (30, 40, 40 and 50) because the total numbers of reads are different in each group. Therefore I normalise them according to the lowest number of reads in each row:
(30/100)*80=24, (40/120)*80=32, (40/90)*80=35.6; (50/80)*80=50
Then I use these "normalised" reads (24, 32, 35.6 and 50) as input for edgeR and performed the analysis like in manual. So far so good. But now I want to use the data from all pools.
I basically tried two setups:
- Take the sum of the lanes per group and pool "normalise" over all 10 values and group them in edgeR like replicates. This gives rather poor values, because the number of reads within each of these 10 groups is rather low and so the normalised values are low too and so they have a rather low significance within edgeR.
- Sum up all the reads of all pools within each group. This results in four groups (like in the example above) with rather high numbers of reads, so the output values from edgeR look significantly better. But of course I loose the replicates and therefore the ability to detect outliers.
These two setups had (except from the top element) very different results. I also tried to simply scale up my normalised values by multiplying them by e.g. 10 and the results were totally different again.
So I think that my results are in fact more the results from my "normalisation" procedure than from the experiment.
Do you have any ideas to get rid of the normalisation? Do you think that I should treat differential RNA editing like DE at all or is there an other model that might fit better?
Regards,
Thomas
Comment