Tweet
Hi everyone,
I'm currently a masters student studying bioinformatics and we have been tasked to reproduce an article which was written in 2014 and looked at the concordance between RNA-seq and Microarrays for discovery of differentially expressed genes.
We are running a pipeline which is tophat -> cufflinks -> cuffmerge -> cufflinks -> limma analysis using the paired-end reads which the authors of the paper had uploaded, we are also using all versions of programs which were available around the time of the article being written.
We are aware of the issues about using Limma for FPKM, and also for using cufflinks twice and not cuffdiff or Kalista or better programs, but we can't chage this so lets not focus on it.
For pipeline 5 we are having a problem with the ERCC values. We are expecting that if the ERCC mix was added correctly to each sample then we would be able to normalise across samples based on these, but this isn't something we've been able to do.
I have now gone back to the tophat results and gathered the raw read counts for all ercc's and the total read counts for the sample and calculated the FPKM for each ERCC across the five samples (across four different groups so 20 samples in total). I was hoping/expecting that the FPKMS for ERCC's across the 5 samples in one group would be comparable, as in if one ERCC showed a variance of around 10% most of the other ERCC's would show a variance of 10% as well, or one group could be double the others so the results should be halfed (we can't assume the authors just put in double the mix by mistake and have to take their work as perfect in everyway) but the FPKM's I have calculated show a variance of ~350% between the ERCC FPKM's.
Can you let me know if my thinking is right, where I'm believing this pipeline has failed and we should re-do it with different options/arguments?
If it is expected for the variance to be so high could you let me know how you would normalise these?
I have attached two documents which show the variability, there are hidden columns which show all the raw values and I'd appreciate it if you could confirm my calculations are correct! I compared to one sample arbitrarily btw, it happened to be the sample in the middle of my file as well as the sample with the median FPKM for the first ERCC.
Alot of the ERCC's seem to have very few read matches and I was wondering if it would be best to use a subset of the ERCC's which show the greatest number of matches.
I'm currently a masters student studying bioinformatics and we have been tasked to reproduce an article which was written in 2014 and looked at the concordance between RNA-seq and Microarrays for discovery of differentially expressed genes.
We are running a pipeline which is tophat -> cufflinks -> cuffmerge -> cufflinks -> limma analysis using the paired-end reads which the authors of the paper had uploaded, we are also using all versions of programs which were available around the time of the article being written.
We are aware of the issues about using Limma for FPKM, and also for using cufflinks twice and not cuffdiff or Kalista or better programs, but we can't chage this so lets not focus on it.
For pipeline 5 we are having a problem with the ERCC values. We are expecting that if the ERCC mix was added correctly to each sample then we would be able to normalise across samples based on these, but this isn't something we've been able to do.
I have now gone back to the tophat results and gathered the raw read counts for all ercc's and the total read counts for the sample and calculated the FPKM for each ERCC across the five samples (across four different groups so 20 samples in total). I was hoping/expecting that the FPKMS for ERCC's across the 5 samples in one group would be comparable, as in if one ERCC showed a variance of around 10% most of the other ERCC's would show a variance of 10% as well, or one group could be double the others so the results should be halfed (we can't assume the authors just put in double the mix by mistake and have to take their work as perfect in everyway) but the FPKM's I have calculated show a variance of ~350% between the ERCC FPKM's.
Can you let me know if my thinking is right, where I'm believing this pipeline has failed and we should re-do it with different options/arguments?
If it is expected for the variance to be so high could you let me know how you would normalise these?
I have attached two documents which show the variability, there are hidden columns which show all the raw values and I'd appreciate it if you could confirm my calculations are correct! I compared to one sample arbitrarily btw, it happened to be the sample in the middle of my file as well as the sample with the median FPKM for the first ERCC.
Alot of the ERCC's seem to have very few read matches and I was wondering if it would be best to use a subset of the ERCC's which show the greatest number of matches.