Rebasing gene lengths when combining the FPKMs of several studies

[SamCurt]

We're planning to combine FPKMs of several GEO datasets for our project. However, these projects uses different gene annotations on GRCh37--some use hg37 and some use hg19--for their gene counting.

Do you think, in such a case, that it is necessary do "rebase" the FPKMs using gene lengths from the same annotation?

[Brian Bushnell]

I'd suggest regenerating the FPKMs from the raw data with a uniform pipeline, as the processing methodologies obviously differ, which can have an unexpectedly large effect on results.

[SamCurt]

I'd suggest regenerating the FPKMs from the raw data with a uniform pipeline, as the processing methodologies obviously differ, which can have an unexpectedly large effect on results.

The biggest problem is some of the data used has non-public SAM/BAMs, so I have to live with count-level data. This is not to mention some of these sets use Ensembl IDs, some use EntrezGene IDs and some use symbols...

[Brian Bushnell]

Personally, I am not very interested in unexplainable or unreproducible data. I think that reproducibility is essential to science; using black-box software or secret data is not explainable or reproducible, and thus not scientific, in my opinion. I suggest that you pursue something that is reproducible.

[WhatsOEver]

In addition to what Brian wrote: Currently, >50% of the public data I'd like to use for comparison to my data are not reproducible with the published analyses pipelines. This is either because there are different 1-man-1-time-0-comments-custom-scripts that simply do not work or because essential information (e.g. program parameters!) are not provided anywhere. However, I can not simply ignore this existing data as it is sometimes also the only fitting reference in terms of sample prep, sequencing, etc...

Hence, I currently take the most annoying and time consuming ways: Start at fastq, contact the authors, dig into the code. If the data is published in a peer-reviewed journal, you should be able to get access to the raw data, so you are able to reproduce the published results.

[SamCurt]

The complete code for one of those sets--the one I mentioned to currently only have "open" counts data" is available open source (I already saw the code). I just want to know how many days would I need to re-align those, since SRX toolkit is slow in my institution for some reason...