suhrig
commented
3 years ago Hi Mazlina,
The second warning (subsampling of fusions) is harmless. It just means that there are more supporting reads for a fusion than necessary. Collecting more reads would not increase the confidence any further. Summarizing the reads for each fusion is very resource-consuming (especially in terms of memory), so subsampling is performed. You can increase the subsampling threshold using the parameter -U if you want. But there is no point in doing so other than the supporting read count becoming more accurate for highly expressed fusions. There is a risk that your machine runs out of memory under some conditions, though, in particular when you process multiple myeloma samples.
The first warning has nothing to do with the size of the sample. I have processed samples with 1 billion reads successfully. It is a bit of a concern. At the same time, it is normal. It depends on the severity whether you should be worried about this. 14 million reads looks serious. You have to put this number in relation to the number of chimeric reads that Arriba reports. Look at the line Reading chimeric alignments from /path/to/file.bam (total=???). The number of reads reported as malformed is the number of reads that could be used for fusion detection, but that are lost because they have an invalid SAM format. For example, if the number of malformed reads is 2 million and the number of successfully loaded chimeric reads (total) is 1 million, it means you are losing 2/3 of the information and thus 2/3 of the detection power. In practice, many of the malformed reads are actually useless (they are mostly clipped adapter sequences), so the problem is not actually that severe; you are not losing 2/3 of the useful reads. However, if a large fraction of the reads is affected by this problem, then fusion detection can suffer. The vast majority of the malformed alignments is caused by a bug in STAR, which affects paired-end alignments with small insert size such that the reads fully overlap. I suspect that your sample also has a small insert size.
I have opened a ticket with the developer a while ago. Hopefully, it will be fixed soon. You can have a look at https://github.com/alexdobin/STAR/issues/1135 to learn more about the nature of the bug & bug fixing progress. And issue #97 - notably this comment - tells you possible workarounds, namely: trimming adapters, and possibly aligning fully overlapping reads in single-end mode and non-overlapping reads in paired-end mode. I am not sure if you want to invest the effort of fully implementing this workaround, or if you simply want to wait for the bug fix to be released. Possibly, the fusion detection is not even impaired that much, despite quite a lot of reads going lost (because most of them are useless adapter sequences). I run Arriba regularly on samples with small insert size, and it still picks up the important fusions. If you notice that Arriba misses a fusion in your samples, you could try the workaround and see if the fusion is then detected.
Regards, Sebastian
MazlinaIsmail
Hello,
I have been getting a couple of warnings when I run the 'STAR in conjunction with Arriba' method. I would like to understand a bit better how much they affect the fusions output file.
Below is an example:
WARNING: 14146239 SAM records were malformed and ignored WARNING: some fusions were subsampled, because they have more than 300 supporting reads
Not sure if helpful, but here are some additional info about the input file. There is ~125M reads and when I run STAR separately (for other rnaseq analysis), I get 85% uniquely mapped reads.
I guess the second warning is nothing to be too concerned about (?). With the first, I guess it is the same underlying issue in that my input file is quite big, and so the second warning is simply a case of it being too may alignment records? My log files are indicating that Arriba managed to process them (it gets to Done in the log file, and I get the fusions and discarded fusions output files) but I was just wondering if the warnings may indicate some other issues.
Thank you.
Best, Mazlina