minimum depth of coverage for transfrag extension

[apadr007]

Hello,

I am wondering why cufflinks is creating transcript model extensions based on low depth of coverage (see the red highlighted box in the attached image). So the reference annotation is in orange and the cufflinks transcript model is in blue.

As you can clearly tell, there is good coverage up to a certain point and suddenly the coverage drops down to 4 reads and cufflinks continues to extend its model to include those 4 reads. Does anyone know how to prevent cufflinks from generating these false extension? And also, if you notice, there is a gap between the area with good coverage and the 4 reads, yet cufflinks thinks this low coverage regions belongs to the previous region with good coverage. Does anyone know why this is occurring and are there parameters that can control for this? Here are the parameters I have used in TopHat and Cufflinks:

 #tophat
tophat -o tophat_files7 -p 10 --initial-read-mismatches 3 -i 40 -I 50000 -F 0.30 --no-coverage-search -a 10 -g 20 chromosomes_all_4.fastq
#cufflinks
cufflinks -o cufflinks7 -b CHROMOSOMES_names_ed.fa -I 50000 --min-intron-length 40 --trim-3-avgcov-thresh 3 --trim-3-dropoff-frac 0.4 -p 10 --no-update-check -j 0.05 accepted_hits.bam

Thanks,

[apadr007]

...Anyone?

[cswarth]

Have you checked out the -j/--pre-mrna-fraction <0.0-1.0> option to cufflinks?
I haven't used it myself but it looks like that might be useful.

[Ron12345]

I am having this identical problem and tried playing with the 3' end trimming '--trim-3-dropoff-frac <int>' and '--trim-3-avgcov-thresh <int>', but have not solved the problem of extension into low coverage regions. Has anyone found a solution yet? Thanks!

[apadr007]

Hi Ron, I also attempted to optimize certain parameters, however, if you go back to the cufflinks article you'll notice authors mention that they base their categorization of transcripts simply on splice junctions, since there is variability of sequencing at both the 3' and 5'. Also, in the same paper they mention that their algorithm attempts to span 'transcript island' without necessarily having any read overlap; this allows Cufflinks to generate much larger models. You can experimentally test this -- design primers for the smaller section of your transcript without very many reads and run a real time.

[Ron12345]

Great - thanks for the info, 007. To solve the problem addressed in this thread, I'll write a script to reassign the boundary of Cufflinks-detected transcripts and can be contacted for help with this.

[CHLin]

This might help...

I think this option may help to change the parameter:
--overlap-radius [#nt]
If the gap between two units is larger than the #, they will be merged.