Ratios shifted upwards in CN segment plot

[lbeltrame]

I have a sample that is supposed to be mostly copy number neutral (no large events detected in a sWGS study). However, when I run the same sample (done with WES in this case, ~180X coverage) with TitanCNA using the output from GATK (sample vs its matched normal), I see graphs like this one for some solutions:

which at least in my case it is not correct. It looks like the "baseline" for the log2ratio has been shifted to around ~0.1-0.2. It doesn't help that the data is quite noisy.

Notice that I might be running an older version of TitanCNA (whatever is available in bioconda).

[gavinha]

Hi @lbeltrame

Is this a selected solution made by selectSolutions.R? If so, then some changes in issue #61 might help.
Is this a single run with ploidy initialized to 2? If so, then can I also see genome wide LOH plot of allelic fractions?

Thanks,
Gavin

[lbeltrame]

Yes, it's made via selectSolutions, from a run that tested ploidies 2,3,4and clusters from 1 to 5. The only difference is that bins come either from CNVkit or GATK CNV rather than readCounter.

Here is the LOH plot (note that it is from a different run, because I was tweaking the segmentation parameters, but this occurs in all cases):

Note that another solution, which is not selected, is closer to the actual sample (mostly CN neutral, some LOH).

[gavinha]

Thanks, @lbeltrame.

I think the commits associated with #61 might help.
Also, what value are you setting for --alphaK in titanCNA.R? For WES data, try using a value between 1000 and 3000.

Alternatively, you can adjust the --threshold in selectSolutions.R so that it weights the ploidy 2 solution more.

[lbeltrame]

I'm using 2500 for --alphaK as per the recommendations for WES. I found out that the original sample had a very noisy reference for CN, so I switched it and I'm trying again. I'll then try with the latest master to see if it improves.

[lbeltrame]

I've had the opportunity to test this again. selectSolution selects the only "wrong" solution (2 clusters, but reports only 1 as numClust in the output) which has this shift (tested with the latest code). Other solutions (1 cluster, 3-5 clusters) all correctly report the whole genome as without CNAs.

From other evidence, the solution with one cluster seems the most correct biologically, but it has a slightly lower statistic than the one with 2, which gets picked.