cellular prevalence estimation

[efa...@gmail.com]

Hi Andrew,

I can't find any explanation for what I'm seeing in my result table.

I explain about my input, first. The two samples are from the same patient and 100% pure.

Because of low coverage of my data, I used pyclone-binomial with total_copy_number as prior.

And since, these two mutations are in copy neutral region I gave 0 as minor_cn and 2 as major_cn.

As it can be seen in the following, the estimation of cellular prevalence for the first mutation is as expected.

VAF for the present sample is more than VAF for relapse sample. So cellular prevalence for present sample is more than relapse sample.

However, for second mutation, result is strange. VAF for the present sample is less than relapse sample. But the estimated cellular prevalence for present sample is more than relapse sample. How is it possible?

mutation_id    sample_id    cluster_id    cellular_prevalence    cellular_prevalence_std    variant_allele_frequency

6:139226217    relapse    1    0.604121507553    0.21166604889    0.4
6:139226217   present  1    0.64760552495    0.210427272162    0.536312849162

mutation_id    sample_id    cluster_id    cellular_prevalence    cellular_prevalence_std    variant_allele_frequency
5:154300933    relapse    3    0.621620959735    0.14462790712    0.652173913043
5:154300933   present  3    0.677807436148    0.222299764311    0.539682539683

[Andrew]

Without seeing that data I can't say specifically what is happening here. However, it is perfectly reasonable for the cellular prevalence of mutation 1 to be lower than mutation 2, even if the VAF is higher for mutation 1 than mutation 2. This would occur if mutation 1 had the genotype BB, in which case the cellular prevalence ~ VAF. If mutation 2 has the genotype AB then the celular prevalence ~ 2 x VAF. Thus mutation 2 could have a lower VAF but a higher cellular prevalence.

There is a common misconception that PyClone clusters the VAFs. It does not. I uses the count data and corrects for the mutational genotypes using the prior copy number information. Thus looking at VAFs as surrogates of the cellular prevalence is misleading.

With all that said, your data sounds a bit ill posed. Specifically why is the tumour content 1? Do you have copy number information? If by copy neutral you mean no copy number change, the appropriate copy number is major_cn 1, minor_cn 1. This corresponds to the presence of both parental alleles. If you use this with the 'major_copy_number' (default) prior, PyClone will perform much better as it can assume the mutational genotype is AB. The case of major_cn 2 and minor_cn 0 is copy neutral LOH, that is loss of one parent chromosome and the doubling of another. This is a much harder state for PyClone as it has to consider whether the mutation has the AB, or BB genotype.

Cheers,

Any

[efa...@gmail.com]

Hi Andrew,

Thanks for reply. In our study we use FACS to seperate tumour cells from normal cells and then estimated the tumor content which showed it is 1 for this sample. And I did whole exam sequencing on the samples. By using a few programs like sequenza and excavator I obtained copy number information which their result were very different in some cases. Sequenza report parental copy number information, but excavator doesn't. I used sequenza result for minor and major copy number. And Pyclone  estimated 3 clusters for this sample. However, because sequenza didn't work well on some of my samples (some samples have deletions or duplication based on karyotype but can't be detected by sequenza) I'm now using excavator. 

Saying all that, what is your suggestion regarding copy number information  for Pyclone input as I couldn't get precise parental copy number information? I'd like to compare the result I'm obtaining from Pyclone to my result from expands and reach a conclusion about tumour evolution.

Thank you

[Andrew]

Without parental copy number there will be more uncertainty in the results. This will manifest as multi-modal posterior distributions and/or high variances in the posteriors. If the copy number data is low quality, that is all that can realistically be done though.

Cheers,

Andy

[Stefano Cheloni]

Hi Andrew, 

I'm posting here as I think I have a similar question.

In particular, for which reason when vaf of a mutation is 0, its inferred cellular prevalence may become different from 0?

I have use pyclone_binomial with total_copy_number as prior.

Here I reported two mutations as example:

mutation_id sample_id cluster_id cellular_prevalence cellular_prevalence_std variant_allele_frequency

chr17:7578406 TUM 2 0.00584390778941 0.00681520333213 0.0

chr17:7578406 RELAPSE 2 0.457623675792 0.152280877152 0.352112676056

chr16:52473642 TUM 2 0.00632783528219 0.00764891100874 0.0

chr16:52473642 RELAPSE 2 0.431547165701 0.1481020977 0.255319148936

Thank you

Stefano

[Andrew Roth]

Hi Stefano,

There are two answers:

1) Just because the VAF is zero does not imply the prevalence is zero. Unless you have sequenced every cell in a tumour, there is always a chance you miss rare clones. If your depth is 100s-1000s of reads this is actually quite likely.

2) There is also a technical issue with the PyClone model in that it will never set a value to 0, it will always be some small value. This is just a manifestation of the priors.

In practice just set very low cellular prevalence clusters to zero if you need to. In your example zero is within one SD of the mean, so this would be reasonable.

Cheers,

Andy

--
You received this message because you are subscribed to the Google Groups "Pyclone User Group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pyclone-user-gr...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/pyclone-user-group/a50abc95-cb25-464d-b7be-662155f1c6fa%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.