Normal filtering in TIDDIT #1127
Replies: 5 comments 2 replies
-
|
Very nice write-up! A question about the 4 different kind of variants. Is the normal always from buffy coats? In that case, would one expect to see less of this kind where the tumor comes from a blood samples vs say a solid tumor? What are the types of tumors in these 3 cases? |
Beta Was this translation helpful? Give feedback.
-
|
I don't know if I'm ignorant or buffy coats = blood sample! I'll give myself some credit and assume it means blood sample. It is generally known that haemotlogical cancers have a greater tumor contamination than solid tumors at least, and in that case I would expect that variants in group 1 and 4 would be more pushed towards the center. I tried to read up on what the tumor in normal contamination fraction usually is earlier today, and from what I've seen so far it seems quite variable. But that at least around 38% of saliva samples from AML and in MPN 91% of sorted CD3-T cells had Tumor In Normal (TIN) values above 1%. For sarcoma samples only 2% of samples were contaminated with TIN above 2%. https://www.biorxiv.org/content/10.1101/2022.03.09.483623v1.full.pdf But I found it difficult to find what the actual fraction commonly is. I don't actually know exactly but at least:
Perhaps the people taking these samples are aware of the contamination problem and is avoiding taking the normal from blood in the case of haematological cancers |
Beta Was this translation helpful? Give feedback.
-
|
Buffy coat is derived from blood and contains the white blood cells and platelets. Regarding contamination of Normal sample with Tumour... it is possible that this happens, and it is good that we consider this. Skin: The reason for using skin in hematological malignancies is to reduce the quantity of tumor in normal. A skin biopsy per se can still contain some blood, hence a small risk still remains. If the lab has cultured fibroblasts from skin biopsies, then there are no more tumour cells. This would be the safest normal to use, but requires both an invasive skin punch + 4-6 days of culturing. Saliva: Does contain tumor DNA and at variable frequency. In many cases low one-digit percentage numbers. >10% is rare afaik. We can bring this up with clinical experts for further discussion. Blood/Buffy coat for solid tumours: Not much contamination expected, but can never be entirely ruled out. |
Beta Was this translation helpful? Give feedback.
-
|
Here's an update on this filtering. Issues regarding the filtering:
Filtering based on AF_T and AF_N, accounting for 10% TINThe earlier mentioned filters was fine for the 3 cases investigated, and reported earlier. But these cases were not good examples of tumor contamination. Purely based on logic it was possible to see that the allowing for a 10% of the AF_T in the normal was a bit too stringent if the goal was indeed to allow for 10% TIN, as while the average AF_N of these variants would be AF_T * 0.1, there should of course be a spread based on randomness. I did some simulations of the presence of tumor variants in the normal data to have a better idea of what we might expect, and at what thresholds to put the filtering. Given preset list of AF_T, sequencing depth in the normal, and a TIN of 10% a scatterplot of the AF_N and AF_T can be seen below:
For 40% TIN the presence of the AF in the normal is predictably much greater.
With a filter of AF_N / AF_T > 0.25 = IN_NORMAL, and thereby allowing for a maximum of 25% TIN, and accounting for the spread. The result on this simulated data can be seen below for a 10% TIN:
Which comes down to: 1.19% filtered out somatic variants However, the average coverage of these variants does not come to 30x normal, but 89.42x. And if we limit the coverage to exactly 30x in the normal the results are a bit worse, but still ok in my mind:
With around 2.25% falsely filtered out variants at a TIN of 10%. What does applying this filter look like in the 3 cases (A, B, C)?
And the diagonal line of the filter is somewhat outlined by the separation of the "PASS" and "IN_NORMAL" variants, and you can imagine it going from 0.0 diagonally to 0.25 of AF_N. This looks fine to me as it avoids the mess of variants in the center which should mostly be germline and artifacts, while still allowing for an average of 10% TIN. The average coverage in the normal of the simulated variants filtered out as "in_normal" was a bit lower than those with "PASS", 55.17x in comparison to 89.83x. This trend is expected given that the spread of AF_N should be higher the lower you sequence, and would increase the chance that a TIN of 10% reached a higher level than 25%, which is the maximum allowed here. Below is a scatterplot of the "in_normal" variants, with coverage in the normal on the X-axis.
Perhaps some additional filter could be implemented to rescue these lower coverage variants as well. It should be noted again however that even without this additional rescue filter, the number of falsely filtered out variants at 10% TIN is 2.25%, and a TIN of 10% should be pretty rare. ContigsI brought this up earlier as well... The existence of these variants with assembled contigs make the filtration a little bit more complex as the AF values don't always make sense. For instance, there are in these 3 cases:
For instance, if 3 reads in the normal is sufficient to construct the contig, and the position has been sequenced at 30x, it would have an AF_N of 0.1. If the corresponding variant in the tumor has a frequency of 90% this would be within the allowed TIN of 25%. But we don't have any data on this. In the above example, the AF_N and the AF_T may both be 0, and the consequence of this uncertainty of AFs is that the allowed presence in the normal of variants with and without the assembled contigs may be different. I mentioned this earlier too, and referenced a discussion with Jesper where the conclusion was that a variant with a contig in the normal was most likely a normal variant. Speaking to him again, it was possible according to him that as few as 3 reads may be sufficient in some cases to construct the contig. In a scenario of 10% TIN, this would put quite a number of variants with contigs in the normal at risk of being filtered out. The question is; is it worth accounting for these possibilities and not filter variants with contigs at all? To start with, we can at least filter out all variants without either a contig or any AF in the tumor, beyond that we can set to PASS all variants with neither any contig or AF in the normal.
After applying these most loose of filters there's a significant number of "Uncertain" variants with a presence in the normal and the tumor which we may want to filter. How many of these have a contig?
The variants in the blue box have no contig in either tumor or normal, and can be filtered simply with the AF filters previously discussed, and then we have the remaining variants.
It's clear that there's a lot of variants with contigs assigned in both tumor and normal (thousands in each case), and including these would make the analysis quite noisy. We have three options to deal with these variants with contigs in both samples...:
For a large number of these variants we have AFs, and they do follow the same pattern as those variants without contigs. In the scatterplot below both the tumor and the normal have assembled a contig for the variant (red bars in the barplot above):
The option I personally prefer is option 3, which means apply the AF-based filters to all variants, and when no AFs are available defer to the stringent filter of removing variants if there is a contig in the normal. If we focus only on the Red box in the barplot above, the variants with contigs in both tumor and normal.
For anyone who wants to suffer here's what the combination of different values and the decision of applying filters or not. And here are the barplots for these different groups of variants: Applying these filters would result in this: Where this, previously seen scatterplot, shows the allele_frequencies: And these are only the PASS variants on the scatterplot, labeled according to their combination of properties: And the corresponding barplot, showing that the vast majority of variants with PASS have no presence at all in the normal (as a result of these normal-samples having no TIN probably) BND variantsBND variant in TIDDIT has 2 calls, sometimes 1 of these are found in the normal, and the 2nd not. See below: BND1: 2 218932779 SV_386_1 BND2:2 218933210 SV_386_2 If we applied the filters mentioned earlier to these variants we might cripple the annotation of some of them as both BNDs are required in VEP to correctly annotate the variants. The best way to filter these variants would be to have a script that could look at the dual-BNDs as 1 (which is not possible with BCFtools since they are on different rows). This way we could filter out these variants only if BOTH BNDs are in the normal. The faster option, which does not rely on anything more than bcftools would be to filter out only the BNDs that are ONLY in the normal. The faster option would mean that we would keep the variants in the green box, regardless if they are clearly germline variants. Which would add on average 600 variants extra as noise.
Extra figures:Just for fun, number of SVtypes per case:
|
Beta Was this translation helpful? Give feedback.
-
|
Current state of the filters: normal_variant set if: high_normal_af set if: high_normal_af_fraction set if: in_normal set if:
|
Beta Was this translation helpful? Give feedback.
-
|
As I mentioned earlier, due to BND variants having 1 variants per BreakEnd, and the need for both to exist for proper annotation in VEP, and this combined with the observation that in some cases 1 of the BND variants are filtered as "in normal" and the other "PASS", I implemented this filter in a custom python script rather than using bcftools. From tiddit_sv_tumor_normal rule: The first step annotates the variant with the normal-filters The second step rescues BNDs if at least 1 out of the 2 breakends has PASS Final filter assignment results: Scatterplot, as you can see there are a few variants with not allowed allele-frequencies because they're rescued due to its companion BND having filter PASS: Perhaps more clearly observed here, keeping only the PASS variants: Very few BND-variants are rescued in this approach. PASS variants before and after:
|
Beta Was this translation helpful? Give feedback.
-
Background
In issue #1118 it was brought up that there has been a significant increase in the number of SV calls in the final VCF since version 10 when TIDDIT was introduced, and which is due primarily to the lack of filtering of normal variants in the merged tumor + normal VCF.
This is attempted to be solved in this PR: #1120 by implementation of filters in bcftools which filters variants based on their presence in the normal.
But before implementing the filtering I wanted to report the results of this filtering, and allow for feedback based on this.
This filtering is only relevant for T / N WGS cases, and I selected 3 at random to look at: A, B and C, all with about 13k variants each in the merged vcf from TIDDIT without any filters applied. (See original caseIDs in private CG google drive doc: https://docs.google.com/document/d/1dP1mM-LkuvZUOA6cripXC5I089V-M8knEw26W5hBMEs/edit)
Results
Out of the total 40117 variants looked at, 2 had variant-frequencies that gave me some concern for the viability of calculating allele-frequencies from TIDDIT, as they had allele frequencies in the normal above 1, which I would not expect to be possible.
For downstream plotting and reporting these 2 variants are excluded.
Below is a scatterplot of all variants in all 3 samples, based on allele frequency in normal (Y) and tumor (X).
After applying filters in the PR:
Reasoning for the filter set in this way:
I had preferred to implement a filter where the allowed frequency in the normal was based on a maximum tumor contamination level of 0.1. Such that:
Like this:
However it doesn't seem like bcftools is capable of implementing a filter which depends on multiple samples within the same vcf in this way, and I was forced to rely on a fixed value. So I chose a space of allele-frequencies that would allow for tumor contamination between 25-10%.
There are 4 groups of variants in this merged vcf based on AF_T and AF_N, and I interpret them like this:
Contig variants
It's a bit difficult to see in these plots the overlapping variants, and there are some interesting ones left to mention...
While it appears from the plots above that all variants with AF_N = 0 has PASS, there are variants with the filter "in_normal" that have an allele frequency in the normal sample of 0. Yet they have been called in the normal.
In fact out of 40115 variants, 3311 of them fit that description. The same is true for the tumor variants, where 2499 of the variants have an AF_T of 0 and yet are still called as PASS.
This seems to have to do with how TIDDIT reports certain variants that have successfully been assembled into a contig, where the stats are sometimes lost.
In the scatterplot below I have included only variants with filter "in_normal", and the colors marks if they have a contig in the normal or not.
In addition I can confirm that:
I discussed these variants with Jesper, the creator of TIDDIT, last week and he claimed that it was unlikely that a variant could have an assembled contig without significant read support, and based on that it would be safe to remove them as normal variants.
Some of these variants are still rescued in the contamination filtering, and in the scatterplot below only variants with a contig in the normal are shown, the orange marked variants will be rescued:
Discussion
Is there a significant risk that I'm filtering away true somatic variants that should be taken into account here?
Risk factors identified:
In general I think this is a safe filter to implement to reduce the amount of noise. At least in these cases it does not appear to be a lot of tumor in the normal contamination (TINC), if there was a significant TINC I would expect to see less of a clear separation between the shared and unique variants.
Beta Was this translation helpful? Give feedback.
All reactions