On Mar 5, 2020, at 4:36 PM, Tim Triche, Jr. <tim.triche at gmail.com> wrote:
?
a few thoughts:
1) look into Shearwater (https://bioconductor.org/packages/release/bioc/html/deepSNV.html), then
2) talk to Todd Druley @ WashU, Elli Pappaemanuil @ MSKCC, Ruud & Bob @ Erasmus, the usual suspects
3) plan to validate w/ddPCR (at the absolute very least) and be aware that most MRD in leukemia is done by a combination of BCR/TCR + breakpoint PCR (lymphoid/fusion-driven) or DFN flow (myeloid + normal cyto)
not saying that ML-based methods might not help, but if you've got a 30x-100x genome (or even 1000x FM1) and are trying to compete with existing standard approaches that can detect molecules at 1e-6, it'll be rough. An alternative approach (that has been used repeatedly) is to throw caution to the wind, generate primers for numerous subject-specific somatic variants, and use the ensemble to try and model MRD (speaking of ML). On the one hand, that could give the clinic a "customer for life"; on the other hand, it's not conducive to large-scale automation & deployment. As far as I know, it never got much traction, in leukemia or anywhere else. (Consider that flow cytometry is capable of detecting 1-in-10K to 1-in-a-million cells in most clinical flow labs.)
Best of luck! (and if you're not already working with UMI-tagged reads, please talk to the people in #2 above; the reason most people won't go below 5% VAF is that you get thwacked by error rates at that level, and the reason most NGS-based MRD is based on UMIs is that existing PCR-based methods have 6 logs sensitivity.)
--t
On Thu, Mar 5, 2020 at 4:08 PM Mulder, R <r.mulder01 at umcg.nl> wrote:
Hi,
I was wondering if anyone could help me with a script and support for the above mentioned goal.
For this I have several BAM files for which I want to determine de nucleotide count per region of interest. The latter could be several hotspot mutation sites. I would like to get an overall overview of all the BAM files. Next I want to use these counts to determine for any new BAM file if the count for a particular genomic position is higher than the allowable range, hence could indicate if a mutation is present. For this I would like to use the modified Thompson Tau test. I think machine learning could be used for this. So, why do I want to do all this? Well, normal NGS pipelines only deal with variants at a frequency of 5%. Mutatios below this frequency are often missed. To know if a mutation is present below this level, you showed dive into the alignment and most often manually investigate the base calls. I know that this races some questions regarding call qualities, but then again our conventional assays have actually confirmed some of these low mutations. In addition, NGS can
be used to determine the presence of low frequent mutation which is of great importance for determining the measurable residual disease after treatment.
I am new to r and bioconductor so I would be very thankful if someone could help me in my mission to setting up a script for this purpose.
Thanks,
Rene Mulder
Laboratory Medicine
University Medical Center Groningen
The Netherlands