This is a test version of Biostars. For the public version, visit https://www.biostars.org.
Struggling to detect known partial deletion in NOTCH2 from WES (germline, small cohort, no reference panel) - CNV callers give inconsistent/wrong results

Hi all, looking for advice on tooling/approach for a problem I'm stuck on.

Setup:

  • 5 WES samples (paired-end, Illumina, BWA-MEM aligned, sambamba dedup) from the same family, hg38/GRCh38. Each sample represents an independent patient so these 5 samples are unrelated.
  • Capture kit target BED not available to me (~60Mb on-target footprint per sample, possibly Agilent SureSelect V6 based on size, but unconfirmed)
  • No unrelated normal/control WES samples currently confirmed usable as a reference panel
  • Goal: confirm which sample(s) carry a known partial deletion in NOTCH2 (clinically confirmed by other means in 2 of the 5 patients, but I don't yet know the exact exon(s) or method used for that clinical confirmation)

What I've tried:

  1. DELLY (germline SV workflow, sr/merge/genotype/filter) - no deletion calls anywhere near NOTCH2 in any of the 5 samples
  2. CNVkit batch mode with a flat reference (no matched/pooled normals available) - segmentation collapsed the whole gene into one CN=2 segment for 4/5 samples; per-bin bintest flagged several exons but the same bins were flagged across nearly all samples in the same direction, which reads like shared technical noise rather than patient-specific signal
  3. Manual IGV visual inspection (group-autoscaled coverage tracks) across the whole gene - no obvious dropout found in the samples I was able to review carefully
  4. Control-FREEC, single-sample/no-control mode, restricted to a 34-exon NOTCH2-only BED pulled from UCSC (window=0, maxThreads=1 to avoid a BED-parsing race condition I hit with multithreading) - this called a clean, reproducible heterozygous deletion (CN=1) at the same coordinates in 2 of the 5 samples

The problem: the 2 samples Control-FREEC flagged do NOT match the 2 samples independently confirmed by my PI through other means. So I have an apparent false positive pair and false negative pair from my pipeline.

Questions:

  • For germline partial-gene deletion detection in a small WES cohort with no confirmed-normal reference samples, what's the current best-practice tool/approach? (ExomeDepth? GATK gCNV? something else?)
  • Is there a known issue with Control-FREEC's no-control exome mode producing false positives at specific loci, especially near segmental duplications (part of my deleted region overlaps the NOTCH2NL paralog)?
  • Any advice on validating/troubleshooting a mismatch like this before trying yet another caller - e.g., specific things to check in the BAM/pileup at the clinically-confirmed-positive samples that a depth-based caller might be missing (small intra-exon deletion not removing a whole exon? breakpoints entirely intronic, invisible to WES?)

Appreciate any pointers, trying to land on one standardized, defensible workflow rather than chasing every tool that exists.

wesanalysis detectionofpartialdeletion
• 81 views • link
updated by Chris Miller Chris Miller • written by odi • odi

1 answer

  • Without some kind of matched normal, this is going to be tricky, especially if you don't know that any of the samples have clean NOTCH2 without the deletion. Exome sequencing means variable capture affinities and depths for each probe, so there's a lot of noise. I'd look around for some other samples you can use as a control - even stuff from dbGaP that used the same capture kit/sequencing method could be useful here.

  • If I was trying to confirm the deletion for myself, I'd probably start by tiling 50bp windows over the region of interest, plot them and look for big drops, then zoom in to IGV to look for the telltale signs of a deletion - split or paired reads, softclipping, etc. If you get lucky, maybe the BP is in an exon. If you can get a positive ID in one sample, that makes it way easier to look for that same event in the other family members.

• 0 views • link
written by Chris Miller Chris Miller

Log in to answer this question.

More posts like this