Alt-Depth Reconciliation — VarDict vs Dragen vs BAM

1. The four-way discrepancy, per site

Every number below is measured from this job's own BAM. "BAM truth" = distinct DNA fragments carrying ALT ÷ distinct fragments overlapping the site (mate overlap removed).

VarDict/Dragen/mpileup depths differ because they count different things: VarDict & Dragen report fragment-level depth; mpileup-capped is limited by -d 8000; mpileup-uncapped counts overlap-collapsed fragments. This panel is ~99% overlapping mate pairs (R1 and R2 both span each site), so raw read count ≈ 2× fragment count.

2. Deep dive — ZRSR2 chrX:15,821,873 A>G (your example)

2a. Where the ALT reads actually live (amplicon decomposition)

The main amplicon 15821721–15821977 carries 20 / 10,329 = 0.19% G — ordinary background error. The minor amplicon 15821872–15821977 (starts 1 bp before the variant) carries 69 / 70 = 98.6% G. A genuine heterozygous or somatic variant would appear at the same fraction in every amplicon covering the base. It does not.

2b. The artifact signature (measured on the 69 ALT reads of the hot amplicon)

Every alt read carries G as its first sequenced base, at high quality and high MAPQ, on both strands and both mates, and only in this one short amplicon. That is the fingerprint of a primer / amplicon-edge artifact: the primer's 3′ terminus sits on the variant position, so the synthetic primer base is read into the insert as a constant "allele" independent of the template.

2c. Why each caller reports a different number — reconstructed from the reads

Cap simulation: sorting the 10,899 reads by coordinate and keeping the first 8,000 leaves 12 ALT; keeping the last 8,000 leaves 92 ALT. samtools' -d cap discards the late-starting amplicon where the artifact lives — reproducing Dragen's 11 by pure sampling geometry, not biology.

3. The same pattern across all six sites

Two distinct failure modes emerge. Read-position-0 fraction and the hottest minor amplicon are measured from the BAM.

4. What to trust — per site

5. Remediation — per-variant structural QC (proceeding)

One BAM pass stamps every called variant with the structural evidence and an advisory verdict, joined into the combined Excel as new columns. Annotate, never filter — the reviewer and downstream QC gate decide; no variant is silently removed.

New Excel columns (per variant)

How the verdict is decided (advisory, tunable thresholds)

6. Prototype validated on this sample (700 candidate alleles)

The prototype ran over the full VarDict+FreeBayes candidate union for RUN42-DNA-21 (700 alleles, 61 ms each). Every count reproduces the hand analysis in §1–3.

Every variant flagged PRIMER_EDGE in this sample (15)

Sorted by VAF. Note the spread — from 95% down to 0.6% — and that one is called by both callers. VAF magnitude does not separate these from real variants; structure does.

48 variants were flagged CAP_SUPPRESSED — genuine signal that samtools' default -d 8000 cap under-reports (like ASXL1: 7 alt kept of 270). Any verification step that pileups with the default cap will call these falsely low. 3 low-VAF single-amplicon calls were flagged NEEDS_CONFIRM. Nothing was dropped — all 700 variants carry forward into the Excel with these columns attached.

All six clinical sites classify correctly: ZRSR2 / PHF6 / TET2 → PRIMER_EDGE, ASXL1 → CAP_SUPPRESSED, both PRPF8 → PASS. Full table in variant_structural_flags.tsv; function in variant_structural_qc.py (61 ms/variant, one BAM pass).

7. Methods

Read-level counts computed with pysam 0.24 over the job BAM newpanel-i100val-RUN42-DNA-21_S2_L001_001.bam, pileup with truncate=True, max_depth=10⁷, stepper=nofilter, ignore_overlaps=False, min_base_quality=0 (no hidden filtering). Fragments = distinct query names. Amplicon = distinct (reference_start, reference_end) span. mpileup configs: -A -B -q20 -Q20 (capped, default -d), -d 0 (uncapped), -q0 -Q0 -d 0 (raw). VarDict fields from the job's .vardict.vcf.gz; Dragen from Dragen_Illumina_caller.sites.txt. Every count in this report is reproducible from readlevel_analysis.json, vardict_sites.json, mpileup_rerun.json in this folder.