Cross-Caller Concordance — GATK vs DeepVariant, before and after digging in

GATK HaplotypeCaller and DeepVariant were run independently against the same aligned, deduplicated, fastp-trimmed CAPN3/DMD reads and cross-checked with bcftools isec. The raw numbers look like a real disagreement — 2,314 vs 2,906 calls, only 2,267 shared. Following the flow below shows what that apparent 639-call gap actually is.

86% of the apparent disagreement wasn't disagreement. DeepVariant reports a record for every candidate site it examines, not only confident calls — so most of its 639 "private" records are RefCall/NoCall bookkeeping, not genuine variant calls the two tools actually disagree on. Only 89 of the 639 are real PASS-filtered calls unique to DeepVariant. Both callers' totals reconcile exactly: 2,314 = 47 + 2,267, and 2,906 = 639 + 2,267.

Reading the diagram: the "Concordant calls" node looks larger than its 2,267 label because Sankey diagrams total every flow into a node — and the same 2,267 calls arrive twice, once counted from GATK's side and once from DeepVariant's side. Its true value is 2,267, not the ~4,530 its height alone would suggest.