Skip to content

varlociraptor completeness audit (2026-08-21)

Upstream: snakemake-workflows/dna-seq-varlociraptor @ v6.10.0 (b65c3350) · Port: oxo-flow-varlociraptor (verified; full Tier A run in progress on bioinfo-wsx).

Mode matrix (upstream — config × sample-sheet driven)

axis options
entry points rule all / only_alignment (BROKEN upstream) / benchmark / testcase
calling type variants / fusions / fusions,variants (samples.tsv calling col)
datatype dna (bwa/vg) / rna (STAR+arriba)
callers freebayes / delly (both on → parallel chains) / arriba (fusions)
scenario YTE-rendered per group (default scenario.yaml; per-group override via groups table scenario column)
alignment bwa-mem (default) / vg giraffe pangenome (ref/pangenome/activate; hprc v1.1/v2.0)
reads local fq / SRA auto-download; adapters; UMI branch; per-sample primer panels
scatter calling=16 scatteritems + 4 gather points
downstream tables (vembrane) / oncoprint / datavzrd reports (stratify) / MAF / mutational burden / mutational signatures / population DB update / CHM-eval benchmarking

Gap tiers vs the port

P0: RNA fusion calling (STAR+arriba+splitncigarreads, unannotated FDR path); delly caller chain; UMI/consensus-reads branch; primer trimming/panel mode; per-group scenario overrides; custom alignment properties; benchmarking mode (CHM-eval); testcase generation; mutational signatures + burden + population DB + MAF branches; SRA read source; target_regions/off-target filtering.

P1: none license-wise (MIT; CADD/REVEL/DGIdb/COSMIC free with own terms — DGIdb = live web API dependency).

P2: fdr-control modes (local/global smart/strict), retain-artifacts, infer_genotypes, VEP plugin set (LoFtool/REVEL/CADD/SpliceAI/ AlphaMissense — CADD needs score download; SpliceAI/AlphaMissense user-supplied), report stratification, tables options, scatter count.

Port deviations already documented in its README fidelity table (scatter 16→1, single pre-rendered scenario, wrapper-utils→plain scripts, gather_annotated_calls/filter_odds not ported — benchmarking- only upstream) — the audit now re-labels those as P0 under the full-line mandate (scatter = engine capability; multi-scenario rendering = P0 porting work).

Upstream dead code: only_alignment unsatisfiable; get_haplotype_args/ get_count_group_kmers_input unused; vg2svg producer-less; tables output keys superseded.

Verdict

The live Tier A run covers the DNA-variants default path fully; the full-line gap = fusions (RNA), delly, UMI/primers, scenario matrix, and the downstream branch set.

Live evidence (2026-08-22 — verdict #24, LIVE-PASS)

Full Tier A end-to-end on bioinfo-wsx (64c / 1.4TB, NFS /data): 88-rule DAG completed with exit 0, 0 failed✓ 91 output files verified (45.8GB total) + report snapshot written. Engine v0.13.1 release binary; reference data (Ensembl 111 genome/GTF/VEP cache+plugins, REVEL, known-variants VCFs, 4.4GB pangenome gbz) relayed to the server and pre-placed; every compute rule executed for real including pangenome autoindex (45.8GB), freebayes candidate calling, BQSR, varlociraptor estimate/call/FDR, VEP + dbNSFP annotation, oncoprint, and both datavzrd reports.

Final run: 26 rules executed fresh + 62 checkpoint-verified (rule fingerprints and input manifests confirmed unchanged since their last successful run), 0 failed. Fix chain of the campaign: 16 commits cd54fcb..71bf729 — fixture window re-selection (repeat-rich → uniquely-mapping, MAPQ 60), conda-forge version removals re-pinned (vega-lite-cli/datavzrd), vg env + samtools, freebayes 96→48 threads + region-list sed repair (delimiters, 3-col BED, empty-region guard), oncoprint empty-call guards, yte SimpleNamespace + recursive frame-encoded dict rebuild, plus server-side conda cache/zombie recovery. New failure classes archived in the live-test failure catalog; engine-side notes: none new (resource fast-fail behavior is by design — over-capacity requests clamp).

Re-verification (2026-08-23, 9-mini queue — campaign closure)

Engine: latest main (post-v0.14.1) · Box: bioinfo-wsx · Mode: real CLI run, not dry-run (checkpoint cleared, forced real execution).

First run 17 succeeded + resume 38 succeeded / 50 skipped / 0 failed — full chain green. Coverage (real): preprocessing → vg giraffe pangenome alignment → variant calling → region merge/expand → multiqc. The single failure was the delly exclusion-region download (raw.githubusercontent.com unreachable) — real repo fix pushed (6fda195): ghfast.top mirror fallback (same pattern as mag's build_ale.sh), plus box-side Mac relay pre-staging.