BRCA1/2 large-genomic-alteration clinical-screen redesign

Corrected conclusion and release decision

What successful clinical implementations demonstrate

The recurring design pattern is broad or purpose-built normalization, large assay-matched baselines, target-specific noise modeling, explicit QC/no-calls, and confirmation or otherwise rigorously validated reporting. Performance units differ—events, exons, analytical calls, candidates—and are preserved where their public source exposes them; paper-reported aggregates are labeled as such.

Delivered algorithm: anchored clinical-depth screen

The model deliberately refuses a BRCA-only input matrix. It requires stable non-reportable controls so an event cannot normalize itself away. Every positive is a REFLEX_POSITIVE, never an autonomous diagnosis.

lgasieve-depth fit \
  --targets targets.tsv \
  --fit-counts fit_normals.tsv \
  --calibration-counts calibration_normals.tsv \
  --components 12 --alpha 0.01 \
  --out site_assay_model.json

lgasieve-depth screen \
  --model site_assay_model.json \
  --counts cases.tsv \
  --out screen_results.json

The CLI records input/model hashes and rejects malformed manifests, insufficient controls, target mismatches, and invalid count matrices. A negative result is full only when every reportable target is callable.

Recursive development loop

Each iteration changed one identified failure mode and was preserved as a versioned JSON/model. Quick streams were used for development; the final stream used a new seed offset, larger cohorts, and was not followed by another algorithm change.

Locked full-panel in-silico validation

The simulator contained 550 targets (500 non-reportable controls and 50 BRCA bins), target efficiencies, GC effects, eight latent capture factors, Poisson molecule sampling, sample/target instability, site shift, low depth, mosaic amplitudes, and deliberately injected technical artifacts. Fit, calibration, development, and frozen validation specimens were disjoint.

The displayed Wilson intervals are binomial Monte Carlo intervals conditional on this fixed simulator and frozen parameter set. They do not include simulator misspecification, model-selection uncertainty, batch dependence, assay variation, or clinical-population uncertainty.

Performance by event class

Whole-gene detection was 52/52, but exact localization was only 23/52 because joined windows and masked targets intentionally favor detection over exact boundaries. The correct disposition is reflex confirmation, not a falsely precise exon interval.

Stress test: where the model still fails

Clinical validation must therefore predeclare a minimum depth/input regime, target-specific exclusions, the behavior for mosaic dosage, and whether a no-call triggers repeat sequencing or direct orthogonal dosage testing.

Legacy BAM-simulator baseline: why the decision layer was replaced

The preserved BAM-level simulator used 40 PoN normals, 16 training normals, 20 training carriers, 14 test normals and 26 test carriers. The strict final tier was too conservative even though many true events existed in the candidate stream.

Manual audit showed 24/26 events localized somewhere among candidates, but orthogonal-support and family-max requirements rejected many. One absent BRCA2 exon-6 duplication retained clip/pair evidence suitable for a breakpoint-rescue path; the other absent event involved a deliberately unstable BRCA2 exon-3 probe and should be a target no-call. The redesign therefore uses a sensitivity-first screen plus reflex rather than demanding every evidence channel agree before a candidate can leave the NGS layer.

Rerun on the supplied data

The available matrix has 147 QC-pass specimens × 50 BRCA targets; 150 complete evidence bundles exist, including three low-depth libraries. There are no control genes/off-target bins and no orthogonally confirmed BRCA truth. The broad-panel algorithm correctly refuses this input. A constrained fallback normalized BRCA1 with BRCA2 and vice versa only to test what information remains.

Unconfirmed screening signals from the constrained fallback

Explicit no-calls

Why these candidates cannot be treated as truth

The fallback’s development sensitivity was 64.43% overall and 35.50% for single-exon events; duplication sensitivity was 44.37%. Its 34/37 presumed-negative clean fraction is not specificity because those specimens are not verified negatives.

Legacy unconfirmed screening signals on the same presumed-negative pool

Comparator research and transfer plan

A two-caller voting rule is not recommended: it can veto the single-exon events that a sensitive primary caller is meant to retain. Secondary callers, allele balance, off-target/global dosage, split reads and fragment evidence should rank and review candidates; an independently designed dosage assay should adjudicate them.

The BRACNAC source was downloaded and hash-pinned as a comparator, but not represented as executed. R is absent in the current environment, so DECoN, ExomeDepth and panelcn.MOPS should be run in a pinned R/container environment once full-panel counts exist. GATK-gCNV should likewise wait for a compatible assay-wide cohort.

What is needed to cross from workable software to a clinical pipeline

Missing inputs

Locked validation program

1. Recover the complete assay feature space and process metadata.
2. Freeze a site/chemistry-specific model using disjoint fit and calibration normals.
3. Compare the new caller with GATK-gCNV, DECoN/ExomeDepth, panelcn.MOPS and a transparent Z-score method on identical locked folds.
4. Optimize a sensitive NGS screening threshold subject to a prospectively locked reflex workload, not a fictitious zero-false-positive requirement.
5. Confirm every positive by MLPA, exon-dense aCGH or independently designed ddPCR during validation and initial clinical deployment.
6. Report event sensitivity, exact localization, raw screen specificity, false-reflex workload, final adjudicated specificity, PPV and no-call rate with all denominators and 95% intervals.

Software, evidence and provenance

Verification performed

python -m pytest -q
# 151 passed, 3 skipped, 1 warning, 7 subtests passed

python -m scripts.clinical_panel_simulation \
  --validation-seed-offset 1000 \
  --output analysis/clinical_panel_simulation_final.json \
  --model analysis/anchored_clinical_depth_model_final.json

python -m scripts.cross_gene_reflex_reanalysis

The three skipped tests require real downloaded sequencing libraries that are not present locally. The 75-sample technical lockbox remains sealed; it is a 1000G technical holdout, not an orthogonally validated clinical truth set.

Source integrity

Original retrieved source hashes remain in SOURCE_MANIFEST.json. The prior report is preserved as a superseded audit record. This v2 report corrects its field-level conclusion while retaining its useful diagnosis of the narrow 50-target data limitation.

Selected references

1. FDA myChoice CDx technical information (P190014/S003).
2. Judkins et al. Clinical validation of a hereditary cancer NGS CNV workflow.
3. Fowler et al. DECoN clinical BRCA exon-CNV implementation.
4. Lincoln et al. BRCA1/2 NGS assay validation.
5. Invitae/Labcorp hereditary multigene test validation summary.
6. Povysil et al. panelcn.MOPS.
7. GATK GermlineCNVCaller official documentation.
8. 2025 benchmark of twelve targeted-panel CNV callers.
9. Real-world BRCA1/2 NGS CNV calling and MLPA confirmation study.
10. BRACNAC paper and open-source BRCA CNV software.

All web references were accessed on 2026-07-31. Product and vendor performance is summarized in its published unit and cannot be transferred to LGASieve without assay-specific validation.