LGASieve v2: internal verification of a BRCA1/2 exon deletion/duplication screen

Clinical feasibility is established; transferability is not

Large genomic alterations (LGAs) in BRCA1 and BRCA2 include exon-level and whole-gene deletions and duplications. Published clinical laboratories have detected these events from targeted NGS, often with a large assay-specific normalization set and independent confirmation. Examples include DECoN, a purpose-built BRCA1/2 assay, a 34-gene inherited cancer panel with reflex aCGH, and the FDA-reviewed myChoice CDx tumor assay.

Those validations demonstrate feasibility for their own specimen types, capture designs, laboratories, reference cohorts, algorithms, and confirmation policies. They do not validate LGASieve or make performance numbers transferable to a different assay.

Research question

What measurement design, statistical safeguards, and validation evidence are required for LGASieve to become a useful screen for BRCA1/2 LGAs, and how does the corrected implementation behave when the data distribution departs from its development simulator?

Predeclared hypotheses and decisions

How the experiments were conducted

3.1 Analysis sequence

1. Audit the recovered source and evidence. Reconstruct the pipeline, inventory available inputs, and identify absent BAM/CRAM/FASTQ, capture-design, batch, and orthogonal-truth files.
2. Correct structural defects. Exclude the tested interval from its own baseline, separate model fitting from null calibration, use one joint deletion/duplication search family, prevent windows from bridging uncallable gaps, and enforce strict model identity and provenance.
3. Recursively simulate events. Iterate model structure and thresholds on development streams, then freeze the selected model before same-generator, stress, and changed-assay checks.
4. Reanalyze the supplied data. Apply only a constrained BRCA1-to-BRCA2 fallback because the broad-panel model correctly refuses the 50-target input.
5. Verify the software. Run the full automated test suite, preserve frozen JSON artifacts, and publish hashes and reproduction commands.

3.2 Datasets and cohort separation

3.3 Corrected screening architecture

The broad-panel screen normalizes every sample against a large set of stable, non-reportable controls; estimates robust target centers and latent technical factors from the fit cohort only; scores contiguous callable windows with deletion and duplication alternatives; and uses an untouched calibration cohort to set one joint empirical threshold. A candidate window is excluded from reference matching and nuisance estimation before it is scored.

The output vocabulary is deliberately operational:

• REFLEX_POSITIVE: an NGS screening hypothesis requiring independent dosage confirmation;
• NO_REFLEX_FLAG: no referral from the frozen research model, not a clinical negative;
• NO_CANDIDATE_PARTIAL_COVERAGE: no candidate with incomplete reportable coverage, so no negative interpretation;
• NO_CALL_QC: sample-level depth or noise failure.

3.4 Endpoints and engineering gates

Positive events were evaluated by any referral, at least 80% reciprocal overlap, exact endpoints, and exact target set. Full truth was never clipped to the callable target mask. Results are reported both intention-to-test (all generated events, with failures retained) and completed-case (sample QC passed and all truth targets were callable).

• Completed 80%-reciprocal-overlap recovery at least 98%;
• completed no-event referral rate no greater than 2.5%;
• at least 90% of reportable targets callable;
• positive-event completion at least 85%.

These gates were engineering selection rules, not clinical acceptance criteria.

Why the favorable number is not the main result

The 98.88% completed-case overlap recovery in the same-generator check is real for that fixed simulation, but it is conditional on 3,299 completed events and on a generator closely related to development. Intention-to-test recovery was 83.00%, and the selected model failed its positive-completion and no-event referral gates on that stream.

The stress result is more informative for present decision-making. With lower depth, site effects, mosaics, artifacts, and reduced callability, only 55.21% of positive events completed and 47.50% were recovered by the overlap endpoint in the full intention-to-test denominator. This gap shows that good conditional discrimination does not repair missing or unstable measurement.

The changed-assay experiment produced complete QC refusal. This is preferable to false confidence, but it also establishes that the frozen model cannot be transferred to a different target-efficiency and latent-factor structure. Each assay, chemistry, site, and materially changed process requires refitting and recalibration.

The supplied dataset is useful as an information-limit study, not as a clinical validation cohort. Its five flags require independent confirmation, and its other states cannot be interpreted as negative because targets are excluded and truth is absent.

Published context

Clinical precedent supports the direction of travel: DECoN reported 16/16 BRCA exon CNVs in its evaluation set and confirmed 23/24 later clinical flags by MLPA; a purpose-built BRCA assay reported 79 positive and 3,067 negative exon units concordant with reference testing; a 34-gene panel used a large assay-specific baseline and reflex aCGH; and FDA technical information for myChoice CDx reports 402/402 concordant valid BRCA1/2 large-rearrangement calls in an accuracy study. These are assay-specific results and are not LGASieve performance claims.

Conclusions

1. BRCA1/2 LGA detection by targeted NGS is clinically feasible when the assay and pipeline are purpose-built and locally validated.
2. LGASieve is a functioning research and assay-development scaffold, not a validated clinical pipeline.
3. The primary engineering hypothesis was not fully met: the best development model passed three of four gates, and the held-out same-generator check passed two of four.
4. Same-generator completed-case performance overstates robustness. The within-assay stress stream is the more decision-relevant software result.
5. The corrected model fails safely under a deliberately changed assay, but that means it must be refit and recalibrated for transport.
6. The current supplied export cannot support clinical negatives or performance estimates, and its five flagged specimens remain unconfirmed.
7. Candidate exclusion is methodologically necessary, but the available BRCA-only measurement remains particularly weak for duplications, single-exon events, and mosaics.

What should happen next

1. Recover the complete measurement domain. Obtain assay-wide unique-fragment or UMI counts, capture BED, genome build, transcript definitions, and run/kit/lot/instrument/operator metadata.
2. Define target roles before validation. Freeze invariant-control, ignored, reportable, and sample-specific failed-target masks.
3. Build an orthogonally characterized cohort. Include single-exon, multi-exon, terminal, promoter, mosaic, and whole-gene deletions and duplications, plus verified negatives.
4. Freeze the protocol. Lock code, model schema, target mask, QC rules, thresholds, endpoints, and the MLPA/aCGH/ddPCR confirmation SOP before opening the test cohort.
5. Use a blinded clinical lockbox. Retain every no-call in intention-to-test denominators and adjudicate every referral independently.
6. Benchmark alternatives on identical folds. Compare LGASieve with DECoN/ExomeDepth, GATK-gCNV, panelcn.MOPS, and a transparent dosage baseline.
7. Measure reproducibility and transport. Vary runs, sites, instruments, reagent lots, operators, DNA input, depth, and quality; predefine acceptable failure and referral rates.
8. Separate detection endpoints. Report referral triggering, reciprocal overlap, exact endpoints, and exact target sets independently.