Event sensitivity
3,850 / 3,892 · 95% CI 98.54–99.20%Clinical-screen redesign · 31 July 2026
A candidate BRCA LGA screen, built to fail safely.
BRCA1/2 large genomic alterations can be detected by NGS in clinical practice. This research prototype now implements a candidate architecture for real-assay verification: sensitive dosage screening, explicit no-calls and independent confirmation.
Research and assay-development software. Not for patient reporting.
Decision
Feasible in principle. Not clinically validated here.
Clinically established approach
Published laboratories and an FDA-reviewed tumor assay prove that NGS can detect BRCA exon-level dosage changes.
LGASieve remains research software
The local dataset lacks assay-wide controls and orthogonal truth, so its synthetic metrics cannot become clinical claims.
The correction is not “this cannot be done.” It is: a clinical screen needs broad normalization, calibrated no-call boundaries, assay-matched truth and a confirmation workflow.
01 / Verification
Recursive simulation exposed the weak spots.
Nine development iterations were followed by a larger locked random stream. The final broad-panel model passed its internal gates, then a deliberately harder stress set revealed the remaining failure modes.
Event sensitivity
3,850 / 4,000; positive no-calls count as missedSpecificity
1,957 / 1,987 · 13/2,000 no-calls; 99.35% normal completionIntention-to-test sensitivity
Mosaic, single-exon, low-depth and transport-shift casesClaim boundary
These are internal software-development results from in-silico data. Their 95% intervals are conditional Monte Carlo intervals under the fixed simulator and model—not clinical-population or assay uncertainty. They are not clinical sensitivity, specificity, LoD, PPV or NPV.
02 / Method
The architecture separates screening from diagnosis.
The caller is intentionally sensitive at the first layer and conservative at release. That is the pattern used by successful clinical workflows.
Normalize broadly
Model BRCA targets against a large, assay-matched control domain instead of normalizing an affected gene against itself.
Correct technical structure
Use robust target centers, latent-factor correction, matched references and target-specific callable masks.
Screen with explicit states
Return REFLEX_POSITIVE, NEGATIVE_SCREEN, INCOMPLETE_SCREEN or NO_CALL_QC—never silently convert missing evidence into a negative.
Confirm independently
Send candidate events to MLPA, exon-dense aCGH, ddPCR or another validated dosage method before patient reporting.
03 / Published evidence
Clinical precedent is real.
Public validations differ in assay, truth unit and reporting policy. Their performance is evidence of feasibility—not a transferable performance claim for LGASieve.
Assay / workflowContextPublic result
myChoice CDxFDA-regulated tumor NGS402/402 valid gene-level LR analytical calls concordantQuest 34-gene panelGermline screen + reflex aCGH42/42 positive trials; 448/457 = 98.0% raw screen specificityDECoNClinical BRCA implementation16/16 evaluation events; 23/24 prospective flags confirmedCounsyl BRCA1/2Purpose-built germline assay79/79 positive exon units; 3,067/3,067 negative exonsValues are summarized with their original units. See the full report for caveats, confidence intervals and additional comparators.
04 / Supplied evidence
The new data were rerun—without overstating them.
The broad-panel caller correctly refuses the narrow 50-target input. A constrained BRCA1↔BRCA2 fallback was run only to identify review candidates and to measure how much information remains.
Incomplete screens
No signal, but four excluded targets prevent a negative interpretation
Reflex-positive signals
Unconfirmed screening candidates; truth unknown
QC no-calls
Low-depth evidence bundles
Truth-paired samples
No local orthogonal BRCA LGA truth
05 / Route to clinical validation
What must happen next.
- 1Freeze the assay and caller
Code, panel version, target mask, thresholds, QC rules and confirmation SOP.
- 2Restore the normalization domain
Complete panel-wide counts, actual bait design and matched routine controls.
- 3Build independent truth cohorts
Single-exon through whole-gene DEL/DUP positives plus verified negatives.
- 4Measure transport and precision
Runs, lots, instruments, operators, sites, depth, input and failure cases.
- 5Run a new, blinded clinical lockbox
Use an orthogonally truthed cohort and report intention-to-test and completed-case metrics with confidence intervals; the existing 1000G technical lockbox cannot estimate clinical performance.
Full record
Methods, denominators, limitations and downloadable artifacts.
The technical HTML report contains the complete evidence review, algorithm design, recursive simulation history, stress results, supplied-data rerun and frozen validation plan.