# LGASieve project status — 2026-07-31 ## Decision Exon-level BRCA1/2 deletion and duplication detection has been implemented successfully in specifically designed and locally validated NGS workflows. That feasibility is assay-, laboratory-, and pipeline-specific. LGASieve is now a workable **research software scaffold for assay development**. It is not ready for patient reporting. - **PROCEED:** retrospective, assay-specific software development with complete count data and orthogonal truth. - **HOLD:** prospective clinical validation, patient reporting, autonomous interpretation, or a claim of clinical sensitivity/specificity. - **REFLEX:** every research flag requires confirmation by an independently designed dosage method. - **DO NOT CALL NEGATIVE:** absence of a research flag does not establish absence of a BRCA large genomic alteration. The public, biologist-facing website is the primary readable report: **https://lgasieve-scientific-review.free-washington.chatgpt.site/** Publication checks passed: - production deployment succeeded; - a direct unauthenticated request returned HTTP 200; - an independent external fetch node returned HTTP 200 and found the headline and critical 47.5% stress result; - `/og.png` returned HTTP 200 as `image/png`; - the production page rendered in a browser with its key sections and no production console errors. The earlier `analysis/LGASieve_clinical_screen_report_v2_2026-07-31.html` and its favorable 98–100% headline simulation cards are superseded. They remain only as audit history. ## What the project is about LGASieve attempts to find exon-level deletions and duplications in BRCA1 and BRCA2 from targeted sequencing. Two related development paths are preserved: 1. a BRCA-only matched-reference conditional-count model for the truncated supplied export; and 2. a broad-panel, control-anchored model intended for a complete assay-wide count matrix. The first path can study the information limit of the available data. The second is the route that could support a future assay-specific validation. ## Major defects found and corrected ### 1. A candidate could influence its own baseline The old BRCA-only likelihood path allowed the candidate window to influence reference matching and nuisance estimation. In an injected-event experiment, the old reference set changed in 98.76% of deletion tests and 96.57% of duplication tests. Correction: - exclude the exact candidate window before selecting references or fitting neutral fraction/overdispersion; - bind the scored window to that excluded fit; - use one ordered, target-ID-qualified hypothesis family. This correction does not remove PCR duplicates. Candidate exclusion prevents a candidate from changing its baseline; duplicate removal determines how many independent molecules support the measurement. Both are required. ### 2. The broad-panel calibration cohort trained on itself The former broad model used calibration rows to estimate their own residual centre, target dispersion, callability, covariance, and threshold. A future case was therefore not exchangeable with the calibration samples. Correction: - all nuisance quantities are learned from `fit_counts` only; - the untouched calibration cohort only scores the frozen model; - one maximum across DEL, DUP, and every reportable window sets the empirical family threshold. ### 3. The empirical alpha check was mathematically wrong The former implementation split alpha across window widths but accepted a cohort too small to resolve the resulting tail probability. Correction: - use one joint search family; - require at least `ceil(1 / alpha) - 1` calibration libraries; - fail model fitting if the requested family error rate cannot be attained. ### 4. Uncallable gaps could be bridged The former model could describe a continuous event across an excluded target. Correction: - build windows only inside uninterrupted callable, reportable runs; - uncallable or non-reportable gaps split the event family. ### 5. Model files and cohorts lacked a strict identity contract Correction: - schema upgraded to `lgasieve.anchored-clinical-depth-screen.v2`; - version-1 artifacts are refused and must be refitted; - fit/calibration sample-ID overlap and identical count rows are rejected; - strict dimensions, numeric types, thresholds, masks, windows, and provenance hashes are checked when loading; - model writes are atomic. ### 6. Output language overstated what the model knew Correction: - `REFLEX_POSITIVE`: research candidate requiring independent confirmation; - `NO_REFLEX_FLAG`: no referral flag from the frozen research model, not a clinical negative; - `NO_CANDIDATE_PARTIAL_COVERAGE`: no flag with incomplete target coverage; - `NO_CALL_QC`: sample-level QC failure. The claims boundary is embedded in every result. ### 7. The old simulation was an inverse problem The old “locked validation” changed random seeds but reused the same assay, latent structure, event universe, and exact copy-ratio generator. It also clipped truth to callable targets and emphasized completed cases. Correction: - preserve full truth without clipping; - report trigger, 80% reciprocal overlap, exact endpoints, and exact target sets separately; - report intention-to-test and completed denominators separately; - call negative-sample outcomes referral/no-referral rates, not specificity; - add a genuinely changed assay with aligned target IDs but altered target efficiency, latent loadings, and instability; - label the independent dosage stream as workflow-logic simulation only. ## Real supplied-data rerun: old versus corrected Input: - 150 public 1000 Genomes exome evidence bundles; - a 50-target BRCA1/2 count matrix; - 46 globally usable targets; - no stable off-gene normalization domain; - no MLPA/aCGH-confirmed positives or verified negatives. The constrained BRCA1↔BRCA2 fallback was rerun after the v2 correction. | Outcome on 150 evidence bundles | Previous model | Corrected v2 | | --- | ---: | ---: | | Unadjudicated research flags | 6 | 5 | | Sample QC no-calls | 3 | 21 | | Partial-coverage/no-candidate | 141 | 124 | Only two previous flags were retained (`NA18645`, `NA20530`). Four disappeared, and three new flags appeared. Corrected v2 research flags: - `HG01870`: BRCA2 duplication-like, exons 23–24 - `HG02601`: BRCA1 deletion-like, exon 6; BRCA2 duplication-like, exons 17–18 - `NA18645`: BRCA2 deletion-like, exons 16–17 - `NA19759`: BRCA2 deletion-like, exon 2 - `NA20530`: BRCA1 duplication-like, exons 15–22 These are unadjudicated referral hypotheses, not diagnoses. The correction- induced turnover is evidence against autonomous interpretation. On 2,400 deterministic conditional spike-ins placed on the 37 development backgrounds, corrected completed-case recovery was 59.66% overall, 35.18% for single-exon events, 80.63% for deletions, and 37.70% for duplications. There were 211 positive no-calls. Three of 33 completed presumed-negative libraries were referred. These are internal development measurements, not clinical sensitivity or specificity. Artifact: - `analysis/cross_gene_reflex_reanalysis_v2.json` - `analysis/cross_gene_reflex_model_v2_brca1.json` - `analysis/cross_gene_reflex_model_v2_brca2.json` ## Candidate-exclusion experiment on the real backgrounds With the same 37 host libraries, 259 windows, seeds, and endpoints: | Paired development endpoint | Event-contaminated nuisance fit | Candidate-excluded | | --- | ---: | ---: | | Exact-window recovery | 41.84% | 44.78% | | Deletion recovery | 55.16% | 68.06% | | Duplication recovery | 28.53% | 21.51% | | Presumed-negative libraries without a flag | 33/37 | 35/37 | Candidate exclusion is retained because it removes a structural bias. It is not presented as a universal sensitivity improvement: deletion recovery rose, duplication recovery fell, and the available export still lacks the information needed for a clinically useful operating point. Other one-change tests: - one joint DEL+DUP search at alpha 0.05: 38.54% exact-window recovery; - calibration-only reference-count selection chose `k=10`, with no improvement over the development result; - direct log2 scoring recovered only 0.82% of single-exon duplications at alpha 0.05; - same-superpopulation matching reduced recovery to 19.16%. The failed branches are preserved as evidence and were not promoted. ## Corrected synthetic software verification The full v2 simulation uses 500 synthetic controls plus 50 BRCA bins, 1,300 fit normals, 3,000 untouched calibration normals, and a selected 12-component, alpha 0.01 model. These are simulator settings, not assay acceptance criteria. ### Review first: transport and stress - **Changed assay:** all 2,400 event samples and all 1,200 no-event samples failed sample QC. Completion and ITT recovery were 0%. This is safe refusal and demonstrates that the model is not transportable; the changed assay must be refit and recalibrated. - **Within-assay low-depth/site/mosaic stress:** 1,325/2,400 events completed (55.21%); ITT 80%-reciprocal-overlap recovery was 47.50%; exact-endpoint ITT recovery was 45.92%; the referral rate among 806 completed synthetic no-event samples was 3.97%. ### Same-generator software check - 3,299/4,000 events completed (82.48%); - any referral trigger: 96.73% ITT; - 80%-reciprocal-overlap: 83.00% ITT and 98.88% among completed events; - exact endpoints and exact target set: 81.55% ITT; - no-referral rate: 97.40% ITT and 97.38% among completed no-event samples. The 98.88% conditional result is not clinical sensitivity. It is a same-generator software-verification result after excluding incomplete events. Artifacts: - `analysis/clinical_panel_simulation_final.json` - `analysis/anchored_clinical_depth_model_final.json` ## What “adequate coverage” means There is no universal read-depth cutoff. Coverage is adequate only when: 1. the input represents enough independent fragments or UMI families; 2. the target is stable, uniquely mappable, and sufficiently recovered; 3. the case is compatible with its assay-, chemistry-, and batch-matched reference; 4. observed variance separates CN1, CN2, and CN3 at the declared error rate; 5. failed exons and samples become explicit no-calls. Nominal depth is not independent molecular depth. PCR duplicates, library complexity, molecular recovery, UMI handling, and target efficiency must be measured empirically. ## What published assays demonstrate Published implementations support feasibility, not transferability: - [DECoN](https://pmc.ncbi.nlm.nih.gov/articles/PMC5409526/) detected 16/16 known BRCA CNVs, including eight single-exon events; 23/24 later clinical flags were MLPA-confirmed. Screen-negative clinical specimens were not all orthogonally tested. - [Counsyl BRCA1/2](https://pubmed.ncbi.nlm.nih.gov/27375968/) demonstrated a purpose-built BRCA-only assay with invariant-set normalization in a 114-sample validation. - [Quest 34-gene panel](https://pubmed.ncbi.nlm.nih.gov/32090079/) used a large assay-specific baseline and reflex aCGH for every NGS CNV flag. - [SeqNext six-gene workflow](https://doi.org/10.1016/j.jmoldx.2017.07.003) evaluated 48 MLPA-positive families representing 18 BRCA CNV types and 120 MLPA-negative patients. These studies do not validate LGASieve’s PCA/OAS matched-filter/conformal algorithm, component count, control threshold, or decision cutoff. ## Remaining work before real assay development 1. Reacquire complete assay-wide counts or BAM/CRAM/FASTQ, capture BED, genome build, transcript, and batch/kit/lot/run metadata. 2. Add explicit invariant-control/ignore/reportable target roles. 3. Add sample-specific failed-target masks and an end-to-end locked unique-fragment or UMI count extractor. 4. Assemble orthogonally characterized BRCA1/2 deletions and duplications, including single-, multi-exon, terminal, promoter, and whole-gene classes. 5. Freeze code, manifest, masks, thresholds, and analysis plan before an independent validation. 6. Retain every no-call in intention-to-test denominators and adjudicate every referral with an independent method. 7. Compare the locked caller with DECoN/ExomeDepth, GATK-gCNV, panelcn.MOPS, and a transparent dosage baseline on identical folds. ## Software and verification Main corrected modules: - `lgasieve/countmodel.py` - `lgasieve/countscreen.py` - `lgasieve/clinical_depth.py` - `lgasieve/clinical_depth_cli.py` Main regression tests: - `tests/test_countmodel.py` - `tests/test_countscreen.py` - `tests/test_clinical_depth.py` - `tests/test_clinical_panel_simulation.py` Focused verification completed: - 32 count-model/count-screen/clinical-depth tests passed; - 18 clinical-depth/simulation tests passed; - corrected full synthetic verification completed deterministically; - both corrected cross-gene v2 model files reload through the strict parser. Full-repository verification: - **162 tests passed, 3 skipped, and 7 subtests passed in 45.97 seconds.** - The three skips require real sequencing libraries that are not present in the reconstructed source. - The only warning was a harmless CPU-core-count fallback from `joblib`. ## Source preservation and artifact roles - Dropbox source: `https://www.dropbox.com/home/Yang%20Shao/Personal/Codexhome/Bioinformatics/lgasieve` - `SOURCE_MANIFEST.json` records the reconstructed source archives and hashes. - No source BAM, CRAM, FASTQ, BAI, BED, VCF, sequencing manifest, or orthogonal BRCA dosage truth was recovered from the archives. - Markdown files are the durable project notebook for reviewers. - JSON files are machine-readable model, provenance, and result artifacts. - The public website is the human-readable scientific report for a biologist.