NO-GODo not issue autonomous reportable BRCA CNV calls from this pipeline or this 50-target normalization domain.
The conclusion does not depend on one bad threshold. At the predeclared family alpha of 0.05, the best new development model reached 61.6% deterministic synthetic sensitivity, but only 9.9% for single-exon duplications. With count sampling added, single-exon duplication sensitivity was 10.2%; 50%-mosaic duplication sensitivity was 0.9%. The 147 native libraries are presumed—not verified—negatives, so their clean fraction is not clinical specificity.
903,168stochastic trials; each test host held out from training and calibration
49.6%single-exon germline DEL sensitivity, stochastic
10.2%single-exon germline DUP sensitivity, stochastic
96.9%presumed-negative clean fraction; explicitly not specificity
“Perfect” performance is not a reasonable stopping rule. Repeated tuning on one development collection can make internal numbers look perfect while worsening transport. The valid stopping rule is: freeze a method that passes predeclared development gates, then test it once on independent, orthogonally confirmed specimens. No method tested here earned that lockbox opening.
| Claim | Evidence here | Allowed wording | Not allowed |
|---|---|---|---|
| Sampling power | Known-baseline Poisson oracle | Optimistic information ceiling | Clinical sensitivity |
| Algorithm behavior | Held-out deterministic and stochastic count perturbations | Internal software sensitivity | End-to-end assay sensitivity or LoD |
| False-flag burden | 147 previously examined 1000 Genomes libraries | Presumed-negative clean fraction | Clinical specificity |
| External benchmark | Published callers on different full capture panels | Architecture prior and plausibility check | Transferable performance for this assay |
Reproducible workflow
What I did, in order
- Preserved and reconstructed the source. Seven Dropbox archives were retained with SHA-256 hashes. The legacy
referencepath was restored as a byte-identical copy ofreference_grch37; no algorithm was changed for baseline reproduction. - Reproduced the original implementation. The untouched baseline completed with 107 passed, 3 expected skips, and 7 subtests. The archived decision was already
NO_GO_CLINICAL_OR_SCIENTIFIC_RELEASE. - Located and corrected one proposal-loss defect. The window scan formerly rejected a real multi-exon interval if any internal exon had the opposite residual sign, and then silently retained only the strongest 50 candidates. The corrected scan tests every contiguous window whose two boundaries support the direction; a regression test covers a three-exon deletion with one noisy internal exon.
- Re-ran the frozen native partition. The wrapper can now load and validate the archived 50/60/37 partition instead of reconstructing an unavailable selection file. The caller completed all 97 non-reference libraries.
- Built a candidate-excluded conditional Gaussian experiment. Every one of 147 hosts was outer-held-out; all preprocessing and robust conformal family calibration were fitted inside training folds. All 462 callable windows were scored.
- Recursively swept operating points. Family alpha 0.01 through 0.40, joint two-sided max-statistic calibration, and an exploratory direction-weighted allocation were compared. No high/high operating point appeared.
- Added stochastic LGA mimics. Binomial thinning modeled deletions; coupled Poisson increments modeled duplications. Germline and 50%-mosaic amplitudes, five widths, whole-gene-feasible events, eight common-random-number replicates, depth strata, exact and reciprocal-overlap localization, and specimen-cluster uncertainty were tested.
- Separated physical information from implementation loss. Exact Poisson ceilings, signal geometry, empirical residual floors, target correlation, reference learning, batch/population proxy transport, and calibration resolution were audited.
- Benchmarked against literature and public truth resources. The 2025 12-caller benchmark supplement was preserved locally; ICR96, panelcnDataset, GIAB, simulator packages, professional standards, and MLPA confirmation guidance were evaluated for applicability.
- Kept the lockbox sealed. Since development gates failed, the 75-specimen holdout was not downloaded or analyzed.
| Preserved archive | Bytes | SHA-256 |
|---|---|---|
| exomedepth_development_v2.zip | 401,811 | 527eb7cebfa811fc02cc1cf09ddf1742eedcd7b40e2f9fbdfe3d235986b35852 |
| lgasieve.zip | 1,226,467 | 98c3caacd3a1eaf5b95d13829379e8b2f2623073cd24a2b898ab5e2b69f4b554 |
| reference_grch37.zip | 214,274 | 2665e9fb40e1f3ffc98ed851665639f6dd08cfe241adb05e62bc600d87aec773 |
| scripts.zip | 363,439 | d7629a71e455c8adf097d406880ef22a49b17cb695023c22182814e27c2b23cb |
| tests.zip | 821,210 | 080a7d228473626b3498f9ac662b9f993b2b570493c571260a1c0e20dd1f9c32 |
| validation_corrected.zip | 3,667,430 | 653a8aec481580ba4de8e229cc2339476c607599309825e67a30ef0bd14cd66a |
| validation_v2.zip | 422,371 | 711e32c7937ec947c1fedd358bade94f9a03b8a24a4781e341a007d2a3df28f7 |
Dropbox source folder: authenticated folder. Archive hashes establish the reconstruction boundary; the working tree contains the documented research changes below.
Root-cause analysis
The dominant failure is information recovery, not nominal depth alone
1. Count sampling is often adequate
Under an optimistic known-baseline Poisson oracle with family alpha split over 46 targets and two directions, mean single-exon power is 93.5% for a heterozygous deletion and 87.5% for a single-copy duplication. In libraries with median depth at least 100, oracle duplication power is 98.9%.
This is an optimistic sampling ceiling, not clinical or modeled sensitivity. It proves only that low observed duplication sensitivity cannot be blamed entirely on shot noise.
2. The model discards most duplication information
The older nested predictive model converted only 6.3% of the high-depth oracle duplication power into detections. The new conditional model improves multi-exon behavior, but still detects only 10.2% of stochastic single-exon germline duplications at alpha 0.05.
This gap points to residual capture/batch variation, reference transport, target-specific calibration, and multiple-testing cost—not simply “sequence deeper.”
0.585log2 signal for 3 copies vs −1.000 for a heterozygous deletion
1.70×Poisson Bhattacharyya depth multiplier for equal DUP/DEL information
2.92×equal-z Gaussian depth multiplier; model-dependent upper comparison
46/50callable targets; the archived workflows also contain a conflicting 48-target mask
Measured residual structure
| Diagnostic | Observed | Interpretation |
|---|---|---|
| Median target residual MAD | 0.182 log2 | Too large for reliable single-copy duplication separation at strict family thresholds. |
| Median Poisson delta-method SD | 0.113 log2 | Empirical residual noise is 1.62× the Poisson component at the median target. |
| Quadrature systematic floor | 0.140 log2 | Extra depth cannot remove this estimated capture/model component. |
| Count vs target residual MAD | Spearman ρ = -0.818 | Depth helps, but does not eliminate the residual floor. |
| Adjacent target correlation | median r = 0.199 | Multi-exon windows contain less independent information than their width suggests. |
| Leave-super-population-out transport | 4.7% worse | A weak proxy only; actual site/run/kit/lot/operator metadata are absent. |
| 60-calibrator two-direction p-value floor | 3.28% | Resolution and threshold stability are inadequate for a 1% family false-positive claim. |
Whole-gene non-identifiability. BRCA2 occupies 27 of 50 targets. A uniform BRCA2 deletion changes the panel median to 0.5; median normalization makes affected BRCA2 targets exactly 1.0 and makes unaffected BRCA1 targets appear doubled. A duplication similarly makes BRCA2 neutral and BRCA1 appear deleted. This is an identifiability failure: no amount of depth or additional references can recover an absolute scale that was normalized away. Whole-gene and panel-global events must be no-called unless external invariant anchors are added.
Mask ambiguity is a release blocker. The source matrix contains 50 targets; this audit finds 46 callable (BRCA2_E01, BRCA1_E23, BRCA1_E11, BRCA1_E01 excluded), while archived validation paths sometimes use 48. One versioned mask must be frozen and propagated through training, calibration, evaluation, reporting, and truth matching.
Recursive in-silico testing
More realistic simulation confirms the failure—and reveals instability
The primary simulation perturbs only the outer-held-out host. Reference selection, covariance estimation, robust score scaling, and the empirical sample-family null are learned without that host. Eight common-random-number replicates make method comparisons less noisy. The specimen is the bootstrap cluster; treating 903,168 correlated trials as independent would give a misleadingly narrow CI.
| Event width | Germline DEL | Germline DUP | 50% mosaic DEL | 50% mosaic DUP |
|---|---|---|---|---|
| 1 exon | 49.6% | 10.2% | 3.2% | 0.9% |
| 2 exons | 80.7% | 23.0% | 6.0% | 1.7% |
| 3 exons | 91.6% | 40.3% | 10.6% | 2.6% |
| 5 exons | 98.8% | 71.8% | 28.5% | 8.2% |
| 7 exons | 100.0% | 85.6% | 42.1% | 14.1% |
Uncertainty decomposition. Overall stochastic overlap sensitivity was 35.88%. Its specimen-cluster 95% interval was 34.61%–37.07%, whereas the finite Monte Carlo interval conditional on the fixed grid was only 35.849%–35.905%. Host variation, not simulation count, controls uncertainty.
Transport instability. One outer fold achieved only about 21.6% overlap while the other four were approximately 39.2–39.9%. The same fold was weak in deterministic testing. This reproducible fold effect is a red flag for reference and calibration transport.
Operating-point recursion
Increasing alpha raises sensitivity only by accepting more native flags. Direction-weighted Bonferroni calibration (20% DEL / 80% DUP) did not repair the imbalance: at alpha 0.05 it sacrificed most deletion power and still failed to make single-exon duplication acceptable.
| Family α | Native clean | Overall synthetic | Single DEL | Single DUP |
|---|---|---|---|---|
| 0.01 | 97.3% | 15.7% | 5.7% | 0.4% |
| 0.02 | 96.6% | 34.4% | 19.3% | 3.4% |
| 0.05 | 95.9% | 61.6% | 48.1% | 9.9% |
| 0.10 | 90.5% | 78.4% | 75.9% | 24.1% |
| 0.20 | 76.9% | 88.2% | 91.3% | 45.5% |
| 0.40 | 55.8% | 93.6% | 96.7% | 65.2% |
Simulation coverage boundary. These count-level perturbations test normalization, scoring, localization, calibration, and sampling variability. They do not reproduce hybrid-capture allele dropout, GC/mappability changes, pseudogene misalignment, breakpoint position, duplicate marking, UMI consensus, insert-size effects, library preparation, or variant truth ascertainment. BAM/FASTQ injection with Bamgineer/SECNVs and physical contrived specimens are the next layers, not substitutes for each other.
Code change and fresh rerun
The proposal defect was real, but it was not the main bottleneck
| Behavior | Before | After | Verification |
|---|---|---|---|
| Internal noisy exon | Any wrong-sign internal z-score vetoed the exact multi-exon interval | Only both boundaries must support the direction; internal noise is permitted | New BRCA2 E12–E14 deletion regression with a wrong-sign E13 |
| Candidate family | Sorted and silently truncated to 50 | All boundary-qualified contiguous windows retained | Complete small-panel family; max-statistic remains downstream |
| Frozen partition | Rerun depended on a missing selection file | Archived 50/60/37 JSON accepted after exact membership, size, uniqueness, and overlap validation | Positive load test and duplicate-membership rejection test |
97/97native non-reference libraries completed
94/97presumed-negative libraries without final flags
35/97libraries with more proposals after the scan fix
3libraries with 4 total final flags; unchanged from the archived rerun
The candidate set increased in 35 libraries (range +1 to +23 candidates), yet the final four flags were unchanged. That A/B result is valuable: proposal loss existed, but normalization, residual modeling, corroboration, and calibration are the limiting layers now.
The observed clean fraction is 96.91% (exact 95% interval 91.23%–99.36%). These samples lack orthogonal negative confirmation and were previously examined; this cannot be reported as specificity.
Iteration ledger
| Iteration | Design | Internal result | Decision |
|---|---|---|---|
| Prior corrected caller | 516 derivatives, only 8 host libraries | sensitivity 52.3%; clean fraction 75.0% | Rejected: low sensitivity and heuristic score bypass |
| Beta-binomial, reused PoN null | 95 PoN libraries reused for LOO null; 4 dev hosts | sensitivity 55.4%; clean fraction 100.0% | Rejected as optimistic reuse bias |
| Beta-binomial, strict split | 50 fit / 60 calibration / 8 dev hosts; 4,704 spikes | sensitivity 7.7%; clean fraction 100.0% | Rejected: sensitivity 7.7% |
| Predictive residual | Nested 5×5 cross-fit; 147 hosts; 24,108 spikes | sensitivity 47.6%; clean fraction 96.6% | Rejected: single-DUP sensitivity 7.5% |
| Official ExomeDepth default | v1.1.16; 4 hosts; 544 spikes | sensitivity 6.6%; clean fraction 100.0% | Rejected: default HMM sensitivity 6.6% |
| Official ExomeDepth tuned | transition=0.01; development-tuned; 4 native hosts | sensitivity 43.0%; clean fraction 100.0% | Rejected: sensitivity 43.0%; negative n=4 |
| Patched LGASieve native rerun | 50 PoN fit; 97 non-reference development libraries | clean fraction 96.9% | Improved false-flag burden; still below ≥99% gate |
| Conditional OAS / conformal model | Nested 5×5; 147 outer-held hosts; 56,448 exact-factor events | overall 61.6%; single DUP 9.9% | Rejected: no high/high operating point; exploratory only |
| Stochastic conditional model | 903,168 trials; 147 hosts; 8 replicates | overall 35.9%; single DUP 10.2%; mosaic single DUP 0.9% | Rejected: poor DUP/mosaic sensitivity and fold transport |
Published evidence and truth data
Successful callers use much more normalization context than this slice provides
The strongest external comparison is the 2025 benchmark of 12 germline CNV callers on four real, MLPA-prevalidated targeted-panel datasets (495 samples, 231 CNVs). Its supplementary archive is preserved locally. At aggregate gene level, the best tools performed well—but they normalized using complete capture panels with many more regions of interest. Those numbers are architecture evidence, not performance claims for LGASieve.
| Caller | TP | FP | FN | Aggregate sensitivity | Why it does not transfer |
|---|---|---|---|---|---|
| CoNVaDING | 228 | 150 | 3 | 98.70% | High sensitivity, lower precision; full panel context |
| GATK-gCNV | 224 | 27 | 7 | 96.97% | Cohort latent model and thousands of intervals |
| DECoN | 220 | 62 | 11 | 95.24% | Optimized reference sets over the captured panel |
| CODEX2 | 219 | 93 | 12 | 94.81% | Latent-factor/batch structure estimated from broad coverage |
| ExomeDepth | 205 | 38 | 26 | 88.74% | Designed to build an aggregate reference from many exome/panel bins |
A recurring design rule: analyze all usable on-target and off-target regions to establish library scale and latent technical factors, then mask output to the reportable genes. panelcn.MOPS explicitly excludes the gene of interest when selecting correlated references. Restricting normalization itself to BRCA1/2 removes the very anchors needed for gene-wide events.
Public truth-set audit
| Resource | Truth | Access | Use here |
|---|---|---|---|
| ICR96 | 96 TruSight Cancer v2 samples; 66 positive, 30 negative; 68 validated CNVs (25 single-exon, 43 multi-exon; 51 DEL, 17 DUP); MLPA confirmation | FASTQ/BAM data under EGA controlled access EGAS00001002428 / EGAD00001003335 | Valuable external algorithm benchmark after authorization, but different capture assay and not clinical validation of LGASieve |
| panelcnDataset | 161 analyzed samples from a 170-sample TruSight Cancer collection; 41 CNVs (13 single-exon, 28 multi-exon; 36 DEL, 5 DUP) | EGA controlled access EGAS00001002481 / EGAD00001003400 | Useful for method transport research; not assay-matched and not openly downloadable |
| NIST/GIAB nstd175 | Genome-wide structural-variant benchmark regions/calls | Public truth; matching raw assay libraries still required | Useful to test truth matching and some large events, but not an exon-level BRCA capture truth set |
| Current 1000 Genomes count matrix | No orthogonal CNV-negative confirmation in supplied files | Local | Development residuals and false-flag burden only |
| ifCNV repository data | TSCA and Juno somatic tumor panels with gene-level aCGH labels and no exon breakpoint truth | Downloaded and preserved at commit cf11324af004c31624f0668d77c88a89bcbb4b77 | Rejected as a BRCA germline exon-level validation set; retained only for possible domain-shift research |
No openly accessible, assay-matched, orthogonally confirmed BRCA1/2 positive-and-negative FASTQ cohort was found that can turn this run into clinical validation. Downloading a different panel’s dataset can test software transport; it cannot validate this library preparation, capture design, laboratory process, or reportable range.
Recommended solution
Redesign the measurement and calibration domain, then simplify the caller
| Layer | Required change | Why | Release control |
|---|---|---|---|
| Assay inputs | Retain all capture-panel on-target bins, validated off-target bins, explicit invariant control probes, and/or molecule-count spike-ins | Restores absolute scale and latent technical context | Whole-gene calls are no-call until external anchors are validated |
| Metadata | Record site, run, flow cell, kit, lot, capture batch, read length, operator, extraction, input mass, and UMI metrics | Allows batch-aware reference selection and true transport tests | Out-of-domain batches no-call automatically |
| Reference cohort | Build an assay/batch-matched pool; estimate learning curves inside training data; current data suggest roughly 100 candidates before plateau, not a universal constant | Small reused PoNs make empirical tails coarse and unstable | No patient/replicate derivative may cross train/calibration/test |
| Normalization | Candidate-excluded hierarchical negative-binomial/beta-binomial latent-factor model with target, library, batch, GC, insert-size, duplicate, and capture effects | Prevents the candidate from defining its own baseline and models overdispersion directly | Version all masks, factors, and hyperparameters |
| Statistic | Joint DEL/DUP window likelihood over all callable runs; covariance-aware matched filter; no per-exon hard veto | Uses coherent multi-exon evidence without silently losing exact boundaries | One frozen family definition and one max-statistic calibration |
| Calibration | Batch-aware cross-conformal or held-out max-statistic calibration with at least ~300 verified negatives for a perfect-run lower bound near 99% | 60 calibrators cannot resolve a 1% two-direction family error reliably | Report confidence intervals and no-calls; never tune on the lockbox |
| Orthogonal evidence | BAF from validated heterozygous sites with site-specific beta-binomial noise; split/discordant reads, local assembly, and UMI evidence as secondary support | Independent channels can distinguish dosage from capture artifacts | Absence of breakpoint evidence must not veto exon-scale events |
| Reporting | Use NGS as a high-sensitivity screen; reflex every candidate to MLPA or an independently designed dosage assay | False positives are tolerable only when they cannot become final clinical calls | Single-probe/isolated exon findings require a second kit or different technique |
Pragmatic near-term product. A sensitivity-first screening layer with mandatory reflex confirmation is achievable before an autonomous caller. Its validated claim is “selects specimens for orthogonal dosage testing,” not “clinically confirms a BRCA CNV.”
Why this is more likely to work
ExomeDepth, DECoN, panelcn.MOPS, GATK-gCNV, CODEX2, CNVkit, SavvyCNV, ClearCNV, and Cobalt differ statistically, but the successful implementations share broad normalization context, reference/QC selection, explicit latent or count-noise modeling, and validation on physical truth. None supplies evidence that a BRCA-only 50-target median-scaled slice can identify whole-gene dosage autonomously.
Verification plan
A validation ladder that prevents simulation from becoming a clinical claim
- Unit and invariant tests. Freeze target ordering/mask, coordinate conventions, candidate exclusion, boundary behavior, joint family size, p-value monotonicity, duplicate handling, deterministic seeds, and atomic artifacts.
- Analytic limits. Re-run exact Poisson/negative-binomial power, compositional identifiability, calibration resolution, and reference learning curves for every assay revision.
- Count-level stochastic simulation. Use hierarchical negative-binomial libraries with fitted target/batch covariance, germline and mosaic DEL/DUP, heterogeneous exon amplitudes, GC/insert perturbations, missingness, contamination, and out-of-domain batches. Hold the host and all derivatives out.
- BAM/FASTQ simulation. Inject CNVs with Bamgineer, SECNVs, or an independently verified simulator before alignment; include pseudogene/homology and breakpoint placement. Compare expected and observed molecule/read-count shifts.
- Physical analytic specimens. Use reference materials, cell-line mixtures, engineered controls, and dilution series spanning single exon, multi-exon, whole gene, DEL, DUP, mosaic fraction, input mass, and depth. Include low-quality and interference conditions.
- Clinical truth cohort. Orthogonally confirmed positives and verified negatives across the reportable range; MLPA or independent dosage testing establishes truth. Do not label unflagged population samples as negatives.
- Repeatability and reproducibility. Replicate operators, instruments, sites, kits, lots, captures, runs, days, and bioinformatics deployments. Cluster intervals by patient and original library, not by synthetic derivative.
- Frozen prospective holdout. Lock code, container, model, masks, gates, and truth-matching rules; open once. If it fails, the set becomes development data and a new holdout is required.
Predeclared gates for an autonomous reportable caller
| Gate | Minimum | Denominator / interval rule | Current status |
|---|---|---|---|
| Analysis completion | ≥99% | All eligible physical specimens; no silent exclusions | 97/97 dev only |
| Overall positive agreement | ≥95%; one-sided patient-cluster lower bound ≥90% | Orthogonally confirmed positives, stratified by type/width | Not met internally |
| Single-exon DEL | ≥95% | Physical truth; sufficient n for interval, not repeated derivatives | 49.6% stochastic |
| Single-exon DUP | ≥90% | Physical truth; target- and depth-stratified | 10.2% stochastic |
| Multi-exon DEL and DUP | ≥95% each | Width, gene, breakpoint, and whole-gene strata | Not met in each stratum |
| Negative agreement | ≥99%; one-sided lower bound ≥95% | Verified negatives; patient/batch clustering; no-call included transparently | Not assessable |
| Reproducibility | ≥95% call and localization agreement | Lots/sites/operators/runs; predefined equivocal handling | Archived 12.5–25% |
| No-call rate | Predeclared by use case | Included in completion and clinical workflow impact | Must be designed |
How many negatives?
If every tested negative is clean, exact one-sided 95% lower bounds require at least 59 specimens to exceed 95%, 299 to exceed 99%, and 2995 to exceed 99.9%. These are ideal independent-specimen counts; batch clustering and any failure increase the requirement.
The AMP/API/CAP in-silico recommendations support simulation as a supplement and gap-filler, not as a replacement for physical clinical specimens. EMQN, ACMG, and CCMG guidance likewise requires validation across the reportable laboratory workflow and clear limitations.
Holdout governance
The lockbox remains sealed
SEALED_NOT_DOWNLOADED_NOT_ANALYZEDNo lockbox specimen was downloaded, inspected, scored, or used for tuning
75 specimenssame-project technical holdout; not a verified clinical cohort
The canonical specimen-list hash is 37f1939ddff65d89b969a8308cf089d28d517d062ea09b9f9a8d1d499bdc550a. Opening this set now would spend the only holdout on a method that already fails development gates. The correct action is to keep it sealed, build the broader assay/reference model, freeze the resulting pipeline, and open the lockbox once only if all development gates pass. Even a passing result would be technical evidence, not clinical validation, because the truth status is not verified.
Audit trail
Artifacts, hashes, and exact rerun commands
| Artifact | Bytes | SHA-256 |
|---|---|---|
| SOURCE_MANIFEST.json | 2,675 | 6b0c886cf24d9e293b617eed8277c96fd41895ec6f991f5225a2ed2b6b4a9e77 |
| lgasieve/segment.py | 13,701 | 54ec5b510986f9e5fe1a1ddd3a784e6ec261a60d3f61a4d303d1b39c5fdf482c |
| lgasieve/config.py | 16,593 | c552a1e2e0dd25972a2e8aec250d6bdbe8fb6e9a3008b99c9d4d115e84c98bfa |
| scripts/native_reanalysis.py | 9,131 | cc2579b9ea90927ccb470283759e5711efecfbaf94481dfc08adfd30db3a2921 |
| scripts/conditional_gaussian_experiment.py | 39,652 | ef3468e1bbd1036c0ad98d469c43a3a0f49429ef714461ca578d6c298fefd77a |
| scripts/stochastic_conditional_experiment.py | 38,485 | d28c2e9c94f3b6296a2451518701285e11e06c6b420877a6faa020114a637e17 |
| scripts/information_limit_audit.py | 12,582 | e5dc458dbcbf75ecfdf1114c8b94f7cbfb5983490c9a00495379ae7a3af288b4 |
| scripts/residual_structure_audit.py | 38,552 | 3a444996013feece91d2330d98c743eeeb6115900136fcbf3c022a50c7b2bbb5 |
| scripts/build_research_report.py | 78,291 | 6bc391a28a3fc33f3d44a220d67a1cc07e2f70c9037c7e4b60cd4f86ccfbab49 |
| tests/test_science_regressions.py | 13,622 | 348bf7e53bdc6552e11da4498db06e1a81e7394cc3c2b80a4efe2fe1855401e0 |
| tests/test_conditional_gaussian.py | 6,813 | 4df355440acce566379f6ce5279a3c067bcb36b60a04c1abda1c1b6800276d53 |
| tests/test_stochastic_conditional.py | 4,954 | 2a7be41359cc433059be4ca01501a681d66451104c0d5a65c635e17fac5a8fd6 |
| tests/test_information_limit_audit.py | 1,437 | c8752dd2b82d1442d5c207973794df472883a7f4543403d5dddc21f6a9de5ad7 |
| tests/test_residual_structure_audit.py | 6,024 | c02b0e39a13d01aaf22ef0d5e918e52019466aaffd909db62b79af6e03bf644c |
| analysis/native_reanalysis_complete_scan.json | 26,850 | 9d50c0738632143994c03632d2691ed054a70ea25dbdb274cbad5c8ae6c4a369 |
| analysis/conditional_gaussian_nested.json | 67,606 | adeb39cee448cfd4abe3e1840b7d0a8af78484c0cd42cf2ea10482ceadf9fceb |
| analysis/stochastic_conditional_nested.json | 612,956 | 3d8d4ed7aef2e8724f2a20780988f782dc18a735bc99133be81837ff9333e841 |
| analysis/information_limit_audit.json | 6,595 | 0e47fc7ccaf458d1b3502dc6b4e8cc79876d6dea6d6b8222683d7e3358e26c93 |
| analysis/residual_structure_audit.json | 40,973 | f1043dac1fb0bab13310bc8cd9157821f5ee822468cd87f421ac175cb96327c7 |
| PROJECT_STATUS.md | 7,753 | 53dac296f54f07c28472f3e7e5fc864c267f9e208be7f3aec005524c074c7cfa |
| validation_v2/lockbox_protocol.json | 2,856 | c275179159c412f0ada761934f21f3fd8c62dc0813010d35dea1e2154550a4df |
| validation_v2/revalidation_summary.json | 11,843 | ba33139b0751ae252745b6197ea701c8abcb61b9e56555bd6456fab80463b63f |
| external_data/archives/supplementary_data_bbae645.zip | 2,891,428 | 28daa22ff7f4cac99b05ec37df851f03e5ff7e5dc9454e79e3d10ca1d76f0cc1 |
Exact principal commands
python -m pytest -q
python -m scripts.native_reanalysis --partition-json work\exomedepth_development_v2\partition.json --output analysis\native_reanalysis_complete_scan.json
python -m scripts.conditional_gaussian_experiment --output analysis\conditional_gaussian_nested.json
python -m scripts.stochastic_conditional_experiment --output analysis\stochastic_conditional_nested.json
python -m scripts.information_limit_audit --output analysis\information_limit_audit.json
python -m scripts.residual_structure_audit --output analysis\residual_structure_audit.json
python -m scripts.build_research_report --test-summary "128 passed, 3 skipped, 1 warning, 7 subtests passed in 48.74s"
Reproducibility boundary. Source archives and key inputs are hashed. JSON artifacts embed source-script and input hashes where implemented. Timestamps and elapsed runtime can vary; scientific summaries, fixed-seed simulations, partitions, and hash-linked source should not.
Primary sources
Literature and standards consulted
All web sources accessed 2026-07-31. Links point to the paper, regulator/standards page, archive record, or manufacturer documentation rather than a search result.
- Munté et al. (2025), systematic comparison of 12 germline CNV callers across four real targeted-panel datasets.
- Plagnol et al. (2012), ExomeDepth: read-depth CNV calling with an optimized aggregate reference.
- Fowler et al. (2016/2017), DECoN development and validation.
- Povysil et al. (2017), panelcn.MOPS for targeted panels.
- Babadi et al. (2023), GATK-gCNV scalable probabilistic CNV calling in 7,962 exomes.
- Talevich et al. (2016), CNVkit and the use of on- and off-target reads.
- Roca et al. (2019), Atlas-CNV targeted-panel validation.
- Laver et al. (2022), SavvyCNV and off-target coverage.
- Jiang et al. (2018), CODEX2 and latent-factor/batch modeling.
- Hartmann et al. (2022), ClearCNV.
- Bergmann et al. (2022), Cobalt clinical exome CNV caller.
- Bamgineer: controlled CNV/allele-specific BAM simulation.
- SECNVs: simulation of CNVs in exome-sequencing data.
- AMP/API/CAP recommendations for in-silico approaches in clinical NGS validation.
- EMQN best-practice guidelines for hereditary breast and ovarian cancer testing.
- ACMG technical standard for BRCA1/BRCA2 testing.
- CCMG laboratory guideline: end-to-end validation and quality assurance for NGS.
- Puget et al., BRCA1 pseudogene homology and assay-design risk.
- MRC Holland SALSA MLPA P002 BRCA1/BRCA2 product guidance.
- ICR96 hereditary cancer panel CNV benchmark and MLPA truth set.
- European Genome-phenome Archive study EGAS00001002428 (ICR96 reads; controlled access).
- European Genome-phenome Archive dataset EGAD00001003335 (ICR96; DAC approval required).
- European Genome-phenome Archive dataset EGAD00001003400 (panelcnDataset; DAC approval required).
- NIST/GIAB structural-variant benchmark in dbVar.
Bottom line. Literature supports a broader normalization domain, candidate-excluded reference modeling, joint calibration, orthogonal confirmation, and physical validation. It does not support promoting the present BRCA-only count slice by choosing a more permissive threshold.