Open tools for esophageal cancer research.

Model improvement: nested selection experiment

We specified five logistic-regression candidates before running: 25, 100 or 500 selected probes and three regularization strengths. Each outer test fold stayed out of feature selection, model selection and threshold selection. Inner grouped folds chose the candidate by specificity at the 98% sensitivity target, then AUROC; no candidate was chosen using its outer test score.

Three fewer false positives; no sensitivity improvement. Sensitivity remained 91.3%, and pooled AUROC fell. This is a mixed development result. The baseline exactly reproduced its earlier metrics and test partitions. We retained all candidate diagnostics, but did not select the apparent best candidate after examining outer test results.

The same 45 patients have informed earlier experiments, so even nested testing is not independent confirmation. No candidate was promoted to the frozen release. The 98% sensitivity goal remains unmet, and clinical review still needs to determine an acceptable false-positive burden.

Are different models missing different cancers?

No. All 5 candidates missed the same 2 cancer specimens; the other 21 were caught by every candidate. Source stage labels for the two misses are II and IV. These small counts cannot establish stage-specific sensitivity, and the labels have not been independently adjudicated.

The audit checked specimen order, labels, fold membership, candidate scores and each training-selected cutoff. All recorded decisions matched those cutoffs. Repeated errors across related models are correlated evidence, not independent confirmations. We have not excluded the missed cases, relabeled them or adjusted thresholds to recover them.

This experiment gives no sensitivity benefit from changing regularization or the selected probe count within the tested range. It does not prove that another model cannot help. The next model proposal needs a distinct rationale and a fixed evaluation plan; improving the score on these repeatedly examined cases alone would not demonstrate progress on new patients.

How stable are the selected probes?

We reconstructed the five outer training fits and verified their held-out scores and decisions against the saved experiment. The fixed 100-probe model used 208 different probes across the folds, with 44 appearing in every fold.

Jaccard measures shared features divided by the union of the two feature sets. The adaptive procedure uses fewer probes in one fold, which affects its overlap. These training sets overlap, and correlated probes can substitute for one another; selection frequency is neither biomarker confidence nor independent replication.

We have not built a consensus panel from all folds and tested it on those same folds—that would let held-out information influence feature selection. This audit changes no prediction or threshold. Any future stability-based selection must happen inside training data.

A different representation: genes instead of probes

We mapped 16,190 probes to 13,882 HGNC genes and calculated each gene's median within each specimen. Ambiguous or unresolved identifiers were excluded before modeling. The fixed classifier selected 100 genes inside each training fold and chose its threshold using inner training folds only.

The gene model flagged 21/23 cancers and 14/22 controls; sensitivity was 91.3%, specificity 36.4% and internal AUROC 0.955. The matched probe baseline flagged 21 cancers and 15 controls. One fewer false positive, unchanged sensitivity and slightly lower AUROC do not establish a better model. These are reused development patients; no model was promoted.

Pathway-level expression

A fixed model using 1,395 Reactome pathways caught 22/23 cancers and flagged 11/22 Barrett controls: sensitivity 95.7%, specificity 50.0%, internal AUROC 0.966. Each feature averages member genes' ranks within a specimen. Pathways were included by fixed size and measurement coverage rules, without selecting against cancer labels.

This missed one fewer cancer and produced four fewer false positives than the all-gene rank comparator on the same folds. The 98% target remains unmet; this is development, not independent validation. These patients have been reused across experiments. No model was promoted. Pathway scores do not establish pathway activation or treatment targets. Independent donor identity, control eligibility and assay compatibility remain the next evidence requirements.

Relative gene expression: all-gene ranks

We tested each gene's rank within a specimen across all 13,882 mapped genes, with no feature selector. The fixed model caught 21/23 cancers and flagged 15/22 controls: sensitivity 91.3%, specificity 31.8%, internal AUROC 0.958. Thresholds were set inside training folds. It missed two cancers and produced one more false positive than the gene-median model. This does not support replacing the model or meet the 98% target; these remain development results.

Did training-only stability selection help?

We tested a fixed stability procedure that rebuilt probe selection inside every training fit, including the inner fits used to select thresholds. Five grouped subsamples ranked probes by selection frequency and mean ANOVA rank; the classifier used 100 probes and fixed regularization C=0.1. The shared probe set from the earlier descriptive audit was not an input.

This experiment did not improve detection. It missed two cancers and added one false positive compared with the matched classifier. We retained the negative result and did not replace the model release. No settings were searched after the result. Independent validation remains outstanding.

First trained model: GSE28302

We joined expression measurements with source patient identities and tissue labels: 23 cancer and 22 nondysplastic Barrett specimens. Nine normal squamous specimens were excluded. A logistic regression selects 100 probes inside each training fold.

This baseline flags almost everyone. It is not a useful detector yet. Observing 100% sensitivity in this small development cohort does not demonstrate our 98% sensitivity goal in an independent population.

GSE28302 is now development data and will not be reused as an untouched test set. Only one cancer was labeled stage I; age, sex, site effects and the deposited measurement scale limit interpretation. Source preprocessing may already involve all samples. No other study was pooled into this model.

Next: resolve raw-data preprocessing, compare demographic and expression baselines, and evaluate a prespecified gene-level model across compatible studies. Clinical review will determine acceptable false positives. Source dataset · Source paper

Age and sex comparison: original deposited inputs

All three models used the same 45 patients and outer test splits, with thresholds chosen inside training folds. Age and sex alone ranked higher by AUROC. Adding expression did not show an advantage in this experiment; the difference is descriptive, not a significance test or proof that expression lacks value.

The demographic model missed one cancer. Both expression models flagged nearly every control. This reused development cohort cannot establish clinical performance. The following experiment examines source preprocessing before any increase in model complexity or pooling of studies.

Rebuilding from non-normalized intensities

We mapped all 54 specimens by source barcode and retained the same 25,037 probes. For this fixed follow-up, we applied log2 and subtracted each specimen’s own median. This uses no cross-patient normalization and is an alternative to the deposited processing, not an exact reconstruction of the authors’ Lowess procedure.

The alternative improved ranking in this development experiment, but missed 2 cancers. These decisions use training-only cutoffs on identical outer patient splits: sensitivity 91.3%, specificity 54.5%. The 98% sensitivity target remains unmet.

The final release cutoff was selected afterward from pooled development predictions; its apparent result catches 23/23 cancers but falsely flags 17/22 controls. That is a separate, optimistic calculation—not independent performance at a validated threshold.

The deposited numeric scale remains unresolved. The probe set was inherited from the source analysis, instrument-level processing remains as supplied, and age, sex, site and stage limitations persist. We will retain both results; no external validation or clinical claim is established. Original assay export and source record

Holding barcode groups out together

The 45 development specimens span 20 barcode-prefix groups; 15 contain both cancer and Barrett specimens. We kept each group together in both the outer test folds and the inner folds used to select thresholds. These prefixes are a putative chip grouping, not confirmed laboratory batches or collection sites.

Ranking performance remained similar, but the grouped test missed 2 cancers and had more false positives. Sensitivity was 91.3% and specificity 31.8%. This does not establish batch independence, external validity or the 98% sensitivity target.

Assay quality audit: the median fraction of retained probes with detection P below 0.01 was 38.8% for cancer specimens and 36.6% for Barrett specimens. Individual specimens ranged from 16.2% to 42.4%. No retained measurement had fewer than three beads. These are descriptive measurements, not pass/fail criteria; no specimens or probes were removed using these metrics.

Next: review the low-detection specimens and obtain verified batch/site metadata. Any quality filtering must be specified and evaluated as a new development experiment, not used to erase an unfavorable result.

Age and sex comparison: alternative expression inputs

With log2 and within-specimen median preprocessing, expression ranked higher than age and sex alone under both split designs. All models share the same outer and inner partitions within each design. Cutoffs were selected only inside training data; grouped folds keep barcode prefixes apart at both levels.

Adding age and sex to expression reduced false positives from 15 to 13 in the grouped test, but still missed two cancers. In patient folds, it added one false positive. These small differences do not establish added clinical value, and no model was promoted. The expression-only results reproduce the earlier experiments.

The 98% sensitivity goal remains unmet. These are repeated experiments on the same 45 development patients. Independent evaluation, verified site and processing metadata, and assessment of early-stage disease remain outstanding.

Are errors concentrated in low-detection specimens?

We joined assay quality to the existing barcode-group predictions without retraining or changing thresholds. The lowest quarter of detection fractions contained 12 specimens and 2 prediction errors. The other 33 specimens contained 15 errors. These groups have different cancer/Barrett proportions and are descriptive, not a controlled comparison.

The two missed cancers had detection fractions between 36.8% and 38.9%, both above the first-quartile boundary of 35.9%. Low detection alone does not explain these misses. No specimen or probe was removed. Detection P measures assay signal, not the probability of cancer; technical and biological causes remain unresolved.

What do the model’s probes measure?

A probe is an assay feature designed to measure a particular nucleic-acid sequence. Probe-to-gene annotation links that feature’s identifier to the gene or transcript listed by its source platform. It is a mapping step, not evidence that a gene detects cancer.

We matched all 100 targets in the saved development model to the original platform annotation using exact identifiers. Each has one historical symbol, producing 100 distinct symbols. The audit retains source probe IDs and transcript accessions; no conflicting row was silently selected.

The current-identifier review below follows this historical audit. Unique sequence alignment remains unchecked. This audit covers the final development model, not how consistently each feature was selected across folds. Matching names across platforms does not make the model transferable. No external expression data were opened, and model predictions are unchanged.

Current identifiers and candidate assay coverage

Using a saved HGNC download, we resolved 93 of the 100 historical names to one approved gene identifier. 3 names have conflicting matches and 4 remain unmatched. Approved names, previous names and aliases are checked together; ambiguous names are not silently assigned.

Of those resolved targets, 91 have unambiguous symbol-level coverage on GPL21185, the candidate Agilent platform. 2 resolved targets lack such coverage. Platform controls were excluded, and all candidate probe matches were retained for review.

Annotation coverage is not model compatibility. Probe sequence, transcript coverage and measurement processing still need review. We cannot substitute these probes into the frozen model. No candidate expression measurements were accessed and no performance result changed.

Actual probe sequences: ADCY3 and MCM2

We checked the original 50-base probes and candidate 60-base probes against both the historical and current returned RefSeq records. Every probe matched exactly once within each corresponding transcript, on the forward strand. The source and candidate probes occupy separate, non-overlapping regions.

Coordinates are one-based and include both endpoints. This is exact matching within four specified transcript records, not a genome-wide specificity check. Sharing a transcript does not establish equivalent hybridization, calibration or expression measurements. No model coefficients or thresholds changed, and candidate patient expression remains untouched.

Goal: independently test cancer detection

Our current endpoint is cancer versus Barrett tissue classification, targeting 98% sensitivity while reporting false positives. This does not establish population screening, early detection, or treatment benefit.

The barcode-group expression and combined models each flagged 21 of 23 cancers (91.3%). With 23 cancer cases, even one miss puts observed sensitivity below 98%. Catching all 23 would still not establish performance in new patients. These cases remain development data; we will not tune against them and call the result independent validation.

Candidate cohorts and unresolved evidence

Source	Evaluation eligibility	Evidence gaps and limitations
GSE28302	excluded due to reuse	Label discrepancy resolved using explicit tissue characteristics; 22 Barrett samples are nondysplastic.; Audit patient and specimen overlap; EAC biopsies overlap GSE10506 copy-number data.; Check assay compatibility, collection sites, treatment and stage metadata before defining inclusion.; The paper also combines prior studies; distinguish this deposited cohort from reused cohorts.; Only one stage-I cancer specimen; poor precision for an early-stage sensitivity claim.
GSE1420	requires evidence review	Verify specimen-to-patient mapping and assay metadata.; Barrett tissue from cancer patients does not represent cancer-free screening controls.; Cross-study independence remains unverified; direct page access returned a browser check.
GSE94869	requires evidence review	Non-malignant is not an established Barrett control label.; RNA-seq requires a prespecified assay-transfer and preprocessing strategy.; Verify patient mapping, overlap and source methods; direct page access returned a browser check.
GSE37203	excluded due to reuse	A new accession for a parent series does not create an independent dataset.
GSE299535	requires evidence review	No explicit patient IDs in characteristic fields; repeated title prefixes suggest pairing but are not verified identities.; Cancer-free status of Barrett controls is not established; potentially tissue from cancer-bearing patients.; Different Agilent assay; processing header identifies Feature Extraction but does not establish a compatible normalization protocol.; Early-cancer grade/stage and overlap need confirmation; only 13 cancer specimens and no female specimens.; No linked publication in GEO at review.

No candidate is approved merely because it appears here. A completed protocol form also requires scientific review. Cohort descriptions reflect the recorded source review, not a live registry search.

32 of 48 measurable genes agree across all three comparisons. These are exploratory directions, not detection accuracy. The two GSE13898 analyses reuse specimens and are not independent replications. Two of the original 50 genes lacked unique mappings in that cohort.

Discovery source · First comparison source · Third cohort source

New data under review

A metadata check found all 47 sample titles match between the long-read plasma study GSE271389 and short-read GSE271390. Different BioSample identifiers do not prove different patients. We will not count these as 94 independent patients or use one accession as independent validation of the other without a verified donor mapping.

GSE271390 contains 47 plasma RNA samples: 19 adenocarcinoma, 12 Barrett and 16 controls. This could support a separate blood-based detection study. Patient identities, stage, Barrett dysplasia grade, control eligibility and collection batches still need review. It is not an approved independent test set, and a tissue model cannot be applied directly to plasma.

GSE210647 studies subsequent progression from nondysplastic Barrett to HGD or cancer, a different outcome from detecting cancer already present. Only metadata has been reviewed for these candidates; no expression rows were parsed or used to choose genes.

Transcript evidence for the nine annotation gaps

We compared source transcript accession families with the saved HGNC table and candidate platform annotations. Four targets have a single accession-supported gene identity; five lack an HGNC accession match in this snapshot. Two targets have a candidate-platform probe listing the same accession family.

MCM2 and SPP1 accession evidence supports one gene among their conflicting symbol matches. H2AC18 and CCL4 identities gain accession support, but the candidate platform still lacks an exact single-field accession-family match under this rule. ADCY3 has a platform accession-family match but no match in the saved HGNC accession field.

Version suffixes were removed only for accession-family comparison. This does not verify exact sequence versions or equivalent probe binding. Blank, composite and conflicting platform accession fields were not guessed. These are separate evidence findings, not nine resolved transfer gaps; the model and candidate expression data remain unchanged.

Published-analysis reproduction

Our reconstruction of the DEG-overlap step in Yi et al. (2023) found 21 shared upregulated and 84 downregulated genes, compared with the published 27 and 104. The counts were not reproduced. We used deposited matrices, historical annotations and mean probe aggregation; the reviewed methods do not specify enough batch-correction detail for an exact reconstruction. This is a partial reproduction attempt, not a detection result.

Supplement audit: both published gene tables contain 25 numeric entries where gene labels are expected. We preserved these as unresolved labels. The remaining source symbols overlap our historical annotation in 19,832 cases, with differences still under review. These WGCNA tables do not supply the DEG list or executable batch-correction steps, so the reproduction discrepancy remains unresolved.

New cohort under metadata review: GSE299535 includes 13 early-cancer and 11 Barrett specimens. Patient linkage, comparator suitability and assay compatibility remain unresolved. No expression measurements from this candidate have been analyzed.