Corpus 1,272 assessed · 1,173 scored · 643 reproduced ≥75 · 168 flagged ·∅ 74.1/100
← New search

Machine learning-based DNA microarray analysis for disease detection using the MICRO-AI framework.

Sci Prog · 2026
L2 22/100 PQI 68
⚑ Flagged for review — a reproduced result did not match the reported value

Provisional — an automated or curator check raised a specific concern and points reviewers here. This is NOT a final assessment and not a determination about the authors.

Why this verdict

The main result did not reproduce in this reproduction attempt. Where our recomputation produced values that differ from the published ones, those discrepancies are listed below. This is a single automated attempt — not peer review and not a finding of error or misconduct — and differences can also arise from data access, undocumented parameters or the computing environment. The verdict can be contested via “report an error”.

Reproduced on the brainbox compute brainarbeit.com
Scoring basis — itemised

Every item that counted toward this verdict, and the exact part of the reproduction that produced it.

Main result did not reproduce
Decisive
From: Q5 · Derivability / plausibility 🔴
Main result did not reproduce
Decisive
From: Q7 · Core claim 🔴
Main result did not reproduce
Decisive
From: Q8 · Severity of the miss (overall human judgment) 🔴
✓ What held up
  • Same input data as the authors
  • Reported values were directly comparable
What did not (or only partly)
  • 🔴A deviation arose in the data or preprocessing
  • 🔴A deviation was attributed to the published material
  • 🔴Reported values were not (fully) derivable from the shared data
  • 🔴The deviation was non-trivial in magnitude
  • 🔴The central claim did not (fully) hold under reproduction
  • 🔴Overall, the reproduction showed a material discrepancy
How its reproducibility compares
22/100
Reproducibility score
3.0 SD below mean
vs. all fields · 1173 studies
🎯 Scores higher than 1% of all assessed papers rank 1165 of 1173 scored

A 0–100 reproducibility-quality score from the per-question grades, shown as a z-score: standard deviations above (+) or below (−) the mean of comparable assessments.

Reproduction agent’s raw note

MISMATCH with high fabrication concern, confirmed on TWO datasets with COMPLETE 5-seed x 5-fold leakage-free CV runs of the AUTHORS' OWN micro_ai.py (commit ad27cab) on the REAL GEO data. Dataset dimensions match exactly for both (C1,C9). Every performance claim fails by huge margins: GSE2034 Acc 72.7% vs 95.8%, AUC 0.691 vs 0.978, F1 0.481 vs 0.955, MCC 0.301 vs 0.914; GSE7390 Acc 63.1% vs 97.5%, AUC 0.573 vs 0.986, F1 0.285 vs 0.973, MCC 0.052 vs 0.949. Both reproduced accuracies sit AT/BELOW the majority-class baseline (75.9% / 68.7%); GSE7390 per-seed AUC reaches 0.52 and per-seed MCC is negative (worse than chance). Two independent fabrication indicators beyond performance: (a) Table 6 lists MAD-filter survivors 8,456/7,892 but the authors' own MAD filter passes 16,712 for BOTH datasets; (b) the reported fixed 127/98 selected genes are not produced -- the shipped RFECV varies 127-5000 / 98-4500 per fold. The repo ships NO script reproducing any table (main runs on np.random synthetic data). Reported 0.978/0.986 AUC exceed every published result on these well-studied cohorts (literature ceiling ~0.65-0.70). Labels: GSE2034 'bone relapses' (only public binary phenotype), GSE7390 'e.dmfs' distant-metastasis event -- both honest public prognosis labels; conclusion robust to label choice. NOT attempted (80/20): TCGA-BRCA, GSE62254, E-MTAB-365, GSE9891, Table-7 external baselines. All grades AUTOMATED and PROVISIONAL; a human reviewer decides.

💻 Code ↗ 🗄 Data: GSE2034

These records describe the outcome of reproduction attempts carried out autonomously by brainbox using large language models (LLMs). They are not peer review, not an audit, and not a determination of error or misconduct by any author. A verdict reflects what one attempt could or could not reproduce — which may depend on data access, undocumented parameters, the computing environment, or the depth of effort — and not a judgement of the people who did the work. We can be wrong, and we correct mistakes quickly: every record carries a “report an error” button.

Assessment versions

Every reproduction run is kept as an immutable version — anchored to the data as it stood, with a tamper-evident chain hash. A rerun (e.g. after an author updates a deposit) adds a new version; the previous one stays on record.

  1. v1 current initial assessment Score 22
    assessed: 2026-06-16 ⛓ 043d97a33831
✎ I am an author of this paper

Updated or fixed a deposit, or is there an erratum? Ask us to re-run the metrics. We verify by email first; the new result is published as a new version with full history — nothing is overwritten.

Reason for the rerun

We email you a confirmation link first. The rerun is an objective re-measurement — it cannot change the verdict in your favour, only ask us to look again.

Provenance — full disclosure

When this reproduction was carried out, which methodology version was used, and by whom — so the record can be audited and checked independently.

Reproduced
2026-06-22
Rubric version
v1.0
Assessed by
👤 1 human curator(s) · Level L2 2026-06-16
Last updated
2026-08-05

Provisional, curator- or AI-assessed, and independently checkable. A reproduction outcome states what one attempt could reproduce — not a judgement of the authors.

Deep full-text extraction

Model: sonnet
Founding hypothesis

The paper investigates whether an integrated, end-to-end machine learning framework (MICRO-AI) combining attention-weighted feature selection with heterogeneous ensemble classification can overcome the high dimensionality, small sample size, batch effects and class imbalance that limit DNA microarray-based disease diagnosis, achieving higher and more clinically deployable diagnostic accuracy than existing single-component or deep learning approaches.

Core claims
  • MICRO-AI's attention-weighted feature fusion reduces dimensionality by over 99% (from ~20,000 to ~127 genes) without loss of biological significance finding
  • Heterogeneous ensemble (GBM+RF+SVM) with SLSQP-based adaptive weight optimisation achieves 5-7% higher accuracy than individual classifiers and 0.8% improvement over equal weighting finding
  • The integrated preprocessing pipeline (quantile normalisation, ComBat correction, weighted KNN imputation) contributes a 1.7% accuracy improvement over uncorrected data finding
  • MICRO-AI outperforms 10 state-of-the-art methods by 1.2-7.5% accuracy across six benchmark datasets from GEO, TCGA and ArrayExpress finding
  • MICRO-AI trains 2.4-6.0x faster than deep learning alternatives with an average training time of 52.3 s finding
  • MICRO-AI (Microarray Classification and Recognition using Artificial Intelligence) is a modular four-module framework: preprocessing, attention-weighted feature selection, ensemble classification and clinical integration resource
  • Attention scores for gene prioritisation are computed from deterministic statistical moment embeddings (mean, variance, skewness, kurtosis) and mutual information with class labels, not end-to-end gradient learning method
  • The framework produces calibrated probability estimates via isotonic regression with expected calibration error minimisation for clinical decision support mechanism
Experimental setups
Assay System Perturbation Readout Platform
DNA microarray gene expression profiling / classification Human samples: breast cancer, gastric cancer, ovarian cancer, leukaemia (6 datasets from GEO, TCGA, ArrayExpress) none (disease diagnosis/classification, no experimental perturbation) disease class prediction / classification accuracy, sensitivity, specificity, F1, MCC, AUC-ROC
Key results
  • Average classification accuracy across six datasets 96.8%
  • Average sensitivity 95.2%
  • Average specificity 97.4%
  • F1-score and Matthews correlation coefficient F1=96.0%, MCC=0.928
  • Area under ROC curve 0.983
  • Dimensionality reduction from feature selection >99% (~20,000 to ~127 genes)
  • Accuracy gain from adaptive ensemble weighting vs individual classifiers / equal weighting 5-7% / 0.8%
  • Accuracy improvement from preprocessing pipeline vs uncorrected data 1.7%
Key statistics
  • mean 96.8% accuracy (average classification accuracy across six benchmark datasets)
  • mean 95.2% sensitivity, 97.4% specificity, 96.0% F1-score (average diagnostic performance metrics)
  • other MCC = 0.928 (Matthews correlation coefficient, average across datasets)
  • other AUC-ROC = 0.983 (average area under ROC curve)
  • fold_change 1.2-7.5% higher accuracy (comparison vs 10 state-of-the-art methods)
  • fold_change 2.4-6.0x faster execution (training time vs deep learning alternatives (avg 52.3 s))
  • other >99% dimensionality reduction (~20,000 to ~127 genes) (attention-weighted feature selection result)
  • count 2321 total samples across 6 datasets, 3 repositories, 4 cancer types (overall validation cohort size)

Statistical methods review

Model: sonnet

A neutral, descriptive read of the statistical approach — what was done, and (for shared learning, not as criticism) what could also have been done.

MICRO-AI is a machine learning classification framework evaluated entirely through classification performance metrics (accuracy, sensitivity, specificity, F1-score, MCC, AUC-ROC) on six pre-existing benchmark genomic datasets totalling 2321 samples across four cancer types. No traditional hypothesis tests are reported; comparative benchmarking against ten published methods is conducted by direct point-estimate metric comparison. Results are presented as averages across datasets without measures of dispersion, confidence intervals, or statistical significance assessment of cross-method differences.

Replicationunclear Sample sizeSix benchmark datasets drawn from GEO, TCGA, and ArrayExpress; per-dataset sample sizes stated as 155–1097, 2321 total; no formal power analysis or sample-size justification mentioned GroupsCancer subtypes vs. controls or among cancer types (breast, gastric, ovarian, leukaemia) across six independent datasets Pairingna Randomization/blindingnot stated Dispersionnone Exact p-valuesno Confidence intervalsno Multiplicity correctionnone stated
Statistical tests used
Test Applied to n Assumptions
Recursive feature elimination with cross-validation (RFECV) with attention-weighted scoring Gene/feature subset identification across all six datasets 155–1097 per dataset; 2321 total across six datasets not stated
Gene Ontology (GO) enrichment analysis Biological validation of selected gene signatures (mentioned in abstract contributions) not stated
Cross-validation-based classification performance evaluation (fold count not specified in available text) Primary model evaluation for all six datasets across GBM, RF, SVM ensemble 155–1097 per dataset not stated
Approaches that could also have been used
  • Cross-method performance comparisons report point-estimate accuracy differences (1.2–7.5% improvement over ten methods) without any significance testing
    Could also: A Wilcoxon signed-rank test or McNemar's test applied to per-dataset paired performance scores could also assess whether observed differences are statistically distinguishable from chance variation — Significance tests on paired per-dataset results distinguish systematic gains from dataset-specific fluctuation, which is especially informative when only six datasets are available
  • Classification metrics are reported as means across six datasets with no measure of spread
    Could also: Reporting per-dataset results alongside their mean ± SD, or computing bootstrap confidence intervals on the aggregate metrics, could also convey variability — Dispersion measures help readers assess consistency: a high mean with wide variance across datasets signals different generalisability than a high mean with narrow variance
  • Ten simultaneous cross-method comparisons are made without adjusting for the number of comparisons
    Could also: A Friedman test followed by post-hoc Nemenyi or Holm-corrected Wilcoxon pairwise tests would also control the family-wise error rate across the full comparison set — Controlling for multiple simultaneous comparisons reduces the probability that at least one apparent superiority claim arises by chance when ten methods are evaluated together
  • Feature selection uses RFECV; the stability of the selected gene set across cross-validation folds is not reported
    Could also: Stability selection (Meinshausen & Bühlmann, 2010) or bootstrap-based selection frequency scores could also quantify how consistently each gene is chosen across resampling iterations — Selection-frequency metrics complement the dimensionality-reduction result by indicating whether the ∼127 selected genes form a reproducible signature or vary substantially across folds
  • The nesting structure of cross-validation (outer evaluation loop vs. inner hyperparameter/feature-selection loop) is not specified in the available methodology text
    Could also: Nested cross-validation — with a separate inner loop for feature selection and hyperparameter tuning and an outer loop for unbiased performance estimation — is also a standard approach when selection and tuning are coupled — Explicitly nested CV prevents optimistic bias that can arise when the same fold structure is used for both model selection and performance reporting
  • Model calibration is addressed via isotonic regression with ECE minimisation, but no calibration visualisation (e.g., reliability diagrams) or per-dataset calibration scores are mentioned in the available text
    Could also: Reliability diagrams paired with the Brier score or Expected Calibration Error reported per dataset could also illustrate and quantify calibration quality — Visual calibration curves identify probability regions where the model is systematically over- or under-confident, which is directly relevant for clinical threshold-based decision support
Software: ComBat (empirical Bayes batch correction) · SLSQP — Sequential Least Squares Programming (ensemble weight optimisation) · Isotonic regression (probability calibration module)

Citation network

Where this publication sits in the reproducibility-weighted citation graph — what it is built on, and what is built on it. Citation data from OpenAlex.

Authors · 1
Citations
0
Impact: low
Foundation confidence
None of its references are in our reproducibility record yet — its foundation cannot be assessed.
Topics

No assessed neighbours yet — the network grows as more papers are assessed.

Data lineage

The datasets this paper uses (text-mined from the full text via Europe PMC), and which other assessed papers stand on the same data. A shared dataset is a factual link — not a judgement.

GSE7390 GEO in Acknowledgments (http://purl.org/orb/Acknowledgments)
also used by 2 papers:
GSE2034 GEO in Acknowledgments (http://purl.org/orb/Acknowledgments)
also used by 1 paper:
GSE62254 GEO in Acknowledgments (http://purl.org/orb/Acknowledgments)
also used by 1 paper:
GSE9891 GEO in Acknowledgments (http://purl.org/orb/Acknowledgments)
also used by 1 paper:
E-MTAB-365 ArrayExpress in Acknowledgments (http://purl.org/orb/Acknowledgments)
no other assessed paper uses this yet
GSE20347 GEO in Discussion (http://purl.org/orb/Discussion)
no other assessed paper uses this yet

Downstream reach in the literature

465 downstream papers · 6 datasets

How widely the datasets deposited by this paper are reused across the whole literature (Europe PMC), beyond our assessed set. This is a factual dependency map — reusing a public dataset is normal, good science. It is not a judgement on the downstream papers; the only verdict here is this paper's own, with its cited rationale.

This paper is currently under reproducibility review (see the verdict above). The map below shows where the data in question has propagated — so reuse can be traced, not so the downstream work is presumed affected.

What was reproduced

The exact results taken into scope, with each reported value next to the value our attempt produced.

Scope — pmid-41925147 (MICRO-AI)

Paper: Othman MA. "Machine learning-based DNA microarray analysis for disease detection using the MICRO-AI framework." Sci Prog 2026. PMID 41925147 / PMC13051084 / DOI 10.1177/00368504261436834.

Code: https://github.com/Manal-oth/MICRO-AI (commit ad27cab01ab83695980b7bed62459ef325508d7a, pushed 2026-02-21). One file micro_ai.py (1080 lines) — a self-contained scikit-learn-style framework class MICROAI (quantile-norm → ComBat → KNN-impute → attention feature selection → RFECV → adaptive ensemble of GB/RF/SVM with isotonic calibration). README is one sentence. NO data loader and NO driver that reproduces the paper's tables: the __main__ block only runs the class on synthetic random data (np.random.randn). License: none declared.

Data: GEO GSE2034 (Wang et al. 2005) — 286 lymph-node-negative primary breast cancer patients, Affymetrix HG-U133A (GPL96), 22,283 probes. Binary label relapse (1=True): 106 relapse (distant metastasis) vs 180 relapse-free. Publicly downloadable series matrix. This is the dataset the RU brief points to.

In scope (pipeline-derived → attempted)

The paper evaluates MICRO-AI on six datasets (Table 5). We reproduce the GSE2034 row — the brief's named dataset — by running the authors' own micro_ai.py on the real GSE2034 data with the described protocol (5-fold stratified CV).

id reported (Table) value
C1 dataset dims (Table 2) GSE2034 = 286 samples × 22,283 probes, U133A, 2 classes
C2 Accuracy (Table 5) 95.8%
C3 AUC (Table 5) 0.978
C4 F1 (Table 5) 95.5%
C5 MCC (Table 5) 0.914
C6 Sensitivity/Specificity (Table 5) 94.2% / 97.1%
C7 # selected genes (Table 6) 127 genes (99.4% reduction)

Protocol (Experimental Setup): "5-fold stratified cross-validation with inner 3-fold CV for hyperparameter tuning. Performance averaged over 10 random seeds."

Out of scope (not attempted, with reason)

  • Other 5 datasets (GSE7390, TCGA-BRCA, GSE62254, E-MTAB-365, GSE9891): 80/20 — one clean, fully-specified dataset (GSE2034) is sufficient for a 1:1 honesty check; the framework and protocol are identical across rows, so GSE2034 is representative.
  • Table 7 comparison methods (SVM-RFE, XGBoost, GeneFormer, scBERT, …): these are external baselines, not part of the shipped MICRO-AI code; reproducing 10 other tools is out of scope and not the paper's own contribution.
  • Training-time column (52.3 s): hardware-dependent, not a scientific result.

Reproduction note / fabrication watch

GSE2034 distant-metastasis prediction is a notoriously hard prognosis task: the published literature (incl. the 76-gene Wang signature and many follow-ups) reports AUCs ≈ 0.6–0.7 and accuracies ≈ 0.65–0.70 on this exact cohort. A reported 95.8% accuracy / 0.978 AUC is far above anything in the field and is the kind of value that warrants a careful 1:1 check. We run the authors' own code on the authors' own data with an honest, leakage-free outer CV and report whatever it gives.

Figures / tables: Table
C1
Reported
GSE2034 = 286 x 22,283 probes, GPL96, 2 classes (Table 2)
Reproduced
286 x 22,283, GPL96, 2 classes (label=bone relapses 69/217)
exact
C2
Reported
GSE2034 Accuracy 95.8% (Table 5)
Reproduced
72.7% (mean 5 seeds, sd 1.2)
did not match
C3
Reported
GSE2034 AUC-ROC 0.978 (Table 5)
Reproduced
0.691 (sd 0.019)
did not match
C4
Reported
GSE2034 F1 95.5% (Table 5)
Reproduced
48.1% (sd 2.2)
did not match
C5
Reported
GSE2034 MCC 0.914 (Table 5)
Reproduced
0.301 (sd 0.026)
did not match
C6
Reported
GSE2034 Sens/Spec 94.2%/97.1% (Table 5)
Reproduced
52.8% / 79.0%
did not match
C7
Reported
GSE2034 127 selected genes (Table 6)
Reproduced
127-5000/fold (mean 1816); RFECV not fixed at 127
did not match
C8
Reported
GSE2034 MAD-filter survivors 8,456 (Table 6)
Reproduced
16,712 (authors' own MAD filter)
did not match
C9
Reported
GSE7390 = 198 x 22,283 probes, GPL96, 2 classes (Table 2)
Reproduced
198 x 22,283, GPL96, 2 classes (label=e.dmfs 62/136)
exact
C10
Reported
GSE7390 Accuracy 97.5% (Table 5)
Reproduced
63.1% (sd 2.7)
did not match
C11
Reported
GSE7390 AUC-ROC 0.986 (Table 5)
Reproduced
0.573 (sd 0.044); per-seed down to 0.52
did not match
C12
Reported
GSE7390 F1 97.3% (Table 5)
Reproduced
28.5% (sd 8.4)
did not match
C13
Reported
GSE7390 MCC 0.949 (Table 5)
Reproduced
0.052 (sd 0.093); per-seed negative
did not match
C14
Reported
GSE7390 Sens/Spec 96.8%/98.2% (Table 5)
Reproduced
23.9% / 81.0%
did not match
C15
Reported
GSE7390 98 selected genes (Table 6)
Reproduced
98-4500/fold (mean 1563); not fixed at 98
did not match

Assessments & scoring basis

Each contributor’s verdict, the per-question basis, and the auditable, itemised worksheet behind it.

👤 Schlein Lab (curation team) L2 25/100
🟢1. Data identity
🟢2. Endpoint comparability
🔴3. Location of the main deviation
🔴4. Cause of the deviation
🔴5. Derivability / plausibility
🔴6. Severity of the deviation
🔴7. Core claim
🔴8. Severity of the miss (overall human judgment)
Scoring basis — itemised

Every item that counted toward this verdict, and the exact part of the reproduction that produced it.

Main result did not reproduce
Decisive
From: Q5 · Derivability / plausibility 🔴
Main result did not reproduce
Decisive
From: Q7 · Core claim 🔴
Main result did not reproduce
Decisive
From: Q8 · Severity of the miss (overall human judgment) 🔴
🤝
Reproduced automatically — and fairly

Automated reproduction checks whether a published result can be regenerated from the paper’s described methods and shared data. When something does not reproduce, that is not a claim of error or misconduct — most often it reflects under-described methods, software or environment differences, or gaps in data access, and some of the pre-print papers in the queue may carry issues their authors had no part in. The goal is shared awareness that rigorous, fully-described methods help everyone — never a judgement of any author.

Are you an author? We would genuinely like to hear from you — to clarify the record, add data or code, re-run the pipeline after an accession update, and publish your response right next to the assessment. Everything here is open and auditable.

🚩 Report an error in this record

Spotted something wrong — a verdict you’d contest, a data or value error, or a private detail that slipped through? Tell us, with a short justification. Authors and readers are equally welcome to write in; we review every report.

Prefer email, or the form below not working? Contact us at support@doesitreproduce.com.

Reproduction footprint

claude-opus-4-8

Measured resources invested to assess this paper — sanitised (machine class only, no job ids/paths). Compute = HPC accounting (SLURM); tokens = the AI agent's session.

497.5 k
tokens (I/O) · 41.7 M incl. cache
215 min
runtime · 3.08 CPU-h
4.2 GB
peak RAM
5 (1 failed)
HPC jobs
hummel
machine