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

Integrative bioinformatics and artificial intelligence analyses of transcriptomics data identified genes associated with major depressive disorders including <i

Neurobiol Stress · 2023
L1 66/100 3/4
Why this verdict

The main results reproduced, with only marginal, non-material deviations.

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.

Supporting (toward a concern)
Content-critical question only partially held
+2 pts
From: Q5 · Derivability / plausibility 🟡
Content-critical question only partially held
+2 pts
From: Q7 · Core claim 🟡
Content-critical question only partially held
+2 pts
From: Q8 · Severity of the miss (overall human judgment) 🟡
Minor / cosmetic deviation
+1 pts
From: Q3 · Location of the main deviation 🟡
Minor / cosmetic deviation
+1 pts
From: Q4 · Cause of the deviation 🟡
Minor / cosmetic deviation
+1 pts
From: Q6 · Severity of the deviation 🟡
Concordant (toward reproduced)
Code + data deposited & functional
-2 pts
From: Data & code availability Available & functional
Total score +7
✓ What held up
  • Reported values were directly comparable
What did not (or only partly)
  • 🟡Could not use the authors’ exact input data
  • 🟡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
66/100
Reproducibility score
0.5 SD below mean
vs. all fields · 1173 studies
🎯 Scores higher than 28% of all assessed papers rank 830 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

DESCRIBED WELL ENOUGH? Partly. autoNeuro is a generic tabular-ML grid-search tool (LR/SVM/RF/XGB/LGBM x SelectKBest/SelectFromModel, 10-fold StratifiedKFold) applied to GSE98793; the repo ships ONLY the ML stage. Its training inputs (per-batch gene-symbol CSVs) are NOT in the repo, and the GEO->matrix preprocessing (MAS5>50 & CV>10% -> 1446 genes; 157-sample subset) cannot be reconstructed from the public GCRMA series matrix. RESULT: running the unmodified autoNeuro grid on a faithful-as-feasible reconstruction (top-1446-variance gene panel, native 96/96 batch split, full 192 samples) reproduces the SAME performance regime as Table 4 - batch2 within tolerance and slightly exceeding the paper (acc 0.81 vs 0.75, AUC 0.80 vs 0.79, F1 0.76 vs 0.70), batch1 and merged lower by 0.06-0.14 (acc 0.70/0.72 vs 0.79/0.80) but every std band overlaps the paper's. No hard mismatch, no fabrication signal; the shortfalls are explained by the two unreconstructable preprocessing steps that would plausibly lift the paper's numbers. 1:1 vs different: DIFFERENT in exact point values, SAME in regime -> partial. NOT attempted: GSEA gene lists, the NRG1/10-gene biomarker panel, transfer-learning external validation, wet-lab qPCR.

💻 Code ↗ 🗄 Data: GSE98793

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 66
    assessed: 2026-06-20 ⛓ 7ee86da4537f
✎ 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
2026-06-20
Rubric version
v1.0
Assessed by
🤖 AI curator · claude (ai-curator room) · v1.0 · run #1 2026-06-20
no human curator yet
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 study tests whether integrating bioinformatics (GSEA) and machine learning analysis of transcriptomic data from MDD patients versus healthy controls can identify reliable, non-invasive diagnostic biomarkers for major depressive disorder.

Core claims
  • Differentially expressed genes in MDD patients are enriched in immune response, inflammatory response, neurodegeneration, and cerebellar atrophy pathways. finding
  • Feature selection combined with ML algorithms produced predictive models distinguishing MDD from healthy controls with ≥75% accuracy. finding
  • Integrative bioinformatics and ML analysis identified ten key MDD-related biomarkers: NRG1, CEACAM8, CLEC12B, DEFA4, HP, LCN2, OLFM4, SERPING1, TCN1 and THBS1. finding
  • NRG1 was the most robust and reliable biomarker distinguishing MDD patients from healthy controls across independent external datasets of mixed populations. finding
  • NRG1 is upregulated in saliva samples of MDD patients compared to healthy controls in an independent Kazakhstan cohort. finding
  • NRG1 shows high expression in subcortical limbic brain regions implicated in depression, per functional mapping to human brain regions. finding
  • An integrative pipeline combining absolute/normal GSEA, feature selection methods (PCA, SelectKBest, LR, RF), and multiple ML classifiers (LR, RF, XGBoost, SVM, KNN) was developed to identify MDD biomarkers. method
Experimental setups
Assay System Perturbation Readout Platform
GSEA (gene set enrichment analysis) whole blood microarray, human MDD patients and healthy controls (GSE98793) none (case-control) enriched cellular pathways and differentially expressed leading genes Affymetrix Human Genome U133-Plus 2.0 gene-chip
Machine learning classification (LR, RF, XGBoost, SVM, KNN with feature selection) whole blood transcriptomic data, human MDD patients and healthy controls (GSE98793) none (case-control) classification accuracy, F1 score, ROC-AUC for MDD vs HC prediction SkLearn v1.2.1, Python v3.9
Differential gene expression analysis (Limma) transcriptomic datasets from four external cohorts (GSE99725, GSE76826, GSE38206, GSE32280; French/Caucasian, Japanese, Chinese populations) none (case-control) differentially expressed genes between MDD and HC Limma package, R software
Functional mapping and annotation of gene expression to brain regions postmortem human brain tissue (six neurotypical adult brains, ~3700 tissue samples) none regional brain expression of biomarker genes, including NRG1 Allen Human Brain Atlas (AHBA)
qRT-PCR saliva samples from MDD patients and healthy controls, Kazakhstan population none (case-control) NRG1 mRNA relative/fold expression change (normalized to 18S rRNA) QuantStudio3 system with Maxima SYBR Green/ROX qPCR Master Mix
Key results
  • DEGs in MDD patients were significantly enriched in immune response, inflammatory response, neurodegeneration and cerebellar atrophy pathways. p < 0.05
  • ML models predicted MDD status based on MDD-altered genes. ≥75% accuracy
  • Ten key MDD-related biomarkers identified: NRG1, CEACAM8, CLEC12B, DEFA4, HP, LCN2, OLFM4, SERPING1, TCN1, THBS1.
  • NRG1 best distinguished MDD patients from healthy controls across independent external datasets with mixed populations.
  • NRG1 expression was upregulated in saliva of MDD patients compared to healthy controls.
  • NRG1 showed high expression in main subcortical limbic brain regions implicated in depression.
Key statistics
  • count 170 MDD patients and 121 healthy controls (total samples mined from publicly available transcriptomic datasets)
  • count 128 MDD patients and 64 healthy controls (discovery dataset GSE98793)
  • count 42 MDD patients and 57 healthy controls (combined cohort across four external confirmation datasets)
  • count 12 MDD patients and 8 healthy controls (saliva sample validation cohort from Kazakhstan)
  • pvalue p < 0.05 (significance cut-off for GSEA and DEG identification)
  • fold_change FC > 1.5 (upregulated); FC < 0.5 (downregulated) (thresholds defining differentially expressed genes in GSEA)
  • other ≥75% accuracy (performance of ML models predicting MDD vs healthy control status)
  • count ~25,828 annotated cellular pathways across seven gene sets (MSigDB gene sets used for absolute GSEA analysis)

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.

The study combined bioinformatics and machine-learning analyses of microarray transcriptomic data (GSE98793: 128 MDD patients, 64 healthy controls) to identify MDD-associated genes. Differential pathway enrichment was assessed with Gene Set Enrichment Analysis (GSEA, permutation-based, p<0.05, fold-change cutoffs), and differential expression in four independent external cohorts was assessed with the Limma R package (p<0.05). Multiple machine-learning classifiers (logistic regression, random forest, XGBoost, SVM, KNN) with feature selection and 10-fold cross-validation were used to build and evaluate predictive models, reported via accuracy, F1-macro, and ROC-AUC; the top candidate gene (NRG1) was further evaluated by qRT-PCR in an independent saliva cohort.

Replicationmixed Sample sizeSample sizes are stated for the discovery dataset (128 MDD/64 HC), pooled external validation datasets (42 MDD/57 HC across four cohorts, detailed in Table 1), and an independent Kazakhstan saliva cohort (12 MDD/8 HC); no formal statistical power calculation is described GroupsMDD patients vs healthy controls (transcriptomic profiles across discovery and external datasets; NRG1 saliva expression in an independent cohort) Pairingunpaired Randomization/blindingnot stated Dispersionunclear Effect sizesyes
Statistical tests used
Test Applied to n Assumptions
Gene Set Enrichment Analysis (absolute GSEA followed by normal/standard GSEA, permutation-based) Identification of enriched/activated pathways and differentially enriched genes, MDD vs healthy controls, discovery dataset 128 MDD patients, 64 healthy controls not stated
Limma (linear models for microarray data, moderated statistics) Differential gene expression between MDD patients and healthy controls in each of four external validation datasets varies by dataset per Table 1 (e.g., 18 vs 15; 12 vs 10; 9 vs 9; 8 vs 8) not stated
ANOVA F-value (used as a feature-scoring/selection method, SelectKBest) Feature (gene) selection prior to machine-learning classification discovery dataset (GSE98793) not stated
10-fold cross-validated classification (logistic regression, random forest, XGBoost, SVM, KNN) evaluated by accuracy, F1-macro, ROC-AUC Classification of MDD patients vs healthy controls based on selected gene features discovery dataset merged batches; independent evaluation on 42 MDD/57 HC across four external datasets na
Approaches that could also have been used
  • GSEA pathway significance is described using a p<0.05 threshold across thousands of annotated gene sets.
    Could also: Reporting the FDR q-value that GSEA computes by default — would explicitly convey the expected proportion of false discoveries when testing thousands of gene sets simultaneously, which a raw p-value threshold alone does not capture
  • Differential expression between MDD and healthy controls in the external datasets was assessed with Limma at p<0.05.
    Could also: Reporting Benjamini-Hochberg (or similar FDR) adjusted p-values, which Limma calculates by default — would help control the false discovery rate when testing many genes per array, a standard consideration in microarray/transcriptomic differential expression work
  • Classifier performance (accuracy, F1-macro, ROC-AUC) is summarized as point estimates from 10-fold cross-validation.
    Could also: Reporting confidence intervals or a bootstrap distribution around these cross-validated metrics — would convey the uncertainty/variability of performance estimates, which is often informative when comparing multiple models or datasets of modest size
  • Feature selection (ANOVA F-value, PCA, embedded LR/RF importance) was performed before model training and cross-validation.
    Could also: Embedding feature selection within each cross-validation fold (nested cross-validation) — would also address the well-known overfitting risk when the number of features greatly exceeds the number of samples, keeping feature selection and performance evaluation fully separated
  • Four independent external transcriptomic datasets from different populations were each analyzed separately with Limma.
    Could also: A formal meta-analytic approach combining effect sizes or p-values across the four cohorts (fixed- or random-effects meta-analysis) — would allow pooled evidence and a single combined estimate of effect across ethnically diverse cohorts, in addition to per-dataset results
  • NRG1 saliva expression was compared between a small MDD cohort (n=12) and healthy controls (n=8), though the specific statistical test for this comparison is not shown in the provided excerpt.
    Could also: A nonparametric test such as the Mann-Whitney U test — is often preferred for small sample comparisons where normality of the expression/fold-change data has not been established
Software: R (in-house pipeline; MAS5 and GCRMA, Bioconductor) · Limma (R package) · Python / scikit-learn (SkLearn) scikit-learn v1.2.1 (Python 3.9); GridSearchCV noted separately as sklearn version 0.24.2 · GSEA / MSigDB (Broad Institute)

What was reproduced

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

Figures / tables: Table
C1_merged_acc
Reported
0.80 ± 0.10
Reproduced
0.72 ± 0.11
partial
C2_merged_auc
Reported
0.82 ± 0.12
Reproduced
0.73 ± 0.12
partial
C3_merged_f1
Reported
0.76 ± 0.11
Reproduced
0.67 ± 0.13
partial
C4_batch1_acc
Reported
0.79 ± 0.11
Reproduced
0.70 ± 0.12
partial
C5_batch1_auc
Reported
0.88 ± 0.13
Reproduced
0.74 ± 0.18
partial
C6_batch1_f1
Reported
0.73 ± 0.17
Reproduced
0.67 ± 0.12
within tolerance
C7_batch2_acc
Reported
0.75 ± 0.13
Reproduced
0.81 ± 0.10
within tolerance
C8_batch2_auc
Reported
0.79 ± 0.15
Reproduced
0.80 ± 0.19
within tolerance
C9_batch2_f1
Reported
0.70 ± 0.16
Reproduced
0.76 ± 0.13
within tolerance
C10_n_samples
Reported
192 (128 MDD / 64 HC)
Reproduced
192 (128 MDD / 64 HC)
exact
C11_n_genes_filter
Reported
1446 genes (MAS5>50 & CV>10%)
Reproduced
uncheckable - MAS5 intensities not in deposit
partial
C12_n_subset
Reported
157 (106 MDD / 51 HC)
Reproduced
uncheckable - undocumented exclusion of 35 samples
partial

Assessments & scoring basis

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

🤖 AI curator · claude (ai-curator room) · v1.0 L1 66/100

An automated assessment. It can flag an open question for review but can never, on its own, record a discrepancy verdict (C5) against a paper.

🟡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.

Supporting (toward a concern)
Content-critical question only partially held
+2 pts
From: Q5 · Derivability / plausibility 🟡
Content-critical question only partially held
+2 pts
From: Q7 · Core claim 🟡
Content-critical question only partially held
+2 pts
From: Q8 · Severity of the miss (overall human judgment) 🟡
Minor / cosmetic deviation
+1 pts
From: Q3 · Location of the main deviation 🟡
Minor / cosmetic deviation
+1 pts
From: Q4 · Cause of the deviation 🟡
Minor / cosmetic deviation
+1 pts
From: Q6 · Severity of the deviation 🟡
Concordant (toward reproduced)
Code + data deposited & functional
-2 pts
From: Data & code availability Available & functional
Total score +7

Running the unmodified autoNeuro grid on a faithful-as-feasible reconstruction of GSE98793 reproduces the same performance regime as Table 4 (acc 0.70–0.81, AUC 0.73–0.80, F1 0.67–0.76): batch2 within tolerance and slightly above the paper, batch1/merged lower by 0.06–0.14 but with overlapping std bands and no hard mismatch. The deviations sit on the input/preprocessing side — the unreconstructable MAS5>50 & CV>10% 1446-gene filter and the undocumented 157-sample subset — split between our forced method choices and the authors' underspecification, not a computation error. No fabrication signal; magnitudes are achievable. The biomarker/NRG1 panel that anchors the paper was out of scope, so the central biological claim itself remains untested here.

🤝
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.

224.4 k
tokens (I/O) · 13.3 M incl. cache
54 min
runtime · 0.37 CPU-h
2 GB
peak RAM
1
HPC jobs
hummel
machine