← New search

PredTAD: A machine learning framework that models 3D chromatin organization alterations leading to oncogene dysregulation in breast cancer cell lines.

Comput Struct Biotechnol J · 2021

PubMed 34093998 ↗ DOI 10.1016/j.csbj.2021.05.013 ↗ pmid-34093998

L1 77/100 ★ ☆ ★ ★ 3/4

Why this verdict

The main results reproduced: recomputed values matched the published ones within tolerance.

Reproduced on the brainbox compute

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: Q8 · Severity of the miss (overall human judgment) 🟡

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 +3

✓ What held up

✓Reported values were directly comparable
✓The central claim held under reproduction

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
🟡Overall, the reproduction showed a material discrepancy

How its reproducibility compares

77/100

Reproducibility score

at the mean

vs. all fields · 1187 studies

🎯 Scores higher than 50% of all assessed papers rank 574 of 1187 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 and reproduced 1:1 for the pipeline's headline outputs. PredTAD's repo ships 5 trained H2O GBM binary models (not the raw 192-feature matrix). H2O binaries embed cross-validation + training metrics, so on «our HPC» I loaded each model in its matching H2O version (1x 3.26.0.2, 4x 3.28.0.2; versions read from the binaries) and read back the stored 5-fold CV AUC + variable importances. 4 of the 4 directly-stated CV AUCs reproduce EXACTLY to 4 decimals (0.7101->0.71007, 0.7860->0.78609, 0.7860->0.78601, 0.7537->0.75365). Feature importance corroborates the paper's CTCF/ATAC/TSS biology. One discrepancy flagged: the chr-split all-features model stores CV AUC 0.7835 (== authors' own GBM_model_R_list.txt) but the paper TEXT reports 0.7166 for it (the test AUCs in the same sentence DO match) -> likely a manuscript typo/stale value, flagged for human, not fabrication. NOT attempted (the heavy ~80%): re-deriving test AUCs (test feature matrix not shipped), TAD boundary counts (raw Hi-C not shipped), DEG/estrogen/RET analyses (8+ GEO + TCGA accessions), boundary-alteration accuracy 0.7545 and GM12878 cross-cell AUCs (derived .Rdata not shipped), and full retraining from raw multi-omic data through scripts 1-2.

💻 Code ↗ 🗄 Data: GSE66733

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.

v1 current initial assessment Score 77

assessed: 2026-06-15 ⛓ 62292b5fa97d

✎ 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.

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: 2026-06-15
Rubric version: v1.0
Assessed by: 🤖 AI curator · claude (ai-curator room) · v1.0 · run #1 2026-06-15
no human curator yet
Last updated: 2026-09-19

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 tests whether epigenomic and genomic features from readily accessible data types (e.g., ChIP-seq) can be used to predict TAD boundaries, and whether such a tool can uncover large-scale 3D chromatin structural changes that contribute to breast cancer disease state.

Core claims

★ PredTAD, a Gradient Boosting Machine model using epigenomic and genomic ChIP-seq-derived features plus neighboring-bin information, classifies 10 kb genomic bins as TAD boundary or non-boundary method
★ CTCF, cohesin subunits (RAD21 and SMC3), and chromosome number are among the most important features for predicting TAD boundaries and boundary alterations finding
★ Genes near altered TAD boundaries in breast cancer are involved in important signaling pathways such as Ras, Jak-STAT, and estrogen signaling finding
★ A TAD boundary alteration contributes to overexpression of the RET oncogene in breast cancer cells finding
PredTAD can successfully predict TAD boundary changes in other conditions and diseases beyond breast cancer resource
DNA methylation is negatively correlated with CTCF transcription factor binding mechanism

Experimental setups

Assay	System	Perturbation	Readout	Platform
Hi-C	MCF10A, MCF7, T47D, GM12878 cell lines	none (normal vs breast cancer cell lines)	TAD calling and TAD boundary identification (insulation square analysis)	—
ChIP-seq	MCF10A, MCF7, T47D, GM12878 cell lines	none	histone modification marks (9 types) and DNA binding protein/TF occupancy (15 proteins) narrow peaks per 10 kb bin	BWA-MEM alignment, MACS2 peak calling
ATAC-seq	MCF10A, MCF7, T47D, GM12878 cell lines	none	DNA chromatin accessibility signal per 10 kb bin	MACS2 narrow peak calling
DNA methylation array	MCF10A, MCF7, T47D, GM12878 cell lines	none	average CpG methylation signal per 10 kb bin	Illumina Methylation 450K BeadChip
RNA-seq	MCF10A vs MCF7 cell lines	cancer vs normal cell line comparison	differentially expressed genes (TPM via RSEM)	HiSeq 2000, RSEM v.1.2.7
RNA-seq	T47D cell line	none	gene expression profile	NovaSeq 6000, TruSeq Stranded Total RNA LT Sample Prep Kit
Microarray and RNA-seq	MCF7 cell line	17β-estradiol (E2, 10 nM) vs vehicle control treatment	estrogen-responsive differentially expressed genes	Affymetrix U133 A/B GeneChips (RMA normalization); Bowtie2 + DEseq v.1.14.0
RNA-seq	TCGA Luminal A (ER+/Her2-) breast cancer patient tumor vs matched normal breast tissue	disease state (tumor vs normal)	differentially expressed genes	—

Key results

– PredTAD achieved AUC of 0.7954 for MCF10A TAD boundary prediction AUC=0.7954
– PredTAD achieved AUC of 0.7993 for MCF7 TAD boundary prediction AUC=0.7993
– PredTAD achieved AUC of 0.7487 for T47D TAD boundary prediction AUC=0.7487
– Top predictive features across MCF10A, MCF7, and T47D included chromosome number, TFBS/ChIP-seq signal of CTCF, RAD21, and SMC3, histone marks (H3K9ac, H3K27me3, H3K9me3), gene density, relative chromosomal distance, and DNA methylation
▼ DNA methylation was negatively correlated with CTCF binding
– Genes near altered TAD boundaries were enriched in Ras, Jak-STAT, and estrogen signaling pathway involvement
▲ A TAD boundary alteration was identified that contributes to RET oncogene overexpression
– 6271 genes were differentially expressed between MCF10A and MCF7 cell lines n=6271 genes, log2FC>1, p<0.05

Key statistics

other AUC=0.7954 (PredTAD boundary prediction performance in MCF10A)
other AUC=0.7993 (PredTAD boundary prediction performance in MCF7)
other AUC=0.7487 (PredTAD boundary prediction performance in T47D)
count 3305 / 3273 / 3166 / 3097 TAD boundaries (Number of TAD boundaries used for MCF10A, MCF7, T47D, and GM12878 respectively)
count 296,651 genomic regions total; 70%/30% train/test split (Total 10 kb genomic bins used for PredTAD model training/testing)
count 6271 differentially expressed genes (DEGs between MCF10A and MCF7 (log2FC>1, p<0.05, n=3 biological replicates))
count 3947 differentially expressed genes (DEGs between TCGA Luminal A tumor samples (n=315) and matched normal breast samples (n=40), log2FC>1, p<0.05)
count 521 unique estrogen-related genes (Curated estrogen-responsive gene list from two independent E2-treatment datasets)

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.

PredTAD uses a Gradient Boosting Machine (GBM) classifier trained on multi-omics epigenomic and genomic features binned at 10 kb resolution across the human genome to classify regions as TAD boundaries or non-boundaries, evaluated by AUC on a held-out 30% test set. Differential gene expression between cell lines was assessed using log2FC and p-value thresholds, with one dataset processed via DESeq and one via microarray with RMA normalization. A negative correlation between DNA methylation and CTCF binding was reported with a p-value threshold. Model performance and gene counts meeting threshold criteria were the primary outputs.

Replicationbiological Sample sizeTAD boundary counts per cell line stated (MCF10A: 3305, MCF7: 3273, T47D: 3166, GM12878: 3097); RNA-seq replicates stated (n=3 or n=2 per group); no formal power analysis described GroupsNormal breast epithelial (MCF10A) vs. breast cancer cell lines (MCF7, T47D); cell lines vs. TCGA Luminal A tumors vs. matched normal breast tissue; estrogen-treated vs. vehicle-control MCF7 Pairingmixed Randomization/blindingnot stated Dispersionnone Exact p-valuesno Effect sizesyes Confidence intervalsno Multiplicity correctionAdjusted p-value < 0.2 applied in the Lin CY microarray dataset; no correction described for the other differential expression comparisons, for the correlation test, or across GBM feature importances

Statistical tests used

Test	Applied to	n	Assumptions
Gradient Boosting Machine (GBM) binary classification evaluated by AUC (area under the ROC curve)	TAD boundary vs. non-boundary classification for MCF10A, MCF7, T47D, and GM12878 cell lines (Fig. 3)	~296,651 genomic 10 kb bins per cell line; 70% training, 30% test	not stated
Differential gene expression: log2FC > 1, p-value < 0.05 (statistical test not explicitly named; RSEM v1.2.7 used for quantification)	MCF10A vs. MCF7 RNA-seq comparison	n=3 biological replicates per group	not stated
DESeq (v1.14.0) differential expression, log2FC > 1, p-value < 0.05	Nagarajan et al. (GSE55922): estrogen-treated MCF7 vs. vehicle control RNA-seq	n=2 biological replicates	not stated
Microarray differential expression with RMA normalization, log2FC > 1, adjusted p-value < 0.2	Lin CY et al. (GSE11352): estrogen-treated MCF7 vs. vehicle control (Affymetrix U133 A/B GeneChips)	n=3 biological replicates	not stated
Differential gene expression: log2FC > 1, p-value < 0.05 (statistical test not stated)	TCGA Luminal A breast cancer tumor vs. matched normal breast tissue	315 tumor samples, 40 matched normal samples	not stated
Correlation between DNA methylation and CTCF binding (correlation type not specified), p < 0.05	Supplementary Fig. 1	—	not stated

Approaches that could also have been used

Model generalization was estimated from a single 70:30 random train/test split of genomic bins

↳ Could also: k-fold cross-validation or a chromosome-held-out cross-validation scheme could also estimate generalization performance — Cross-validation reduces variance in AUC estimates; a chromosome-level holdout would additionally test whether the model transfers to chromosomal contexts not seen during training, which is particularly relevant given that chromosome number was identified as a top predictive feature
AUC (area under the ROC curve) was the sole reported classification metric on a dataset with an approximately 1:3.5 positive-to-negative class ratio

↳ Could also: The area under the precision-recall curve (AUPRC) or the F1-score could also quantify classifier performance under class imbalance — ROC-AUC can be optimistic when the negative class dominates; AUPRC focuses on the minority positive class and is often recommended as a complementary metric for imbalanced binary classification tasks
Differential gene expression across multiple independent comparisons was assessed using a nominal p-value threshold of < 0.05 (or < 0.2 for one dataset), with thousands of genes tested simultaneously

↳ Could also: A false discovery rate (FDR) correction (e.g., Benjamini-Hochberg) applied to all gene-level tests could also control the rate of false discoveries — With genome-wide testing, the expected number of false positives under a nominal p < 0.05 threshold is substantial; FDR correction is a standard approach that limits the proportion of declared differentially expressed genes that are false positives
AUC values per cell line are reported as single point estimates with no measure of uncertainty

↳ Could also: Bootstrap confidence intervals around AUC could also quantify uncertainty in classification performance — Confidence intervals allow readers to assess the precision of AUC estimates and to make more informed comparisons across cell lines (e.g., MCF10A AUC 0.7954 vs. T47D AUC 0.7487)
The correlation between DNA methylation and CTCF binding is reported only as p < 0.05 without a stated correlation coefficient

↳ Could also: Reporting the Pearson r or Spearman rho alongside the p-value could also convey the strength and direction of the association — A p-value alone does not quantify the magnitude or practical significance of a correlation; the effect size (correlation coefficient) provides additional interpretive context, especially given that genome-wide sample sizes can make even weak correlations statistically significant
GBM was selected as the sole classifier, with hyperparameters (500 trees, max depth 10) set without a described tuning or validation procedure

↳ Could also: Systematic hyperparameter tuning via grid search or random search with cross-validation could also be used to select GBM parameters; Random Forest or penalized logistic regression (LASSO/elastic net) are also standard classifiers for high-dimensional tabular multi-omics data — Documented hyperparameter selection and comparison with at least one alternative classifier are common practices for establishing that a chosen model and its settings are well-suited to the task; Random Forest requires less sensitivity tuning for depth, and penalized regression yields directly interpretable, regularized feature weights

Software: GBM (Gradient Boosting Machine; specific package/library not named) · RSEM v1.2.7 · DESeq v1.14.0 · Bowtie2 v2.1.0 · BWA-MEM · MACS2

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 · 4

1Jacqueline Chyr 2Zhigang Zhang 3Xi Chen 4Xiaobo Zhou

Citations

Impact: low

Foundation confidence

None of its references are in our reproducibility record yet — its foundation cannot be assessed.

Topics

Biology Computational biology Genetics Gene

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.

GSE63525 GEO in Methods (http://purl.org/orb/Methods)

also used by 1 paper:

Identity rather than 3D position informs splic... L1 98/100

AAA17888 in Supplementary material (http://purl.obolibrary.org/obo/IAO_0000326)