Inference of RNA polymerase II transcription dynamics from chromatin immunoprecipitation time course data.
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.
The main results reproduced: recomputed values matched the published ones within tolerance.
Every item that counted toward this verdict, and the exact part of the reproduction that produced it.
- Nothing in this column.
- 🟡Could not use the authors’ exact input data
- 🟡Reported values were only indirectly comparable
- 🔴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
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
Partial, honest reproduction. The released PyPol_II code (third-party-independent, authors' own GitHub repo) was successfully run on «our HPC» for its 10 shipped example genes (the exact Table 5 top-10 gene list) and its self-contained synthetic benchmark (Toy_data.py). Point-estimate transcription speeds (n=20 random restarts) landed within-CI or exact for 8/10 genes vs the paper's reported median+95%CI; 2 genes (TIPARP, PDLIM7) missed the CI on both n=1 and n=20 runs. Using the README's exact documented n=1 command only reproduced 5/10 within-CI, revealing that the paper's point-estimate procedure is materially sensitive to the number of ML random restarts - a genuine methodological finding, not a bug. The synthetic benchmark exactly reproduced the paper's qualitative ranking (GP-Conv best, kernel method breaks down under convolution) with same-order-of-magnitude MNSE values. A documentation bug was found and flagged: the code/README label output as 'kilobases per second' when it is actually kilobases per minute (matches Table 5 units; a literal per-second reading would be physically implausible for Pol II elongation). Genome-wide claims (2623 candidate genes, 383 well-fit genes, 12 clusters) were exactly verified via the author-deposited PLOS File S1 supplementary gene lists (cheap, exact, file-count verification) without needing to redo the un-released raw-ChIP-seq-to-RPM-matrix pipeline. NOT attempted (explicitly out of scope, no code released): genome-wide re-derivation of the 383-gene RPM matrices from raw GEO ChIP-seq; the Bayesian HMC procedure that actually produces Table 5/6's median+CI (repo only has ML point estimation); TF-binding enrichment (Table 9) and GO enrichment analyses; the Singh & Padgett comparison dataset (Table 4). GSE44800 (8 ChIP-seq samples, MCF-7, E2 time course) is complete and open but only delivers per-sample BED files, not the per-gene segment-RPM matrices the released code actually consumes.
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.
✎ 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
- 2026-08-07
- Rubric version
- v1.0
- Assessed by
-
🤖 AI curator · claude (ai-curator headless) · v1.0 · run #1 2026-08-07no human curator yet
- Last updated
- 2026-08-07
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: opusCan a probabilistic (convolved Gaussian process) model of RNA polymerase II movement, fitted to low-temporal-resolution pol-II ChIP-seq time course data, be used to estimate transcription speed genome-wide and infer promoter activity profiles that reveal co-regulated, rapidly responding genes? The paper tests this by modelling the transcription wave of estrogen-responsive genes in MCF7 cells.
- ★ A convolved Gaussian process model of pol-II occupancy across gene segments captures the transcription wave and yields estimates of transcription speed and promoter-proximal pol-II activity. method
- ★ Model parameters can be inferred by maximum likelihood or by Bayesian inference via Hamiltonian Monte Carlo/MCMC sampling, the latter giving confidence intervals and allowing informative priors on transcription speed. method
- ★ Because the model does not assume constant transcription speed, it can identify variable/non-uniform transcription rates such as those arising from pol-II pausing. mechanism
- ★ Fitting the model genome-wide identifies transcriptionally engaged genes and filters out transcripts where pol-II does not travel down the gene body, providing a principled way to find responsive (early estrogen-responsive) genes. finding
- ★ Clustering the inferred promoter activity profiles groups genes that respond quickly to stimuli and that may be co-regulated. finding
- ★ Rapidly induced genes are enriched for both estrogen receptor alpha (ERα) and FOXA1 binding in their proximal promoter regions. finding
- ★ Transcription speeds obtained genome-wide agree with previously published estimates that were derived for small numbers of long genes. finding
- The method is released as a freely available Python package, PyPol-II (https://github.com/ciiram/PyPol_II). resource
| Assay | System | Perturbation | Readout | Platform |
|---|---|---|---|---|
| pol-II ChIP-seq time course | MCF7 human breast cancer cells | estradiol (E2) stimulation | pol-II occupancy per gene segment (gene divided into 5 segments, 20% each) over time points from 0 to 320 min; inferred delays, transcription speed, promoter activity profile | — |
| Clustering of inferred promoter activity profiles | MCF7 pol-II ChIP-seq derived latent promoter activity profiles | estradiol (E2) | gene clusters of shared temporal activity / rapidly responding genes | — |
| Transcription factor binding site analysis and public TF ChIP-seq profile analysis | Promoter regions of clustered genes; MCF7 (same system) public ChIP-seq TF datasets including ERα and FOXA1 | none | cluster-specific TF binding patterns; proximal promoter TF binding enrichment | Genomatix (author affiliation) / publicly available ChIP-seq datasets |
| Canonical pathway enrichment analysis | Gene clusters from MCF7 pol-II ChIP-seq | none | enriched canonical pathways per cluster | — |
| Synthetic/simulated benchmark data | In silico signals generated as a sum of N=20 Gaussian kernels with random delay and additive Gaussian noise | varying number of observations; added convolution to distort signal shape | Median Normalised Square Error (MNSE) of estimated delay over 50 random realisations | — |
| Reanalysis of published pre-mRNA/exon-intron junction delay dataset | Dataset from Singh and Padgett (9 genes, all >100 kbp) | reversible transcription inhibition (as performed in the original study) | delay in appearance of pre-mRNA signal at exon-intron junctions, used to compute transcription rate | — |
| Alternative time-delay estimation methods (benchmark comparators) | Synthetic and real time series | none | delay estimates via cross-correlation, discrete correlation function (DCF), and kernel-based method with kernel width set by leave-one-out cross-validation | — |
| GRO-seq (stated as applicable data type for the model) | not specified | none | nascent transcript markers of transcriptional activity used as model input | — |
- – For the TIPARP gene, pol-II ChIP-seq shows a transcription wave moving down the gene and the model estimates transcription speed approximately 2 kilobases per minute
- – Model-derived genome-wide transcription speeds are consistent with previously published rates obtained for limited sets of long genes
- ▲ Rapidly induced gene clusters show enrichment of ERα and FOXA1 binding in proximal promoter regions
- – Inferred pol-II time profiles with 95% confidence intervals and delay histograms were obtained for top-ranked genes MYH9, TIPARP and RAB10 95% confidence intervals shown
- – Time series of pol-II occupancy at successive gene segments show the transcription wave spreading out ('diffusion'), possibly due to loss of synchrony in the cell population
- – Delay estimation accuracy improves with number of observations, quantified as MNSE of the estimated delay on synthetic data
- other approximately 2 kilobases per minute (Estimated transcription speed for TIPARP from this model (gene split into 5 segments))
- other 4.3 kilobases per minute (Prior transcription speed from Darzacq et al., mechanistic model of nascent RNA at a single locus)
- other 3.1 kbp min−1 (Prior average transcription rate from Wada et al. for genes >100 kbp using intronic RNA signal)
- other 3.79 kbp min−1 (Prior average transcription rate from Singh and Padgett (2009) for 9 genes all >100 kbp)
- other 95% (Confidence interval shown for inferred pol-II profiles in Figure 2)
- count 5 segments (20%, 40%, ... of the gene) (Number of segments each gene is divided into for modelling)
- count 50 random realisations; N = 20 Gaussian kernels (Synthetic benchmark: realisations used for MNSE computation and number of kernels forming the underlying function)
- other 0 to 320 min (Time range of raw pol-II ChIP-seq reads for TIPARP after estradiol addition; prior study [5] sampled every 7.5 min)
Statistical methods review
Model: sonnetA neutral, descriptive read of the statistical approach — what was done, and (for shared learning, not as criticism) what could also have been done.
This is a computational-methods paper that fits a convolved Gaussian process (GP) model to pol-II ChIP-seq time course data to estimate transcription dynamics. Model parameters (segment delays, transcription speed, promoter activity) are estimated either by maximum likelihood optimization of the GP marginal likelihood or by Bayesian inference using Hamiltonian Monte Carlo (HMC) MCMC sampling, with the latter used to derive confidence intervals. Average transcription speed per gene is obtained via linear regression through the origin of delay versus gene position, and the delay-estimation approach is benchmarked against classical alternatives (cross-correlation, discrete correlation function, kernel-based methods) on synthetic data (50 random realisations) and a previously published real dataset, using the Median Normalised Square Error (MNSE) as the performance metric.
| Test | Applied to | n | Assumptions |
|---|---|---|---|
| Maximum likelihood estimation (gradient-based optimization of GP marginal likelihood) | Fitting the convolved Gaussian process model to pol-II occupancy time series for each gene | — | not stated |
| Bayesian inference via Hamiltonian Monte Carlo (MCMC/HMC) | Parameter and delay estimation, including derivation of confidence intervals for transcription speed/promoter activity | — | not stated |
| Linear regression through the origin | Estimating average transcription speed from delay (minutes) versus gene position (bp) across gene segments (e.g. 5 segments per gene as in Figure 1) | 5 gene segments per gene | not stated |
| Cross-correlation, discrete correlation function (DCF), and kernel-based delay estimation | Benchmark comparison of alternative time-delay estimation methods on synthetic data and the Singh & Padgett real dataset | 50 random realisations (synthetic benchmark); real dataset n not stated | not stated |
| Median Normalised Square Error (MNSE) | Assessing effect of number of observations on quality of delay inference in the synthetic benchmark | 50 random realisations | na |
-
Model parameters (including transcription speed and delays) can be estimated via a single point maximum likelihood estimate.↳ Could also: Using the fully Bayesian MCMC/HMC approach (already applied elsewhere in the paper) as the default for all reported estimates rather than switching between ML and Bayesian fitting. — Propagating full posterior uncertainty into downstream analyses (e.g., clustering of promoter activity profiles) can capture estimation uncertainty that a single ML point estimate does not convey.
-
Benchmark performance of delay-estimation methods is summarized using the Median Normalised Square Error (MNSE) across 50 random realisations.↳ Could also: Reporting the full distribution of errors (e.g., via bootstrap confidence intervals or interquartile range) alongside the median. — Would convey the spread and variability of benchmark performance across realisations, not just a central-tendency summary.
-
Average transcription speed is computed by ordinary linear regression through the origin of delay versus segment position.↳ Could also: A weighted or Bayesian regression that incorporates per-segment delay uncertainty (e.g., posterior variance from MCMC) as regression weights. — Could account for heteroscedastic uncertainty across segments, potentially refining the precision of the speed estimate.
-
Alternative delay-estimation methods (cross-correlation, DCF, kernel-based) are compared to the proposed GP approach primarily via an error metric (MNSE).↳ Could also: A formal Bayesian model comparison (e.g., marginal likelihood or Bayes factors) across the candidate methods. — Provides a probabilistic basis for comparing modelling approaches in addition to point-estimate error metrics.
-
Genes are classified as 'transcriptionally engaged' based on how well the model fits genome-wide ChIP-seq data.↳ Could also: An explicit multiple-testing correction (e.g., Benjamini-Hochberg FDR) if a likelihood-based threshold or test statistic is used across many genes simultaneously. — Would control the expected proportion of false-positive gene classifications when screening thousands of genes genome-wide.
-
Uncertainty in inferred promoter activity and delay estimates is reported as 95% confidence intervals derived from MCMC samples.↳ Could also: Presenting full posterior density plots or posterior predictive checks alongside the intervals. — Could reveal skewness or multimodality in the posterior that a single symmetric interval summary may not capture.
What was reproduced
The exact results taken into scope, with each reported value next to the value our attempt produced.
Assessments & scoring basis
Each contributor’s verdict, the per-question basis, and the auditable, itemised worksheet behind it.
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.
Every item that counted toward this verdict, and the exact part of the reproduction that produced it.
What deviates: With n_try=20 restarts, 8/10 of the repo's shipped example genes reproduce inside the paper's reported 95% CIs, but TIPARP (2.55 vs 1.9 [1.4-2.4]) and PDLIM7 (1.45 vs 3.5 [1.7-6.4]) miss consistently regardless of restart count, and the synthetic-benchmark MNSEs are same-order but up to ~7x off (GP-Conv 1.5e-05 vs 2.1e-06) while the full method ranking GP-Conv < GP-NoConv < Corr < DCF << Kern reproduces exactly. Whose side: predominantly the authors' - the HMC sampler that actually produced Table 5/6's medians+CIs is not in PyPol_II (so the comparison is necessarily ML-point-estimate vs posterior summary), the raw-BED-to-RPM-segment pipeline for the other 373 genes is unreleased, and the code mislabels kb/min as kb/s; our side contributes only environment drift and the restart-count choice. Severity: moderate, not critical - direction and magnitude hold everywhere, and File S1 matches every independently checkable headline count exactly (2623 candidate genes, 383 well-fit genes, 12 clusters), which argues strongly against fabrication. This is a legitimate partial reproduction with named, bounded gaps rather than silent omissions.
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.