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

Computationally scalable regression modeling for ultrahigh-dimensional omics data with ParProx.

Brief Bioinform · 2021
L1 79/100 PQI 93
Why this verdict

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

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: Q6 · Severity of the deviation 🟡
Minor / cosmetic deviation
+1 pts
From: Q2 · Endpoint comparability 🟡
Concordant (toward reproduced)
Code + data deposited & functional
-2 pts
From: Data & code availability Available & functional
Total score +7
✓ What held up
  • Nothing in this column.
What did not (or only partly)
  • 🟡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
How its reproducibility compares
79/100
Reproducibility score
0.3 SD above mean
vs. all fields · 1173 studies
🎯 Scores higher than 55% of all assessed papers rank 514 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 to reproduce 1:1 on the SHIPPED artifacts. ParProx.jl is the authors' own Julia tool; the repo ships the GSE25006 'set2' breast-cancer demo data IN-REPO. I re-ran the authors' documented Example-2 logistic overlapping-group-lasso tutorial verbatim on «our HPC» (Julia 1.6.7 chosen deliberately for MersenneTwister RNG parity with the paper-era Julia 1.5; CPU path — gene selection is numerically identical to GPU). EXACT 1:1 reproduction of every checkable shipped/documented number: group matrix (1043, 21963); CV-selected lambda=1.2589254117941663e-9 (seed 222, the RNG-sensitive step reproduced exactly); final fit 1045 nonzero (557 pos / 488 neg) at both the documented and CV-selected lambda; cohort 469 patients (95 pCR / 374 RD, matches paper). A spot-checked coefficient (RFC2) matches to 8 significant figures (|diff|=2.6e-9, within solver tolerance). This is a strong positive integrity signal: the tool behaves exactly as documented. THE DIVERGENCE: the paper's HEADLINE real-data number ('830 genes, 489+/341-' from 12307 genes / 11734 groups) is NOT reproducible from shipped artifacts, because the repo ships only a REDUCED 1043-gene demo matrix (exprdata_set2_cut.txt.gz), not the full 12307-gene GSE25006 training matrix used for that figure; at the demo's CV-chosen lambda the solution is essentially dense (1045/1043 nonzero), nothing like a sparse 830-gene signature. Flagged per rule 5 as a value not derivable from the shipped data/code — most likely benign UNDER-SHIPPING (a size-reduced demo), NOT fabrication: the full-data analysis is plausible, simply not auditable from the release. Hence overall PARTIAL. NOT ATTEMPTED (80/20): simulation GPU-vs-CPU timing (Fig 2, hardware-specific), the Cox/GBM example (same engine already verified), the ROC-vs-RF/SVM comparison (non-ParProx comparators), and biological interpretation. All grades provisional for human audit (see AUDIT.md).

💻 Code ↗ 🗄 Data: GSE25006

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 79
    assessed: 2026-06-15 ⛓ b58d6889da7b
✎ 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
2026-06-15
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
2026-08-05

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: opus
Founding hypothesis

Can multivariable group-lasso regression for ultrahigh-dimensional omics data be made computationally scalable—via parallel/distributed (GPU) computing and a latent variable group representation—so that interpretable, biologically-grouped models can be fit routinely where existing implementations are intractable?

Core claims
  • ParProx implements overlapping and non-overlapping (latent) group lasso regression for time-to-event (Cox) and classification (logistic) analysis with variables grouped by biological priors. resource
  • A latent variable group representation enables parallel/distributed model fitting of ultrahigh-dimensional data on GPUs or cloud environments natively in Julia. method
  • ParProx uses an easily parallelizable proximal gradient descent algorithm rather than inherently sequential (block) coordinate descent, allowing simultaneous coordinate/group updates. method
  • Group-penalized estimation yields interpretable regression models consistent with known biological relationships (genomic regions, pathways/GO) during model estimation rather than post hoc. finding
  • Simulation studies demonstrate the scalability of ParProx with GPUs compared to existing CPU-based implementations. finding
  • Computing the entire regularization path in parallel takes little more time than solving a single penalized regression problem of the same size, without a screening rule. mechanism
Experimental setups
Assay System Perturbation Readout Platform
Somatic mutation count / Cox proportional hazards survival regression TCGA pan-cancer patients (9707 patients) none (observational genomic variants) overall survival; mutation counts of ~56,000 DNA sequence segments grouped by gene / interacting gene pairs GPU-based ParProx (Julia)
Gene expression-based logistic regression (classification) Breast cancer patients (neoadjuvant chemotherapy) none (observational) pathological complete response (pCR) as binary outcome; ~12,000 mRNA gene-level measurements grouped by overlapping pathways/GO terms
DNA methylation-based Cox proportional hazards survival regression Liver cancer patients (377 patients), TCGA methylation array none (observational) survival; methylation status of CpG-island probes (289,509 probes) grouped by adjacent genes (overlapping groups) TCGA methylation array (>865,000 probes original)
Simulation / benchmarking study synthetic ultrahigh-dimensional data none computation time / scalability vs existing implementations (glmnet, grpreg, etc.) GPU vs CPU
Key results
  • ParProx scales to ultrahigh-dimensional data on GPUs, outperforming existing CPU implementations in simulation benchmarks.
  • Cox model of somatic mutation counts fit for overall survival across TCGA patients using gene/gene-pair variable groups.
  • Logistic model for pCR to neoadjuvant chemotherapy fit using overlapping pathway/GO gene groups.
  • Cox model of CpG-island DNA methylation fit for liver cancer survival with overlapping adjacent-gene groups.
Key statistics
  • count 9707 (TCGA patients in Cox somatic mutation survival analysis)
  • count ~56,000 (DNA sequence segments as independent variables (mutation counts))
  • count ~12,000 (mRNA gene-level measurements in breast cancer logistic regression)
  • count 377 (liver cancer patients in DNA methylation Cox analysis)
  • count 289,509 (methylation probes used (trimmed from >865,000 original))
  • other up to 4.3 gigabytes (memory footprint of each of the three trimmed data sets)

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.

ParProx is a methods and software paper presenting a parallelized proximal gradient descent algorithm for fitting group lasso penalized regression models — Cox proportional hazards for survival outcomes and logistic regression for binary classification — on ultrahigh-dimensional omics data. The central statistical approach uses structured sparsity-inducing penalization (overlapping and non-overlapping group lasso) informed by biological priors such as genomic regions and pathways, with the regularization parameter λ selected via cross-validation. The paper demonstrates scalability through simulation benchmarks comparing wall-clock time across implementations, and applies the method to three TCGA datasets; classical null hypothesis significance testing is not the primary mode of inference.

Replicationunclear Sample sizeSample sizes stated for each application (n=9707 TCGA survival; n=377 liver cancer; breast cancer n not stated in excerpt); predictor counts stated (~56,000 mutation segments; ~12,000 genes; 289,509 methylation probes); no formal power analysis described GroupsParProx vs. grpreg/grpregOverlap/glmnet for computational benchmarks; survival vs. censored and pCR vs. no-pCR as model outcomes in applications Pairingna Randomization/blindingnot stated Dispersionnone Exact p-valuesno Effect sizesno Confidence intervalsno Multiplicity correctionnone stated; group lasso penalization jointly selects or zeroes out entire variable groups but no post-selection or post-hoc multiplicity correction for inference is described
Statistical tests used
Test Applied to n Assumptions
Group lasso penalized Cox proportional hazards regression (partial likelihood with group L1/L2 penalty, optimized via proximal gradient descent) Application 1: TCGA somatic mutation counts vs. overall survival; Application 3: liver cancer DNA methylation vs. overall survival 9707 (Application 1); 377 (Application 3) not stated
Group lasso penalized logistic regression (binary outcome, group L1/L2 penalty via proximal gradient descent) Application 2: gene expression vs. pathological complete response (pCR) to neoadjuvant chemotherapy in breast cancer not stated
K-fold cross-validation for regularization parameter (λ) selection along the full regularization path All three real-data applications na
Descriptive wall-clock time benchmarking for scalability assessment Simulation studies comparing ParProx (GPU/CPU) against grpreg, grpregOverlap, and glmnet na
Approaches that could also have been used
  • The group lasso penalty zeroes out entire variable groups jointly; all variables in a selected group receive non-zero coefficients
    Could also: Sparse group lasso (SGL) could also be used, adding an L1 term within groups so that only a subset of variables in a selected group are retained — SGL is appropriate when within-group sparsity is expected — for example, when only a subset of a pathway's genes is likely active — yielding solutions that are simultaneously group-sparse and within-group-sparse, which may improve interpretability in large pathways
  • Cross-validation is used to select the regularization parameter λ along the full regularization path
    Could also: Extended BIC (EBIC) or stability-based criteria could also be used for λ selection in high-dimensional settings — EBIC penalizes model complexity more aggressively than standard BIC and is well-suited to high-dimensional sparse models; it avoids the additional computational cost of cross-validation, which may be substantial at the scale of problems this paper targets
  • Cox proportional hazards regression is used for all time-to-event outcomes
    Could also: Penalized accelerated failure time (AFT) models or Fine–Gray competing-risks models could also be applied to these survival endpoints — AFT models provide a different parameterization (covariate effects on log survival time rather than the hazard), which some find more directly interpretable; competing-risks models would additionally account for non-cancer death as a competing event in heterogeneous cancer cohorts such as TCGA
  • Computational scalability is assessed descriptively via single-run wall-clock time comparisons across implementations
    Could also: Repeated timing measurements with reported dispersion (e.g., via microbenchmark or BenchmarkTools.jl) could also quantify timing variability alongside point estimates — Replicated benchmarks with confidence intervals or standard deviations allow readers to assess whether observed speedups are consistent across runs, hardware states, and problem instances, strengthening the empirical support for scalability claims
  • Model performance in the real-data applications is conveyed through the selected (non-zero) coefficients after penalization
    Could also: Nested cross-validation or evaluation on a held-out external cohort could also be used to report out-of-sample predictive performance (e.g., C-index, AUC) — Using the same data for both λ selection (inner CV) and performance reporting can yield optimistic estimates; a separate outer evaluation loop or an independent validation cohort provides a less biased estimate of how models generalize to new patients
  • Variable selection stability is assessed implicitly through the single penalized solution at the chosen λ
    Could also: Stability selection or bootstrap inclusion frequencies could also be applied to assign each variable a selection probability across data perturbations — Penalized regression can yield different sparse solutions under small data perturbations; stability selection aggregates over many subsamples to identify features that are consistently selected, providing a complementary measure of which variables are robustly informative beyond the single-fit solution
Software: Julia (ParProx — this paper's contribution) · R/glmnet · R/grpreg · R/grpregOverlap · R/snpnet · Julia/MendelIHT.jl

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.

Citations
5
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.

GSE25006 GEO in Methods (http://purl.org/orb/Methods)
no other assessed paper uses this yet
GSE32646 GEO in Methods (http://purl.org/orb/Methods)
no other assessed paper uses this yet

Downstream reach in the literature

66 downstream papers · 2 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.

What was reproduced

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

Scope — pmid-34254998 (ParProx)

Paper: Ko S, Li GX, Choi H, Won JH. Computationally scalable regression modeling for ultrahigh-dimensional omics data with ParProx. Brief Bioinform 2021. PMID 34254998 · PMC8575036 · DOI 10.1093/bib/bbab256. Code: https://github.com/kose-y/ParProx.jl (commit pinned at run, master HEAD 7ee1d04, 2023-07-24). Data: ships INSIDE the repo data/.

ParProx.jl is the authors' own Julia package implementing proximal-gradient regularized regression (L1 / overlapping-latent group lasso) for logistic, Cox, and multinomial models, with optional CUDA acceleration. The paper's value claim is computational scalability (GPU vs CPU timing), demonstrated on simulations plus two real omics analyses.

Reported results and in/out-of-scope classification

Result Pipeline In scope? Note
R1 GSE25006 logistic overlapping group-lasso gene selection — "selected 830 genes (489 positive, 341 negative)" from 12 307 genes / 11 734 groups ParProx logistic group-lasso + 5-fold CV PARTIAL The repo ships only a reduced demo matrix exprdata_set2_cut.txt.gz with 1043 genes, not the full 12 307. The exact 830/489/341 headline used the FULL data, which is not shipped. We reproduce the shipped pipeline end-to-end and compare its documented demo output 1:1; the full-data headline is graded as not-reproducible-from-shipped-artifacts (honest caveat, possible under-shipping flag — NOT fabrication).
R2 ParProx shipped-tutorial demo output (docs/src/index.md Example 2, CPU path): group matrix (1043, 21963); CV-selected λ=1.2589e-9; final fit 1045 nonzero vars (557 pos / 488 neg); per-gene coefficients ParProx logistic group-lasso, seed 222 YES (primary) Fully specified, deterministic (final fit RNG-independent; CV uses seed 222). This is the clean 1:1 reproduction target — re-running the authors' shipped code+data must reproduce their own documented numbers.
R3 Simulation timing, p=2e5..2e6, GPU (Titan V) vs CPU grpregOverlap (Fig 2) synthetic data generator + timing OUT (80/20) Timing is hardware-specific (their Titan V vs our H100/EPYC); not a 1:1 numeric claim. Not attempted.
R4 Cox example (GBM TCGA somatic mutations, Example 1) ParProx Cox group-lasso OUT (80/20) Second worked example; same engine as R2. Skipped to keep one clean focused data point per the 80/20 rule.
R5 ROC/AUC comparison vs RF/SVM/lasso (Fig 4B) external tools OUT Comparator methods (random forest, SVM) are not ParProx; out of scope.
R6 Biological interpretation ("DNA replication / mitotic cell cycle" pathways) manual OUT Wet-lab/manual interpretation.

Primary reproduction target

R2 (and, by extension, structural comparison to R1): run the authors' ParProx.jl Example-2 logistic overlapping-group-lasso pipeline on the shipped GSE25006 set2 demo data and check the documented numbers reproduce exactly:

  • group matrix size (1043, 21963) — deterministic, RNG-free;
  • CV-selected λ = 1.2589254117941663e-9 (seed 222, 5-fold, AUC criterion);
  • final fit 1045 nonzero original variables (557 positive / 488 negative);
  • reference coefficient RFC2 = 0.017828370354349202.

All compute on «our HPC» (Julia 1.6.7 LTS for RNG parity; CPU path). The GPU/CUDA path is not required — the gene-selection result is numerically identical on CPU.

Figures / tables: Fig 2
R2_grpmat
Reported
group matrix (1043, 21963)
Reproduced
(1043, 21963)
exact
R2_lambda
Reported
CV-selected lambda = 1.2589254117941663e-9 (seed 222)
Reproduced
1.2589254117941663e-9 (idx 30/31, match=true)
exact
R2_nnz
Reported
1045 nonzero original vars (557 positive, 488 negative)
Reproduced
1045 nonzero (557 positive, 488 negative)
exact
R2_coef_RFC2
Reported
RFC2 = 0.017828370354349202
Reproduced
0.017828372906321225 (8 sig figs, |diff|=2.6e-9)
within tolerance
R1_cohort
Reported
469 training patients
Reproduced
469 (95 pCR / 374 RD)
exact
R1_genes_selected
Reported
830 genes selected (489 positive, 341 negative)
Reproduced
NOT REPRODUCIBLE FROM SHIPPED DATA: repo ships reduced 1043-gene demo, not full 12307-gene matrix; demo selects 1045/1043 at the tiny CV lambda
partial
R1_dims
Reported
12307 genes, 11734 groups
Reproduced
1043 genes, 5719 groups (shipped demo)
partial
R3_sim_timing
Reported
GPU sub-linear vs CPU super-linear (Fig 2)
Reproduced
NOT ATTEMPTED (80/20; hardware-specific)
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 79/100

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: Q6 · Severity of the deviation 🟡
Minor / cosmetic deviation
+1 pts
From: Q2 · Endpoint comparability 🟡
Concordant (toward reproduced)
Code + data deposited & functional
-2 pts
From: Data & code availability Available & functional
Total score +7

The authors' own ParProx.jl reproduces its shipped, documented Example-2 pipeline 1:1 — CV-selected λ to all digits, 1045 nonzero (557+/488−), and RFC2 to 8 significant figures (|Δ|=2.6e-9) — a strong positive integrity signal that the tool behaves exactly as published. The single deviation is that the paper's headline real-data number (830 genes, 489+/341−) used the full 12307-gene/11734-group GSE25006 matrix, which was never deposited; the repo ships only a reduced 1043-gene demo, so the headline is not auditable (the demo yields a dense ~1045-gene solution at its CV λ). This sits on the authors'/data-availability side (under-shipping), not our methodology, and is most plausibly benign rather than fabrication. Overall yellow: solid reproduction with one clearly explained, non-suspect gap.

🤝
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.

161.7 k
tokens (I/O) · 10.5 M incl. cache
20 min
runtime · 0.44 CPU-h
2.2 GB
peak RAM
3 (1 failed)
HPC jobs
hummel
machine