pith. sign in

arxiv: 2607.01307 · v1 · pith:TYSMWRO2new · submitted 2026-07-01 · 💻 cs.LG · q-bio.GN

A Novel Machine Learning Approach for Central Nervous System Tumor Classification from DNA Methylation

Paulo R. Ferreira Jr. , Lucas Coutinho Freitas , La\'is dos Santos Gon\c{c}alves , William Borges Domingues , Lucas Petitemberte de Souza , Mariana B. Michalowski , Vinicius F. Campos This is my paper

Pith reviewed 2026-07-03 21:46 UTC · model grok-4.3

classification 💻 cs.LG q-bio.GN
keywords CNS tumor classificationDNA methylationmachine learningsparse random projectionmultinomial logistic regressioncancer diagnosticsmethylation profilingtumor subtype identification
0
0 comments X

The pith

A sparse random projection plus multinomial logistic regression model classifies central nervous system tumors from DNA methylation profiles with higher accuracy than the prior reference method.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether a straightforward machine learning pipeline can classify central nervous system tumors more reliably from DNA methylation data than an established reference classifier. It reduces the high-dimensional methylation profiles with sparse random projection and then applies multinomial logistic regression to assign one of 91 tumor classes or their broader families. On a held-out clinical cohort of 1,104 samples the new model reaches 86 percent class-level accuracy and 93 percent family-level accuracy, beating the reference figures of 82 percent and 88 percent. The authors argue that these gains matter because more accurate subtype calls can change treatment choices. The work keeps the same evaluation protocol and data splits as the reference to make the comparison direct.

Core claim

The combination of sparse random projection for dimensionality reduction and multinomial logistic regression for classification yields 96 percent mean accuracy under stratified 3-fold cross-validation on the 2,801-sample reference cohort and, on the independent 1,104-sample clinical cohort, delivers 86 percent accuracy at the 91-class level and 93 percent at the methylation class family level, exceeding the reference classifier's 82 percent and 88 percent concordance by roughly four and five percentage points.

What carries the argument

Sparse random projection for dimensionality reduction followed by multinomial logistic regression for classification.

If this is right

  • Higher accuracy on independent data indicates better cross-cohort transferability for methylation-based CNS tumor classification.
  • A five-point gain at the family level can directly affect cancer subtype assignment and influence treatment selection.
  • Grounding the model in stronger methodological practice produces more reliable CNS tumor classification across evaluation settings.
  • The approach can materially improve the reliability of methylation-based tumor classification in diagnostic use.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the accuracy edge holds in additional independent cohorts, the pipeline could serve as a practical upgrade for pathology labs already running methylation arrays.
  • The same dimensionality-reduction-plus-logistic-regression pattern may transfer to other high-dimensional molecular profiling tasks without needing deep learning.
  • Re-testing the method against a wider range of modern classifiers would clarify whether logistic regression remains the best choice after the projection step.

Load-bearing premise

The independent clinical evaluation cohort follows exactly the same preprocessing steps, class definitions, and evaluation protocol as the reference classifier without any unstated differences that could inflate the reported gains.

What would settle it

Reprocessing the same 1,104-sample independent cohort with any deviation in preprocessing or class labeling that eliminates or reverses the four-to-five-point accuracy gains would falsify the claim of consistent improvement.

Figures

Figures reproduced from arXiv: 2607.01307 by La\'is dos Santos Gon\c{c}alves, Lucas Coutinho Freitas, Lucas Petitemberte de Souza, Mariana B. Michalowski, Paulo R. Ferreira Jr., Vinicius F. Campos, William Borges Domingues.

Figure 1
Figure 1. Figure 1: t-SNE embedding of the 2,801-sample reference cohort after SRP, colored by the 91 methylation classes. 5.1 Dimensionality Reduction Given the extremely high dimensionality of DNA methylation profiles, we first applied Sparse Random Projection (SRP) to map each sample to a substantially lower-dimensional space while approximately preserving pairwise distances. To qualitatively assess whether the SRP space p… view at source ↗
Figure 2
Figure 2. Figure 2: Confusion matrix for the 91-class (3-fold CV). Capper et al. reported cross-validated error rates of 4.89% (raw) and 4.28% (calibrated) for the same 2,801-sample / 91-class reference cohort using their classifier. In our stratified 3-fold cross-validation, the proposed approach yields an average error rate of 3.42%, which is comparable to—and slightly lower than — those values. While our study uses a diffe… view at source ↗
Figure 3
Figure 3. Figure 3: Clinical evaluation of the proposed approach vs. Capper et al. (state-of-the-art) 6 Conclusion This article proposed a simple, fully reproducible, and computationally efficient pipeline for DNA methylation-based brain tumor classification, designed to be transparent and practical for clinical translation. The method combines Sparse Random Projection for dimensionality reduction with a multinomial logistic … view at source ↗
read the original abstract

NA methylation profiling has become a powerful approach for central nervous system (CNS) tumor classification, yet important challenges remain regarding cross-cohort transferability, methodological correctness, and robust multiclass evaluation. In this work, we propose a novel and methodologically rigorous machine-learning approach for methylation-based CNS tumor classification that combines Sparse Random Projection for dimensionality reduction with multinomial logistic regression for classification. We evaluate the proposed approach in the same general experimental setting established by a widely used reference classifier. On the 2,801-sample reference cohort, our method achieves a mean accuracy of 96\% under stratified 3-fold cross-validation. On the independent 1,104-sample clinical evaluation cohort, it reaches 86\% accuracy at the 91-class level and 93\% when predictions are evaluated at the methylation class family level. These results improve upon the corresponding state-of-the-art reference figures of 82\% class-level concordance and 88\% family-level concordance, yielding absolute gains of approximately 4 and 5 percentage points, respectively. This improvement is clinically relevant: in a diagnostic setting, a 5-point increase in correct tumor classification can directly affect cancer subtype assignment and, in turn, influence treatment selection and downstream clinical decision-making. Our results show that the proposed model, grounded in stronger methodological practice in machine learning, consistently outperforms the previous state of the art across evaluation settings and can materially improve the reliability of CNS tumor classification.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper proposes combining Sparse Random Projection for dimensionality reduction with multinomial logistic regression for multiclass CNS tumor classification from DNA methylation profiles. It reports 96% mean accuracy under stratified 3-fold cross-validation on the 2,801-sample reference cohort and, on an independent 1,104-sample clinical cohort, 86% accuracy at the 91-class level and 93% at the family level, outperforming the reference classifier's 82% and 88% by 4-5 points in the same general experimental setting.

Significance. If the independent-cohort protocol exactly replicates the reference preprocessing, taxonomy, and inclusion criteria, the gains would constitute a clinically relevant improvement (a 5-point increase in correct subtype assignment can affect treatment selection) achieved with an interpretable linear model after explicit dimensionality reduction. The external clinical cohort evaluation is a methodological strength that grounds the claims beyond internal cross-validation.

major comments (3)
  1. [Abstract / Methods (independent cohort)] Abstract and evaluation protocol description: the central claim of 4-5 point superiority on the 1,104-sample cohort is load-bearing only if preprocessing (probe selection, normalization, missing-value handling), the exact 91-class taxonomy, family groupings, and sample inclusion criteria are identical to the reference. The manuscript states only 'the same general experimental setting'; without explicit confirmation or a side-by-side protocol table, the observed delta cannot be unambiguously attributed to Sparse Random Projection + multinomial logistic regression.
  2. [Methods] Hyperparameter handling: the free parameters (random projection dimension and sparsity, logistic regression regularization strength) are listed but no section describes the selection procedure, cross-validation grid, or sensitivity analysis. This omission directly affects reproducibility of the reported 96% CV accuracy and the independent-cohort figures.
  3. [Results (independent cohort evaluation)] Statistical comparison: no p-values, confidence intervals, or McNemar-style test is reported for the 86%/93% vs. 82%/88% differences on the independent cohort, leaving the clinical-relevance claim without quantified uncertainty.
minor comments (1)
  1. [Introduction] The 91-class count is introduced only in the abstract; a brief parenthetical in the introduction would improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments and for recognizing the value of the external clinical cohort evaluation. We address each major comment below and have revised the manuscript to improve clarity, reproducibility, and statistical support for the claims.

read point-by-point responses

  1. Referee: Abstract and evaluation protocol description: the central claim of 4-5 point superiority on the 1,104-sample cohort is load-bearing only if preprocessing (probe selection, normalization, missing-value handling), the exact 91-class taxonomy, family groupings, and sample inclusion criteria are identical to the reference. The manuscript states only 'the same general experimental setting'; without explicit confirmation or a side-by-side protocol table, the observed delta cannot be unambiguously attributed to Sparse Random Projection + multinomial logistic regression.

    Authors: We agree that explicit confirmation is essential to attribute the observed gains unambiguously to the proposed method. In the revised manuscript we have added a new supplementary table (Table S1) that provides a side-by-side comparison of all preprocessing steps, probe selection criteria, normalization procedures, missing-value handling, the precise 91-class taxonomy, family groupings, and sample inclusion/exclusion criteria used in both our study and the reference classifier. The table confirms that these elements are identical, thereby grounding the 4–5 point improvement in the same experimental setting. revision: yes

  2. Referee: Hyperparameter handling: the free parameters (random projection dimension and sparsity, logistic regression regularization strength) are listed but no section describes the selection procedure, cross-validation grid, or sensitivity analysis. This omission directly affects reproducibility of the reported 96% CV accuracy and the independent-cohort figures.

    Authors: We acknowledge the omission. The revised Methods section now includes a dedicated subsection on hyperparameter selection. We describe a grid search over random-projection dimension (100–2000), sparsity level (0.01–0.2), and logistic-regression regularization strength (C ∈ {0.001, 0.01, 0.1, 1, 10, 100}), performed via stratified 5-fold cross-validation on the reference cohort to maximize mean accuracy. We also report the selected values and include a brief sensitivity analysis (Figure S2) showing that performance remains stable within a broad neighborhood of the chosen hyperparameters. revision: yes

  3. Referee: Statistical comparison: no p-values, confidence intervals, or McNemar-style test is reported for the 86%/93% vs. 82%/88% differences on the independent cohort, leaving the clinical-relevance claim without quantified uncertainty.

    Authors: We agree that quantified uncertainty strengthens the clinical-relevance claim. In the revised Results section we now report 95% bootstrap confidence intervals (1,000 resamples) for all accuracy figures on the independent cohort. We additionally apply McNemar’s test to the paired predictions and report p-values (class-level p=0.012; family-level p=0.008), confirming that the observed improvements are statistically significant at the conventional 0.05 threshold. revision: yes

Circularity Check

0 steps flagged

No significant circularity; evaluation uses external reference and independent cohort.

full rationale

The paper introduces Sparse Random Projection + multinomial logistic regression and reports accuracies on a 2,801-sample reference cohort (via 3-fold CV) and a separate 1,104-sample clinical cohort. It compares these to an external state-of-the-art reference classifier using the same general experimental setting. No equations or steps reduce by construction to fitted inputs, self-definitions, or self-citations. The central performance claims rest on standard ML training and an independent hold-out cohort rather than any renaming, ansatz smuggling, or uniqueness theorem from the authors' prior work. The comparison protocol concern is a potential correctness issue, not a circularity reduction.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central performance claims rest on the assumption that the reference experimental setting is faithfully reproduced and that the two cohorts are directly comparable; no free parameters are explicitly listed in the abstract but the projection sparsity and logistic regression regularization are implicit tuning choices.

free parameters (2)
  • random projection dimension and sparsity
    Sparse random projection requires choice of target dimension and sparsity level; these are not stated in the abstract and are presumed tuned on the reference cohort.
  • logistic regression regularization strength
    Multinomial logistic regression typically includes a regularization hyperparameter whose value is not reported.
axioms (1)
  • domain assumption The reference classifier's experimental setting is exactly replicated without unstated differences in preprocessing or class handling.
    The abstract states evaluation occurs in the same general experimental setting established by the reference classifier.

pith-pipeline@v0.9.1-grok · 5826 in / 1400 out tokens · 32276 ms · 2026-07-03T21:46:50.264666+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

56 extracted references · 56 canonical work pages

  1. [1]

    DNA methylation-based classification of central nervous system tumours

    Capper, David and others. DNA methylation-based classification of central nervous system tumours. Nature. 2018. doi:10.1038/nature26000

    work page doi:10.1038/nature26000 2018

  2. [2]

    and others , title =

    Sheng, H. and others , title =. Oncogene , year =

    work page

  3. [3]

    and others , title =

    Manoranjan, B. and others , title =. Cancer Letters , year =

    work page

  4. [4]

    and others , title =

    Korshunov, A. and others , title =. Neuro-Oncology , year =

    work page

  5. [5]

    Yoon, J. W. and others , title =. International Journal of Cancer , year =

    work page

  6. [6]

    and others , title =

    Ray, S. and others , title =. Diagnostics , year =

    work page

  7. [7]

    and others , title =

    Slika, H. and others , title =. Cancers , year =

    work page

  8. [8]

    and others , title =

    Das, S. and others , title =. Sensors , year =

    work page

  9. [9]

    Louis, D. N. and others , title =. Neuro-Oncology , year =

    work page

  10. [10]

    and others , title =

    Chen, J. and others , title =. Frontiers in Immunology , year =

    work page

  11. [11]

    Guo, J. U. and others , title =. Nature Neuroscience , year =. doi:10.1038/nn.3602 , note =

    work page doi:10.1038/nn.3602

  12. [12]

    Capell, B. C. and Berger, S. L. , title =. Journal of Investigative Dermatology , year =

    work page

  13. [13]

    Greer, E. L. and Shi, Y. , title =. Nature Reviews Genetics , year =

    work page

  14. [14]

    and Lin, Y.-C

    Chen, D.-P. and Lin, Y.-C. and Fann, C. S. J. , title =. Briefings in Functional Genomics , year =

    work page

  15. [15]

    Irizarry, R. A. and others , title =. Nature Genetics , year =

    work page

  16. [16]

    and others , title =

    Geissler, F. and others , title =. Therapeutic Advances in Medical Oncology , year =

    work page

  17. [17]

    and others , title =

    Song, M.-A. and others , title =. PLOS ONE , year =

    work page

  18. [18]

    and Packer, R

    Jackson, K. and Packer, R. J. , title =. Current Neurology and Neuroscience Reports , year =

    work page

  19. [19]

    2024 , note =

    Medulloblastoma: Diagnosis and Treatment , howpublished =. 2024 , note =

    work page 2024

  20. [20]

    2020 , note =

    What is Exploratory Data Analysis , howpublished =. 2020 , note =

    work page 2020

  21. [21]

    2019 , url =

    Infinium MethylationEPIC BeadChip Data Sheet , howpublished =. 2019 , url =

    work page 2019

  22. [22]

    2022 , url =

    Infinium MethylationEPIC v2.0 BeadChip Data Sheet , howpublished =. 2022 , url =

    work page 2022

  23. [23]

    2015 , month = nov, number =

    Infinium HD Methylation Assay Protocol Guide , organization =. 2015 , month = nov, number =

    work page 2015

  24. [24]

    iScan System: Generated Files (IDAT intensity output) , year =

    work page

  25. [25]

    2010 , month = nov, number =

    GenomeStudio Methylation Module v1.8 User Guide , organization =. 2010 , month = nov, number =

    work page 2010

  26. [26]

    and Baggerly, Keith A

    Smith, Mike L. and Baggerly, Keith A. and Bengtsson, Henrik and Ritchie, Matthew E. and Hansen, Kasper D. , title =. F1000Research , year =. doi:10.12688/f1000research.2-264.v1 , url =

    work page doi:10.12688/f1000research.2-264.v1

  27. [27]

    and Jaffe, Andrew E

    Aryee, Martin J. and Jaffe, Andrew E. and Corrada-Bravo, Hector and Ladd-Acosta, Christine and Feinberg, Andrew P. and Hansen, Kasper D. and Irizarry, Rafael A. , title =. Bioinformatics , year =

    work page

  28. [28]

    and Lindenstrauss, Joram , title =

    Johnson, William B. and Lindenstrauss, Joram , title =. Contemporary Mathematics , volume =

    work page

  29. [29]

    Journal of Computer and System Sciences , year =

    Achlioptas, Dimitris , title =. Journal of Computer and System Sciences , year =

    work page

  30. [30]

    and Church, Kenneth W

    Li, Ping and Hastie, Trevor J. and Church, Kenneth W. , title =. Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD) , year =

    work page

  31. [31]

    Hastie, Trevor and Tibshirani, Robert and Friedman, Jerome , title =

    work page

  32. [32]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Defazio, Aaron and Bach, Francis and Lacoste-Julien, Simon , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

    work page

  33. [33]

    Proceedings of the International Conference on Machine Learning (ICML) , year =

    Niculescu-Mizil, Alexandru and Caruana, Rich , title =. Proceedings of the International Conference on Machine Learning (ICML) , year =

    work page

  34. [34]

    Proceedings of the National Academy of Sciences , year =

    Selection bias in gene extraction on the basis of microarray gene-expression data , author =. Proceedings of the National Academy of Sciences , year =

    work page

  35. [35]

    BMC Bioinformatics , year =

    Bias in error estimation when using cross-validation for model selection , author =. BMC Bioinformatics , year =

    work page

  36. [36]

    Journal of Machine Learning Research , year =

    On Over-fitting in Model Selection and Subsequent Selection Bias in Performance Evaluation , author =. Journal of Machine Learning Research , year =

    work page

  37. [37]

    International Conference on Database Theory (ICDT) , year =

    When Is ``Nearest Neighbor'' Meaningful? , author =. International Conference on Database Theory (ICDT) , year =

    work page

  38. [38]

    Journal of the Royal Statistical Society: Series B , year =

    Geometric representation of high dimension, low sample size data , author =. Journal of the Royal Statistical Society: Series B , year =

    work page

  39. [39]

    International Joint Conference on Artificial Intelligence (IJCAI) , year =

    A Study of Cross-Validation and Bootstrap for Accuracy Estimation and Model Selection , author =. International Joint Conference on Artificial Intelligence (IJCAI) , year =

    work page

  40. [40]

    IEEE Transactions on Knowledge and Data Engineering , year =

    Learning from Imbalanced Data , author =. IEEE Transactions on Knowledge and Data Engineering , year =

    work page

  41. [41]

    Information Processing & Management , year =

    A systematic analysis of performance measures for classification tasks , author =. Information Processing & Management , year =

    work page

  42. [42]

    Journal of Machine Learning Research , year =

    Visualizing Data using t-SNE , author =. Journal of Machine Learning Research , year =

    work page

  43. [43]

    Proceedings of the 12th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , year =

    Very Sparse Random Projections , author =. Proceedings of the 12th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , year =

    work page

  44. [44]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    SAGA: A Fast Incremental Gradient Method With Support for Non-Strongly Convex Composite Objectives , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

    work page

  45. [45]

    Proceedings of the International Conference on Machine Learning (ICML) , year =

    On Calibration of Modern Neural Networks , author =. Proceedings of the International Conference on Machine Learning (ICML) , year =

    work page

  46. [46]

    Dataset Shift in Machine Learning , editor =

    work page

  47. [47]

    IEEE Transactions on Information Theory , year =

    On Optimum Recognition Error and Reject Tradeoff , author =. IEEE Transactions on Information Theory , year =

    work page

  48. [48]

    , title =

    Guo, Chuan and Pleiss, Geoff and Sun, Yu and Weinberger, Kilian Q. , title =. Proceedings of the International Conference on Machine Learning (ICML) , year =

    work page

  49. [49]

    Advances in Large Margin Classifiers , year =

    Platt, John , title =. Advances in Large Margin Classifiers , year =

    work page

  50. [50]

    Dataset Shift in Machine Learning , publisher =

    work page

  51. [51]

    Chow, C. K. , title =. IEEE Transactions on Information Theory , year =

    work page

  52. [52]

    and Wegkamp, Marten H

    Bartlett, Peter L. and Wegkamp, Marten H. , title =. Journal of Machine Learning Research , year =

    work page

  53. [53]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Geifman, Yonatan and El-Yaniv, Ran , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

    work page

  54. [54]

    Journal of Machine Learning Research , year =

    van der Maaten, Laurens and Hinton, Geoffrey , title =. Journal of Machine Learning Research , year =

    work page

  55. [55]

    International Conference on Learning Representations (ICLR) , year =

    Hendrycks, Dan and Gimpel, Kevin , title =. International Conference on Learning Representations (ICLR) , year =

    work page

  56. [56]

    Ortiz, H. G. and Rechenmacher, C. and Domingues, W. B. and others , title =. Clinical and Translational Oncology , year =

    work page