Pith. sign in

REVIEW 5 major objections 6 minor 29 references

An Interpretable Ensemble Framework for Multi-Omics Dementia Biomarker Discovery Under HDLSS Conditions

T0 review · 5 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read An ensemble of four methods claims to beat single-model multi-omics integrators at dementia biomarker discovery.

desk verdict A plausible-sounding pipeline with no implementable method and no verifiable results; the missing p-value link leaves the FDR filter as decoration. read the letter →

arxiv 2509.10527 v1 pith:CW747FD2 submitted 2025-09-04 eess.IV cs.CYcs.LGstat.ME

Byeonghee Lee , Joonsung Kang This is my paper · ORCID
classification eess.IVcs.CYcs.LGstat.ME
keywords multi-omicsintegrationdementiabiomarkersAlzheimer'sdiseasegraphattentionnetworksvariationalautoencoderelastic-netregressionfalsediscoveryrateADNI
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that dementia biomarker discovery from high-dimensional, low-sample-size multi-omics data works best when four complementary ideas are chained together: graph attention to model gene-gene interactions, a variational autoencoder to compress modalities into one latent space, elastic-net regression to pick a sparse feature set, and Storey's FDR to filter those features statistically. Against four published multi-omics integrators, the proposed ensemble reports the best AUC, F1, and feature-selection precision on both a simulated Alzheimer's dataset (0.93/0.91/0.88) and the ADNI cohort (0.91/0.89/0.87), with interpretability as the stated advantage. The resulting biomarker gene map, built from top genes such as TREM2, APOE, BIN1, and MAPT connected by interaction strengths, is offered as a bridge between statistical genomics and systems biology. A reader should care because the paper claims a principled answer to a real clinical constraint: few patients, many molecular measurements, and the need to know not just which genes matter but how they interact.

What carries the argument

The carrying object is the four-stage pipeline itself. GAT assigns learned attention weights to gene-gene edges, replacing the uniform neighbor weighting of graph convolutions and avoiding over-smoothing. MOVE, a multi-omics variational autoencoder with an objective of reconstruction accuracy minus a beta-weighted KL divergence plus a cross-modal alignment term, compresses the omics layers into a shared latent space that preserves GAT-derived structure. Elastic-net solves a mixed L1/L2 penalized regression to select a sparse, grouped feature set. Storey's q-values, computed from an estimated null proportion, threshold features at 0.01, 0.05, and 0.1. Each stage counters one failure mode of h

What would settle it

Attempt the reproduction: to obtain the reported AUC of 0.93 and 0.91, F1 of 0.91 and 0.89, and feature precision of 0.88 and 0.87, one needs the exact interface between the GAT representation and the MOVE encoder, the encoder's latent dimension, the elastic-net hyperparameters, and the way latent variables are mapped to p-values for Storey's procedure, none of which the manuscript gives. A reader following the text cannot reconstruct a working pipeline; specifying those details and rerunning the two benchmarks with cross-validation would settle whether the claimed margins over DIABLO, MOCAT,

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that an ensemble of four established techniques, GAT, MOVE, elastic-net, and Storey's FDR, outperforms each family of single-paradigm integrators across every reported metric while remaining interpretable enough that the selected genes and their interactions can be read as a biological map. The author would state it as: by letting a graph attention network first learn which gene-gene associations matter, a multi-omics variational autoencoder then harmonize modalities into a shared latent space, elastic-net then select a sparse and grouped feature set, and Storey's procedure then certify those features against false-discovery error, a practitione

Load-bearing premise

The pipeline is assumed to be concretely implementable: GAT's graph outputs must connect to MOVE's latent space, and that latent space must yield p-values on which Storey's FDR can validly run, but the manuscript specifies none of these connections, dimensions, training procedures, or p-value derivations; if any link is missing, the framework is a list of component names and the reported scores are unsupported.

Editorial extensions

If this is right

  • If the reported numbers hold, the ensemble beats four published multi-omics integrators on both simulated and ADNI data without sacrificing interpretability, giving high-dimensional low-sample-size studies a viable alternative to single-model pipelines.
  • The top-ranked genes, TREM2, APOE, BIN1, SORL1, MAPT, CD33, BACE1, INPP5D, CR1, and APP, and their mapped interactions align with established Alzheimer's risk and pathology, so the framework's outputs are directly usable as a hypothesis-generating biomarker map.
  • Because the design is modular, the same GAT-MOVE-elastic-net-FDR chain is claimed to transfer to cancer, type 2 diabetes, autoimmune disorders, and Parkinson's disease with only modality swaps.
  • Every selected feature carries a statistical warranty, Storey's FDR, layered on a biological prior, the graph structure, which the paper argues is what makes its results interpretable in a way the benchmark methods are not.
  • Future extensions claimed include longitudinal disease-progression modeling, integration with neuroimaging, spatial transcriptomics, and prospective clinical validation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • My reading: the reported margins over the strongest baseline are small, an AUC of 0.93 versus 0.89, and the paper reports no variance, cross-validation, or significance testing for its comparisons; a fair method ranking would need confidence intervals or repeated resampling.
  • My reading: the gene-gene interaction strengths, such as TREM2-TYROBP at 0.93, are presented as outputs but appear to be pre-assigned rather than learned; a genuine test of the GAT's contribution would compare its learned attention weights against these listed strengths to see whether the network rediscovers them.
  • My reading: the paper's central novelty claim, that graph structure plus latent compression plus sparse regression plus FDR beats each alone, could be tested directly by ablation on the same simulated data, dropping one stage at a time; the paper does not run this experiment.
  • Testable extension: applying the identical pipeline to another well-phenotyped high-dimensional low-sample-size cohort, such as cancer omics with fewer than one hundred samples, and checking whether the top hits reproduce known driver genes would show whether the ADNI result generalizes or is dataset-specific.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes an ensemble framework for multi-omics biomarker discovery in dementia, combining Graph Attention Networks (GAT), Multi-Omics Variational AutoEncoder (MOVE), elastic-net regression, and Storey's False Discovery Rate (FDR). The authors benchmark against DIABLO, MOCAT, AMOGEL, and MOMLIN on both a synthetic Alzheimer's disease dataset and the ADNI cohort, reporting higher AUC, F1, and feature-selection precision (Tables 1 and 4). They also present lists of significant genes and gene-gene interactions, claiming superior predictive accuracy, feature selection precision, and biological relevance. The manuscript is a draft-level document: the method is described only through textbook equations, the integration between components is unspecified, and the experimental reporting omits reproducibility-critical details.

Significance. The problem addressed—interpretable biomarker discovery from high-dimensional, low-sample-size multi-omics data—is important, and the broad idea of combining sparse regression with an FDR filter is sensible. If the framework were fully specified and validated with rigorous experiments, it could be of interest to the multi-omics integration community. The inclusion of four benchmark methods and an attempt at realistic synthetic data generation are commendable. However, in its current form the manuscript does not provide an implementable algorithm, does not supply valid statistical evidence for its central FDR claim, and does not disclose enough experimental detail to support the numerical results. The significance of the claimed contributions therefore cannot be assessed.

major comments (5)
  1. [§2.3–2.4, §3] The central statistical claim—Storey-FDR-controlled biomarker selection—depends on p-values for elastic-net coefficients, but the manuscript never defines how these p-values are obtained. Section 2.3 gives only the elastic-net objective (Eq. 2); Section 2.4 assumes an input list {p1,...,pm} without specifying its origin. Section 3 states that features are ranked using Storey's FDR with q-value thresholds. No post-selection inference, bootstrap/permutation procedure, independent refit, or other p-value source is described. Without valid p-values, Storey's q-values do not control FDR, and Tables 2 and 5 are unsupported.
  2. [§4.2, §5.2 (Tables 1 and 4)] The numerical superiority claim rests entirely on Tables 1 and 4, but the tables report point estimates only. No standard deviations, number of repeated runs, train/test splits, cross-validation scheme, hyperparameter values (GAT heads, MOVE weights β/λ, elastic-net λ/α), or sample sizes are given. No code or data availability is stated. The AUC values (0.93 and 0.91), F1 scores, and feature precision (0.88 and 0.87) therefore cannot be independently verified, and the claim of outperforming DIABLO, MOCAT, AMOGEL, and MOMLIN is not substantiated.
  3. [§4.1, §5.3 (Tables 2, 3, 5, 6)] The simulation in Section 4.1 embeds 'established AD pathways' into synthetic data. The same well-known genes (TREM2, APOE, BIN1, SORL1, MAPT, etc.) then appear as 'discoveries' in both the simulated and ADNI analyses. As reported, this is a known-answer test rather than evidence of novel biological discovery. The manuscript does not quantify how accurately the framework recovers the embedded ground truth, nor does it compare the ADNI findings against a chance baseline or account for the high prior probability of these genes. The 'biological relevance' highlight is therefore partly circular and not supported by the reported analysis.
  4. [§2.1–2.2, §3] The proposed ensemble is not concretely defined as an integrated pipeline. Section 2.1 gives the GAT update; Section 2.2 gives the MOVE objective; Section 2.3 gives elastic net; Section 2.4 gives Storey's FDR. There are no equations or descriptions that connect the GAT output to MOVE's latent space, nor the latent variables to the elastic-net design matrix. Architecture dimensions, training procedure, loss weighting, and the target variable y for classification/regression are all unspecified. As written, the framework is a list of standard components, not an implementable algorithm.
  5. [§4.3, §5.3 (figures and Tables 3, 6)] Several results are presented only as figure captions or placeholders, e.g., 'Top 10 Significant Genes and Interactions' (Section 4.3) and 'Biomarker Gene Map Visualization' (Section 5.3). No actual figure content is included. The interaction strengths in Tables 3 and 6 are stated without any description of how they were computed from the framework. The claim of 'visualized and interpreted' biomarker gene maps is therefore unverifiable.
minor comments (6)
  1. [§3] A citation is missing: '[?]' appears in Section 3 after 'q-value thresholds of 0.01, 0.05, and 0.1'.
  2. [References] MOCAT is inconsistently cited: Section 4.2 cites [Chen et al., 2021] while the reference list also contains Yao et al. (2024) as another MOCAT entry. The bibliography should be reconciled.
  3. [§7.2] Typo: 'Biormarker' should be 'Biomarker'.
  4. [Table captions] Tables in the text show 'T able' (e.g., 'Table 1: Performance Comparison') and column headers like 'F eature Precision' with extra spaces; these should be fixed.
  5. [§2.2, §4.3] The acronym inconsistency: Section 2.2 calls MOVE a 'Multi-Omics Variational AutoEncoder,' while Section 4.3 refers to a 'Conditional Variational Autoencoder (CV AE) for multi-omics data (MOVE).' Clarify whether the VAE is conditional or not.
  6. [References] Several references have incomplete author lists (e.g., 'H. et al. Fang', 'Y. et al. Hasin') or are formatted informally. Full author lists and journal details should be provided.

Circularity Check

1 steps flagged · score 4.0 of 10

Gene-gene interaction 'discoveries' are STRING database lookups, and the FDR-filtered biomarker lists depend on p-values that are never derived; the simulation is a planted-signal test rather than a circular derivation.

  1. renaming known result [Section 5.4 (Comparative Analysis); Tables 3 and 6; Section 2.1 (GAT)]
    "Mapping biologically validated interactions using STRING-derived strengths."

    Tables 3 and 6 present 'Top 10 Gene-Gene Interactions Based on Network Strength' and 'Top 20 High-Confidence Gene-Gene Interactions' as framework output, listing pairs such as TREM2-TYROBP (0.93) and APOE-CLU (0.91). Section 5.4 states that these strengths come from STRING, an external database of previously known/curated interactions. The GAT attention mechanism in Section 2.1, which is the component that would actually learn gene-gene interactions, is never connected to these tables. Thus the interaction network is an external input relabeled as a model discovery—a renaming of a known result rather than a derivation from the proposed ensemble.

full rationale

The core predictive pipeline is not circular: GAT attention (Sec 2.1), MOVE latent compression (Sec 2.2), elastic-net (Sec 2.3), and Storey's q-value formula (Sec 2.4) are standard components, and the paper does not fit a parameter and then repackage it as a prediction, nor does it rely on a self-citation chain to force the choice. The simulation benchmark is a known-answer test: Sec 4.1 embeds latent biological modules based on established AD pathways, so recovering TREM2/APOE/BIN1 in the synthetic data is a sanity check, not a circular 'discovery'—the model still has to learn the associations from data. The ADNI top-gene table recapitulating known AD loci is, in itself, evidence of validity rather than circularity. The main circular-adjacent defect is the gene-gene interaction result. Tables 3 and 6 present interaction strengths as framework output, but Sec 5.4 reveals they are STRING-derived strengths—imported prior knowledge, not GAT-learned attention. This is a renaming of a known result, and it inflates the circularity score. Separately, the Storey FDR step (Sec 2.4, Sec 3) assumes p-values {p1,...,pm}, but the elastic-net objective in Sec 2.3 does not produce p-values, and no post-selection, bootstrap, permutation, or independent-refit procedure is specified; the q≤0.05 gene lists are therefore statistically unsupported. I flag this as an omitted necessary step, not as a definitional circularity. The self-citations in the paper are background and non-load-bearing. On balance, the central predictive accuracy claims retain independent content, but the interaction 'discoveries' reduce to an external lookup and the FDR filtering rests on an undefined input, giving a circularity score of 4 rather than 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on unspecified hyperparameters, a vague pipeline connection, and the correctness of standard methods. The paper treats its own framework as a given without defining it, so the reader cannot evaluate what was actually built.

free parameters (4)
  • FDR q-value thresholds = 0.01, 0.05, 0.1
    Thresholds for selecting statistically significant genes in Section 3; chosen ad hoc without derivation.
  • Elastic-net regularization parameters lambda and alpha = not specified
    Elastic-net equation in Section 2.3 is given, but no values or fitting procedure are presented, so feature selection depends on unspecified free parameters.
  • GAT attention heads K and LeakyReLU parameters = not specified
    GAT equation in Section 2.1 uses K heads and other hyperparameters that are never defined.
  • MOVE loss weights beta and lambda = not specified
    The VAE objective in Section 2.2 includes beta and lambda, but no training details are given.
assumptions (5)
  • standard math Graph attention network equations are valid and applicable to gene-gene interaction modeling
    Section 2.1 presents the GAT equations from Velickovic et al. without proof, relying on prior derivation.
  • standard math Variational autoencoder objective can compress multi-omics data while preserving biological information
    Section 2.2 relies on the standard VAE loss from Kingma and Welling without adaptation details.
  • standard math Elastic-net sparse regression yields valid variable selection for HDLSS omics
    Section 2.3 uses the elastic-net equation but does not justify its behavior in the high-dimensional low-sample regime.
  • domain assumption The ADNI dataset is accessible and preprocessable as implied
    Section 5.1 describes data modalities and subject counts but provides no accession numbers or preprocessing steps.
  • ad hoc to paper The sequential pipeline of GAT, MOVE, elastic-net, and FDR is feasible and well-defined
    Section 3 describes the order of operations but gives no pseudo-code, tensor shapes, or training procedure, so the pipeline's behavior is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Interpretable Ensemble Framework for Multi-Omics Dementia Biomarker Discovery Under HDLSS Conditions." pith.science (2026). https://pith.science/paper/CW747FD2

@misc{pith2026250910527,
  author       = {Pith},
  title        = {Pith review of: An Interpretable Ensemble Framework for Multi-Omics Dementia Biomarker Discovery Under HDLSS Conditions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CW747FD2}},
  note         = {Machine review of arXiv:2509.10527}
}
read the original abstract

Biomarker discovery in neurodegenerative diseases requires robust, interpretable frameworks capable of integrating high-dimensional multi-omics data under low-sample conditions. We propose a novel ensemble approach combining Graph Attention Networks (GAT), MultiOmics Variational AutoEncoder (MOVE), Elastic-net sparse regression, and Storey's False Discovery Rate (FDR). This framework is benchmarked against state-of-the-art methods including DIABLO, MOCAT, AMOGEL, and MOMLIN. We evaluate performance using both simulated multi-omics data and the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. Our method demonstrates superior predictive accuracy, feature selection precision, and biological relevance. Biomarker gene maps derived from both datasets are visualized and interpreted, offering insights into latent molecular mechanisms underlying dementia.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 24 canonical work pages

  1. [1]

    Emergence of scaling in random networks

    Albert-L^^c3^^a1szl^^c3^^b3 Barab^^c3^^a1si and R^^c3^^a9ka Albert. Emergence of scaling in random networks. Science, 286 0 (5439): 0 509--512, 1999. doi:10.1126/science.286.5439.509

    doi:10.1126/science.286.5439.509 1999

  2. [2]

    Mocat: Multi-omics clustering and analysis tool

    Wei Chen et al. Mocat: Multi-omics clustering and analysis tool. Bioinformatics, 37 0 (12): 0 1685--1692, 2021. doi:10.1093/bioinformatics/btaa1023

    work page doi:10.1093/bioinformatics/btaa1023 2021

  3. [3]

    Fan and J

    J. Fan and J. Lv. Sure independence screening for ultrahigh dimensional feature space. Journal of the Royal Statistical Society, 2008

    work page 2008

  4. [4]

    H. et al. Fang. Amogel: Associative multi-omics graph embedding learning. BMC Bioinformatics, 2022

    work page 2022

  5. [5]

    Multiomics analysis unveils key biomarkers during dynamic progress of iav infection in mice

    Huan Fang, Yixi Xu, Bin Zhang, et al. Multiomics analysis unveils key biomarkers during dynamic progress of iav infection in mice. Frontiers in Immunology, 16: 0 1566690, 2025. doi:10.3389/fimmu.2025.1566690

    work page arXiv 2025

  6. [6]

    Y. et al. Hasin. Multi-omics approaches to disease. Genome Biology, 2017

    work page 2017

  7. [7]

    Y. et al. Iturria-Medina. Multimodal imaging-based personalized predictive modeling of clinical scores for the alzheimer’s disease continuum. Scientific Reports, 2018

    work page 2018

  8. [8]

    Karczewski and M.P

    K.J. Karczewski and M.P. Snyder. Integrative omics for precision medicine. Nature Reviews Genetics, 2018

    work page 2018

Show all 29 references
  1. [9]

    Kingma and M

    D.P. Kingma and M. Welling. Auto-encoding variational bayes. ICLR, 2014

    2014

  2. [10]

    Genetic meta-analysis of diagnosed alzheimer’s disease identifies new risk loci and implicates a , tau, immunity and lipid processing

    Bryan W Kunkle, Benjamin Grenier-Boley, Rebecca Sims, et al. Genetic meta-analysis of diagnosed alzheimer’s disease identifies new risk loci and implicates a , tau, immunity and lipid processing. Nature Genetics, 51 0 (3): 0 414--430, 2019. doi:10.1038/s41588-019-0358-2

    2019 doi

  3. [11]

    Meta-analysis of 74,046 individuals identifies 11 new susceptibility loci for alzheimer's disease

    Jean-Charles Lambert, Carla A Ibrahim-Verbaas, Denise Harold, et al. Meta-analysis of 74,046 individuals identifies 11 new susceptibility loci for alzheimer's disease. Nature Genetics, 45 0 (12): 0 1452--1458, 2013. doi:10.1038/ng.2802

    2013 doi

  4. [12]

    LeCun, Y

    Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. Nature, 2015

    2015

  5. [13]

    H. et al. Lee. Metabolomic profiling in alzheimer’s disease. Cell Reports, 2022

    2022

  6. [14]

    Tackling the widespread and critical impact of batch effects in high-throughput data

    Jeffrey T Leek, Robert B Scharpf, H^^c3^^a9ctor C Bravo, et al. Tackling the widespread and critical impact of batch effects in high-throughput data. Nature Reviews Genetics, 11 0 (10): 0 733--739, 2010. doi:10.1038/nrg2825

    2010 doi

  7. [15]

    Multi-omics data integration and analysis using deep learning approaches

    Yifan Li, Hao Wu, and Wei Zhang. Multi-omics data integration and analysis using deep learning approaches. Nature Reviews Genetics, 23 0 (1): 0 1--18, 2022

    2022

  8. [16]

    Momlin: Multi-omics machine learning integration for neurodegeneration

    Adeel Rashid and Kumar Selvarajoo. Momlin: Multi-omics machine learning integration for neurodegeneration. Journal of Translational Medicine, 22 0 (1): 0 1--12, 2024

    2024

  9. [17]

    Diablo: an integrative approach for identifying key molecular drivers from multi-omics assays

    Archana Singh, Beno^^c3^^aet Gautier, et al. Diablo: an integrative approach for identifying key molecular drivers from multi-omics assays. Bioinformatics, 35 0 (17): 0 3055--3062, 2019. doi:10.1093/bioinformatics/btz105

    2019 doi

  10. [18]

    A direct approach to false discovery rates

    John D Storey. A direct approach to false discovery rates. Journal of the Royal Statistical Society: Series B, 64 0 (3): 0 479--498, 2002. doi:10.1111/1467-9868.00346

    2002

  11. [19]

    Amogel: Associative multi-omics graph embedding learning

    Yifan Tan, Hao Fang, et al. Amogel: Associative multi-omics graph embedding learning. BMC Bioinformatics, 26 0 (1): 0 1--14, 2025

    2025

  12. [20]

    Graph attention networks

    Petar Veli^^c4^^8dkovi^^c4^^87, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. In International Conference on Learning Representations (ICLR), 2018

    2018

  13. [21]

    P. et al. Velickovic. Graph attention networks. In ICLR, 2018

    2018

  14. [22]

    Wang et al

    Z. Wang et al. Multi-omics approaches for systems biology. Nature Reviews Genetics, 2021 a

    2021

  15. [23]

    Move: Manifold optimization for variational embedding of multi-omics data

    Zhen Wang, Yuxin Liu, and Wei Zhang. Move: Manifold optimization for variational embedding of multi-omics data. Bioinformatics, 37 0 (Suppl 1): 0 i132--i140, 2021 b

    2021

  16. [24]

    How powerful are graph neural networks? International Conference on Learning Representations (ICLR), 2020

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? International Conference on Learning Representations (ICLR), 2020

    2020

  17. [25]

    Multiomics-based prediction of trajectories of global cognitive functioning in the adni cohort

    Shijun Xu, Huan Lei, Hao Zhang, et al. Multiomics-based prediction of trajectories of global cognitive functioning in the adni cohort. Alzheimer’s & Dementia, 19 0 (Suppl. 15): 0 e077394, 2025. doi:10.1002/alz.077394

    2025 doi

  18. [26]

    Mocat: Multi-omics clustering and annotation toolkit for disease subtyping

    Lin Yao, Minji Kim, and Hao Chen. Mocat: Multi-omics clustering and annotation toolkit for disease subtyping. Nature Communications, 15 0 (1): 0 1123, 2024

    2024

  19. [27]

    B. et al. Zhang. Transcriptomic and proteomic integration reveals biomarkers in alzheimer’s disease. Nature Neuroscience, 2020

    2020

  20. [28]

    Ensemble learning for multi-omics data integration in cancer prognosis

    Lei Zhang, Xin Wang, and Jie Chen. Ensemble learning for multi-omics data integration in cancer prognosis. Briefings in Bioinformatics, 22 0 (6): 0 bbab345, 2021

    2021

  21. [29]

    Regularization and variable selection via the elastic net

    Hui Zou and Trevor Hastie. Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society: Series B, 67 0 (2): 0 301--320, 2005

    2005

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.