Cancer : et si le piège venait de la méthode de mesure ?

Hypothèse formelle

If microarray gene expression data, which is compositional (i.e., each sample’s total measured intensity is constrained to a constant sum), is preprocessed using a centered log-ratio (CLR) or isometric log-ratio (ILR) transformation before training a deep learning classifier (CNN or LSTM), then the false positive rate (FPR) of feature selection will be significantly lower and the balanced accuracy of cancer subtype classification will be significantly higher compared to standard normalization (RMA, quantile), because the log-ratio transformation removes the spurious negative correlation bias induced by the closure constraint, thereby revealing the true covariance structure among genes.

Titre donné par l’affûteur : Log-Ratio Transformation of Microarray Gene Expression Data Reduces False Positive Feature Selection and Improves Cancer Subtype Classification Accuracy in High-Dimensional, Low-Sample-Size Settings

Contre-preuves

1. This paper concludes that data normalization techniques are not helpful in classification problems with homogeneous data. This could be interpreted as a minor counter-argument if gene expression data is considered “homogeneous” in terms of its compositional structure. However, the paper deals with sensor data, not gene expression, and the hypothesis specifically argues that gene expression is NOT homogeneous due to the closure problem.

   Gravité mineureHomogeneous Data Normalization and Deep Learning: A Case Study in Human Activity Classification (2020)

2. This paper introduces synthetic RNA standards for normalization of HTS data, suggesting that external spike-ins can provide absolute quantification, which could circumvent the compositional problem. This is a minor counter-point because the hypothesis focuses on standard normalization methods (RMA, TPM) that do not use spike-ins, and the paper does not invalidate the compositional nature of data without spike-ins.

   Gravité mineureComputational and analytical framework for small RNA profiling by high-throughput sequencing. (2009)

Inconnues et conditions de validité

Inconnues identifiées

• We do not know whether the optimal orthonormal basis for ILR transformation is dataset-dependent.
• We do not know whether the stability of deep learning classifiers to the singular covariance matrix after CLR is sufficient without explicit regularization.
• We do not know whether the magnitude of the FPR reduction is consistent across different cancer types and microarray platforms.

Conditions de validité

• The total intensity per sample must have a coefficient of variation (CV) less than 15% across samples in the dataset.Justification : If the CV exceeds 15%, the closure constraint is weak, and the compositional artifact is negligible; the proposed mechanism would not apply.
• The number of samples (n) must be at least 30, and the number of features (p) must be at least 10 times n (p/n > 10).Justification : The HDLSS regime is defined by p >> n. Below p/n = 10, standard regularization methods may already suffice, and the compositional artifact is less dominant.
• The deep learning classifier must use a regularization technique (e.g., dropout, weight decay, or early stopping) to handle the singular covariance matrix after CLR transformation.Justification : Without regularization, the classifier may fail to converge or overfit severely due to the rank-deficient covariance matrix.
• Genes with zero expression in more than 20% of samples must be filtered out or imputed before log-ratio transformation.Justification : Log-ratio transformations are undefined for zeros. A pseudo-count of 0.5 is used for remaining zeros, but high sparsity can bias results.

Mécanisme proposé

Chaîne causale

1. Step 1: Raw microarray probe intensities for a given sample are normalized by standard methods (RMA, quantile) which do not enforce a constant sum constraint, or by CoDA methods (CLR, ILR) which explicitly map the data from the simplex to real space via log-ratios.
2. Step 2: For standard normalization, the closure constraint (sum of intensities per sample is approximately constant) remains implicit, inducing a negative correlation bias among all genes (Pearson’s paradox). For CoDA, the log-ratio transformation removes this bias by making the covariance structure independent of the total sum.
3. Step 3: A feature selection step (e.g., ANOVA F-test or mutual information) is applied to the normalized data. Under standard normalization, many genes appear differentially expressed due to the spurious negative correlation, inflating the FPR. Under CoDA, the true differentially expressed genes are ranked higher.
4. Step 4: A deep learning classifier (CNN or LSTM) is trained on the top-k selected features. With CoDA preprocessing, the input feature space contains fewer false positives, reducing overfitting and improving generalization, measured as higher balanced accuracy on held-out test sets.
5. Step 5: The stability of the selected feature set across cross-validation folds is quantified via the Jaccard index. CoDA preprocessing is expected to yield higher stability because the selection is driven by true biological signal rather than compositional artifacts.

Postulats clés

• Assumption 1: The total measured intensity per microarray sample (after background correction) is approximately constant across samples within a given dataset (coefficient of variation < 10%).
• Assumption 2: The compositional nature of the data is the dominant source of spurious correlations in the high-dimensional, low-sample-size (HDLSS) regime.
• Assumption 3: The deep learning classifiers (CNN, LSTM) are capable of learning non-linear decision boundaries from the log-ratio transformed features without requiring the covariance matrix to be full rank.
• Assumption 4: Zero or near-zero expression values can be handled by a pseudo-count replacement (e.g., 0.5) or a filter (e.g., remove genes with mean expression < 1 count per million) without introducing significant bias.

Cadre théorique

Compositional Data Analysis (CoDA) and the Aitchison geometry of the simplex, applied to high-dimensional biological data.

Variables

Prédictions falsifiables

1. 1

   The FPR of feature selection (at a nominal alpha=0.05) will be at least 15% lower for CLR/ILR-preprocessed data compared to RMA-preprocessed data.

   Borne quantitative

   Reduction of 15-30% in FPR, measured over 100 Monte Carlo simulations of synthetic data with known ground truth and fold-changes of 1.2-1.5.

   Méthode de mesure

   Synthetic microarray data generated from a known covariance structure with a subset of truly differentially expressed genes. FPR computed as (number of false positives) / (total number of non-differentially expressed genes).Test statistique Paired t-test (or Wilcoxon signed-rank if normality violated) on FPR values from 100 simulation replicates, alpha=0.05, one-sided.

   Hypothèse nulle

   H0: The FPR for CLR/ILR is equal to or greater than the FPR for RMA (one-sided test).

2. 2

   The balanced accuracy of cancer subtype classification will be at least 5 percentage points higher for CLR/ILR-preprocessed data compared to RMA-preprocessed data, when using a CNN classifier on the top 200 selected features.

   Borne quantitative

   Increase of 5-12 percentage points in balanced accuracy, measured on 3 public microarray datasets (e.g., GSE2034, GSE1456, GSE3494) using repeated 10-fold cross-validation (5 repeats).

   Méthode de mesure

   Balanced accuracy = (sensitivity + specificity) / 2, computed on held-out test folds. Hyperparameters of CNN are tuned on a separate validation set in a blinded manner.Test statistique Paired t-test on the 50 cross-validation accuracy values (10 folds x 5 repeats) for each dataset, alpha=0.05, one-sided, with Bonferroni correction for 3 datasets.

   Hypothèse nulle

   H0: The balanced accuracy for CLR/ILR is equal to or less than the balanced accuracy for RMA (one-sided test).

3. 3

   The Jaccard index of the top 200 selected features across cross-validation folds will be at least 0.10 higher for CLR/ILR-preprocessed data compared to RMA-preprocessed data.

   Borne quantitative

   Increase of 0.10-0.25 in mean Jaccard index, measured over 10-fold cross-validation.

   Méthode de mesure

   For each fold, select top 200 features by ANOVA F-test. Compute pairwise Jaccard index between all fold pairs. Average over all pairs.Test statistique Paired t-test on the mean Jaccard indices from 5 repeated cross-validation runs, alpha=0.05, one-sided.

   Hypothèse nulle

   H0: The mean Jaccard index for CLR/ILR is equal to or less than the mean Jaccard index for RMA (one-sided test).

Protocole expérimental

Références

8 références, toutes issues de Semantic Scholar. Une référence vérifiée est un article qui existe et qui est indexé par Semantic Scholar. Cela ne veut pas dire qu’il confirme l’idée.

1. 1
   T. Quinn, T. Crowley, M. Richardson (2018). Benchmarking differential expression analysis tools for RNA-Seq: normalization-based vs. log-ratio transformation-based methods.appui direct · 68 citations · doi:10.1186/s12859-018-2261-8Ce qu’en retient le documentaliste Log-ratio transformation-based methods (e.g., ALDEx2) identify differentially expressed genes with high precision, outperforming standard normalization-based methods in some contexts, confirming that compositional principles apply to sequencing data.Pertinence Directly supports the core mechanism: demonstrates that RNA-Seq data (like gene expression) is compositional and that log-ratio transformations (CoDA) can replace standard normalization for differential expression analysis. This validates the mathematical bridge between compositional data analysis and gene expression.
2. 2
   Antoine Godichon-Baggioni, C. Maugis-Rabusseau, A. Rau (2017). Clustering transformed compositional data using K-means, with applications in gene expression and bicycle sharing system data.appui direct · 58 citations · doi:10.1080/02664763.2018.1454894Ce qu’en retient le documentaliste Clustering on log-ratio transformed gene expression data (using the coseq package) yields meaningful groups of co-expressed genes, validating the compositional approach for gene expression analysis.Pertinence Directly applies CoDA (log-ratio transformations) to gene expression data for clustering, demonstrating that compositional data analysis is valid and useful for gene expression profiles. Supports the idea that gene expression data can be treated as compositional.
3. 3
   Waleed Ali, Faisal Saeed (2023). Hybrid Filter and Genetic Algorithm-Based Feature Selection for Improving Cancer Classification in High-Dimensional Microarray Data.appui indirect · 71 citations · doi:10.3390/pr11020562Ce qu’en retient le documentaliste High-dimensional microarray data suffers from slow performance and noisy features; hybrid filter-genetic feature selection improves classification accuracy, but the underlying normalization issue (compositionality) is not addressed.Pertinence Indirectly supports the hypothesis by highlighting the feature selection paradox in high-dimensional microarray data. The paper proposes a hybrid feature selection method to address this, but does not consider compositional normalization. This underscores the need for the CoDA-based approach proposed in the hypothesis.
4. 4
   Francielly Morais-Rodrigues, Rita Silvério-Machado, R. Kato et al. (2019). Analysis of the microarray gene expression for breast cancer progression after the application modified logistic regression..appui indirect · 43 citations · doi:10.1016/j.gene.2019.144168Ce qu’en retient le documentaliste Microarray data is characterized by low sample size and high dimensionality, creating a non-square matrix and incomplete rank, which complicates classification. This motivates the need for better preprocessing (like CoDA).Pertinence Indirectly supports by describing the HDLSS problem in microarray data (low sample size, large feature number, non-square matrix). This is the exact setting where the hypothesis claims CoDA can resolve artifacts. The paper does not use CoDA, highlighting the gap.
5. 5
   Joseph M. De Guia, M. Devaraj, C. Leung (2019). DeepGx: Deep Learning Using Gene Expression for Cancer Classification.appui indirect · 78 citations · doi:10.1145/3341161.3343516Ce qu’en retient le documentaliste Transforming gene expression values into 2D images for CNN-based cancer classification achieves good performance, but does not account for compositional artifacts in the input data.Pertinence Indirectly supports by showing that deep learning (transforming 1D gene expression to 2D images) can be effective for cancer classification. The hypothesis proposes a different preprocessing (CoDA) before deep learning, which could potentially improve upon this approach.
6. 6
   Can Chen, Scott T. Weiss, Yang-Yu Liu (2022). Graph convolutional network-based feature selection for high-dimensional and low-sample size data.appui indirect · 39 citations · doi:10.1093/bioinformatics/btad135Ce qu’en retient le documentaliste Feature selection under HDLSS is challenging due to overfitting; GRACES uses graph convolutional networks to select important features, but does not address the compositional normalization issue.Pertinence Indirectly supports by addressing the HDLSS feature selection challenge with a deep learning method (GRACES). The hypothesis argues that CoDA preprocessing can reduce false positive feature selection, which is complementary to this approach.
7. 7
   F. Gaidies, Y. E. Morneau, D. Petts et al. (2020). Major and trace element mapping of garnet: Unravelling the conditions, timing and rates of metamorphism of the Snowcap assemblage, west‐central Yukon.appui par analogie · 32 citations · doi:10.1111/jmg.12562Ce qu’en retient le documentaliste Compositional data from garnet (major and trace elements) are used to model metamorphic conditions, showing that compositional data analysis is a mature and effective approach in geochemistry.Pertinence Analogous support from geochemistry: demonstrates the use of compositional data (major and trace elements in garnet) to infer geological processes. This is the domain from which CoDA principles originate, supporting the validity of the methodological transfer proposed in the hypothesis.
8. 8
   J. Qian, Y. Kluger, Haiyuan Yu et al. (2003). Identification and correction of spurious spatial correlations in microarray data..appui indirect · 31 citations · doi:10.2144/03351BM03Ce qu’en retient le documentaliste Microarray data suffer from systematic spatial artifacts that can be corrected by normalization methods like SNOMAD, highlighting the importance of proper preprocessing to avoid false correlations.Pertinence Indirectly supports by showing that microarray data contains systematic errors (spurious spatial correlations) that require correction. This is analogous to the “spurious correlations” problem in compositional data (Pearson’s paradox), which the hypothesis aims to address via CoDA.

Nouveauté

Score de nouveauté : 0,85 sur 1 · Verdict : jugée nouvelle

Ce score est attribué par un agent à partir des travaux trouvés. C’est une estimation, pas une mesure. Comment ce score est produit

Travaux les plus proches

• Benchmarking differential expression analysis tools for RNA-Seq: normalization-based vs. log-ratio transformation-based methods (2018)Différence avec l’idée : This paper applies log-ratio transformations (CoDA) to RNA-Seq count data for differential expression analysis, but does not extend to cancer classification with deep learning (LSTM-DBN, CNN) nor address the HDLSS feature selection paradox. The hypothesis proposes a novel pipeline combining CoDA preprocessing with deep learning classifiers for cancer subtype classification, which is not explored here.
• Clustering transformed compositional data using K-means, with applications in gene expression and bicycle sharing system data (2017)Différence avec l’idée : This paper applies CoDA (log-ratio transformations) to gene expression data for clustering, not classification. The hypothesis specifically targets cancer classification using deep learning and addresses the normalization paradox in feature selection, which is a distinct methodological contribution.
• DeepGx: Deep Learning Using Gene Expression for Cancer Classification (2019)Différence avec l’idée : This paper uses deep learning (image-based transformation) for cancer classification from gene expression, but does not apply CoDA preprocessing or address the compositional nature of the data. The hypothesis proposes a fundamentally different preprocessing step (log-ratio transformation) that could resolve artifacts from standard normalization.

Lacunes et données

Lacunes identifiées

• No paper in the provided list directly applies CoDA preprocessing (CLR/ILR) to deep learning classifiers (LSTM-DBN, CNN) for cancer classification. This remains an unexplored gap.
• No paper in the list provides a direct comparison of CoDA-based normalization vs. standard normalization (RMA, TPM) for cancer subtype classification accuracy. This is a key experimental gap.
• The biological interpretability of log-ratio transformed features (whether they correspond to known pathways) is not addressed in any of the provided papers. This remains an epistemic gap.

Données disponibles

• Paper 3e150fcc (Quinn et al., 2018) provides a benchmarking framework and code for comparing normalization vs. log-ratio transformation methods for RNA-Seq data, which could be adapted for the proposed hypothesis.
• Paper 416d657d7e (Ali & Saeed, 2023) uses microarray datasets for cancer classification, but specific dataset identifiers (GSE numbers) are not provided in the abstract. Would need full text.
• Paper 6bc5941d (De Guia et al., 2019) uses RNA-seq from the Pan-Cancer Atlas, which is a publicly available dataset that could be used to test the hypothesis.

Synthèse du panel

Note de consensus : 6,29/10 Moyenne des cinq notes, pondérée par la confiance que chaque relecteur déclare.

Verdict du méta-relecteur : publier

Points d’accord

• The hypothesis identifies a fundamental and often overlooked problem: the compositional nature of microarray data and the negative correlation bias induced by the constant-sum constraint (Pearson’s paradox).
• The proposed causal chain (standard normalisation → correlation bias → elevated FPR → poor generalisation) is logical, well-articulated, and theoretically grounded in Aitchison geometry.
• The three-phase protocol with explicit GO/NO-GO/PIVOT criteria is rigorous and permits objective decision-making, avoiding exploratory drift.
• The use of synthetic data with known ground truth (Phase 1) before proceeding to real data considerably strengthens the internal validity of the protocol.
• The addressable market (molecular pathology laboratories, CROs) is clear and quantifiable, with a net competitive advantage (simple software modification, no hardware change).

Points de désaccord

• Disagreement regarding the validity of the central hypothesis (Assumption 1): the Methodologist and the Domain Expert consider that the constant-sum constraint is partially artificial after RMA normalisation, whereas the Contrarian asserts that the coefficient of variation of total intensity is very likely >10% in real data, which would collapse the predicted effect.
• Disagreement regarding the dominance of compositional bias over other confounders: the Contrarian argues that batch effects, technical noise, and cross-hybridisation generate spurious correlations far stronger than the simple sum constraint, rendering the reduction in FPR potentially undetectable.
• Disagreement regarding generalisation: the Methodologist and the Funding Strategist note that the scope is too narrow (three breast-cancer datasets) and that the hypothesis must be tested on other cancers and technologies (RNA-seq), whereas the Domain Expert recommends focusing on a specific context to validate the mechanism.
• Disagreement regarding TRL and maturity: the Funding Strategist considers that the current TRL (2–3) is too low for competitive funding, whereas the Industrialist sees immediate commercial potential for the new signatures under development.
• Disagreement regarding the handling of zeros and negative values: the Domain Expert and the Contrarian note that applying CLR/ILR to RMA data (which may contain negative values) poses an unresolved methodological problem, with a risk of substantial bias through the use of pseudo-counts.

Chemin critique

The most critical factor determining the success or failure of this hypothesis is the validation of Assumption 1 (coefficient of variation of total intensity < 10%) on real microarray data. If this central hypothesis is false, the predicted effect size (reduction in FPR of 15–30%) collapses, and the log-ratio transformation could introduce more bias than it corrects. Phase 1 (synthetic data) and Phase 2 (first real dataset) must therefore be designed to test this condition specifically, with a clear pivot to an alternative approach (e.g., robust batch-effect normalisation plus standard log transformation) if the CV exceeds 10%.

Recommandation finale

The panel recommends publication of this protocol provided that the following conditions are clarified in the final manuscript: (1) a pre-registration on OSF or AsPredicted with a detailed static analysis plan, including an a priori power calculation for Phase 2; (2) the addition of at least one dataset from another cancer type (e.g., lung, GSE18842) or from RNA-seq (e.g., TCGA-BRCA) to strengthen external validity; (3) a dedicated section on the handling of zeros and negative values in the application of CLR/ILR, with a sensitivity analysis comparing different strategies (pseudo-count, filter, imputation). The protocol is methodologically rigorous, the hypothesis is original and mechanistic, and the addressable market is clear. The risks identified (validity of Assumption 1, dominance of batch effects) are correctly anticipated by the proposed pivots. Publication of this protocol will allow the community to test and falsify the hypothesis in a transparent and reproducible manner.