Cancer: could the background noise of DNA microarrays conceal the true signals?

Formal statement

If microarray gene expression data from cancer samples are preprocessed using a centered log-ratio (CLR) transformation (a CoDA method) instead of standard normalization (e.g., RMA), then the balanced accuracy of cancer subtype classification by both linear (LASSO, SVM) and deep learning (CNN, LSTM-DBN) classifiers will increase by at least 8 percentage points (pp) on held-out test sets, because CLR removes spurious correlations induced by the compositional closure constraint (sum-to-constant) of total mRNA, which standard normalization fails to address.

Title given by the sharpener: Compositional Data Analysis (CoDA) Preprocessing Resolves Normalization Artifacts in High-Dimensional, Low-Sample-Size Cancer Gene Expression Classification: A Cross-Domain Methodological Transfer

Counter-evidence

1. Scanpy, a widely used toolkit for single-cell analysis, does not explicitly use CoDA-based preprocessing (e.g., CLR) as a default or recommended step for normalization before classification. This suggests that the current standard practice does not view compositionality as a critical issue for single-cell data, potentially contradicting the hypothesis’s claim that ignoring it introduces significant artifacts.

   Severity minorSCANPY: large-scale single-cell gene expression data analysis (2018)

2. edgeR, a standard tool for differential expression, uses a model (negative binomial) that accounts for library size (total counts) but does not use log-ratio transformations from CoDA. Its widespread success and validation suggest that the “closure problem” may be adequately addressed by existing normalization methods (e.g., TMM) in many contexts, potentially reducing the perceived need for a full CoDA framework.

   Severity minoredgeR: a Bioconductor package for differential expression analysis of digital gene expression data (2009)

3. Shows that false positives in differential expression (a key concern of the hypothesis) can be substantially reduced by a simple technique (winsorization) rather than a full CoDA transformation. This suggests that the false positive problem may have simpler solutions than the proposed methodological transfer.

   Severity addressableWinsorization greatly reduces false positives by popular differential expression methods when analyzing human population samples (2024)

Unknowns and boundary conditions

Known unknowns

• We do not know whether the optimal imputation method for zeros is the same across all cancer types (e.g., solid tumors vs. liquid tumors).
• We do not know whether the effect size (BAcc gain) is larger for microarray data than for RNA-seq data, as RNA-seq has different noise characteristics.
• We do not know whether the CLR transformation introduces a dependency on the number of genes (D) that affects classification stability for very high-dimensional data (D > 50,000).

Boundary conditions

• The proportion of zero entries in the expression matrix must be less than 20% before imputation.Rationale: Above 20% zeros, imputation methods introduce high bias and variance, and the CLR geometric mean becomes unstable; the hypothesis may not hold for single-cell RNA-seq data where zero inflation is >50%.
• The number of samples (n) must be at least 30 per class for the classification task.Rationale: Below n=30, the variance of BAcc estimates becomes too large (>10 pp) to detect the predicted 8 pp effect with adequate statistical power (β>0.8 at α=0.01).
• The number of genes (p) must be between 10,000 and 50,000.Rationale: Below 10,000 genes, the compositional effect is weak (few spurious correlations); above 50,000, computational instability of the CLR (division by geometric mean of many near-zero values) may introduce artifacts.
• The classification task must be categorical (cancer subtype or condition), not regression (e.g., survival time).Rationale: The hypothesis is formulated for classification accuracy; for regression, the compositional effect on continuous outcomes is not yet characterized.

Proposed mechanism

Causal chain

1. Step 1: Raw microarray gene expression data are subject to the closure constraint: each sample’s total measured mRNA is an arbitrary constant (e.g., total fluorescence), making each gene’s expression a relative proportion, not an absolute quantity.
2. Step 2: Standard normalization (RMA, quantile) assumes that total mRNA is irrelevant and applies linear scaling or rank-based adjustments, which do not break the unit-sum constraint and preserve the simplex geometry, leading to spurious negative correlations between highly expressed genes (Pearson’s paradox).
3. Step 3: CoDA transformation (CLR) maps the data from the simplex (D-1 dimensions) to real space (D dimensions) via log-ratios relative to the geometric mean of all genes, eliminating the closure constraint and enabling standard Euclidean geometry for machine learning.
4. Step 4: The removal of spurious correlations reduces the number of false-positive differentially expressed genes (features) and improves the signal-to-noise ratio for classifiers, leading to higher balanced accuracy on held-out test sets.
5. Step 5: The effect is independent of classifier complexity because it operates at the data representation level, not at the model architecture level; linear and deep learning models benefit equally from the improved geometry.

Key assumptions

• Assumption 1: The total mRNA per cell (or per sample) is not biologically informative for the classification task; i.e., the absolute abundance of mRNA is irrelevant, only relative proportions matter.
• Assumption 2: The number of zeros in the expression matrix is low enough (< 20% of entries) that imputation does not dominate the transformation; if zero proportion > 20%, the CLR may be unstable.
• Assumption 3: The relationship between gene expression and cancer subtype is linear in log-ratio space; i.e., the decision boundary is approximately linear after CLR transformation.
• Assumption 4: The estimated cellularity (from ESTIMATE/CIBERSORT) does not correlate with classification improvement; if |r| > 0.3 between BAcc gain and cellularity, the hypothesis must be revised to account for cell-type composition effects.

Theoretical framework

Compositional Data Analysis (Aitchison geometry on the simplex) and its application to high-throughput molecular biology (Fernandes et al., 2014; Gloor et al., 2017).

Variables

Falsifiable predictions

1. 1

   On the TCGA breast cancer dataset (BRCA, 4 subtypes, n=500), applying robust CLR (with multiplicative replacement for zeros) before LASSO classification will increase balanced accuracy by at least 8 pp compared to RMA-normalized data, on a 20% held-out test set.

   Quantitative bound

   BAcc(CLR+LASSO) ≥ BAcc(RMA+LASSO) + 8 pp, with 95% CI lower bound ≥ 5 pp.

   Measurement method

   5-fold cross-validation repeated 10 times; BAcc computed as (sensitivity + specificity)/2 per subtype, macro-averaged.Statistical test Paired one-sided t-test on the 50 BAcc differences (10 repeats × 5 folds), α=0.01, with Benjamini-Hochberg correction for multiple cancer types.

   Null hypothesis

   H0: μ_BAcc(CLR) - μ_BAcc(RMA) ≤ 0 (no improvement or degradation).

2. 2

   The number of features selected by LASSO (at lambda.1se) will be at least 30% lower for CLR-transformed data than for RMA-normalized data on the same TCGA-BRCA dataset.

   Quantitative bound

   N_features(CLR) ≤ 0.7 × N_features(RMA), with median reduction ≥ 40%.

   Measurement method

   Count of non-zero coefficients in LASSO model at lambda.1se, averaged over 10 cross-validation folds.Statistical test Wilcoxon signed-rank test (paired), one-sided, α=0.01.

   Null hypothesis

   H0: median(N_features(CLR)) ≥ median(N_features(RMA)).

3. 3

   In a simulated dataset with known ground-truth differentially expressed genes (20 true positives, 1000 noise genes), CLR preprocessing will reduce the false positive rate (FPR) of differential expression detection (by limma) by at least 50% compared to RMA preprocessing, at a fixed true positive rate of 80%.

   Quantitative bound

   FPR(CLR) ≤ 0.5 × FPR(RMA), with absolute FPR(CLR) < 0.05.

   Measurement method

   Simulate 100 datasets with n=50 samples, p=1020 genes, 20 true DE genes with log2 fold-change = 2. Apply limma to both RMA and CLR data; compute FPR at the threshold achieving 80% TPR.Statistical test Paired one-sided t-test on logit-transformed FPRs across 100 simulations, α=0.01.

   Null hypothesis

   H0: FPR(CLR) ≥ FPR(RMA).

4. 4

   The gain in balanced accuracy (ΔBAcc = BAcc(CLR) - BAcc(RMA)) will not correlate with estimated sample cellularity (from ESTIMATE) across TCGA cancer types (|r| < 0.3).

   Quantitative bound

   Pearson |r| < 0.3, with 95% CI upper bound < 0.5.

   Measurement method

   For each of 10 TCGA cancer types (e.g., BRCA, LUAD, COAD, etc.), compute ΔBAcc (CLR vs. RMA) using LASSO, and the median ESTIMATE score. Compute Pearson correlation across the 10 cancer types.Statistical test Fisher’s z-transformation test for correlation, one-sided (testing |r| < 0.3), α=0.05.

   Null hypothesis

   H0: |r| ≥ 0.3 (significant correlation indicating cellularity confound).

5. 5

   On a non-cancerous gene expression dataset (Arabidopsis thaliana stress response, n=100, 4 conditions), the BAcc gain from CLR vs. RMA will be at least 5 pp, demonstrating generalizability.

   Quantitative bound

   ΔBAcc ≥ 5 pp, with 95% CI lower bound ≥ 2 pp.

   Measurement method

   Same as Prediction 1, but on the Arabidopsis dataset (GEO: GSEXXXX).Statistical test One-sample one-sided t-test on the 50 ΔBAcc values, α=0.01.

   Null hypothesis

   H0: μ_ΔBAcc ≤ 0.

Experimental protocol

References

10 references, all from Semantic Scholar. A verified reference is a paper that exists and is indexed by Semantic Scholar. It does not mean that the paper confirms the idea.

1. 1
   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.indirect support · 58 citations · doi:10.1080/02664763.2018.1454894What the librarian takes from it CoDA-based clustering (using log-ratio transformations) is effective for identifying groups of co-expressed genes from high-throughput sequencing data.Relevance Demonstrates that gene expression data can be treated as compositional data and that CoDA transformations (e.g., CLR) are applicable, supporting the core premise of the hypothesis.
2. 2
   A. Fernandes, J. N. Reid, Jean M. Macklaim et al. (2014). Unifying the analysis of high-throughput sequencing datasets: characterizing RNA-seq, 16S rRNA gene sequencing and selective growth experiments by compositional data analysis.indirect support · 1,106 citations · doi:10.1186/2049-2618-2-15What the librarian takes from it Standard normalization methods (e.g., RPKM, TPM) do not account for compositionality, and log-ratio transformations (e.g., ALR, CLR) are necessary for valid statistical analysis of sequencing data.Relevance Establishes that RNA-seq and other sequencing data are compositional and that ignoring this leads to spurious results, directly supporting the “closure problem” argument in the hypothesis.
3. 3
   Ionas Erb, G. Gloor, T. Quinn (2020). Editorial: Compositional data analysis and related methods applied to genomics—a first special issue from NAR Genomics and Bioinformatics.indirect support · 10 citations · doi:10.1093/nargab/lqaa103What the librarian takes from it CoDA, originally developed for geochemistry, is increasingly recognized as essential for analyzing genomic data (e.g., RNA-seq, ChIP-seq) due to its compositional nature.Relevance Provides a high-level endorsement of CoDA for genomics, reinforcing the idea that CoDA principles from geochemistry are transferable to genomic data.
4. 4
   N. Štefelová, J. Palarea‐Albaladejo, K. Hron (2021). Weighted pivot coordinates for partial least squares‐based marker discovery in high‐throughput compositional data.indirect support · 8 citations · doi:10.1002/sam.11514What the librarian takes from it Using weighted pivot coordinates (a CoDA approach) enhances the discovery of biological markers in high-throughput compositional data compared to standard normalization.Relevance Shows that CoDA-based transformations (pivot coordinates) improve marker discovery in high-throughput data, supporting the claim that CoDA can reduce false positives in feature selection.
5. 5
   M. Badri, Zachary D. Kurtz, Christian L. Müller et al. (2018). Normalization methods for microbial abundance data strongly affect correlation estimates.support by analogy · 22 citations · doi:10.1101/406264What the librarian takes from it Log-ratio and variance-stabilization transformations provide the most consistent correlation estimates across experiments, while other methods introduce artifacts.Relevance Demonstrates in microbiome data (another compositional domain) that normalization choices (including log-ratio from CoDA) strongly affect correlation estimates, analogous to the proposed effect on gene expression feature selection.
6. 6
   Sophie J. Weiss, Z. Xu, S. Peddada et al. (2017). Normalization and microbial differential abundance strategies depend upon data characteristics.support by analogy · 1,676 citations · doi:10.1186/s40168-017-0237-yWhat the librarian takes from it Different normalization methods (including those based on CoDA) lead to different differential abundance results; the choice depends on data characteristics.Relevance Shows that normalization and differential abundance methods for compositional microbiome data must be chosen carefully, supporting the hypothesis that standard normalization in gene expression may be suboptimal.
7. 7
   Alexander R. Bennett, Jon Lundstrøm, S. Chatterjee et al. (2025). Compositional data analysis enables statistical rigor in comparative glycomics.support by analogy · 13 citations · doi:10.1038/s41467-025-56249-3What the librarian takes from it Applying CoDA to comparative glycomics data eliminates spurious “decreases” and high false-positive rates caused by ignoring compositionality.Relevance Directly analogous: shows that CoDA resolves spurious correlations and false positives in another 'omics field (glycomics), strongly supporting the hypothesis that the same approach would benefit gene expression analysis.
8. 8
   Abdulaziz Alenazi (2021). A review of compositional data analysis and recent advances.indirect support · 22 citations · doi:10.1080/03610926.2021.2014890What the librarian takes from it CoDA provides a rigorous statistical framework for analyzing compositional data, with transformations (CLR, ILR) that enable standard multivariate analysis.Relevance Provides a comprehensive review of CoDA methods, including log-ratio transformations, which are central to the proposed methodological transfer.
9. 9
   E. Grunsky, D. Arne (2020). Mineral-resource prediction using advanced data analytics and machine learning of the QUEST-South stream-sediment geochemical data, southwestern British Columbia, Canada.support by analogy · 21 citations · doi:10.1144/geochem2020-054What the librarian takes from it Multivariate statistical and machine learning methods, applied to compositional geochemical data, can successfully predict mineral resources.Relevance Demonstrates the use of machine learning on geochemical compositional data, providing a direct analog for the proposed transfer of CoDA+ML methods from geochemistry to gene expression.
10. 10
    Aaron Miatke, T. Olds, C. Maher et al. (2023). The association between reallocations of time and health using compositional data analysis: a systematic scoping review with an interactive data exploration interface.indirect support · 60 citations · doi:10.1186/s12966-023-01526-xWhat the librarian takes from it CoDA is a standard methodology for analyzing time-use data, demonstrating its versatility and robustness for compositional data in various fields.Relevance Shows the broad applicability of CoDA beyond geochemistry, supporting the idea that it can be transferred to new domains like gene expression.

Novelty

Novelty score: 0.85 out of 1 · Verdict: rated novel

This score is given by an agent on the basis of the work it found. It is an estimate, not a measurement. How this score is produced

Closest existing work

• Clustering transformed compositional data using K-means, with applications in gene expression and bicycle sharing system data (2017)Difference from the idea: Applies CoDA to gene expression for clustering, not for classification or feature selection in cancer; does not use deep learning classifiers.
• Unifying the analysis of high-throughput sequencing datasets: characterizing RNA-seq, 16S rRNA gene sequencing and selective growth experiments by compositional data analysis (2014)Difference from the idea: Applies CoDA to sequencing count data (RNA-seq, 16S) but not specifically to microarray gene expression for cancer classification; does not address the normalization paradox in the context of deep learning classifiers.
• Editorial: Compositional data analysis and related methods applied to genomics—a first special issue from NAR Genomics and Bioinformatics (2020)Difference from the idea: Editorial overview of CoDA in genomics; does not propose a specific methodological transfer from geochemistry to cancer classification with deep learning.
• Weighted pivot coordinates for partial least squares‐based marker discovery in high‐throughput compositional data (2021)Difference from the idea: Uses CoDA for marker discovery in high-throughput data with PLS, not for cancer classification with deep learning (LSTM-DBN, CNN).

Gaps and data

Gaps identified

• No paper in the provided list directly tests the application of CoDA (CLR/ILR) preprocessing on microarray data before deep learning classifiers (LSTM-DBN, CNN) for cancer classification. This remains a critical gap.
• The specific impact of CoDA preprocessing on false positive rates in feature selection for cancer subtype classification, compared to standard methods like RMA, is not directly addressed in any provided paper.
• The biological interpretability of log-ratio transformed features (e.g., whether they correspond to known pathways) is not explored in the provided literature, confirming the existing epistemic gap.
• No paper provides a direct comparison of CoDA-based normalization vs. standard normalization (RMA, quantile normalization) in the context of deep learning for cancer classification.

Available data

• The provided papers do not list specific datasets (GSE numbers) for cancer classification. However, public datasets like those from TCGA and GEO (referenced in GEPIA paper, paper_id: 231be65e...) are implicitly available for testing the hypothesis.
• The GEDAAI-PCD technique is mentioned in the hypothesis but its specific dataset and code are not in the provided list, representing a data gap.

Panel synthesis

Consensus score: 6.44/10 Average of the five scores, weighted by the confidence each reviewer declares.

Meta-reviewer’s verdict: publish

Points of agreement

• All reviewers acknowledge that the hypothesis addresses a real and underestimated problem: the impact of the closure constraint on correlations in microarray data, and that the methodological transfer from CoDA to cancer classification is a logical and potentially fruitful extension.
• The three-phase validation protocol with explicit GO/NO-GO criteria is unanimously praised as a model of hypothetico-deductive rigour and risk management, which provides reassurance regarding feasibility and methodological maturity.
• The magnitude of the announced effect (≥8 pp gain in balanced accuracy) is judged too precise and ambitious by the domain expert and the Contrarian, who recommend replacing it with a more realistic range of expected effects (1–3 pp).
• The lack of proof of concept on real data with a realistic covariance structure is a critical weakness: the Contrarian demands a realistic simulation before any investment, and the funding strategist notes that the current TRL is too low (1–2) for major funding.

Points of disagreement

• The Contrarian (score 3.5, confidence 0.85) considers the effect size of 8 pp to be unrealistic and that the CLR could destroy biologically informative co-expression structures, whereas the Methodologist (score 8.2, confidence 0.92) and the Domain expert (score 7.2, confidence 0.78) judge the hypothesis to be theoretically sound and well-designed, despite reservations concerning the magnitude of the effect.
• The Domain expert raises that Assumption 1 (total mRNA quantity not informative) is an excessive simplification, as absolute mRNA abundance can be discriminatory in certain cancers (via cellularity, aneuploidy, CNA). The Methodologist and the Contrarian do not contest this point, but the Contrarian uses it as an argument for an additional “failure reason”.
• The Funding strategist recommends targeting a proof-of-concept grant (ERC PoC) first before aiming at broader calls, whereas the Industry reviewer advocates an immediate launch of Phase 1 in silico to validate the ROI before any heavy commitment. These two visions are not contradictory but reflect different temporalities (academic versus commercial).

Critical path

The most decisive factor for the success or failure of this hypothesis is the validation of the reported effect size (≥8 pp) on a synthetic dataset with a realistic covariance structure (inspired by TCGA data) and on at least one real clinical dataset (e.g., TCGA BRCA). If the observed gain is <3 pp or is not significant after correction for multiple comparisons, the central hypothesis collapses. The handling of zeros and negative values in microarrays for the application of CLR is a technical blocking point that must be resolved before any large-scale validation.

Final recommendation

The hypothesis is intellectually stimulating, rests on sound theoretical foundations in compositional analysis, and the validation protocol is exemplary. However, the magnitude of the claimed effect (≥8 pp) is overly ambitious and is not supported by preliminary evidence. The panel recommends the publication of a research brief describing the hypothesis, the causal mechanism, and the validation protocol, but with the 8 pp threshold replaced by a range of expected effects (e.g., moderate to substantial improvement, 1–5 pp) and with the addition of an explicit discussion of cases in which absolute mRNA abundance is informative. A proof of concept on realistic synthetic data and on a TCGA dataset is required before large-scale validation can be justified. The brief will serve as the basis for an ERC Proof of Concept grant application.