{ "brief_id": "SPR-2026-3A8D", "generated_at": "2026-05-11", "domains": [ "Gene expression and cancer classification", "Geochemistry and Geologic Mapping" ], "original_hypothesis": "Compositional data analysis (CoDA) principles from geochemistry can resolve the normalization and feature selection paradox in high-dimensional, low-sample-size gene expression data for cancer classification.", "grounding": { "novelty_assessment": { "score": 0.85, "closest_existing_work": [ { "paper_id": "a139fec18a684e91878b2f0c5f6517b7dacb41e9", "title": "Clustering transformed compositional data using K-means, with applications in gene expression and bicycle sharing system data", "doi": "10.1080/02664763.2018.1454894", "year": 2017, "similarity": "related", "key_difference": "Applies CoDA to gene expression for clustering, not for classification or feature selection in cancer; does not use deep learning classifiers.", "authors": [ "Antoine Godichon-Baggioni", "C. Maugis-Rabusseau", "A. Rau" ] }, { "paper_id": "437a3e209c99c7a3f90b57e4476596cc92382466", "title": "Unifying the analysis of high-throughput sequencing datasets: characterizing RNA-seq, 16S rRNA gene sequencing and selective growth experiments by compositional data analysis", "doi": "10.1186/2049-2618-2-15", "year": 2014, "similarity": "related", "key_difference": "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.", "authors": [ "A. Fernandes", "J. N. Reid", "Jean M. Macklaim", "T. A. McMurrough", "D. Edgell", "G. Gloor" ] }, { "paper_id": "db02a4cf88176e44b5707e26d096a50202efa600", "title": "Editorial: Compositional data analysis and related methods applied to genomics—a first special issue from NAR Genomics and Bioinformatics", "doi": "10.1093/nargab/lqaa103", "year": 2020, "similarity": "related", "key_difference": "Editorial overview of CoDA in genomics; does not propose a specific methodological transfer from geochemistry to cancer classification with deep learning.", "authors": [ "Ionas Erb", "G. Gloor", "T. Quinn" ] }, { "paper_id": "2862b59108e67896f4f5cd4e8ddd30d878ae6efc", "title": "Weighted pivot coordinates for partial least squares‐based marker discovery in high‐throughput compositional data", "doi": "10.1002/sam.11514", "year": 2021, "similarity": "related", "key_difference": "Uses CoDA for marker discovery in high-throughput data with PLS, not for cancer classification with deep learning (LSTM-DBN, CNN).", "authors": [ "N. Štefelová", "J. Palarea‐Albaladejo", "K. Hron" ] } ], "verdict": "novel" }, "evidence_base": [ { "paper_id": "a139fec18a684e91878b2f0c5f6517b7dacb41e9", "title": "Clustering transformed compositional data using K-means, with applications in gene expression and bicycle sharing system data", "doi": "10.1080/02664763.2018.1454894", "year": 2017, "citation_count": 58, "support_type": "indirect", "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.", "key_finding": "CoDA-based clustering (using log-ratio transformations) is effective for identifying groups of co-expressed genes from high-throughput sequencing data.", "authors": [ "Antoine Godichon-Baggioni", "C. Maugis-Rabusseau", "A. Rau" ] }, { "paper_id": "437a3e209c99c7a3f90b57e4476596cc92382466", "title": "Unifying the analysis of high-throughput sequencing datasets: characterizing RNA-seq, 16S rRNA gene sequencing and selective growth experiments by compositional data analysis", "doi": "10.1186/2049-2618-2-15", "year": 2014, "citation_count": 1106, "support_type": "indirect", "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.", "key_finding": "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.", "authors": [ "A. Fernandes", "J. N. Reid", "Jean M. Macklaim", "T. A. McMurrough", "D. Edgell", "G. Gloor" ] }, { "paper_id": "db02a4cf88176e44b5707e26d096a50202efa600", "title": "Editorial: Compositional data analysis and related methods applied to genomics—a first special issue from NAR Genomics and Bioinformatics", "doi": "10.1093/nargab/lqaa103", "year": 2020, "citation_count": 10, "support_type": "indirect", "relevance": "Provides a high-level endorsement of CoDA for genomics, reinforcing the idea that CoDA principles from geochemistry are transferable to genomic data.", "key_finding": "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.", "authors": [ "Ionas Erb", "G. Gloor", "T. Quinn" ] }, { "paper_id": "2862b59108e67896f4f5cd4e8ddd30d878ae6efc", "title": "Weighted pivot coordinates for partial least squares‐based marker discovery in high‐throughput compositional data", "doi": "10.1002/sam.11514", "year": 2021, "citation_count": 8, "support_type": "indirect", "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.", "key_finding": "Using weighted pivot coordinates (a CoDA approach) enhances the discovery of biological markers in high-throughput compositional data compared to standard normalization.", "authors": [ "N. Štefelová", "J. Palarea‐Albaladejo", "K. Hron" ] }, { "paper_id": "eb5c6f6be1acae79b46a10c94746845160a6b2c8", "title": "Normalization methods for microbial abundance data strongly affect correlation estimates", "doi": "10.1101/406264", "year": 2018, "citation_count": 22, "support_type": "analogous", "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.", "key_finding": "Log-ratio and variance-stabilization transformations provide the most consistent correlation estimates across experiments, while other methods introduce artifacts.", "authors": [ "M. Badri", "Zachary D. Kurtz", "Christian L. Müller", "Richard Bonneau" ] }, { "paper_id": "08c7ebde2c2bbb53093d7c5a456424443cb54b7d", "title": "Normalization and microbial differential abundance strategies depend upon data characteristics", "doi": "10.1186/s40168-017-0237-y", "year": 2017, "citation_count": 1676, "support_type": "analogous", "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.", "key_finding": "Different normalization methods (including those based on CoDA) lead to different differential abundance results; the choice depends on data characteristics.", "authors": [ "Sophie J. Weiss", "Z. Xu", "S. Peddada", "Amnon Amir", "K. Bittinger", "Antonio Gonzalez", "C. Lozupone", "Jesse R. Zaneveld", "Y. Vázquez-Baeza", "A. Birmingham", "Embriette R. Hyde", "R. Knight" ] }, { "paper_id": "9c8fef9be693062e91999a04d1b01f53a8a97923", "title": "Compositional data analysis enables statistical rigor in comparative glycomics", "doi": "10.1038/s41467-025-56249-3", "year": 2025, "citation_count": 13, "support_type": "analogous", "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.", "key_finding": "Applying CoDA to comparative glycomics data eliminates spurious 'decreases' and high false-positive rates caused by ignoring compositionality.", "authors": [ "Alexander R. Bennett", "Jon Lundstrøm", "S. Chatterjee", "Morten Thaysen-Andersen", "Daniel Bojar" ] }, { "paper_id": "f943d34e796d7bf56840609655f465ecf28c2c85", "title": "A review of compositional data analysis and recent advances", "doi": "10.1080/03610926.2021.2014890", "year": 2021, "citation_count": 22, "support_type": "indirect", "relevance": "Provides a comprehensive review of CoDA methods, including log-ratio transformations, which are central to the proposed methodological transfer.", "key_finding": "CoDA provides a rigorous statistical framework for analyzing compositional data, with transformations (CLR, ILR) that enable standard multivariate analysis.", "authors": [ "Abdulaziz Alenazi" ] }, { "paper_id": "5f296470146cd649e4227bfc1f5ef0e52f049bb4", "title": "Mineral-resource prediction using advanced data analytics and machine learning of the QUEST-South stream-sediment geochemical data, southwestern British Columbia, Canada", "doi": "10.1144/geochem2020-054", "year": 2020, "citation_count": 21, "support_type": "analogous", "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.", "key_finding": "Multivariate statistical and machine learning methods, applied to compositional geochemical data, can successfully predict mineral resources.", "authors": [ "E. Grunsky", "D. Arne" ] }, { "paper_id": "cdaa30cc5af75892e8ef303991eebdfe787ab528", "title": "The association between reallocations of time and health using compositional data analysis: a systematic scoping review with an interactive data exploration interface", "doi": "10.1186/s12966-023-01526-x", "year": 2023, "citation_count": 60, "support_type": "indirect", "relevance": "Shows the broad applicability of CoDA beyond geochemistry, supporting the idea that it can be transferred to new domains like gene expression.", "key_finding": "CoDA is a standard methodology for analyzing time-use data, demonstrating its versatility and robustness for compositional data in various fields.", "authors": [ "Aaron Miatke", "T. Olds", "C. Maher", "F. Fraysse", "M. Mellow", "Ashleigh E. Smith", "Ž. Pedišić", "J. Grgic", "D. Dumuid" ] } ], "counter_evidence": [ { "paper_id": "7e49aa718a7e5fea9784e5c04c1c853edec9e46a", "title": "SCANPY: large-scale single-cell gene expression data analysis", "doi": "10.1186/s13059-017-1382-0", "year": 2018, "finding": "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": "minor", "authors": [ "F. Wolf", "Philipp Angerer", "Fabian J Theis" ] }, { "paper_id": "8bb6cc4057c0f76e963f19918a79697acbd2bc41", "title": "edgeR: a Bioconductor package for differential expression analysis of digital gene expression data", "doi": "10.1093/bioinformatics/btp616", "year": 2009, "finding": "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": "minor", "authors": [ "M. Robinson", "Davis J. McCarthy", "G. Smyth" ] }, { "paper_id": "25052da76aa0e2df1a52dd69d5206b785cd8be6d", "title": "Winsorization greatly reduces false positives by popular differential expression methods when analyzing human population samples", "doi": "10.1186/s13059-024-03230-w", "year": 2024, "finding": "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": "addressable", "authors": [ "Lu Yang", "Xianyang Zhang", "Jun Chen" ] } ], "gap_manifest_update": { "closed_gaps": [ "The hypothesis that gene expression data is compositional is supported by multiple papers (a139fec..., 437a3e2..., db02a4c...), closing the gap regarding the fundamental mathematical structure.", "The existence of spurious correlations in compositional data (Pearson's paradox) is well-documented in the provided literature (9c8fef9..., eb5c6f6...), closing the gap on the theoretical basis for the problem." ], "new_gaps": [ "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." ], "data_available": [ "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." ] }, "search_queries": [ { "query": "compositional data analysis gene expression classification", "type": "novelty", "rationale": "Direct search for the core hypothesis applying CoDA to gene expression for cancer classification" }, { "query": "centered log-ratio microarray cancer subtype", "type": "novelty", "rationale": "Checks if CLR transformation has been specifically tested on microarray cancer subtype data" }, { "query": "isometric log-ratio high-dimensional low-sample size", "type": "novelty", "rationale": "Searches for ILR application in the HDLSS context typical of gene expression studies" }, { "query": "closure problem gene expression normalization", "type": "novelty", "rationale": "Verifies if the compositional closure problem has been explicitly identified in gene expression normalization literature" }, { "query": "spurious correlation compositional data gene expression", "type": "evidence", "rationale": "Seeks evidence of Pearson's paradox or spurious correlations in gene expression due to compositionality" }, { "query": "log-ratio transformation deep learning cancer", "type": "evidence", "rationale": "Finds evidence of CoDA preprocessing before deep learning classifiers for cancer" }, { "query": "RMA normalization compositional artifact", "type": "evidence", "rationale": "Looks for evidence that standard RMA normalization introduces artifacts due to ignoring compositionality" }, { "query": "false positive differential expression compositionality", "type": "evidence", "rationale": "Searches for evidence linking compositional effects to inflated false positive rates in differential expression" }, { "query": "geochemistry log-ratio machine learning classification", "type": "cross_domain", "rationale": "Finds interdisciplinary precedent of CoDA log-ratio methods used in geochemical classification tasks" }, { "query": "compositional data analysis bioinformatics review", "type": "cross_domain", "rationale": "Searches for existing reviews bridging CoDA from geochemistry to bioinformatics" }, { "query": "simplex constraint high-dimensional classification", "type": "cross_domain", "rationale": "Looks for cross-domain precedents addressing the simplex constraint in high-dimensional classification problems" } ] }, "sharpened": { "title": "Compositional Data Analysis (CoDA) Preprocessing Resolves Normalization Artifacts in High-Dimensional, Low-Sample-Size Cancer Gene Expression Classification: A Cross-Domain Methodological Transfer", "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.", "independent_variables": [ { "name": "Normalization method applied to gene expression matrix", "type": "categorical", "range": "RMA, Quantile, TPM, CLR (robust), ILR", "unit": "N/A (categorical)" }, { "name": "Classifier complexity", "type": "categorical", "range": "LASSO (L1-regularized logistic regression), Linear SVM, CNN (1D convolutional), LSTM-DBN (deep belief network with LSTM)", "unit": "N/A (categorical)" }, { "name": "Imputation method for zero counts (pre-CLR)", "type": "categorical", "range": "Multiplicative replacement, Bayesian imputation (zCompositions), Robust CLR (robCompositions)", "unit": "N/A (categorical)" }, { "name": "Estimated sample cellularity (immune/stromal fraction)", "type": "continuous", "range": "0.0 to 1.0", "unit": "proportion" } ], "dependent_variables": [ { "name": "Balanced accuracy (BAcc) of cancer subtype classification", "type": "continuous", "expected_direction": "increase", "unit": "percentage points (pp)" }, { "name": "Number of selected features (genes) in the final model", "type": "continuous", "expected_direction": "decrease", "unit": "count" }, { "name": "False positive rate (FPR) of differential expression markers (simulated ground truth)", "type": "continuous", "expected_direction": "decrease", "unit": "proportion (0-1)" }, { "name": "Bias of imputation method (mean absolute error vs. known synthetic zeros)", "type": "continuous", "expected_direction": "non-monotonic", "unit": "log-ratio units" }, { "name": "Variance of imputation method (standard deviation across 100 bootstrap replicates)", "type": "continuous", "expected_direction": "non-monotonic", "unit": "log-ratio units" } ], "proposed_mechanism": { "causal_chain": [ "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.", "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).", "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.", "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.", "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." ], "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)." ] }, "falsifiable_predictions": [ { "prediction": "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.", "null_hypothesis": "H0: μ_BAcc(CLR) - μ_BAcc(RMA) ≤ 0 (no improvement or degradation).", "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." }, { "prediction": "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.", "null_hypothesis": "H0: median(N_features(CLR)) ≥ median(N_features(RMA)).", "statistical_test": "Wilcoxon signed-rank test (paired), one-sided, α=0.01." }, { "prediction": "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.", "null_hypothesis": "H0: FPR(CLR) ≥ FPR(RMA).", "statistical_test": "Paired one-sided t-test on logit-transformed FPRs across 100 simulations, α=0.01." }, { "prediction": "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.", "null_hypothesis": "H0: |r| ≥ 0.3 (significant correlation indicating cellularity confound).", "statistical_test": "Fisher's z-transformation test for correlation, one-sided (testing |r| < 0.3), α=0.05." }, { "prediction": "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).", "null_hypothesis": "H0: μ_ΔBAcc ≤ 0.", "statistical_test": "One-sample one-sided t-test on the 50 ΔBAcc values, α=0.01." } ], "boundary_conditions": [ { "condition": "The proportion of zero entries in the expression matrix must be less than 20% before imputation.", "justification": "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%." }, { "condition": "The number of samples (n) must be at least 30 per class for the classification task.", "justification": "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)." }, { "condition": "The number of genes (p) must be between 10,000 and 50,000.", "justification": "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." }, { "condition": "The classification task must be categorical (cancer subtype or condition), not regression (e.g., survival time).", "justification": "The hypothesis is formulated for classification accuracy; for regression, the compositional effect on continuous outcomes is not yet characterized." } ], "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)." }, "protocol": { "protocol_title": "Validation protocol for CLR transformation improving cancer subtype classification accuracy via compositional data analysis", "overall_timeline": "8-14 months", "overall_budget_estimate": "€20k-80k", "phases": [ { "phase_number": 1, "phase_name": "In Silico Validation", "objective": "To rapidly falsify the core hypothesis using public data and computational simulations, at minimal cost, before committing to wet-lab or expensive compute resources.", "methodology": "1. **Data acquisition**: Download TCGA-BRCA (n=500, 4 subtypes), TCGA-LUAD, TCGA-COAD, TCGA-OV, TCGA-LUSC (n≥100 each) from GDC portal using `gdc-client`. Download Arabidopsis thaliana stress dataset (GSEXXXX) from GEO via `GEOquery`. \n2. **Preprocessing**: Apply RMA normalization (via `affy` R package), Quantile normalization (via `preprocessCore`), TPM (for RNA-seq only), and robust CLR with multiplicative replacement (via `robCompositions` R package, `pcaCoDa` function). For zeros >20%, use Bayesian imputation (`zCompositions`). \n3. **Classification pipeline**: For each normalization × classifier combination (LASSO via `glmnet`, Linear SVM via `e1071`, 1D-CNN via `keras`/`tensorflow`, LSTM-DBN via custom `PyTorch` code), run 5-fold cross-validation repeated 10 times. Compute balanced accuracy per subtype, macro-averaged. \n4. **Simulation**: Generate 100 synthetic datasets (n=50, p=1020, 20 true DE genes, log2FC=2, noise=1000) using `splatter` or custom `mvtnorm` with compositional closure. Apply limma (`limma` R package) to RMA vs CLR data; compute FPR at 80% TPR. \n5. **Cellularity correlation**: For each of 10 TCGA cancer types, compute median ESTIMATE score (via `estimate` R package) and ΔBAcc (CLR vs RMA with LASSO). Compute Pearson r and test |r|<0.3 via Fisher z-transformation. \n6. **Generalizability test**: Repeat prediction 1 on Arabidopsis dataset. \n7. **Statistical testing**: All tests as specified (paired t-test, Wilcoxon, Fisher z) with α=0.01 and Benjamini-Hochberg correction for multiple cancer types.", "required_resources": { "equipment": [ "Standard workstation (16 GB RAM, 4+ cores) or cloud instance (e.g., AWS EC2 c5.2xlarge)" ], "software": [ "R (≥4.2) with packages: affy, preprocessCore, robCompositions, zCompositions, glmnet, e1071, limma, estimate, splatter, mvtnorm", "Python (≥3.9) with packages: tensorflow, keras, pytorch, numpy, pandas, scikit-learn, scipy", "GDC client (for TCGA data download)", "GEOquery (for GEO data download)" ], "datasets": [ "TCGA-BRCA (n=500, 4 subtypes: Basal, LumA, LumB, Her2)", "TCGA-LUAD, TCGA-COAD, TCGA-OV, TCGA-LUSC (n≥100 each)", "TCGA-ESCA, TCGA-STAD, TCGA-BLCA, TCGA-KIRC, TCGA-LIHC (n≥100 each, for cellularity correlation)", "Arabidopsis thaliana stress response dataset (GSEXXXX, n=100, 4 conditions)", "Synthetic datasets (100 simulations, n=50, p=1020)" ], "competences": [ "R programming (intermediate)", "Python programming (intermediate)", "Machine learning pipeline implementation (LASSO, SVM, CNN, LSTM-DBN)", "Compositional data analysis theory (Aitchison geometry, CLR, ILR)", "Statistical hypothesis testing (paired t-test, Wilcoxon, Fisher z, multiple testing correction)" ], "estimated_cost": "€500-2000", "estimated_duration": "4-6 weeks" }, "expected_outputs": [ "Output 1: ΔBAcc (CLR vs RMA) for each cancer type and classifier, with 95% CI and p-values.", "Output 2: Feature count reduction (CLR vs RMA) for LASSO, with median and Wilcoxon p-value.", "Output 3: FPR reduction (CLR vs RMA) in simulated DE detection, with t-test result.", "Output 4: Pearson r between ΔBAcc and ESTIMATE cellularity across 10 cancer types, with Fisher z test result.", "Output 5: ΔBAcc on Arabidopsis dataset, with one-sample t-test result.", "Output 6: Sensitivity analysis: effect of zero proportion (5%, 10%, 15%, 20%) on CLR stability and BAcc gain." ], "success_criteria": [ { "metric": "ΔBAcc (CLR vs RMA) for TCGA-BRCA with LASSO", "threshold": "≥ 8 pp, with 95% CI lower bound ≥ 5 pp", "measurement": "Paired one-sided t-test on 50 BAcc differences (10 repeats × 5 folds), α=0.01, BH-corrected" }, { "metric": "Feature count reduction (CLR vs RMA) for LASSO on TCGA-BRCA", "threshold": "Median reduction ≥ 40% (i.e., N_features(CLR) ≤ 0.6 × N_features(RMA))", "measurement": "Wilcoxon signed-rank test (paired, one-sided), α=0.01" }, { "metric": "FPR reduction (CLR vs RMA) in simulated DE detection", "threshold": "FPR(CLR) ≤ 0.5 × FPR(RMA) and absolute FPR(CLR) < 0.05", "measurement": "Paired one-sided t-test on logit-transformed FPRs across 100 simulations, α=0.01" }, { "metric": "Cellularity correlation (|r| between ΔBAcc and ESTIMATE)", "threshold": "|r| < 0.3, with 95% CI upper bound < 0.5", "measurement": "Fisher z-transformation test for correlation, one-sided (testing |r| < 0.3), α=0.05" }, { "metric": "ΔBAcc on Arabidopsis dataset", "threshold": "≥ 5 pp, with 95% CI lower bound ≥ 2 pp", "measurement": "One-sample one-sided t-test on 50 ΔBAcc values, α=0.01" } ], "go_nogo_decision": { "go_if": "All 5 success criteria are met (i.e., all null hypotheses rejected in the predicted direction).", "nogo_if": "Any 2 or more success criteria fail (i.e., null hypothesis not rejected, or effect in opposite direction).", "pivot_if": "Exactly 1 success criterion fails; then investigate why (e.g., cellularity confound, zero proportion >20%, classifier-specific failure). If the failure is due to a specific classifier, drop that classifier and proceed with the others. If failure is due to cellularity, revise hypothesis to include cellularity as a covariate." }, "risks": [ { "risk": "Zero proportion >20% in some TCGA datasets (e.g., lowly expressed genes), causing CLR instability.", "probability": "medium", "mitigation": "Use Bayesian imputation (zCompositions) and test sensitivity; if still unstable, filter genes with >20% zeros and report. If filtering reduces p < 10,000, consider the hypothesis invalid for that dataset." }, { "risk": "Computational cost of LSTM-DBN training on 50,000 genes × 500 samples may exceed workstation capacity.", "probability": "medium", "mitigation": "Use dimensionality reduction (PCA to 100 components) before LSTM-DBN; report if results differ from full gene set. Alternatively, use cloud GPU instances (AWS p3.2xlarge, ~$3/hour)." }, { "risk": "Arabidopsis dataset may not have exactly 4 conditions or n=100; GEO data may be incomplete.", "probability": "low", "mitigation": "Use any plant stress dataset with ≥4 conditions and n≥30 per condition from GEO; if not available, substitute with a mouse brain region dataset (e.g., GSEXXXX)." }, { "risk": "Simulation may not capture real data complexity (e.g., batch effects, non-linear dependencies).", "probability": "medium", "mitigation": "Use `splatter` with parameters estimated from real TCGA data (e.g., dropout rate, library size variability). Also test with a second simulation using a Dirichlet-multinomial model." } ] }, { "phase_number": 2, "phase_name": "Minimal Experimental Validation", "objective": "To confirm the central mechanistic claim (CLR removes spurious correlations) using a controlled in vitro experiment with synthetic RNA mixtures, where ground truth is known.", "methodology": "1. **Design**: Create 4 synthetic RNA mixtures (mimicking cancer subtypes) by pooling purified total RNA from 4 human cell lines (e.g., MCF7, MDA-MB-231, T47D, BT-474 for breast cancer subtypes). Each mixture is composed of known proportions (e.g., 60% dominant subtype + 40% equal mix of others). Prepare 5 biological replicates per mixture (n=20 samples total). \n2. **Microarray hybridization**: Use Affymetrix Human Genome U133 Plus 2.0 arrays (or equivalent) for all 20 samples. Standard protocol. \n3. **Data preprocessing**: Apply RMA (standard) and robust CLR (with multiplicative replacement) to raw CEL files. \n4. **Classification**: Train LASSO and Linear SVM on RMA vs CLR data, using 5-fold cross-validation (leave-one-out per mixture type). Compute balanced accuracy. \n5. **Ground truth comparison**: Since the true proportions are known, compute the true positive rate (TPR) and false positive rate (FPR) for detecting differentially expressed genes between mixtures (using limma). Compare FPR at 80% TPR between RMA and CLR. \n6. **Statistical test**: Paired t-test on BAcc (5 folds) and FPR (across 6 pairwise comparisons), α=0.05.", "required_resources": { "equipment": [ "Affymetrix microarray scanner and hybridization station (or service provider)", "Nanodrop spectrophotometer for RNA quantification", "Thermal cycler for cDNA synthesis (if needed)" ], "software": [ "R (≥4.2) with packages: affy, robCompositions, limma, glmnet, e1071", "Affymetrix Expression Console (for CEL file processing)" ], "datasets": [ "Synthetic RNA mixtures (20 samples, 4 mixtures × 5 replicates)" ], "competences": [ "Cell culture and RNA extraction (intermediate)", "Microarray hybridization (basic, or outsourced to core facility)", "R programming for preprocessing and classification" ], "estimated_cost": "€5k-15k", "estimated_duration": "2-3 months" }, "expected_outputs": [ "Output 1: BAcc(CLR) vs BAcc(RMA) for LASSO and SVM on synthetic mixtures.", "Output 2: FPR(CLR) vs FPR(RMA) at 80% TPR for differential expression detection.", "Output 3: Correlation between known mixture proportions and predicted proportions (via LASSO coefficients) for CLR vs RMA." ], "success_criteria": [ { "metric": "ΔBAcc (CLR vs RMA) for LASSO", "threshold": "≥ 5 pp (relaxed from 8 pp due to smaller n=20)", "measurement": "Paired one-sided t-test on 5 BAcc differences (5 folds), α=0.05" }, { "metric": "FPR reduction (CLR vs RMA) at 80% TPR", "threshold": "FPR(CLR) ≤ 0.6 × FPR(RMA)", "measurement": "Paired one-sided t-test on 6 pairwise FPR differences, α=0.05" } ], "go_nogo_decision": { "go_if": "Both success criteria met (ΔBAcc ≥5 pp and FPR reduction ≥40%).", "nogo_if": "Neither criterion met, or effect in opposite direction.", "pivot_if": "Only one criterion met; then investigate whether the failure is due to classifier (e.g., SVM not benefiting) or due to small sample size. If classifier-specific, proceed with only LASSO in Phase 3." }, "risks": [ { "risk": "Synthetic mixtures may not recapitulate real tumor heterogeneity (e.g., stromal contamination, immune infiltration).", "probability": "medium", "mitigation": "Add a small fraction (10%) of fibroblast RNA (e.g., MRC-5) to each mixture to mimic stroma. This tests robustness to cellularity confound." }, { "risk": "Microarray batch effects between hybridization runs.", "probability": "low", "mitigation": "Randomize sample order across arrays; include a common reference sample on each array for normalization." }, { "risk": "RNA degradation during handling.", "probability": "low", "mitigation": "Use RNA integrity number (RIN) >8 for all samples; exclude any sample with RIN <7." } ] }, { "phase_number": 3, "phase_name": "Full Experimental Protocol", "objective": "To rigorously validate the hypothesis across multiple cancer types, platforms, and classifiers, producing publication-quality results with clinical relevance.", "methodology": "1. **Dataset expansion**: Include 15 TCGA cancer types (n≥100 each) and 5 independent GEO datasets (e.g., GSEXXXX for breast, lung, colon, ovarian, prostate). For each, apply RMA, Quantile, TPM, and CLR (robust with multiplicative replacement). \n2. **Classifier suite**: LASSO, Linear SVM, 1D-CNN, LSTM-DBN, and additionally Random Forest and XGBoost for robustness. Use nested cross-validation (outer 5-fold, inner 3-fold for hyperparameter tuning). \n3. **Zero imputation sensitivity**: Test 3 imputation methods (multiplicative replacement, Bayesian, robust CLR) at zero proportions 5%, 10%, 15%, 20%. Report bias and variance for each. \n4. **Cellularity adjustment**: For each cancer type, compute ESTIMATE and CIBERSORTx immune/stromal scores. If |r| > 0.3 between ΔBAcc and cellularity, include cellularity as a covariate in a mixed-effects model: ΔBAcc ~ cellularity + (1|cancer_type). \n5. **RNA-seq comparison**: For 5 TCGA cancer types with both microarray and RNA-seq data (e.g., BRCA, LUAD), compare CLR effect on RNA-seq (using TPM as baseline) vs microarray. \n6. **Clinical relevance**: For the best-performing normalization (CLR), train a final model on the full TCGA-BRCA dataset and validate on an independent GEO breast cancer dataset (e.g., GSEXXXX, n=200). Report balanced accuracy, sensitivity, specificity, and AUC per subtype. \n7. **Statistical reporting**: All tests with α=0.01, BH correction for multiple comparisons. Effect sizes with 95% CI. Power analysis: for n=500, 80% power to detect 8 pp ΔBAcc at α=0.01 (paired t-test).", "required_resources": { "equipment": [ "High-performance computing cluster (64+ cores, 256 GB RAM) or cloud GPU instances (AWS p3.8xlarge, ~$12/hour)", "Storage: 1 TB SSD for raw data and results" ], "software": [ "R (≥4.2) with packages: affy, preprocessCore, robCompositions, zCompositions, glmnet, e1071, limma, estimate, CIBERSORTx (web tool or R interface), lme4, pROC, caret", "Python (≥3.9) with packages: tensorflow, keras, pytorch, xgboost, sklearn, numpy, pandas, scipy, matplotlib, seaborn", "GDC client, GEOquery, SRA toolkit (for RNA-seq data)" ], "datasets": [ "15 TCGA cancer types (BRCA, LUAD, COAD, OV, LUSC, ESCA, STAD, BLCA, KIRC, LIHC, PRAD, THCA, HNSC, UCEC, CESC) – n≥100 each", "5 independent GEO datasets (breast, lung, colon, ovarian, prostate) – n≥100 each", "5 TCGA cancer types with both microarray and RNA-seq data (BRCA, LUAD, COAD, OV, LUSC)", "Independent validation dataset: GSEXXXX (breast cancer, n=200)" ], "competences": [ "Advanced R and Python programming", "Deep learning (CNN, LSTM, DBN) implementation and tuning", "Compositional data analysis theory and practice", "Biostatistics (mixed-effects models, multiple testing correction, power analysis)", "High-performance computing (Slurm, AWS CLI)" ], "estimated_cost": "€15k-100k+", "estimated_duration": "6-12 months" }, "expected_outputs": [ "Output 1: Comprehensive table of ΔBAcc (CLR vs RMA) for all 15 TCGA cancer types × 6 classifiers, with 95% CI and p-values.", "Output 2: Feature count reduction (CLR vs RMA) for LASSO across all cancer types.", "Output 3: FPR reduction (CLR vs RMA) in simulated and real DE detection (using limma).", "Output 4: Sensitivity analysis of imputation methods (bias, variance) across zero proportions.", "Output 5: Cellularity-adjusted ΔBAcc model (mixed-effects) and correlation plot.", "Output 6: RNA-seq vs microarray comparison: ΔBAcc for CLR vs TPM (RNA-seq) and CLR vs RMA (microarray).", "Output 7: Final validated model (CLR + LASSO) on independent breast cancer dataset, with AUC curves.", "Output 8: Reproducible code and data (GitHub repository, Zenodo archive)." ], "success_criteria": [ { "metric": "Primary: ΔBAcc (CLR vs RMA) averaged across all 15 cancer types and 6 classifiers", "threshold": "≥ 6 pp (conservative, accounting for heterogeneity), with 95% CI lower bound ≥ 4 pp", "measurement": "Mixed-effects model: ΔBAcc ~ 1 + (1|cancer_type) + (1|classifier), one-sided test of intercept > 0, α=0.01" }, { "metric": "Secondary: FPR reduction (CLR vs RMA) in real DE detection (limma on TCGA-BRCA)", "threshold": "FPR(CLR) ≤ 0.5 × FPR(RMA) at 80% TPR (using simulated ground truth embedded in real data)", "measurement": "Paired t-test on logit-transformed FPRs across 100 bootstrap samples, α=0.01" }, { "metric": "Secondary: Independent validation BAcc on GSEXXXX", "threshold": "BAcc(CLR+LASSO) ≥ 0.75, with 95% CI lower bound ≥ 0.70", "measurement": "Binomial test on predicted vs true labels, α=0.01" }, { "metric": "Tertiary: RNA-seq comparison – ΔBAcc (CLR vs TPM) ≥ 4 pp", "threshold": "≥ 4 pp, with 95% CI lower bound ≥ 2 pp", "measurement": "Paired t-test on 5 cancer types, α=0.01" } ], "go_nogo_decision": { "go_if": "Primary success criterion met (ΔBAcc ≥6 pp).", "nogo_if": "Primary criterion not met (ΔBAcc <4 pp), indicating the hypothesis is not generalizable.", "pivot_if": "Primary criterion partially met (4-6 pp) but secondary criteria fail; then investigate whether the effect is specific to certain cancer types or classifiers. If specific, publish as a conditional finding (e.g., 'CLR improves classification in solid tumors with low cellularity')." }, "risks": [ { "risk": "High heterogeneity across cancer types (e.g., some may have zero proportion >20% or n<30 per class).", "probability": "medium", "mitigation": "Exclude cancer types that violate boundary conditions (zero >20%, n<30 per class) and report as 'not applicable'. Use mixed-effects model to account for heterogeneity." }, { "risk": "Computational cost of training LSTM-DBN on 50,000 genes × 500 samples × 15 cancer types may be prohibitive.", "probability": "high", "mitigation": "Use PCA to 100 components before LSTM-DBN; compare with full gene set on a subset (e.g., BRCA only). If results are similar, use PCA for all. Report computational time and cost." }, { "risk": "Independent validation dataset (GSEXXXX) may not have the same subtype distribution as TCGA-BRCA.", "probability": "medium", "mitigation": "Use propensity score matching to balance subtype proportions; report BAcc per subtype separately. If matching fails, use a different validation dataset (e.g., METABRIC for breast cancer)." }, { "risk": "RNA-seq data may have different zero inflation and noise characteristics than microarray, reducing CLR effectiveness.", "probability": "medium", "mitigation": "Pre-filter genes with low counts (e.g., >10 reads in at least 10% of samples) before CLR. Use robust CLR with Bayesian imputation for zeros. If ΔBAcc <4 pp, report as a limitation and discuss in paper." } ] } ], "phase_1_quick_start": { "can_start_today": true, "first_action": "Download TCGA-BRCA microarray data (level 3, normalized and raw CEL files) using GDC client: `gdc-client download -m manifest_brca.txt` and install R packages: `install.packages(c('affy', 'preprocessCore', 'robCompositions', 'zCompositions', 'glmnet', 'e1071', 'limma', 'estimate'))`.", "tools_needed": [ "GDC client (command-line tool)", "R (≥4.2) with required packages", "Python (≥3.9) with tensorflow, keras, pytorch" ], "open_data_sources": [ "TCGA-BRCA (GDC portal: https://portal.gdc.cancer.gov/)", "TCGA-LUAD, TCGA-COAD, TCGA-OV, TCGA-LUSC (same portal)", "Arabidopsis thaliana stress dataset (GEO: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSEXXXX)", "Synthetic data generation via `splatter` R package (Bioconductor)" ] } }, "panel": { "reviews": [ { "reviewer_persona": "methodologist", "overall_score": 8.2, "verdict": "strong_accept", "strengths": [ "Le protocole est remarquablement bien structuré avec une progression logique en trois phases (in silico, expérimental minimal, validation complète), chacune avec des critères GO/NO-GO/PIVOT clairs, ce qui permet une prise de décision objective et une allocation efficace des ressources avant de passer à des étapes plus coûteuses.", "La spécification extrêmement détaillée des prédictions falsifiables avec des bornes quantitatives précises (ex: ΔBAcc ≥ 8 pp, IC 95% ≥ 5 pp), des méthodes statistiques explicites (test t apparié, Wilcoxon, correction de Benjamini-Hochberg), et des hypothèses nulles claires est un modèle de rigueur hypothético-déductive.", "L'identification et la gestion proactive des biais sont exemplaires : biais de sélection adressé par la validation sur 15 types de cancer TCGA et 5 jeux de données GEO indépendants, biais de confirmation traité par l'utilisation de données simulées avec vérité terrain connue, et biais de mesure abordé par l'analyse de sensibilité sur trois méthodes d'imputation des zéros." ], "weaknesses": [ "La puissance statistique pour la prédiction 4 (corrélation avec cellularité) est potentiellement insuffisante : avec seulement 10 types de cancer, la largeur de l'IC à 95% pour un coefficient de corrélation de Pearson sera très grande (environ ±0.6), rendant le critère |r| < 0.3 avec borne supérieure < 0.5 difficile à satisfaire même si la vraie corrélation est nulle.", "Le protocole ne traite pas explicitement le biais de publication potentiel : si les résultats sont négatifs pour certains types de cancer ou classifieurs, y a-t-il un plan pour publier ces résultats nuls ? Le risque de 'file drawer problem' est réel, surtout avec 15 types de cancer × 6 classifieurs = 90 combinaisons.", "La phase 2 (validation expérimentale minimale) utilise des mélanges synthétiques de lignées cellulaires qui, bien que contrôlés, ne reproduisent pas la complexité de l'hétérogénéité tumorale réelle (infiltration immune, stroma, sous-clones). Le passage de la phase 2 à la phase 3 pourrait révéler des facteurs de confusion non anticipés." ], "critical_questions": [ "Comment comptez-vous gérer le risque de surajustement lors de la sélection des caractéristiques (features) par LASSO, sachant que vous utilisez lambda.1se mais que le nombre de gènes (p ≈ 50 000) est 100 fois plus grand que le nombre d'échantillons (n ≈ 500) ? Avez-vous envisagé une validation croisée imbriquée (nested cross-validation) pour estimer l'erreur de généralisation de manière non biaisée ?", "Pour la prédiction 3 sur les données simulées, comment garantissez-vous que la structure de corrélation compositionnelle (contrainte de somme constante) est correctement modélisée ? L'utilisation de mvtnorm avec une matrice de covariance standard pourrait ne pas capturer les dépendances non-linéaires induites par la fermeture compositionnelle. Avez-vous envisagé d'utiliser une distribution logistique normale multivariée ?", "Quel est le plan d'analyse si les résultats montrent une amélioration significative pour certains classifieurs (ex: LASSO, SVM) mais pas pour d'autres (ex: LSTM-DBN) ? Le protocole mentionne un pivot pour 'classifier-specific failure', mais comment distinguerez-vous un échec dû à une réelle absence d'effet d'un échec dû à un manque de puissance statistique pour les modèles plus complexes ?" ], "recommendation": "Ce protocole est exceptionnellement bien conçu et mérite d'être accepté. Je recommande toutefois d'augmenter le nombre de types de cancer dans la prédiction 4 à au moins 20 (en utilisant des datasets GEO supplémentaires) pour obtenir une puissance statistique adéquate pour le test de corrélation. De plus, l'ajout d'une analyse de sensibilité post-hoc sur l'effet de l'hétérogénéité intra-tumorale (en utilisant des données de séquençage en cellule unique) renforcerait considérablement la validité externe des conclusions.", "confidence": 0.92 }, { "reviewer_persona": "domain_expert", "overall_score": 7.2, "verdict": "weak_accept", "strengths": [ "L'hypothèse identifie correctement le problème fondamental de la 'closure constraint' dans les données de microarrays, un point souvent négligé dans la littérature en bioinformatique du cancer. Le transfert méthodologique de la CoDA, bien établi en écologie microbienne et en métabolomique, vers la classification du cancer est une extension logique et potentiellement fructueuse.", "La chaîne causale proposée (Steps 1-5) est cohérente avec les principes de la géométrie d'Aitchison. L'argument selon lequel la CLR élimine les corrélations négatives artefactuelles (paradoxe de Pearson) et améliore le rapport signal/bruit est théoriquement solide et soutenu par la littérature en analyse compositionnelle.", "L'hypothèse est bien délimitée et testable. Les hypothèses sous-jacentes (Assumptions 1-4) sont clairement énoncées, ce qui permet une falsification expérimentale rigoureuse. La mention des 'known unknowns' montre une bonne conscience des limites et des zones d'incertitude du mécanisme proposé." ], "weaknesses": [ "La prémisse centrale (Assumption 1) que la quantité totale d'ARNm par cellule est non informative pour la classification des sous-types de cancer est une simplification excessive. Dans de nombreux cancers, la cellularité tumorale, l'aneuploïdie et les altérations du nombre de copies (CNA) affectent directement l'abondance absolue d'ARNm. Ignorer cette information peut conduire à une perte de signal biologique crucial, ce que la CLR ne peut pas récupérer.", "L'hypothèse d'une amélioration de 'au moins 8 points de pourcentage' (pp) de la balanced accuracy est trop précise et ambitieuse pour une étude de transfert méthodologique. L'ampleur de l'effet dépendra fortement du jeu de données (type de cancer, plateforme de microarray, bruit technique) et du classifieur. Une telle affirmation quantitative sans preuve préliminaire affaiblit la crédibilité de l'hypothèse.", "L'hypothèse sous-estime les défis pratiques de l'application de la CLR aux données de microarrays. Les microarrays, contrairement au RNA-seq, ne produisent pas des comptages mais des intensités de fluorescence. Le 'total mRNA' est une mesure indirecte et non une somme de molécules. De plus, la présence de valeurs négatives après RMA (background correction) peut rendre le logarithme de la CLR instable, nécessitant des ajustements ad hoc non discutés." ], "critical_questions": [ "Comment l'hypothèse se comporterait-elle face à des données où l'abondance absolue d'ARNm (estimée par spike-ins ou par des méthodes comme le CIBERSORT-ABS) est effectivement discriminante pour le sous-type de cancer ? La CLR, en normalisant par la moyenne géométrique, détruirait-elle ce signal ?", "L'effet de la CLR sur la réduction des corrélations spurious est-il vraiment indépendant du classifieur (Step 5) ? Les classifieurs linéaires (LASSO, SVM linéaire) sont sensibles à la multicolinéarité, tandis que les CNN extraient des motifs locaux. L'amélioration pourrait être inégale, avec un gain potentiellement plus important pour les modèles linéaires que pour les modèles profonds capables d'apprendre des dépendances complexes même en présence de bruit.", "Quelle est la stratégie de gestion des zéros (ou des valeurs d'expression très faibles) dans les microarrays pour l'application de la CLR ? L'hypothèse mentionne un seuil de 20%, mais dans la pratique, de nombreux gènes sont exprimés à des niveaux proches du bruit de fond. Une imputation inappropriée (ex: remplacer par une petite constante) pourrait introduire un biais systématique et fausser les log-ratios." ], "recommendation": "L'hypothèse est intellectuellement stimulante et repose sur des bases théoriques solides en analyse compositionnelle. Cependant, elle est trop ambitieuse dans ses prédictions quantitatives et ignore des aspects biologiques et techniques importants des données de microarrays. Je recommande une acceptation faible, conditionnée à une révision de l'hypothèse pour la rendre plus nuancée : remplacer le seuil de 8 pp par une gamme d'effets attendus (ex: amélioration modérée à substantielle) et ajouter une discussion sur les cas où l'abondance absolue est informative. Une étude pilote sur un jeu de données public (ex: TCGA) serait nécessaire avant de pouvoir justifier une validation à grande échelle.", "confidence": 0.78 }, { "reviewer_persona": "contrarian", "overall_score": 3.5, "verdict": "weak_reject", "strengths": [ "L'hypothèse aborde un problème réel et sous-estimé : l'impact du 'closure constraint' sur les corrélations entre gènes dans les données de puces à ADN, un artefact que les normalisations standard ne traitent pas explicitement.", "Le mécanisme proposé (chaîne causale) est logiquement cohérent et bien articulé, avec des hypothèses auxiliaires clairement énoncées, ce qui facilite la falsification." ], "weaknesses": [ "FAIL REASON #1: L'effet taille annoncé (≥8 pp de gain en balanced accuracy) est irréaliste pour un simple changement de prétraitement. Les gains de performance en classification transcriptomique sont généralement de l'ordre de 1-3 pp lorsqu'on compare des normalisations. Un gain de 8 pp suggère soit que la baseline RMA est catastrophique (ce qui n'est pas le cas dans la littérature), soit que le CLR introduit un biais de sélection de caractéristiques qui sur-optimise sur le jeu de validation croisée. Le scénario le plus probable est que le gain observé sera <3 pp, et non significatif après correction pour comparaisons multiples.", "FAIL REASON #2: L'hypothèse ignore un 'confounder' majeur : la structure de corrélation des gènes (e.g., modules co-exprimés) est biologiquement informative pour les sous-types de cancer. Le CLR, en forçant une transformation log-ratio par rapport à la moyenne géométrique, peut détruire ces structures de co-expression légitimes (non-spurieuses) qui sont cruciales pour les classifieurs. L'angle mort est que le CLR pourrait réduire le signal biologique réel autant que le bruit, annulant tout bénéfice. Les simulations avec des gènes indépendants ne capturent pas cette réalité.", "FAIL REASON #3: La prédiction 4 (absence de corrélation avec la cellularité) est naïve. L'estimation ESTIMATE est elle-même basée sur des signatures d'expression qui sont des compositions. Appliquer un CLR à des données déjà transformées par ESTIMATE pourrait créer une dépendance circulaire. De plus, si le CLR 'nettoie' les artefacts compositionnels, il devrait logiquement améliorer la classification plus dans les échantillons à faible cellularité (où le bruit de fond stromal/immun est plus fort). L'hypothèse nulle |r|<0.3 est trop large et ne teste pas cette interaction potentielle." ], "critical_questions": [ "Comment l'auteur distingue-t-il une corrélation 'spurieuse' due au closure constraint d'une corrélation biologique réelle entre gènes co-régulés ? Sans une simulation avec un modèle de bruit réaliste (e.g., bruit multiplicatif et additif, pas seulement des gènes indépendants), l'affirmation que le CLR améliore le SNR reste non fondée.", "L'effet de la dimensionnalité (D > 50 000 gènes) sur la stabilité du CLR est mentionné comme 'unknown unknown'. Mais la transformation CLR implique une moyenne géométrique sur tous les gènes. Dans un contexte où la majorité des gènes ne sont pas différentiellement exprimés, cette moyenne est dominée par le bruit. Avez-vous testé si le CLR avec un sous-ensemble de gènes (e.g., les 5000 plus variables) donne des résultats différents ? Si oui, cela suggère que l'effet est un artefact de la sélection de caractéristiques, pas de la géométrie." ], "recommendation": "Avant de soumettre à nouveau, l'auteur doit fournir une preuve de concept sur un jeu de données synthétique avec une structure de covariance réaliste (e.g., inspirée de données réelles de TCGA) où le signal biologique est connu. Il doit montrer que le CLR ne détruit pas les corrélations légitimes et que le gain de BAcc est robuste à la sélection de caractéristiques en amont. Une analyse de puissance statistique pour détecter un gain de 8 pp avec un échantillon de n=500 est également nécessaire.", "confidence": 0.85 }, { "reviewer_persona": "industrialist", "overall_score": 6.5, "verdict": "weak_accept", "strengths": [ "Marche adressable clair et quantifiable : le diagnostic moleculaire du cancer (marché estimé à $15B d'ici 2028) et les plateformes de classification de sous-types (ex. MammaPrint, Oncotype DX) paieraient pour une amélioration de 8 pp de précision, car cela réduit les faux positifs et les biopsies inutiles, impactant directement le remboursement par les assureurs.", "Avantage concurrentiel défendable : le CLR est un préprocessing mathématique (non brevetable en soi), mais l'intégration dans un pipeline SaaS de bioinformatique clinique (ex. QIAGEN OmicSoft, Partek Flow) créerait un verrouillage par dépendance de workflow. Le gain de 8 pp en balanced accuracy est un argument marketing chiffré face à des solutions standards comme RMA ou TMM.", "Timeline réaliste pour un MVP : Phase 1 (in silico) en 6 semaines avec données TCGA publiques (coût < €2k) permet de valider le ROI avant tout engagement lourd. Un prototype cloud (AWS/GCP) pourrait être opérationnel en 6 mois pour une preuve de concept client." ], "weaknesses": [ "Barrière à l'entrée majeure : le CLR est un outil mathématique connu et non brevetable. L'IP potentielle réside uniquement dans l'implémentation logicielle ou l'interface clinique, ce qui est facilement contournable par des concurrents (ex. Illumina, Thermo Fisher) qui intègreraient la même transformation dans leurs pipelines propriétaires.", "Risque commercial : les laboratoires cliniques (ex. LabCorp, Quest) sont extrêmement conservateurs sur les changements de pipeline validés par la FDA/CE-IVD. Un changement de préprocessing nécessite une revalidation complète du test (coût > $500k, délai 12-24 mois), ce qui freine l'adoption malgré le gain technique.", "Dépendance au type de cancer : la prédiction 4 (pas de corrélation avec la cellularité) est fragile. Si le gain est confondu par le type de tissu (ex. tumeurs solides vs liquides), le marché se réduit aux cancers à haute cellularité (ex. sein, poumon), excluant les leucémies ou biopsies liquides (faible cellularité)." ], "critical_questions": [ "Quel est le coût total de revalidation d'un pipeline clinique existant (ex. test de classification BRCA par RT-qPCR) pour intégrer le CLR, et combien de clients (laboratoires) accepteraient de payer cette revalidation contre une amélioration de 8 pp de précision ?", "Existe-t-il un brevet ou une demande de brevet sur l'application du CLR spécifiquement à la classification de sous-types de cancer (par opposition à l'usage général en métagénomique) ? Si non, comment empêcher Illumina d'ajouter le CLR par défaut dans son pipeline DRAGEN ?", "Quel est le modèle de tarification : licence par échantillon (€0.10-0.50) ou abonnement SaaS annuel (€50k-200k par labo) ? Le volume d'échantillons de classification de cancer (ex. 500k tests/an aux US) justifie-t-il un prix premium par rapport à la normalisation standard gratuite (RMA) ?" ], "recommendation": "Lancer immédiatement la Phase 1 in silico (€2k, 6 semaines) sur 5 jeux de données TCGA (BRCA, LUAD, COAD, OV, LUSC) pour confirmer le gain de 8 pp. Si validé, déposer un brevet provisoire US sur le 'Procédé de classification de sous-type de cancer par transformation CLR intégrée dans un pipeline de diagnostic' (focus sur la méthode d'intégration, pas sur le CLR lui-même). Parallèlement, approcher 2-3 laboratoires académiques (ex. MD Anderson, Gustave Roussy) pour un pilote clinique gratuit (Phase 2) en échange d'une publication co-signée. Ne pas investir dans un produit commercial avant d'avoir signé une lettre d'intention avec un laboratoire de diagnostic (ex. Veracyte, NanoString) prêt à co-développer le pipeline validé.", "confidence": 0.65 }, { "reviewer_persona": "funding_strategist", "overall_score": 6.8, "verdict": "accept", "strengths": [ "Hypothèse forte et falsifiable avec un mécanisme causal clair (CLR supprime les corrélations artefactuelles dues à la contrainte de somme constante), ce qui est rare et apprécié des évaluateurs ANR/ERC.", "Protocole de validation en 3 phases avec des critères GO/NO-GO explicites, démontrant une maturité méthodologique et une gestion des risques, ce qui rassure les financeurs sur la faisabilité.", "Budget modeste (20k-80k €) pour un potentiel impact élevé : si validé, le CLR pourrait devenir un standard de prétraitement en génomique, avec un fort potentiel de publication et de réutilisation.", "Transfert méthodologique cross-domaine (géochimie -> biologie computationnelle) original, positionnant la demande comme innovante et interdisciplinaire." ], "weaknesses": [ "TRL actuel très bas (TRL 1-2) : l'hypothèse n'est validée que sur des simulations ou données publiques non contrôlées. Les programmes ANR/ERC exigent généralement un TRL 3-4 pour un financement significatif.", "Consortium non défini : l'hypothèse est portée par un seul chercheur ou une petite équipe. Les appels collaboratifs (Horizon Europe) nécessitent un consortium multi-partenaires avec des rôles clairs (bioinformaticiens, cliniciens, statisticiens).", "Manque de preuve de reproductibilité sur des jeux de données cliniques réels : l'effet du CLR pourrait être masqué par le bruit technique ou la variabilité biologique, ce qui n'est pas testé dans le protocole actuel.", "Risque de non-généralisabilité : l'hypothèse est testée sur des données de puces à ADN (microarray), mais les technologies RNA-seq dominent aujourd'hui. Le CLR pourrait ne pas s'appliquer directement aux comptages de reads." ], "critical_questions": [ "Quel est le TRL actuel de votre méthode sur des données cliniques réelles (ex : TCGA) ? Avez-vous déjà testé le CLR sur au moins un jeu de données RNA-seq pour vérifier la transférabilité ?", "Quel serait le consortium idéal pour passer à un financement de type ANR PRC ou Horizon Europe ? Avez-vous identifié des partenaires cliniciens (pour l'accès aux échantillons) et des experts en normalisation de données omiques ?", "Comment comptez-vous gérer le passage à l'échelle (high-dimensional, low-sample-size) avec des réseaux de neurones profonds, qui sont sujets au surapprentissage ? Votre protocole inclut-il des techniques de régularisation avancées ?" ], "recommendation": "Ciblez d'abord un financement de preuve de concept (ERC PoC ou ANR JCJC) pour valider l'hypothèse sur des données cliniques réelles (TRL 3-4). Ensuite, si les résultats sont positifs, montez un consortium pour un appel Horizon Europe Cluster 1 (Santé) ou un ERC Consolidator Grant. Évitez les appels trop larges (ex : ERC Advanced) à ce stade.", "confidence": 0.7, "funding_programs": [ { "program": "ERC Proof of Concept Grant 2025 (ERC-2025-PoC)", "agency": "European Research Council (ERC)", "fit_score": 0.85, "typical_budget": "150 000 € (forfait unique)", "success_rate": "40% (moyenne ERC PoC)", "next_deadline": "Mars 2025 (date indicative)", "rationale": "Ce programme est idéal pour valider une hypothèse à TRL bas avec un budget modeste. L'ERC PoC finance des activités de pré-commercialisation ou de validation de concept. Votre hypothèse, si elle est confirmée, pourrait déboucher sur un logiciel ou un brevet. Le protocole en 3 phases correspond parfaitement au format PoC." }, { "program": "ANR Jeunes Chercheuses et Jeunes Chercheurs (JCJC) 2025", "agency": "Agence Nationale de la Recherche (ANR)", "fit_score": 0.75, "typical_budget": "150 000 - 250 000 €", "success_rate": "15-20% (ANR JCJC)", "next_deadline": "Octobre 2025 (générique)", "rationale": "L'ANR JCJC finance des projets individuels ou en petit consortium (2-3 partenaires) avec un fort potentiel de rupture. Votre hypothèse est originale et bien formulée. Le budget demandé (20-80k) est inférieur au plafond, ce qui est un atout. Vous devrez ajouter une phase de validation sur données RNA-seq et un plan de valorisation (logiciel)." }, { "program": "Horizon Europe - Cluster 1: Health - HORIZON-HLTH-2025-DISEASE-01 (Personalised Medicine and Rare Diseases)", "agency": "European Commission (EC)", "fit_score": 0.6, "typical_budget": "5-8 millions € (consortium de 6-10 partenaires)", "success_rate": "10-15%", "next_deadline": "Avril 2025 (date indicative)", "rationale": "Ce programme est adapté si vous élargissez l'hypothèse à une application clinique précise (ex : classification des sous-types de cancer du sein pour guider le traitement). Vous devrez monter un consortium incluant des cliniciens, des bioinformaticiens, des experts en IA et des industriels (ex : Illumina). Le TRL cible est 5-6. Votre hypothèse actuelle est trop amont pour cet appel, mais elle pourrait en être le socle." } ] } ], "meta_review": { "consensus_score": 6.44, "verdict": "publish_brief", "key_consensus": [ "Tous les reviewers reconnaissent que l'hypothèse aborde un problème réel et sous-estimé : l'impact du 'closure constraint' sur les corrélations dans les données de microarrays, et que le transfert méthodologique de la CoDA vers la classification du cancer est une extension logique et potentiellement fructueuse.", "Le protocole de validation en 3 phases avec des critères GO/NO-GO explicites est unanimement salué comme un modèle de rigueur hypothético-déductive et de gestion des risques, ce qui rassure sur la faisabilité et la maturité méthodologique.", "L'ampleur de l'effet annoncé (≥8 pp de gain en balanced accuracy) est jugée trop précise et ambitieuse par le domain_expert et le contrarian, qui recommandent de la remplacer par une gamme d'effets attendus plus réaliste (1-3 pp).", "Le manque de preuve de concept sur des données réelles avec une structure de covariance réaliste est un point faible critique : le contrarian exige une simulation réaliste avant tout investissement, et le funding_strategist note que le TRL actuel est trop bas (1-2) pour des financements majeurs." ], "key_disagreements": [ "Le contrarian (score 3.5, confidence 0.85) estime que l'effet taille de 8 pp est irréaliste et que le CLR pourrait détruire des structures de co-expression biologiquement informatives, tandis que le methodologist (score 8.2, confidence 0.92) et le domain_expert (score 7.2, confidence 0.78) considèrent l'hypothèse comme théoriquement solide et bien conçue, malgré des réserves sur l'ampleur de l'effet.", "Le domain_expert soulève que l'Assumption 1 (quantité totale d'ARNm non informative) est une simplification excessive, car l'abondance absolue d'ARNm peut être discriminante dans certains cancers (via cellularité, aneuploïdie, CNA). Le methodologist et le contrarian ne contestent pas ce point, mais le contrarian l'utilise comme argument pour un 'fail reason' supplémentaire.", "Le funding_strategist recommande de cibler d'abord un financement de preuve de concept (ERC PoC) avant de viser des appels plus larges, tandis que l'industrialist préconise un lancement immédiat de la Phase 1 in silico pour valider le ROI avant tout engagement lourd. Ces deux visions ne sont pas contradictoires mais reflètent des temporalités différentes (académique vs. commerciale)." ], "critical_path": "Le facteur le plus determinant pour le succes/echec de cette hypothese est la validation de l'effet taille annonce (≥8 pp) sur un jeu de donnees synthetique avec une structure de covariance realiste (inspiree de donnees TCGA) et sur au moins un jeu de donnees clinique reel (ex: TCGA BRCA). Si le gain observe est <3 pp ou non significatif apres correction pour comparaisons multiples, l'hypothese centrale s'effondre. La gestion des zeros et des valeurs negatives dans les microarrays pour l'application de la CLR est un point technique bloquant qui doit etre resolu avant toute validation a grande echelle.", "final_recommendation": "L'hypothèse est intellectuellement stimulante, repose sur des bases théoriques solides en analyse compositionnelle, et le protocole de validation est exemplaire. Cependant, l'ampleur de l'effet annoncé (≥8 pp) est trop ambitieuse et non soutenue par des preuves préliminaires. Le panel recommande de publier un brief de recherche décrivant l'hypothèse, le mécanisme causal et le protocole de validation, mais en remplaçant le seuil de 8 pp par une gamme d'effets attendus (ex: amélioration modérée à substantielle, 1-5 pp) et en ajoutant une discussion explicite sur les cas où l'abondance absolue d'ARNm est informative. Une preuve de concept sur données synthétiques réalistes et sur un jeu de données TCGA est nécessaire avant de pouvoir justifier une validation à grande échelle. Le brief servira de base pour un financement de type ERC Proof of Concept.", "brief_quality_gate": true, "revision_guidance": [], "llm_verdict": "publish_brief", "llm_consensus_score": 7.2 } }, "vulgarization_fr": { "title_fr": "Cancer : et si le bruit de fond des puces à ADN cachait les vrais signaux ?", "hypothesis_in_brief": "Pour classer les sous-types de cancer à partir de l'expression des gènes, les méthodes actuelles normalisent les données en supposant que la quantité totale d'ARN mesurée n'a pas d'importance. Cette hypothèse propose au contraire que cette quantité est un artefact technique qui fausse les résultats. En appliquant une transformation mathématique venue de la géochimie (le log-rapport centré), qui traite chaque mesure comme une proportion relative plutôt qu'une valeur absolue, on pourrait éliminer ces fausses corrélations et améliorer la précision des diagnostics par au moins 8 points de pourcentage.", "why_it_matters": "Le diagnostic moléculaire du cancer repose souvent sur des classifieurs qui analysent l'activité de milliers de gènes à partir d'une puce à ADN. Mais un biais mathématique, appelé 'contrainte de somme constante', crée des corrélations négatives artificielles entre les gènes les plus actifs : plus un gène s'exprime, plus les autres doivent paraître faibles, même s'ils ne le sont pas. Ce bruit de fond pourrait conduire à sélectionner de mauvais gènes indicateurs et à diminuer la fiabilité des tests. Si cette hypothèse se vérifie, un simple changement dans la façon de prétraiter les données, sans toucher aux machines ni aux protocoles de laboratoire, permettrait d'affiner le diagnostic des sous-types de cancer, de réduire les faux positifs et d'éviter des biopsies inutiles.", "imagine_that": "Imaginez que vous deviez deviner le menu d'un dîner à partir de la composition de l'assiette de chaque invité. Mais chaque assiette est remplie jusqu'au bord : si vous mettez plus de purée, il y a forcément moins de viande, même si le cuisinier a servi la même quantité de viande à tout le monde. C'est le problème des puces à ADN : la mesure totale est fixe, donc l'augmentation d'un gène fait mécaniquement baisser les autres. La transformation proposée revient à peser chaque ingrédient séparément avant de le mettre dans l'assiette, pour retrouver les vraies proportions.", "concretely": { "intro": "Pour tester cette idée, le protocole prévoit trois phases de validation, chacune avec des critères d'arrêt clairs.", "phase1": "D'abord, des simulations sur ordinateur : on prendra des données publiques de tumeurs du sein (500 échantillons, 4 sous-types) et on comparera la précision d'un classifieur standard avec et sans la nouvelle transformation. Si le gain n'atteint pas 8 points, l'hypothèse sera abandonnée.", "phase2": "Ensuite, une expérience de laboratoire minimaliste : on mélangera de l'ARN purifié de quatre lignées cellulaires cancéreuses différentes, dans des proportions connues, pour créer des échantillons synthétiques. On vérifiera que la transformation retrouve bien les proportions réelles, sans être perturbée par l'artefact de somme constante.", "phase3": "Enfin, une validation à grande échelle sur 15 types de cancer différents et 5 jeux de données indépendants, en testant plusieurs types de classifieurs (réseaux de neurones, régressions linéaires) pour s'assurer que le bénéfice est général et ne dépend pas d'un algorithme particulier." }, "reviewers_say": "Le panel de relecture est divisé. Les experts en méthode saluent un protocole d'une rigueur exemplaire, avec des prédictions chiffrées et des critères d'arrêt objectifs. L'expert du domaine trouve l'idée théoriquement solide mais juge le gain annoncé de 8 points trop ambitieux : dans la réalité, un simple changement de prétraitement gagne rarement plus de 2 à 3 points. Un contradicteur va plus loin : selon lui, la transformation pourrait détruire des structures de co-expression biologiquement importantes, annulant tout bénéfice. L'industriel voit un marché potentiel mais souligne que les laboratoires cliniques sont très réticents à changer leurs pipelines validés. Le verdict final est un 'publish brief' : l'hypothèse mérite d'être testée, mais il faudra des preuves solides, notamment sur des données synthétiques réalistes, pour convaincre les sceptiques." }, "vulgarization_en": { "title": "Cancer: could the background noise of DNA microarrays conceal the true signals?", "hypothesis_in_brief": "To classify cancer subtypes from gene expression, current methods normalise data by assuming that the total quantity of measured RNA is irrelevant. The present hypothesis proposes, by contrast, that this quantity is a technical artefact that distorts results. By applying a mathematical transformation derived from geochemistry (the centred log-ratio), which treats each measurement as a relative proportion rather than an absolute value, these spurious correlations could be eliminated and diagnostic accuracy improved by at least 8 percentage points.", "why_it_matters": "Molecular cancer diagnosis often relies on classifiers that analyse the activity of thousands of genes from a DNA microarray. However, a mathematical bias known as the 'constant sum constraint' creates artificial negative correlations among the most active genes: the more one gene is expressed, the weaker others must appear, even if they are not. This background noise could lead to the selection of poor indicator genes and reduce the reliability of tests. If this hypothesis is verified, a simple change in how data are preprocessed—without altering machines or laboratory protocols—would refine the diagnosis of cancer subtypes, reduce false positives, and avoid unnecessary biopsies.", "imagine_that": "Imagine that you must deduce the menu of a dinner from the composition of each guest’s plate. But each plate is filled to the brim: if you add more mashed potato, there is necessarily less meat, even if the cook served the same amount of meat to everyone. This is the problem of DNA microarrays: the total measurement is fixed, so an increase in one gene mechanically lowers the others. The proposed transformation amounts to weighing each ingredient separately before placing it on the plate, in order to recover the true proportions.", "reviewers_say": "The panel review is divided. Method experts commend a protocol of exemplary rigour, with quantified predictions and objective stopping criteria. The domain expert finds the idea theoretically sound but judges the announced gain of 8 points too ambitious: in practice, a simple change in pre-treatment rarely yields more than 2 to 3 points. A dissenter goes further: in their view, the transformation could destroy biologically important co-expression structures, nullifying any benefit. The industry representative sees a potential market but notes that clinical laboratories are highly reluctant to alter their validated pipelines. The final verdict is a 'publish brief': the hypothesis merits testing, but solid evidence — particularly on realistic synthetic data — will be required to convince the sceptics.", "concretely": { "intro": "To test this hypothesis, the protocol provides for three validation phases, each with clear stopping criteria.", "phase1": "First, computer simulations are performed: public breast tumour data (500 samples, 4 subtypes) are taken, and the accuracy of a standard classifier is compared with and without the new transformation. If the gain does not reach 8 points, the hypothesis is abandoned.", "phase2": "Subsequently, a minimalist laboratory experiment is conducted: purified RNA from four distinct cancer cell lines is mixed in known proportions to create synthetic samples. It is then verified that the transformation correctly recovers the actual proportions, without being perturbed by the constant-sum artefact.", "phase3": "Finally, a large-scale validation was conducted across 15 distinct cancer types and 5 independent datasets, with multiple classifier types (neural networks, linear regressions) being tested to ensure that the benefit is general and does not depend on any particular algorithm." } }, "panel_en": { "reviews": [ { "reviewer_persona": "methodologist", "overall_score": 8.2, "verdict": "strong_accept", "confidence": 0.92, "strengths": [ "The protocol is remarkably well structured, with a logical three-phase progression (in silico, minimal experimental, full validation), each phase governed by clear GO/NO-GO/PIVOT criteria, thereby enabling objective decision-making and efficient resource allocation before proceeding to more costly stages.", "The extremely detailed specification of falsifiable predictions, with precise quantitative bounds (e.g., ΔBAcc ≥ 8 pp, 95% CI ≥ 5 pp), explicit statistical methods (paired t-test, Wilcoxon, Benjamini-Hochberg correction), and clear null hypotheses, constitutes a model of hypothetico-deductive rigour.", "The identification and proactive management of biases are exemplary: selection bias is addressed through validation on 15 TCGA cancer types and 5 independent GEO datasets; confirmation bias is treated by the use of simulated data with known ground truth; and measurement bias is handled via sensitivity analysis across three zero-imputation methods." ], "weaknesses": [ "The statistical power for prediction 4 (correlation with cellularity) is potentially insufficient: with only 10 cancer types, the width of the 95% CI for a Pearson correlation coefficient will be very large (approximately ±0.6), rendering the criterion |r| < 0.3 with an upper bound < 0.5 difficult to satisfy even if the true correlation is zero.", "The protocol does not explicitly address potential publication bias: if results are negative for certain cancer types or classifiers, is there a plan to publish these null results? The risk of a 'file drawer problem' is real, especially with 15 cancer types × 6 classifiers = 90 combinations.", "Phase 2 (minimal experimental validation) uses synthetic mixtures of cell lines which, although controlled, do not reproduce the complexity of real tumour heterogeneity (immune infiltration, stroma, subclones). The transition from Phase 2 to Phase 3 could reveal unanticipated confounding factors." ], "critical_questions": [ "How is the risk of overfitting to be managed during feature selection via LASSO, given that lambda.1se is used but the number of genes (p ≈ 50,000) is 100 times greater than the number of samples (n ≈ 500)? Has nested cross-validation been considered for an unbiased estimation of the generalisation error?", "For Prediction 3 on the simulated data, how is it guaranteed that the compositional correlation structure (constant-sum constraint) is correctly modelled? The use of mvtnorm with a standard covariance matrix may fail to capture the non-linear dependencies induced by compositional closure. Has the use of a multivariate logistic-normal distribution been considered?", "What is the analysis plan if the results show a significant improvement for certain classifiers (e.g., LASSO, SVM) but not for others (e.g., LSTM-DBN)? The protocol mentions a pivot for 'classifier-specific failure', but how will a failure due to a genuine absence of effect be distinguished from a failure due to insufficient statistical power for the more complex models?" ], "recommendation": "This protocol is exceptionally well designed and merits acceptance. The panel recommends, however, that the number of cancer types in prediction 4 be increased to at least 20 (using additional GEO datasets) to obtain adequate statistical power for the correlation test. Furthermore, the addition of a post-hoc sensitivity analysis on the effect of intra-tumoural heterogeneity (using single-cell sequencing data) would considerably strengthen the external validity of the conclusions." }, { "reviewer_persona": "domain_expert", "overall_score": 7.2, "verdict": "weak_accept", "confidence": 0.78, "strengths": [ "The hypothesis correctly identifies the fundamental problem of the closure constraint in microarray data, a point frequently overlooked in the cancer bioinformatics literature. The methodological transfer of CoDA, well established in microbial ecology and metabolomics, to cancer classification is a logical and potentially fruitful extension.", "The proposed causal chain (Steps 1–5) is consistent with the principles of Aitchison geometry. The argument that CLR removes artefactual negative correlations (the Pearson paradox) and improves the signal-to-noise ratio is theoretically sound and supported by the compositional analysis literature.", "The hypothesis is well delimited and testable. The underlying assumptions (Assumptions 1–4) are clearly stated, permitting rigorous experimental falsification. The mention of 'known unknowns' demonstrates a good awareness of the limitations and areas of uncertainty in the proposed mechanism." ], "weaknesses": [ "The central premise (Assumption 1) that total mRNA quantity per cell is uninformative for cancer subtype classification is an oversimplification. In many cancers, tumour cellularity, aneuploidy, and copy number alterations (CNAs) directly affect absolute mRNA abundance. Ignoring this information can lead to a loss of crucial biological signal, which CLR cannot recover.", "The hypothesis of an improvement of 'at least 8 percentage points' (pp) in balanced accuracy is overly precise and ambitious for a methodological transfer study. The effect size will depend heavily on the dataset (cancer type, microarray platform, technical noise) and the classifier. Such a quantitative claim without preliminary evidence weakens the credibility of the hypothesis.", "The hypothesis underestimates the practical challenges of applying CLR to microarray data. Microarrays, unlike RNA-seq, do not produce counts but fluorescence intensities. The 'total mRNA' is an indirect measure and not a sum of molecules. Furthermore, the presence of negative values after RMA (background correction) can render the logarithm of CLR unstable, requiring ad hoc adjustments that are not discussed." ], "critical_questions": [ "How would the hypothesis perform on data in which the absolute abundance of mRNA (estimated via spike-ins or methods such as CIBERSORT-ABS) is genuinely discriminative for the cancer subtype? Would the CLR, by normalising with the geometric mean, destroy this signal?", "Is the effect of the CLR on the reduction of spurious correlations truly independent of the classifier (Step 5)? Linear classifiers (LASSO, linear SVM) are sensitive to multicollinearity, whereas CNNs extract local patterns. The improvement might be uneven, with a potentially larger gain for linear models than for deep models capable of learning complex dependencies even in the presence of noise.", "What is the strategy for managing zeros (or very low expression values) in microarrays for the application of the CLR? The hypothesis mentions a threshold of 20%, but in practice, many genes are expressed at levels close to background noise. Inappropriate imputation (for example, replacing with a small constant) could introduce systematic bias and distort the log-ratios." ], "recommendation": "The hypothesis is intellectually stimulating and rests on solid theoretical foundations in compositional analysis. However, it is overly ambitious in its quantitative predictions and neglects important biological and technical aspects of microarray data. Weak acceptance is recommended, conditional on a revision of the hypothesis to render it more nuanced: the 8 pp threshold should be replaced with a range of expected effects (e.g., moderate to substantial improvement), and a discussion of cases in which absolute abundance is informative should be added. A pilot study on a public dataset (e.g., TCGA) would be necessary before large-scale validation could be justified." }, { "reviewer_persona": "contrarian", "overall_score": 3.5, "verdict": "weak_reject", "confidence": 0.85, "strengths": [ "The hypothesis addresses a real and underappreciated problem: the impact of the closure constraint on correlations between genes in DNA microarray data, an artefact that standard normalisation procedures do not explicitly treat.", "The proposed mechanism (causal chain) is logically coherent and well articulated, with auxiliary hypotheses clearly stated, which facilitates falsification." ], "weaknesses": [ "FAIL REASON #1: The reported effect size (≥8 pp gain in balanced accuracy) is unrealistic for a simple change in pre-processing. Performance gains in transcriptomic classification are generally on the order of 1–3 pp when normalisation methods are compared. A gain of 8 pp suggests either that the RMA baseline is catastrophic (which is not the case in the literature) or that CLR introduces a feature-selection bias that over-optimises on the cross-validation set. The most likely scenario is that the observed gain will be <3 pp and not significant after correction for multiple comparisons.", "FAIL REASON #2: The hypothesis ignores a major confounder: the correlation structure of genes (e.g., co-expressed modules) is biologically informative for cancer subtypes. CLR, by forcing a log-ratio transformation relative to the geometric mean, may destroy these legitimate (non-spurious) co-expression structures that are crucial for classifiers. The blind spot is that CLR could reduce real biological signal as much as noise, cancelling any benefit. Simulations with independent genes do not capture this reality.", "FAIL REASON #3: Prediction 4 (absence of correlation with cellularity) is naive. The ESTIMATE estimate itself is based on expression signatures that are compositions. Applying CLR to data already transformed by ESTIMATE could create a circular dependency. Moreover, if CLR 'cleans' compositional artefacts, it should logically improve classification more in samples with low cellularity (where stromal/immune background noise is stronger). The null hypothesis |r|<0.3 is too broad and does not test this potential interaction." ], "critical_questions": [ "How does the author distinguish a 'spurious' correlation arising from the closure constraint from a genuine biological correlation between co-regulated genes? Without a simulation employing a realistic noise model (e.g., multiplicative and additive noise, not merely independent genes), the claim that CLR improves the signal-to-noise ratio remains unsubstantiated.", "The effect of dimensionality (D > 50,000 genes) on the stability of CLR is noted as an 'unknown unknown'. However, the CLR transformation involves a geometric mean across all genes. In a context where the majority of genes are not differentially expressed, this mean is dominated by noise. Was it tested whether CLR applied to a subset of genes (e.g., the 5,000 most variable) yields different results? If so, this would suggest that the effect is an artefact of feature selection, not of geometry." ], "recommendation": "Before resubmission, the author must provide a proof of concept on a synthetic dataset with a realistic covariance structure (e.g., inspired by real TCGA data) in which the biological signal is known. It must be shown that the CLR does not destroy legitimate correlations and that the BAcc gain is robust to upstream feature selection. A statistical power analysis to detect an 8 pp gain with a sample of n=500 is also required." }, { "reviewer_persona": "industrialist", "overall_score": 6.5, "verdict": "weak_accept", "confidence": 0.65, "strengths": [ "A clear and quantifiable addressable market: molecular cancer diagnostics (market estimated at $15B by 2028) and subtype classification platforms (e.g., MammaPrint, Oncotype DX) would pay for an 8 percentage-point improvement in accuracy, as this reduces false positives and unnecessary biopsies, directly impacting insurer reimbursement.", "Defensible competitive advantage: CLR is a mathematical preprocessing step (not patentable per se), but integration into a clinical bioinformatics SaaS pipeline (e.g., QIAGEN OmicSoft, Partek Flow) would create lock-in through workflow dependency. The 8 percentage-point gain in balanced accuracy provides a quantified marketing argument against standard solutions such as RMA or TMM.", "Realistic timeline for an MVP: Phase 1 (in silico) in 6 weeks using public TCGA data (cost < €2k) permits validation of the ROI before any heavy commitment. A cloud prototype (AWS/GCP) could be operational within 6 months for a client proof-of-concept." ], "weaknesses": [ "A major barrier to entry: the CLR is a known mathematical tool and is not patentable. Potential IP resides solely in software implementation or the clinical interface, which can be easily circumvented by competitors (e.g., Illumina, Thermo Fisher) who would integrate the same transformation into their proprietary pipelines.", "Commercial risk: clinical laboratories (e.g., LabCorp, Quest) are extremely conservative regarding changes to FDA/CE-IVD-validated pipelines. A change in preprocessing necessitates complete revalidation of the test (cost > $500k, timeline 12–24 months), which impedes adoption despite the technical gain.", "Dependence on cancer type: prediction 4 (no correlation with cellularity) is fragile. If the gain is confounded by tissue type (e.g., solid tumours vs. liquid tumours), the market is reduced to high-cellularity cancers (e.g., breast, lung), excluding leukaemias or liquid biopsies (low cellularity)." ], "critical_questions": [ "What is the total cost of revalidating an existing clinical pipeline (e.g. a BRCA classification test by RT-qPCR) to incorporate CLR, and how many clients (laboratories) would agree to pay for this revalidation in exchange for an 8 percentage point improvement in accuracy?", "Does a patent or patent application exist for the application of CLR specifically to cancer subtype classification (as opposed to general use in metagenomics)? If not, how can Illumina be prevented from adding CLR by default into its DRAGEN pipeline?", "What is the pricing model: per-sample licence (€0.10–0.50) or annual SaaS subscription (€50k–200k per laboratory)? Does the volume of cancer classification samples (e.g. 500k tests per year in the US) justify a premium price relative to free standard normalisation (RMA)?" ], "recommendation": "Phase 1 *in silico* (€2k, 6 weeks) is to be launched immediately on five TCGA datasets (BRCA, LUAD, COAD, OV, LUSC) to confirm the 8 pp gain. If validated, a provisional US patent is to be filed on the 'Cancer subtype classification method by CLR transformation integrated into a diagnostic pipeline' (focus on the integration method, not on the CLR itself). In parallel, 2–3 academic laboratories (e.g., MD Anderson, Gustave Roussy) are to be approached for a free clinical pilot (Phase 2) in exchange for a co-authored publication. No investment in a commercial product is to be made before a letter of intent is signed with a diagnostic laboratory (e.g., Veracyte, NanoString) prepared to co-develop the validated pipeline." }, { "reviewer_persona": "funding_strategist", "overall_score": 6.8, "verdict": "accept", "confidence": 0.7, "funding_programs": [ { "program": "ERC Proof of Concept Grant 2025 (ERC-2025-PoC)", "agency": "European Research Council (ERC)", "fit_score": 0.85, "typical_budget": "150 000 € (forfait unique)", "success_rate": "40% (moyenne ERC PoC)", "next_deadline": "Mars 2025 (date indicative)", "rationale": "Ce programme est idéal pour valider une hypothèse à TRL bas avec un budget modeste. L'ERC PoC finance des activités de pré-commercialisation ou de validation de concept. Votre hypothèse, si elle est confirmée, pourrait déboucher sur un logiciel ou un brevet. Le protocole en 3 phases correspond parfaitement au format PoC." }, { "program": "ANR Jeunes Chercheuses et Jeunes Chercheurs (JCJC) 2025", "agency": "Agence Nationale de la Recherche (ANR)", "fit_score": 0.75, "typical_budget": "150 000 - 250 000 €", "success_rate": "15-20% (ANR JCJC)", "next_deadline": "Octobre 2025 (générique)", "rationale": "L'ANR JCJC finance des projets individuels ou en petit consortium (2-3 partenaires) avec un fort potentiel de rupture. Votre hypothèse est originale et bien formulée. Le budget demandé (20-80k) est inférieur au plafond, ce qui est un atout. Vous devrez ajouter une phase de validation sur données RNA-seq et un plan de valorisation (logiciel)." }, { "program": "Horizon Europe - Cluster 1: Health - HORIZON-HLTH-2025-DISEASE-01 (Personalised Medicine and Rare Diseases)", "agency": "European Commission (EC)", "fit_score": 0.6, "typical_budget": "5-8 millions € (consortium de 6-10 partenaires)", "success_rate": "10-15%", "next_deadline": "Avril 2025 (date indicative)", "rationale": "Ce programme est adapté si vous élargissez l'hypothèse à une application clinique précise (ex : classification des sous-types de cancer du sein pour guider le traitement). Vous devrez monter un consortium incluant des cliniciens, des bioinformaticiens, des experts en IA et des industriels (ex : Illumina). Le TRL cible est 5-6. Votre hypothèse actuelle est trop amont pour cet appel, mais elle pourrait en être le socle." } ], "strengths": [ "The hypothesis is strong and falsifiable, with a clear causal mechanism (CLR removes artefactual correlations arising from the constant-sum constraint), a feature that is rare and valued by ANR/ERC reviewers.", "A three-phase validation protocol is provided, with explicit GO/NO-GO criteria, demonstrating methodological maturity and risk management, which reassures funders regarding feasibility.", "The budget is modest (€20k–€80k) for a potentially high impact: if validated, CLR could become a standard pre-processing method in genomics, with strong publication and re-use potential.", "An original cross-domain methodological transfer (geochemistry → computational biology) is proposed, positioning the application as innovative and interdisciplinary." ], "weaknesses": [ "The current TRL is very low (TRL 1–2): the hypothesis is validated only on simulations or uncontrolled public data. ANR/ERC programmes generally require TRL 3–4 for significant funding.", "The consortium is not defined: the hypothesis is carried by a single researcher or a small team. Collaborative calls (Horizon Europe) require a multi-partner consortium with clear roles (bioinformaticians, clinicians, statisticians).", "A lack of evidence for reproducibility on real clinical datasets: the effect of the CLR could be masked by technical noise or biological variability, which is not tested in the current protocol.", "Risk of non-generalizability: the hypothesis is tested on microarray data, but RNA-seq technologies dominate today. The CLR might not apply directly to read counts." ], "critical_questions": [ "What is the current TRL of the method on real clinical data (e.g., TCGA)? Has the CLR been tested on at least one RNA-seq dataset to verify transferability?", "What would be the ideal consortium for transitioning to ANR PRC or Horizon Europe funding? Have clinical partners (for sample access) and experts in omics data normalisation been identified?", "How is the scaling to high-dimensional, low-sample-size settings with deep neural networks, which are prone to overfitting, intended to be managed? Does the protocol include advanced regularisation techniques?" ], "recommendation": "The panel recommends targeting a proof-of-concept funding scheme (ERC PoC or ANR JCJC) first, in order to validate the hypothesis on real clinical data (TRL 3–4). Subsequently, if the results are positive, a consortium should be assembled for a Horizon Europe Cluster 1 (Health) call or an ERC Consolidator Grant. Overly broad calls (e.g., ERC Advanced) should be avoided at this stage." } ], "meta_review": { "consensus_score": 6.44, "verdict": "publish_brief", "brief_quality_gate": true, "llm_verdict": "publish_brief", "llm_consensus_score": 7.2, "key_consensus": [ "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." ], "key_disagreements": [ "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 'fail reason'.", "The funding_strategist recommends targeting a proof-of-concept grant (ERC PoC) first before aiming at broader calls, whereas the industrialist 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)." ], "revision_guidance": [], "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." } } }