Cancer: what if the trap lies in the method of measurement?

The articulation between the mechanistic hypothesis (a negative correlation bias arising from the constant-sum constraint) and the falsifiable predictions, quantified with precise bounds (a 15–30% reduction in FPR and a 5–12 percentage-point improvement in balanced accuracy), is judged to be excellent. This structure permits clear refutation and avoids vague conclusions.

The hypothesis correctly identifies a fundamental and often overlooked problem in gene expression data analysis: the compositional nature of microarray data and the negative correlation bias induced by the constant-sum constraint (Pearson’s paradox). The application of the log-ratio transformation (CLR/ILR) as a solution is theoretically grounded in Aitchison geometry of the simplex.

The identification of the negative correlation problem induced by the constant-sum constraint (Pearson paradox) is theoretically grounded and constitutes a legitimate concern for the analysis of DNA microarray data.

The addressable market is clear and quantifiable: molecular pathology laboratories and CROs (e.g., NeoGenomics, Foundation Medicine, Guardant Health) that perform cancer subtype classification assays from transcriptomic data (Affymetrix, Illumina microarrays). The cost of a false positive in gene selection is high (development of useless biomarker panels, clinical trial failures). A 5–12% gain in balanced accuracy translates directly into a reduction in the failure rate of prognostic signatures, which carries immediate commercial value.

Original and mechanistic hypothesis: the identification of negative correlation bias due to the constant-sum constraint (closure) in microarray data represents a rarely explored angle in biomedical ML, offering strong novelty potential for a reviewer.