Technical-Repeat Sensitivity of Fixed-Reference PAM50 Subtyping

The study asks a narrow question: when public records describe repeated expression measurements linked to the same tumor RNA source, does one frozen computer method return the same PAM50 subtype label?

All 136 linked GSE96058 pairs were primary. The 87 HR-positive/HER2-negative pairs were secondary. A five-pair TCGA-BRCA check and a 16-pair E-TABM-576 cross-platform sensitivity analysis were registered separately and never pooled with the primary result.

GSE96058 and TCGA-BRCA used the fixed 50-gene panel, immutable GSE81538 reference, five fixed PAM50 centroids, Spearman scoring, and no cohort centering. E-TABM-576 used its preregistered reporter collapse and singleton-profile centering contract before the same fixed-centroid scoring.

Every registered pair stayed in its attempted denominator. The study did not impute missing values, inspect outcomes, adapt the method to results, or select a clinical or reliability threshold.

• GSE96058 primary: 131 of 136 matching labels, 96.32%.
• HR-positive/HER2-negative secondary: 83 of 87, 95.40%.
• TCGA-BRCA external check: 5 of 5, with a wide interval.
• E-TABM-576 sensitivity: 13 of 16, 81.25%.

The primary GSE96058 analysis showed high but imperfect agreement. The E-TABM-576 sensitivity estimate was numerically lower, but no between-source test was planned or performed, and platform, processing, reference, cohort, and preservation differences remain entangled.

The five-pair TCGA-BRCA result retained every label but remains imprecise. Its 95% Wilson interval extends down to 56.55%, so it is an external check rather than a broad validation claim.

The public package contains the versioned manuscript, builders, figures, checksum receipts, audit records, source-bound evidence, and deterministic manifest. It excludes raw molecular data, identifiers, outcomes, credentials, and controlled records.

Three committed builds reproduced the same PDF, manifest, and ZIP bytes. Every PDF page and the website's desktop and mobile layouts were visually inspected before release.

Technical-repeat agreement cannot establish diagnostic accuracy, prognosis, treatment response, clinical utility, biological truth, a population agreement rate, or general transportability.

This is a fully computational, public-data-only study. NaS handled no specimen and accessed no clinical outcome or controlled record. The report is not clinical guidance.