Single-Sample PAM50 Reliability and Abstention

PAM50 is widely used in breast cancer research to assign an intrinsic molecular subtype. Some implementations depend on the other samples being analyzed, which means the same tumor can receive a different result when the surrounding cohort changes.

NaS is investigating a stricter research contract. Instead of returning only one subtype, the proposed output would preserve the leading and second centroid scores, their margin, technical repeatability, data-quality state, reliability state, and an explicit reason to abstain when the evidence is insufficient.

The proposed population is clinically HR-positive and HER2-negative primary invasive breast tumors with interpretable receptor metadata and expression measurements that satisfy locked PAM50 input requirements.

The method would operate on one sample without reference to the other test samples. Gene mappings, transformations, centroids, numerical tolerances, perturbations, quality rules, margin rules, repeatability rules, and abstention logic would be frozen before molecular values or outcomes are analyzed.

• Report explicit missing-gene, invalid-mapping, and transformation failures.
• Estimate the proportion of samples in reliable, unstable, insufficient-data, and abstain states.
• Attempt the complete frozen procedure in GSE96058 without tuning it to validation outcomes.
• Separate analytical repeatability from biological truth, prognosis, treatment response, and clinical utility.

The initial evidence review found substantial prior work on PAM50 perturbation, uncertainty, and single-sample classification. That overlap required NaS to narrow the question rather than make a premature novelty claim.

The current work asks whether the exact reliability contract, calibration strategy, abstention states, and unchanged cross-platform transport remain scientifically differentiated and defensible. A no-go decision is a valid outcome if prior methods already resolve the question or if thresholds cannot be justified without using outcomes.

Public TCGA-BRCA expression files and processed GSE96058 artifacts are available, but source availability alone does not establish receptor completeness, PAM50 gene coverage, alias resolution, or compatibility with a locked cross-platform transformation.

The governed feasibility audit therefore remains limited to evidence appraisal and bounded metadata checks. Cohort construction, molecular analysis, outcome analysis, classifier execution, and threshold selection are not authorized at the current gate.

A repeatable centroid assignment is not necessarily a biologically correct assignment. Cross-platform differences can be confused with biological differences, and retrospective public cohorts cannot establish clinical utility.

This study is designed as a nonclinical research method. It will not direct patient testing, diagnosis, prognosis, or treatment. Any future biological or clinical claim would require a separate study, independent evidence, and its own review and release record.