Work

Decoupling Discrimination from Calibration: A Platform-Resilient, Uncertainty-Aware Transcriptomic Biomarker for Estrogen Receptor Phenotyping in Breast Cancer

Year: 2026
Type: preprint
Abstract: Background: Estrogen receptor alpha (ERα, encoded by ESR1) guides endocrine therapy eligibility for roughly 70–80% of breast cancer patients worldwide. Although immunohistochemistry (IHC) remains the clinical reference standard, pre-analytical tissue handling, cold ischemia delays, differences in antibody affinity, and subjective visual scoring contribute to a 5–10% diagnostic discordance rate. High-throughput transcriptomics provides an objective diagnostic alternative, yet existing machine-learning classifiers frequently suffer from cross-platform performance decay, unstable feature selection in high dimensions, and a critical vulnerability we define as the discrimination-calibration paradox—where models maintain high rank discrimination (>0.95 AUROC) while outputting severely distorted, uncalibrated risk probabilities under domain shift. Methods: We developed and evaluated the RAPID-ER framework across 4,218 rigorously audited breast cancer profiles from three independent cohorts: TCGA-BRCA (N = 1,044, primary discovery and 5-fold cross-validation), SCAN-B / GSE96058 (N = 3,073, external RNA-seq validation), and GSE42568 (N = 101, cross-platform Affymetrix microarray stress-test). Across a shared 17,367-gene feature space, we constructed the Stability-Selected Multilayer Perceptron (SS-MLP), combining 100-iteration repeated Elastic-Net bootstrap subsampling with a regularized deep neural network (256–64 GELU architecture). Two domain transfer strategies were benchmarked: parametric frozen training standardization (Strategy A) and label-free within-sample relative rank normalization (Strategy B). Probability calibration was assessed via Sanders/Murphy Brier score decomposition, Expected Calibration Error (ECE), and selective classification abstention. Results: Under internal 5-fold cross-validation on TCGA-BRCA, SS-MLP achieved an out-of-fold AUROC of 0.9560 [95% CI: 0.938, 0.970], AUPRC of 0.9851 [0.977, 0.991], Sensitivity of 92.07%, Specificity of 93.67%, and Brier score of 0.0675. In independent external RNA-seq validation (GSE96058, N = 3,073), univariate ESR1 models failed under frozen standardization: despite an AUROC of 0.9604, probability calibration degraded severely (ECE = 0.4290, Intercept = 5.2361), causing sensitivity to drop to 58.44% and misclassifying 1,177 true ER-positive patients. Conversely, SS-MLP preserved both high discrimination (AUROC 0.9500 [0.933, 0.965], Sensitivity 97.74%) and tight calibration, reducing calibration error by more than 10-fold (ECE = 0.0396, Brier = 0.0396). On cross-platform microarrays (GSE42568, N = 101), frozen standardization failed completely due to probe dynamic range compression (ESR1 Sensitivity 0.0000, Brier 0.6188). Within-sample rank normalization (Strategy B) restored cross-platform utility without retraining, lifting SS-MLP AUROC to 0.7943 [0.701, 0.874] and AUPRC to 0.8773. An in silico ESR1-exclusion ablation demonstrated complete downstream transcriptomic redundancy (AUROC 0.9608, AUPRC 0.9863). Finally, automated confidence triage with selective abstention achieved 99.6% autonomous coverage with a 3.92% error rate (falling to 2.06% at τ = 0.95), safely routing equivocal borderline cases for confirmatory pathology review. Conclusions: Strong rank discrimination can mask severe probability miscalibration under genomic domain shift. By integrating stability-selected multi-gene networks with label-free rank normalization and uncertainty-based abstention, RAPID-ER resolves the discrimination-calibration paradox and establishes an open, platform-resilient decision-support benchmark for precision oncology.
Language: English
Subfield: Cancer Research
Open Access status: green
OpenAlex ID: W7236751948
Cites:
Cited by:
Related to: 10