An Interpretable Machine Learning Framework for Early- versus Late-Stage Classification of Breast Cancer from Gene Expression, with Co-expression Network Analysis
- 1CHANDIGARH UNIVERSITY
Abstract
Anatomical stage is a principal determinant of treatment and prognosis in breast cancer, yet it is unclear how much information about stage is carried by the transcriptome of the primary tumour. We asked whether an a priori panel of 65 cancer-related genes can distinguish early-stage (AJCC Stage I-II) from late-stage (Stage III-IV) breast cancer, and which genes contribute to any signal. RNA-seq expression and pathologic stage for The Cancer Genome Atlas breast invasive carcinoma cohort (TCGA-BRCA) were obtained from the UCSC Xena platform. After restricting the data to primary tumours with a known stage, 1,075 tumours were analysed (805 early, 270 late) and split into stratified training (n = 860) and held-out test (n = 215) sets. Logistic regression, a support vector machine, random forest and gradient boosting were tuned by stratified five-fold cross-validation, with median imputation, standardisation and SMOTE oversampling fitted on training data only and classification thresholds chosen on out-of-fold training predictions. On the test set, Logistic Regression had the highest ROC-AUC, 0.618 (95% CI 0.530-0.704), with balanced accuracy 0.578, sensitivity 0.741 and specificity 0.416; differences in AUC between models were not statistically significant (DeLong p = 0.199-0.585). Permutation importance in the Gradient Boosting model ranked POU5F1, FLT3 and NOTCH1 highest, but this model's AUC confidence interval included 0.5, so the ranking is exploratory. No gene was differentially expressed between stage groups after false-discovery-rate (FDR) correction (smallest FDR = 0.057, CD44). A co-expression network built from training tumours identified REL, EP300 and ESR1 as the most connected genes, a property that reflects co-expression rather than regulation or association with stage. Limitations include a single cohort, no external validation, no clinical covariates and a small number of late-stage test tumours. Expression of this gene panel carries limited but measurable information about anatomical stage; expression-based stage models should be combined with clinicopathological variables and validated in independent cohorts before any clinical application is considered.
Related
- An Interpretable EEG Machine Learning Prototype for Alzheimers Disease Classification
- Learning Epidemic Threshold Equations from Simulation: REEBLRA, an Interpretable Bayesian Classification Framework for Reproduction Numbers
- Subject-independent classification of ADHD from pediatric EEG: a leakage-aware deep learning framework with interpretable neurophysiological mapping
- Explainable computational framework for subtype classification and biomarker identification in head and neck squamous cell carcinoma using machine learning
- An integrated machine learning framework for cerebral aneurysm rupture status classification using clinical, morphological, and hemodynamic features
BibTeX
@misc{biswas2026interpretable,
title = {{An Interpretable Machine Learning Framework for Early- versus Late-Stage Classification of Breast Cancer from Gene Expression, with Co-expression Network Analysis}},
author = {S. Biswas and T. Huda},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.18.759962},
url = {https://www.biorxiv.org/content/10.64898/2026.10.18.759962v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - An Interpretable Machine Learning Framework for Early- versus Late-Stage Classification of Breast Cancer from Gene Expression, with Co-expression Network Analysis
AU - S. Biswas
AU - T. Huda
PY - 2026
DA - 2026/10/18
PB - bioRxiv
DO - 10.64898/2026.10.18.759962
UR - https://www.biorxiv.org/content/10.64898/2026.10.18.759962v1
AB - Anatomical stage is a principal determinant of treatment and prognosis in breast cancer, yet it is unclear how much information about stage is carried by the transcriptome of the primary tumour. We asked whether an a priori panel of 65 cancer-related genes can distinguish early-stage (AJCC Stage I-II) from late-stage (Stage III-IV) breast cancer, and which genes contribute to any signal. RNA-seq expression and pathologic stage for The Cancer Genome Atlas breast invasive carcinoma cohort (TCGA-BRCA) were obtained from the UCSC Xena platform. After restricting the data to primary tumours with a known stage, 1,075 tumours were analysed (805 early, 270 late) and split into stratified training (n = 860) and held-out test (n = 215) sets. Logistic regression, a support vector machine, random forest and gradient boosting were tuned by stratified five-fold cross-validation, with median imputation, standardisation and SMOTE oversampling fitted on training data only and classification thresholds chosen on out-of-fold training predictions. On the test set, Logistic Regression had the highest ROC-AUC, 0.618 (95% CI 0.530-0.704), with balanced accuracy 0.578, sensitivity 0.741 and specificity 0.416; differences in AUC between models were not statistically significant (DeLong p = 0.199-0.585). Permutation importance in the Gradient Boosting model ranked POU5F1, FLT3 and NOTCH1 highest, but this model's AUC confidence interval included 0.5, so the ranking is exploratory. No gene was differentially expressed between stage groups after false-discovery-rate (FDR) correction (smallest FDR = 0.057, CD44). A co-expression network built from training tumours identified REL, EP300 and ESR1 as the most connected genes, a property that reflects co-expression rather than regulation or association with stage. Limitations include a single cohort, no external validation, no clinical covariates and a small number of late-stage test tumours. Expression of this gene panel carries limited but measurable information about anatomical stage; expression-based stage models should be combined with clinicopathological variables and validated in independent cohorts before any clinical application is considered.
ER -