Case study
Deep-Learning Breast-Cancer Subtyping
1D convolutional network classifying five PAM50 molecular subtypes from TCGA-BRCA RNA-Seq (1,082 tumours) — leakage-safe gene selection, weighted-loss training with early stopping, 90.2% test accuracy (macro-F1 0.884) beating four classical baselines, and integrated-gradient gene attribution.
Problem
Classifying the five PAM50 molecular subtypes of breast cancer directly from RNA-Seq expression data.
Method
A 1D convolutional network trained on TCGA-BRCA RNA-Seq from 1,082 tumours, with leakage-safe gene selection, weighted-loss training with early stopping, and integrated-gradient gene attribution.
Result
90.2% test accuracy (macro-F1 0.884), beating four classical baselines.


