Assessing recurrent neural network accuracy and interpretability in breast cancer classification

Abstract

Ribonucleic acid sequencing-based gene expression profiling enables molecular-level cancer diagnosis. However, most machine learning approaches lack the interpretability required for clinical adoption. This investigation examined how accurately a recurrent neural network could classify gene expression data to distinguish malignant from healthy breast tissue, determine molecular subtypes, and reveal key gene expression patterns. It was hypothesised that a multi-task recurrent neural network would achieve high diagnostic accuracy while providing gene-level interpretability. A bidirectional long short-term memory model was trained on The Cancer Genome Atlas Breast Cancer dataset (1,109 malignant and 109 solid tissue normal samples) following extensive preprocessing, including reducing 20,530 genes to 5,000 features. The model achieved 96.5% accuracy for malignancy classification, with 100% sensitivity and 70.6% specificity, and 75.4% accuracy for molecular subtype classification. Explainability analysis identified transferrin, adenosine diphosphate ribosylation factor interacting protein 2, and moesin as key genes, highlighting roles of iron metabolism and cytoskeletal regulation. These findings demonstrate that recurrent neural networks can achieve clinically applicable accuracy for transcriptomic breast cancer diagnosis whilst providing interpretable, gene-level insights, thereby addressing critical barriers to artificial intelligence adoption in oncology.