Comparative Study of Deep Learning Architectures for TP53 Mutation Classification in Breast Cancer Genomics

TP53 mutations are among the most common and clinically significant genetic alterations in breast cancer, often associated with poor prognosis and resistance to therapy. In this study, we leveraged RNA-Seq gene expression data and somatic mutation annotations from The Cancer Genome Atlas (TCGA-BRCA) to classify TP53 mutation status using deep learning models. Gene expression data were normalized using a log2(FPKM + 1) transformation, and the top informative genes were selected based on statistical variance. Samples were labeled as mutant or wild-type based on the presence of nonsynonymous TP53 mutations. We implemented and compared 5 deep learning architectures: Fully Connected Neural Network (FCNN), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Bidirectional Long Short-Term Memory (BiLSTM), and DNABERT, a Transformer-based model pretrained on genomic sequences. All models were trained under identical experimental conditions using PyTorch and TensorFlow, and evaluated using accuracy, precision, recall, F1-score, and AUROC. DNABERT achieved the highest classification accuracy and AUROC, followed by RNN, BiLSTM, and CNN, while FCNN demonstrated the lowest performance. These findings highlight the effectiveness of advanced deep learning architectures—particularly Transformer-based models—in accurately predicting TP53 mutation status and advancing AI-driven genomic diagnostics in breast cancer. Based on our findings, we recommend leveraging Transformer-based models such as DNABERT for genomic mutation detection, integrating multiomics data to improve accuracy, extending the approach to other cancer types for generalizability, and employing transfer learning to enhance performance in low-data scenarios.