Transfer Learning for Low-resource Genomic Mutation Detection

The detection of genomic mutations plays a central role in understanding cancer biology, enabling more precise diagnostic, prognostic, and therapeutic strategies. However, accurate mutation detection is often hindered in low-resource settings due to the limited availability of high-quality sequencing data, computational infrastructure, and expert annotation. These constraints are particularly problematic for key cancer-related genes such as TP53, which is frequently mutated in breast cancer and strongly associated with disease progression and treatment resistance. In recent years, deep learning approaches have shown significant promise in various genomics tasks, including variant calling, sequence classification, and gene expression analysis. Nevertheless, such models typically require large, well-annotated datasets to achieve reliable performance—a condition that is not always feasible in clinical or resource-constrained environments. Transfer learning, a machine learning technique in which knowledge learned from a large source dataset is transferred to a related target task, offers a potential solution to this limitation. It allows for the reuse of pretrained models, reducing the need for large amounts of labeled data while improving generalization. In this study, we propose and evaluate a deep learning framework for detecting TP53 mutations in breast cancer using either gene expression data or DNA sequences. We benchmark the performance of several architectures—namely, fully connected neural networks (FCNN), convolutional neural networks (CNN), bidirectional long short-term memory networks (BiLSTM), Transformers trained from scratch, and DNABERT models both trained from scratch and fine-tuned via transfer learning. All models were trained and tested on a curated subset of TCGA-BRCA samples labeled with TP53 mutation status. Our results show that DNABERT with transfer learning achieved the highest performance across all evaluation metrics, with an accuracy of 92%, F1-score of 0.90, and AUROC of 0.95. In contrast, traditional models such as FCNN and CNN using gene expression data yielded moderate performance (accuracy: 0.81–0.83), while models trained from scratch—including BiLSTM and Transformer—performed better when applied to DNA sequences (accuracy: 0.85–0.87). The consistent superiority of DNABERT highlights the value of pretrained genomic language models and transfer learning in resource-constrained scenarios. These findings underscore the promise of transformer-based models for scalable and accurate mutation detection, especially in clinical settings with limited-data availability. On the basis of our findings, we recommend adopting transfer learning approaches such as DNABERT in clinical genomics applications, integrating gene expression and sequence data to boost accuracy, and extending this framework to additional cancer-related genes to improve model generalizability.