DNA-Based Deep Learning and Association Studies for Drug Response Prediction in Leiomyosarcoma

Leiomyosarcoma (LMS) is a rare and aggressive soft tissue sarcoma with limited treatment options and poor prognosis. Standard therapies, including doxorubicin, gemcitabine, trabectedin, and pazopanib, demonstrate variable efficacy across patients, underscoring the need for predictive biomarkers and computational models to inform personalized therapy. We developed a deep learning framework using DNA mutation and expression data to predict multi-task binary drug responses in LMS. Feedforward neural networks (FNNs) and transformer-based models were trained with binary cross-entropy (BCE) and weighted BCE (WBCE) loss functions to address class imbalance. In addition to predictive modeling, we conducted statistical association studies to identify links between genomic alterations and drug sensitivity, and performed Kaplan-Meier survival analyses to assess the prognostic relevance. Transformer models outperformed FNN baselines, achieving an overall F1-score of 0.87. Association studies revealed biologically meaningful links: TP53 mutations correlated with doxorubicin resistance, RB1 deletions with gemcitabine non-response, ATRX mutations with poor pazopanib outcomes, and MDM2 amplification with trabectedin resistance. This study demonstrates the utility of DNA-driven deep learning combined with association studies for predicting drug responses in LMS. Our framework not only provides multi-task binary predictions but also yields biologically interpretable associations for the targeted DNAs, highlighting key genomic drivers of therapy resistance. These findings support the development of precision oncology strategies for this rare and challenging cancer.