Abstract:
Breast cancer is one of the most prevalent cancers worldwide and a leading cause of mortality among women. Personalized therapeutic strategies remain a pressing need due to the molecular heterogeneity of the disease, which significantly affects treatment outcomes. Drug response prediction has thus become a central task in precision oncology, enabling clinicians to identify effective drugs for individual patients. In this thesis, we propose a Deep Learning–based drug response prediction framework that utilizes drug sensitivity information from the Genomics of Drug Sensitivity in Cancer (GDSC) database together with multi-omics profiles (mutation, CNA, expression) from both GDSC cell lines and The Cancer Genome Atlas (TCGA) breast cancer patients. While GDSC provides experimentally measured IC50 values along with cell line omics data, TCGA offers patient-level multi-omics profiles. By leveraging both resources, the framework predicts patient-specific drug responses and identifies the most effective compounds. A multilayer perceptron (MLP)–based deep learning model was developed. This model achieved performance MSE: 0.3235, R2: 0.8376 and Pearson: 0.9158 and showed predictive capacity for clinically relevant drugs such as Olaparib, Staurosporine, and Mirin. Beyond prediction, this work incorporates a pseudo-labeling strategy, where the framework applies trained models to TCGA profiles as an external dataset. Predicted IC50 values are then used to identify potential effective drugs for each patient, thus extending drug response modeling toward personalized drug Identification with effective IC50 value. To ensure interpretability, SHAP (SHapley Additive exPlanations) was integrated, providing feature-level insights into genomic drivers of drug sensitivity. Finally, an interactive Gradio-based application was developed to allow clinicians to upload patient data and receive ranked drug Identification alongside interpretability plots. This system bridges predictive modeling with clinical usability, ensuring transparency and decision support in real-world oncology settings. This work contributes to the growing field of AI-driven personal genomics designed medicine by demonstrating how deep learning and multi-omics integration can enable individualized drug response prediction and personalized treatment design
