Application of explainable artificial intelligence to precision oncology

This doctoral dissertation, entitled "Application of Explainable Artificial Intelligence to Precision Oncology," explores the development and evaluation of interpretable machine learning models for drug response prediction in personalized cancer therapy. It highlights the critical importance of model interpretability in facilitating the clinical translation of AI-driven solutions in oncology. First, it introduces Optimal Decision Trees (ODT), a novel tree-based model that enhances transparency in treatment guidance through intuitive graphical decision trees. The creation of the ODT R CRAN package ensures its accessibility for researchers and clinicians alike. The model was evaluated against established methods such as Random Forest and XGBoost, enhanced by the SEATS (Systematic Efficacy Assignment with Treatment Seats) method, specifically designed to optimize drug assignment solutions. Through a state-of-the-art review, the dissertation establishes a framework for developing Drug Response Prediction (DRP) models, emphasizing high-quality data, precise normalization, and rigorous validation methods, such as cell-blind and drug-blind cross-validations. These methodologies ensure that DRP models remain accurate, clinically relevant, and adaptable across various datasets. Central to this research is the introduction of SparseGO, a Visible Neural Network (VNN) that leverages the Gene Ontology (GO) hierarchy to enhance computational efficiency and facilitate the discovery of drug mechanisms of action (MoA). The Explainable AI (XAI) method, DeepMoA, employs DeepLIFT and Support Vector Machines (SVMs) to predict drug MoAs, supported by strong computational validation. The research extends to Patient-Derived Cell-lines (PDCs), a crucial step in aligning DRP modeling more closely with patient-specific tumor characteristics. SparseGO and ODT models were trained on this data. This work presents a comprehensive exploration of explainable AI in precision oncology, emphasizing the critical role of interpretability in advancing personalized cancer treatment. Future research will focus on validating these models across diverse datasets, verifying predictions experimentally, and exploring advanced techniques to improve model eneralization and interpretability. This work not only introduces novel methodologies and accessible tools but also paves the way for AI-driven insights to seamlessly integrate into personalized cancer care.