Graph neural network methods for representation and generation in drug discovery

Drug discovery is a time-consuming and expensive process, often spanning over a decade and costing billions of dollars. This thesis advances graph-based machine learning approaches to accelerate this process, making three main contributions. First, we provide a comprehensive review of graph neural networks for conditional molecular generation, establishing a framework for understanding and comparing different methods. Building on these insights, we introduce AMCG (Atomic-Molecular Conditional Generator), a novel generative framework that achieves state-of-the-art performance while offering one-shot generation capability and effective property optimization via gradient ascent. Motivated by the heterophilic nature of molecular graphs — where connected atoms often have dissimilar features — we then develop MaxCutPool, a differentiable graph pooling technique based on the MAXCUT problem. By combining graph-theoretical principles with deep learning, MaxCutPool demonstrates superior performance on heterophilic graphs while remaining competitive on standard benchmarks and maintaining computational efficiency. Together, these contributions advance both the theoretical foundations of graph representation learning and provide practical tools for accelerating drug discovery.