Onzekerheidsquantificatie voor betrouwbare voorspellingen van interacties tussen geneesmiddelen en doelwitten

The development of new therapeutic agents is a time- and resource-consuming process, characterized by high failure rates and development spans of over a decade until a compound can be put on the market. The use of machine learning (ML) approaches can contribute to easing these problems by using the extensive amount of data produced in the drug discovery pipeline to train computational models that can effectively support future projects with their knowledge. Quantitative structure-activity relationship (QSAR) models are well-established in computer-aided drug discovery for identifying compounds with desired features. They enable the prediction of biological activities or properties of chemical compounds based on their molecular structure. Due to the time-consuming and expensive nature of the drug discovery workflow, the reliability of machine learning predictions is essential. In this framework, uncertainty quantification (UQ) has emerged as a powerful tool that allows model users to assess the risk associated with a model prediction before deciding how to progress in the experimental pipeline. To optimally support an informed decision-making process, it is essential to accurately quantify the uncertainty in machine learning predictions. Previous work has concluded that ML models, especially modern neural networks, often fail to provide realistic estimates of the uncertainty associated with predictions in classification tasks, leading to poorly calibrated models. Since uncertainty estimates are only meaningful when they accurately reflect the true uncertainty of a prediction, this work focuses on achieving well-calibrated uncertainty estimation. The primary objective of this work is to identify practices that yield reliable uncertainty estimates for drug-target interaction predictions. Using both real-world and public datasets, we evaluate the performance of various uncertainty quantification approaches and highlight challenges that arise in realistic settings. This thesis is structured into two parts, each based on distinct datasets and addressing different research questions. In the first part, we use a real-world pharmaceutical dataset to conduct a comprehensive, large-scale evaluation of uncertainty estimation methods, including ensemble-based and Bayesian approaches. Furthermore, we use this realistic setting to systematically assess the distribution shifts in label and descriptor space and their impact on the capability of the uncertainty estimation methods. Finally, we adapt uncertainty quantification regression models to enable learning from censored labels, which are commonly available in pharmaceutical datasets but can rarely be used by conventional machine learning methods. The second part of this work focuses on using Hamiltonian Monte Carlo (HMC) sampling for UQ in QSAR models. Although HMC is regarded as the gold standard for posterior approximation, its high computational cost makes it impractical for most real-world applications. We adopt a low-complexity approach that enables HMC-based uncertainty quantification by sampling only parts of the machine learning model. Furthermore, we aim to develop a ground truth for uncertainty quantification in drug-target interaction predictions by applying HMC sampling across entire QSAR models. An additional focus is on identifying strategies that yield well-behaved HMC chains that can effectively sample from the posterior. By applying various UQ methods across different datasets, this work provides a comprehensive study of uncertainty estimation in drug-target interaction predictions. The findings provide valuable insights into how reliable uncertainty estimates can be achieved, ultimately supporting well-informed decision-making and contributing to a more resource-and time-efficient drug discovery pipeline. Furthermore, they highlight the challenges introduced by real-world data, emphasizing the complexity of identifying effective strategies for uncertainty estimation in QSAR models.