Sequential Decision-Making for Drug Design: Towards closed-loop drug design

Sammanfattning: Drug design is a process of trial and error to design molecules with a desired response toward a biological target, with the ultimate goal of finding a new medication. It is estimated to be up to 10^{60} molecules that are of potential interest as drugs, making it a difficult problem to find suitable molecules. A crucial part of drug design is to design and determine what molecules should be experimentally tested, to determine their activity toward the biological target. To experimentally test the properties of a molecule, it has to be successfully made, often requiring a sequence of reactions to obtain the desired product. Machine learning can be utilized to predict the outcome of a reaction, helping to find successful reactions, but requires data for the reaction type of interest. This thesis presents a work that combinatorially investigates the use of active learning to acquire training data for reaching a certain level of predictive ability in predicting whether a reaction is successful or not. However, only a limited number of molecules can often be synthesized every time. Therefore, another line of work in this thesis investigates which designed molecules should be experimentally tested, given a budget of experiments, to sequentially acquire new knowledge. This is formulated as a multi-armed bandit problem and we propose an algorithm to solve this problem. To suggest potential drug molecules to choose from, recent advances in machine learning have also enabled the use of generative models to design novel molecules with certain predicted properties. Previous work has formulated this as a reinforcement learning problem with success in designing and optimizing molecules with drug-like properties. This thesis presents a systematic comparison of different reinforcement learning algorithms for string-based generation of drug molecules. This includes a study of different ways of learning from previous and current batches of samples during the iterative generation.