Machine Learning for low-data drug discovery

The discovery of new drugs is essential for human well-being, as effective treatments are crucial for combating diseases, addressing emerging health threats, and improving overall quality of life. The continuous development of novel therapeutics is particularly important in response to evolving pathogens, drug resistance, and unmet medical needs. However, drug discovery is a complex and resource-intensive process, given the intricacy of biological systems and the vast chemical space of potential drug candidates. Traditional wet-lab experiments are costly and time-consuming, making the search for novel drugs akin to finding a needle in a haystack.

In this thesis, we investigate how machine learning systems can support early-stage drug discovery in scenarios where experimental data is scarce. First, in a practical study addressing a real-world low-data challenge - namely, the search for potential drug candidates during the COVID-19 pandemic - we apply and evaluate a zero-shot learning method. Furthermore, the thesis introduces a novel few-shot learning architecture, MHNfs, which enhances molecular representations through a memory-based mechanism that we refer to as context enrichment, where contextual information is retrieved from an external memory. The work also connects recent few-shot learning approaches in drug discovery with the concept of in-context learning, originally introduced for large language models. Finally, to support real-world applications, the thesis presents an interactive interface that enables chemists and drug designers to easily provide prompts to and interact with the MHNfs model, facilitating the integration of few-shot learning into their workflows.

Overall, the thesis contributes to the advancement of few-shot learning methods for drug discovery and their practical applications in real-world scenarios.