Structure-based Design of Biomolecular Interactions with Geometric Deep Learning
Arne Schneuing · Infoscience (Ecole Polytechnique Fédérale de Lausanne) · 2026
Rationally engineering the molecules of life has major implications for the development of new materials, biosensors and medicines. To modulate their function, we typically need to create ligands that bind selectively to predetermined target molecules. Deep learning has recently emerged as a well suited approach for tackling the complex and probabilistic nature of biological phenomena. Building on this alignment, the latest computational protein and ligand design algorithms achieve remarkable successes, promising to substantially reduce reliance on high-throughput experimental testing and in vitro optimization. However, the fine-grained design of biomolecular interactions at the atomic level, a prerequisite for the design of small-molecule ligands, as well as higher-order interactions involving more than two molecules, remain challenging. In addition, the development of small-molecule design methods is slowed by limited experimental feedback due to costly chemical synthesis. This thesis explores generative models and geometric deep learning concepts for the design of biomolecular interactions. We show that diffusion and flow matching models provide a robust and versatile framework for molecular distribution learning, enabling the efficient generation of high-dimensional and multi-modal structural data. Combined with appropriate sampling and fine-tuning strategies, these models can be readily applied to a range of structure-based drug discovery tasks. Addressing the chemical synthesis bottleneck, this work further investigates the role of virtual chemical spaces and reaction templates in generative small-molecule design. We present advances in both top-down (retrosynthesis planning) and bottom-up (synthesis-constrained design) strategies, with the aim of making de novo approaches more accessible to experimental validation. Finally, we demonstrate how deep learning methods can be applied in low-data regimes by leveraging generalizable data representations. By adapting only the featurisation step to small molecules, we extend the applicability of a neural network trained exclusively on proteins beyond its original use case. Using this approach, we successfully design novel ternary interactions in which protein dimerisation depends on a small molecule that modulates the presented surface features. Overall, these findings showcase the versatility of deep learning in molecular design tasks and underscore the importance of selecting appropriate data representations and learning frameworks. While this dissertation primarily focuses on applications in structure-based drug discovery, many of the underlying concepts are directly applicable to other biotechnology problems and adjacent domains.