Meital Bojan

Nuclear Magnetic Resonance (NMR) spectroscopy is widely used in structural biology, yet analyzing NMR data remains largely manual, and especially assigning peaks to individual atoms in large biomolecules remains a laborious task that often constitutes the bottleneck of biomolecular NMR studies. This project aims to develop machine learning models that predict NMR chemical shifts directly from molecular structures, which will largely facilitate not only this task, but also allows interpreting NMR chemical shifts in terms of structure and dynamics. Using graph-based and equivariant neural networks, the work will focus on learning representations suitable for proteins and other complex molecules. A key objective is to extend these models towards automated peak assignment by linking predicted spectral data with experimental results. The project will explore approaches to improve model accuracy and reliability, including uncertainty estimation and the use of geometric information. By advancing automated spectral analysis, the project seeks to simplify and accelerate NMR-based structural studies.