Physics-Informed Neural Network for Conformational State Prediction from NMR Spectra

David Joseph (Germany)

LinkedIn: David Joseph; X: @DaJo_1729; Bluesky: @dajo-1729.bsky.social

Abstract: Determining conformational ensembles—and thereby the underlying conformational landscape—of peptides and proteins remains a central challenge in structural biology. Conventional NMR approaches rely on iterative, expert-driven interpretation of chemical shifts, NOEs, and scalar couplings, which can be time-consuming and difficult to scale. We introduce a physics-informed neural network for conformational state prediction directly from NMR spectra. The proposed model is a transformer-based architecture that maps raw 2D NMR experiments—including NOESY, HSQC, and TOCSY spectra—to multi-state conformational ensembles. For each state, the network predicts backbone torsion angles, side-chain dihedrals, inter-residue distances, and corresponding Boltzmann populations. Training is performed on a fully synthetic dataset generated by combining physics-based conformational sampling with a differentiable NMR pulse-sequence simulator, enabling the creation of paired (spectrum, ensemble) data without reliance on experimentally determined structures. Preliminary results on held-out validation data indicate that the model accurately recovers backbone geometries and population distributions from spectral inputs alone. Ongoing work focuses on extending the approach to longer sequences, variable ensemble sizes, and application to experimental NMR datasets.