Learned NMR Observables for Experiment-Conditioned Generative Prediction of Protein Conformational Ensembles

Overhauser effect (NOE) contacts encode ensemble-averaged interatomic distances directly. That information has remained hard to couple to generative predictors, which are trained to emit structures rather than to consume the observables an experiment provides. The team proposes to bridge NMR and generative structure prediction by learning the experimental observable itself. Using molecular-dynamics ensembles from mdCATH and a differentiable r⁻⁶ NOE map as physically-grounded supervision, they will train a generator that predicts a NOESY contact map directly from sequence. This learned observable serves as a prior where no measurement exists, and it is trained against the same physical quantity a spectrometer reports.

The researchers will then condition a diffusion-based structure generator on these maps to produce conformational ensembles, and, crucially, condition it equally on real experimental NOESY maps, so that measured data constrain the predicted ensemble rather than merely agreeing with it. The conditioning interface is modality-general: structure factors from X-ray crystallography enter the same way, which we will pursue as a demonstration that the approach extends beyond NMR.

The innovation is to treat an experimental observable as a learnable, sequence-predictable quantity that is physically anchored during training yet replaceable by real measurements at inference. This unifies two regimes that are usually separate: de novo ensemble prediction when no data are available, and experiment-conditioned prediction when they are. The work offers a general and extensible route to conformational ensembles for flexible and disordered proteins, to a means of interpreting sparse NMR data through a generative model, and to a template for incorporating other structure-determining experiments.