Development of Deep Learning Agents Transforming Raw NMR Data to Three Dimensional Protein Structure

The spectacular success of AlphaFold prompts us to propose analogous training of neural networks that transform raw lists of biomolecular NMR observables into structural coordinates. Two frameworks, ShiftFold and NOEfold, would learn to extract spatial constraints directly from chemical shifts and NOEs, respectively. Both would generate structural coordinates directly from spectra of proteins up to ≈ 10kDa, extendable by optional isotope labeling. They would be trained, evaluated, and benchmarked initially separately, then compared, confronted, and reconciled in a unified agent, bypassing the need for sequence-specific assignment and manual or computer-assisted inspection of spectra. While AlphaFold yields predictions, these agents would deliver a data-driven reconstruction of experimental structures. When unified, a powerful AI system trained on both experimental and sequence-based evolutionary information would emerge. AlphaFold’s most probable conformation would thus be refined in light of the dynamic, environment-dependent structural behavior that biomolecular NMR is uniquely positioned to detect.