Chemical Shift Prediction Beyond the Electronic Structure Limit

Abstract

Nuclear magnetic resonance (NMR) is a critical tool for developing pharmaceuticals, studying proteins, and identifying natural products. However, many diagnostic chemical shift variations remain too small to be interpreted by computational methods, making costly multidimensional NMR analyses unavoidable. Here, we show that existing shift prediction methods have reached the electronic structure limit but that the resulting accuracy plateau can be overcome by explicitly modeling rovibrational effects and solute-solvent interactions. Our approach is greatly accelerated by a newly developed family of equivariant neural networks, MagNET, and increases frontier prediction accuracy significantly. In more routine cases, where only implicit solvation is needed, MagNET reduces the cost of high-quality predictions from hours to seconds. These models were trained on a new dataset we release representing the largest collection of computed shieldings ever assembled, encompassing millions of diverse solutes and solvated poses. Our study establishes the machine-learning-accelerated modeling of electronic structure, rovibrational, and solvent effects as a physically grounded path towards chemical accuracy and automated, shift-only structural elucidation.