IMPRESSION Generation 2 – Accurate, fast and generalised neural network model for predicting NMR parameters in place of DFT.

Predicting 3D-aware Nuclear Magnetic Resonance (NMR) properties is critical for determining the 3D structure and dynamics, both stereochemical and conformational, of molecules in solution. Existing tools for such predictions are limited, being either relatively slow quantum chemical methods such as Density Functional Theory (DFT), or niche parameterised empirical or machine learning methods that only predict a single parameter type, often across only a limited chemical space. We present here IMPRESSION-Generation2 (G2), a transformer-based neural network which can be used as a direct replacement for high level DFT calculations in computational workflows of multiple classes of NMR parameter simultaneously, with time-savings of several orders of magnitude. IMPRESSION-G2 is the first system that simultaneously predicts all NMR chemical shifts, as well as scalar couplings for ¹H, ¹³C, ¹⁵N and ¹⁹F nuclei up to 4 bonds apart, in a single prediction event starting from a 3D molecular structure. The accuracy of this multi-parameter predictor in reproducing DFT-quality results for molecules containing C, H, N, O, F, Si, P, S, Cl, Br (~0.07ppm for ¹H chemical shifts, ~0.8ppm for ¹³C chemical shifts, <0.15Hz for ³J_HH scalar coupling constants) exceeds that of existing state-of-the-art empirical or machine learning systems and, critically, it also demonstrates generalisability when tested against molecules from sources that are completely independent of its own training data. Rapid NMR predictions take <50ms to predict on average ~5000 chemical shifts and scalar couplings per molecule, which is approximately 10⁶-times faster than DFT-based NMR predictions starting from a 3D structure. When combined with fast GFN-xTB geometry optimisations to generate the 3D input structures themselves in just a few seconds, a complete workflow for NMR predictions on a new 3D structure is 10³-10⁴ times faster than a complete DFT-based workflow for this. The accuracy and speed of IMPRESSION-G2 coupled to GFN-xTB suggests that it can be used to simply replace DFT for predicting 3D-aware NMR parameters.