JAI: Physically Constrained Prediction of Conformer-Averaged Proton–Proton Coupling Constants in Solution

Machine-learning predictors of proton–proton coupling constants are usually trained on, and evaluated against, couplings computed for single geometries rather than the conformer-averaged values measured in solution. We describe JAI (J-coupling prediction with Artificial Intelligence), a graph transformer for 2J–5J(H,H) of small organic molecules trained on scaled Fermi-contact DFT couplings of about 3700 molecules, computed on geometries from the universal machine-learned potential UMA and averaged over conformers in six solvents, and on more than 25 000 curated experimental couplings annotated with their uncertainty, sign provenance and assignment. A fast mode predicts from the structure diagram; a conformer mode supplies, for each conformer of a reproducible search, the coefficients of Karplus-type expressions that are averaged with predicted populations; and a weighted combination of the two gives the final output, with 68 % and 90 % intervals calibrated on cross-validation errors. Against experimental couplings, the combined output reaches a mean absolute error of 0.43 Hz in cross-validation and 0.55 Hz when all compounds from the same literature source are held out together, our most realistic estimate for new chemistry. It reaches 0.35 Hz on 335 held-out compounds (fast mode: 0.53 Hz in cross-validation and 0.43 Hz on these compounds) and 0.57 Hz on 132 compounds added after the model had been finalized, 37 of them with environments that have very few DFT training examples. For vicinal couplings between sp3 carbons, it lowers the cross-validated error from 0.85 Hz with the fast mode to 0.66 Hz. The intervals hold for compounds like those of the training data but are too narrow for rare chemistry. Ablations on a previous version of the data set show that the physical expressions lower the error by 0.04 Hz overall, with a larger but less certain estimate for coupling types with few training exampl