NMR spectra now yield molecular skeletons directly: inverse-IMPRESSION solves 11 of 19 experimental molecules

NMR spectra can now be inverted directly into a molecule's two-dimensional connectivity, without first enumerating large candidate sets and scoring them one by one. Traditional computer-assisted structure elucidation relies on logical-combinatorial algorithms that derive fragments and assemble them exhaustively; candidates grow exponentially with molecular size, ranking is costly and full automation of complex molecules is limited. On a simulated test set using only hydrogen and carbon chemical shifts plus COSY, HSQC and HMBC correlations, the platform with structure correction reaches 77.8% Top-1, 86.4% Top-3 and 88.7% Top-10 accuracy.

The platform splits bond assignment into three linked stages: one-shot prediction, stepwise bond filling and noise-augmented multishot sampling. The one-shot model predicts all bonds at once; edge features pass through the message-passing of a graph transformer network and a sigmoid layer outputs a bond probability for each atom pair, with 0.5 as the cutoff. On simulated data with only hydrogen and carbon NMR information, the one-shot model reaches just 5.3% single-prediction accuracy but 99.0% edge-level accuracy; the stepwise model raises single-prediction accuracy to 49.6%, and structure correction lifts it to 53.4%.

Adding noise comparable to experimental variability to the input spectra and resampling produces an ensemble of candidate structures that is then ranked by consistency. Nitrogen, oxygen and fluorine atoms carry no usable node or edge features under this spectral condition, so the one-shot model assigns these atoms to the wrong carbons. The stepwise model first removes bonds to these degenerate atoms and reassigns them one subvalent atom at a time; carbons whose predicted chemical shifts deviate by more than 10 ppm, together with their neighboring heteroatoms, have bonds removed and reconnected, raising Top-1 from 74.2% to 77.8%.

The demonstrated scope is molecules containing hydrogen, carbon, nitrogen, oxygen and fluorine with no more than 30 non-hydrogen atoms, with simulated test data drawn from PubChem and ChEMBL; experimental validation covers 19 molecules of 143.2–608.7 g/mol and nSPS 9.6–59.2, with spectra taken from the literature or measured in this work.

Among the experimental cases, the 11 successes span 143.2–480.7 g/mol and nSPS 9.6–56.5, while the 8 failures span 303.4–608.7 g/mol and nSPS 18.4–59.2; accuracy falls below 60% for molecules with more than ten heteroatoms, and removing HSQC and HMBC data leaves no structure correctly elucidated.