A low-intensity machine learning model has begun to identify an intuitive relationship between NMR spectral peaks and chemical reactivity, bypassing structural determination completely. This simplified approach addresses a wider question that asks how else we can use existing data within chemistry and adjacent fields to ultimately open the door to predicting or designing molecular properties such as toxicity or optical absorption without reference to structure.
NMR is the gold standard tool used by chemists to elucidate molecular structure. Through a combination of deduction and rule-of-thumb reasoning, chemists convert chemical shift data into functional groups and atom connectivity, using this structure-first thinking to then rationalise chemical behaviour. But while this is the conventional application of this data, it need not be the only one, and Julie Zimmerman and her team at Yale University, US, instead sought to investigate how spectral information maps onto other properties.
In particular, the energy gap (ΔE) between the highest occupied molecular orbital (Homo) and the lowest unoccupied molecular orbital (Lumo) is a key determinant, not just of chemical behaviour, but physical properties like optical absorption and conductivity too. Chemical shift data essentially provide a readout of local electron density and the wider interactions of these environments over a given molecule, both key influences on this Homo-Lumo energy gap.
Zimmerman’s team therefore developed a straightforward machine learning model to predict ΔE directly from the NMR spectral data of over 100,000 small molecules. Rather than working with the full complexity of each spectrum, they broke the chemical shift data down into simpler features such as the number of peaks, the mean, and maximum or minimum chemical shift values, which they then fed into a mathematical model.
Broadly, the model proved generalisable and gave a reasonable estimate of ΔE across the molecules tested. ‘The most downfield carbon peak seemed to have a pretty large role in predicting the reactivity and that can be traced to this idea of the most electrophilic site,’ explains Tobias Muellers, first-author on the work. ‘However, we saw a pretty clear decrease in model performance for molecules containing sulfur and boron, and we also saw a lot of problems with zwitterions.’ Encouragingly, the model was able to distinguish between molecules which absorbed visible and UV light, showing its potential as a screening tool or an aid to molecule design, adds Muellers.
The system is still in its infancy and was principally trained on computational NMR data owing to a lack of clean experimental samples. Consequently, its wider accuracy still lags behind more conventional methods such as DFT simulations or structure-based models, but state-of-the-art performance was never the principal goal at this stage. ‘We find that we’re not the best performing in this area, but we’re using way fewer variables. Some of these models are using 400 features, whereas we have eight,’ Muellers explains. ‘We’re trying to build something that’s interpretable by people and to show there are rational, intuitive and understandable ways to use NMR that directly relate to chemical behaviour.’
For Kevin Jablonka, a digital chemist at Friedrich Schiller University Jena in Germany, the concept of using experimental data to directly predict molecular properties is interesting, but he questions whether ΔE is the most important parameter for chemists to probe. ‘NMR is super rich: it has constitution, structure, dynamics,’ he says. ‘[This type of methodology] could be useful in lab automation, where one uses this as a cheap heuristic for predicting an experimental property one cares about.’
However, user-friendly, chemistry-adjacent applications are the specific focus of the Zimmerman group’s future research in this area. ‘We’re at the interface of chemistry and toxicology in our work so we’re thinking about mechanism-specific descriptors for reactivity and how we can localise those to NMR,’ says Muellers.
