NMR chemical shift predictions from first principles significantly enhance experimental interpretation and molecular structure identification. The gold-standard method, CCSD(T) with a complete basis set (CBS), provides unparalleled accuracy but at a computational cost that is often prohibitive. On the other hand, while Machine Learning (ML) offers computational efficiency, it frequently struggles with generalizing beyond its training dataset. To bridge this gap, our research introduces innovative ML approaches for NMR chemical shift predictions of H, C, N, and O. Our pivotal innovations encompass:

1. The introduction of tensor environment vectors (TEVs) as a feature representation. Derived from atomic chemical shielding tensors within a molecular context and obtained through cost-effective density functional theory, this rotationally-invariant feature serves as input to approximate the precision of high-level composite theories close to CCSD(T)/CBS.

2. The deployment of a progressive active learning methodology, curtailing the need for resource-intensive calculations while iteratively enhancing the model's accuracy on previously unseen data.

3. The integration of an error estimator, based on the standard deviation of the model ensemble, alerting users to potentially dubious predictions.

Our developed model, termed iShiftML, demonstrates great predictive capacity across varied benchmarks. These include unseen molecules in the NS372 dataset, gas-phase experimental chemical shifts for small organic molecules, and significantly larger and more complex natural products. Impressively, iShiftML accurately distinguishes between subtle diastereomers based on chemical shift assignments, highlighting its invaluable application in refining experimental NMR interpretations and molecular structure identification.

Jiashu Liang

Graduate Researcher, University of California, Berkeley

Appearances