Two-model ML pipeline predicts solution NMR shifts far faster

bravo_abad · x · 2026-07-20

This paper chains two ML models to predict solution-phase NMR chemical shifts more accurately and much faster than traditional approaches.

What the pipeline does

Uses MACE and the newer UMA foundation model as an ML potential to run realistic molecular dynamics at low cost.
Feeds the resulting explicit solute-solvent snapshots into ShiftML3, a model trained on crystal shieldings.
Averages predictions over hundreds of thousands of frames to obtain quantitative chemical shifts.

Why it matters

Standard DFT with implicit solvation struggles with exchangeable protons such as O-H and N-H because it washes out solvent and hydrogen-bond effects. The ML pipeline preserves those interactions and therefore handles difficult cases much better.

Results reported

Tested on water in seven solvents, alcohols, hydrogen-bonded nucleobase pairs, glucose anomers, and alkylated acetamides.
Correlation with experiment improves from R² = 0.68 for implicit-solvent DFT to R² = 0.96 for the ML pipeline.
The method is said to be 3–4 orders of magnitude faster than the DFT-MD baseline.

Takeaway

The key idea is reusing models outside their original domain and gluing them into a physically faithful, low-cost workflow. The author argues this could turn NMR from a confirmation step into a screening tool for drug discovery and materials characterization.

Original post →

Two-model ML pipeline predicts solution NMR shifts far faster

More from Apps