Wed 29 Jan 14:30: NMR Prediction Uncertainty Enables DFT-Free Structural Confirmation
While density functional theory (DFT) remains the standard for accurate simulation of nuclear magnetic resonance (NMR) spectra, its computational cost remains prohibitive. Use of DFT for structural confirmation is only justified where it offers substantial time savings over the experiment, such as total synthesis of natural products. Neural networks are a promising solution for simpler molecules, but published examples cannot estimate the prediction uncertainty.
By incorporating uncertainty estimation into an existing neural network, we can confirm the structure from its NMR spectrum 100,000 times faster than using DFT , with calculations completed in milliseconds rather than hours. Large-scale combinatorial studies show that our approach matches accuracy of DFT -based DP5 analysis and exceeds the sensitivity of simple error analysis. Analysis of 24 misassigned natural product structures demonstrates the generalisability of the method and equal performance to that of DFT .
We are now exploring the potential of the new method for automated structure revision and interpretation of 1H NMR spectra.
- Speaker: Ruslan Kotlyarov, University of Cambridge
- Wednesday 29 January 2025, 14:30-15:00
- Venue: Unilever Lecture Theatre, Yusuf Hamied Department of Chemistry.
- Series: Theory - Chemistry Research Interest Group; organiser: Lisa Masters.
Wed 29 Jan 15:00: Halogenation Site-Selectivity Prediction Just Got Faster
Predicting aromatic substitution sites for new molecules remain a challenge with large industry demand as its products have a myriad of applications. Classical methods involve rule-based approaches to ab initio methods that scale in computational time for more complex scenarios of heteroaromatic and multi-substituted systems. Previous works have explored ab initio, as well as hybrid methods with bespoke descriptors for each reaction site (86% accuracy, average 2,899 ms/inference). Here, we explore a data-driven model for halogenation site-selectivity achieving 80% accuracy with average 43 ms/inference. Our architecture combines machine learning with molecular fingerprints and algorithmic manipulation of chemical scaffolds. We also present an exploration of how different datasets – chlorination, bromination, and iodination – can be combined into a superset to increase prediction power of the final model. Finally, model performance is higher when compared to chemist, as they have through knowledge of scaffolds they have previously worked with. This model compared to chemists. Although the sample size is small, those working on the chemical industry have deep knowledge on certain molecular scaffolds while fast and accurate models can extend their reach to new areas.
- Speaker: Henrique Magri Marçon, University of Cambridge
- Wednesday 29 January 2025, 15:00-15:30
- Venue: Unilever Lecture Theatre, Yusuf Hamied Department of Chemistry.
- Series: Theory - Chemistry Research Interest Group; organiser: Lisa Masters.