This Week In Cheminformatics: Issue #011

Model for Pure Shift Proton NMR, Uncertainty Quantification, What Happens in Successful Optimizations and a long list of papers.

Mar 09, 2026

Highlights

Deep Learning Assisted Proton Pure Shift NMR Spectroscopy

Kakita and Hansen detailed a practical deep learning approach in JACS for generating homonuclear decoupled pure shift 1H NMR spectra without the steep sensitivity penalties typical of broadband decoupling methods like PSYCHE. Their model, FID-Net-PS, processes a tailored input of five 1D spin-echo spectra acquired at different evolution times and transforms them into high-resolution singlets. What makes this implementation particularly relevant for analytical workflows is that it was trained entirely on synthetic data and explicitly outputs point-to-point predicted uncertainties alongside the transformed spectra, providing a reliable proxy for noise during quantitative mixture analysis. While the architecture still faces limitations with strongly coupled spin networks, its ability to recover high-sensitivity pure shift data from complex, overlapping multiplets, especially for low-concentration samples or systems with exchangeable protons, makes it an interesting read.

Uncertainty Quantification for In Silico Chemistry

We constantly deal with imperfect data and approximations in our modeling workflows, yet we often treat the outputs as absolute truth. Frömbgen and co-workers have published this highly relevant review in Chemical Reviews detailing Uncertainty Quantification (UQ) for in silico chemistry that provides a much-needed formalization of this problem. Instead of treating error analysis as a simple benchmark, they systematize UQ across quantum chemistry, molecular dynamics, and ML. The paper breaks down a formal taxonomy, separating representation, level-of-theory, and numerics uncertainties and rigorously details the mathematical tooling, from forward and inverse UQ via Bayesian inference to surrogate models like Gaussian processes. If you are building predictive models or constructing automated workflows where error propagation can drastically alter your conclusions, this paper offers the rigorous methods required to properly quantify and manage those uncertainties.

What Happens in Successful Optimizations? A Survey of 2018–2024 Literature

Leeson’s recent survey analyzed 478 start-to-candidate optimization pairs from the 2018–2024 literature. It provides a rigorous, data-driven baseline of contemporary multiparameter optimization without overstating the trends. The analysis shows that while average lipophilicity remains flat, molecular weight reliably climbs by an average of 81 Da to drive potency. This weight increase is systematically achieved by introducing higher sp3 carbon fractions, H-bond acceptors, and natural product fragments, alongside a measurable shift away from carboaromatic rings. Despite these structural shifts, optimization remains remarkably conservative; candidates retain an average of 70% of their starting structural framework, and 86.3% of the resulting oral candidates still cleanly comply with the Rule of 5. Good read !