The ability to predict electronic absorption spectra reliably is critical for the rational design of photoactive molecules. However, common ab initio methods possess the accuracy but are too expensive for high-throughput applications, while data-driven models lack interpretability.
We present ChromoPredict, a Python based, digitizable and generalizable version of empirical Woodward-Fieser and Fieser-Kuhn rules to predict values of λmax from SMILES strings. Starting with a curated TD-DFT dataset of α,β-unsaturated carbonyl derivatives, we statistically validated classical increments to the rules and achieved mean absolute errors within ~8 nm, comparable to and more accurate than other machine learning and TD-DFT trained predictions. Furthermore, we statistically derived a new increment set for 3-, 4- and 6-substituted coumarins, systems for which ChromoPredict successfully reproduced the experimental trend for substitution, and the systematic error observed by TD-DFT was corrected.
This link from empirical, historical heuristics in chemistry to contemporary cheminformatics advances a rapid, interpretable and inexpensive method of prediction of spectra and establishes an ability to assess and design more efficiently.
Connor Forster