Machine Learning Predictions of Experimental Mössbauer Properties of Fe-Containing Molecular Complexes and Protein Models

Abstract

Mössbauer spectroscopy is among the best techniques for probing molecular structures of iron complexes and proteins. Extracting structural insights from experimental spectra often necessitates accurate DFT predictions of isomer shift (IS) and quadrupole splitting (QS) parameters. However, substantial computational cost and sensitivity to the choice of functional and basis set limit their utility for high-throughput structural screening. To address this gap, various traditional machine learning (ML) and deep learning (DL) models were investigated, using a curated experimental data set of 522 diverse compounds. After engineering models and descriptors, the DL model Chemprop has the highest coefficient of determination (R2) for the test data set (0.92 for IS; 0.73 for QS), with R2 = 0.98/0.91 for IS/QS in the overall data set. Validation of this model corroborates its high, close-to-DFT accuracy for IS predictions, and also shows a significant improvement room for QS predictions. Its predictions of several iron sites in both heme and nonheme proteins and both first- and second-coordination sphere models with challenging and/or unique novel structural features replicated previous DFT structural predictions that are in excellent agreement with experiment, which is a significant application given that no protein data was used to train the ML model. This approach achieved a 218-fold reduction in computational time compared to DFT calculations with same geometries. Results from this first ML work on Mössbauer data of molecular iron complexes will facilitate rapid structural screening previously impractical for large sets of iron complexes and iron-protein binding site structures.