Data-driven ligand field exploration of Fe(iv)-oxo sites for C-H activation

Abstract

High-valent Fe(iv)-oxo intermediates, found in enzyme active sites, are excellent targets for biomimetic design of molecular catalysts for C-H bond activation. C-H bonds in inert aliphatic hydrocarbons, such as methane, possess strong bonds that are resistant to chemical functionalization. To aid in the screening of potential catalysts for C-H bond activation, computational methods, such as density functional theory (DFT) and machine learning (ML), are valuable tools for performing high-throughput virtual searches of the vast chemical compound space. In this study, we have designed a database of 50 Fe(iv)-oxo species with varying coordination environments which are further functionalized for a total of approximately 181k structures. DFT calculations are then performed on a subset of the molecular database to determine spin states and C-H bond activation energies. The collected data are then curated based on a series of chemically informed criteria. To avoid performing 181k DFT calculations on the total chemical compound space, we developed ML models that utilize a novel molecular representation based on persistence homology, called persistence images (PIs). In particular, we have developed a novel similarity search algorithm, followed by training a regression model to predict C-H activation energies and a classification model to predict the spin states. The priority is to provide high-fidelity predictions for C-H activation barriers. For this purpose, we divided the full database into low- and high-fidelity structures and introduced a metric (δΔG‡) which evaluates the effect of a specific ligand modification with respect to the parent, unsubstituted structure. A validation step that included additional DFT calculations on 15 structures demonstrated the credibility of the proposed methodology.