Assessment of Strong-Correlation Corrected Range-Separated Local Hybrid Functionals for Metalloenzyme Reactions

Accurately modeling the complex electronic structure of metalloenzyme active sites remains a significant challenge for density functional theory (DFT). The MME55 test set [J. Chem. Theory Comput. 2023, 19, 8365] is a recent benchmark providing high-level DLPNO–CCSD(T)/CBS reference values for large, practically relevant metalloenzyme models with up to 116 atoms. So far, the best-performing rung-4 hybrids were MPW1B95-D3(BJ) with a mean absolute error (MAE) of 2.67 kcal/mol and ωB97M-V (2.78 kcal/mol). This work focuses on range-separated local hybrids (RSLHs), a recent class of rung-4 functionals that use both a real-space dependent and interelectronic distance dependent admixture of exact (Hartree–Fock-like) exchange. Strong-correlation corrected RSLHs (scRSLHs) based on real-space nondynamical correlation models are also assessed. The best-performing functional is the scRSLH ωLH23tdB-D4 with an MAE of 2.46 kcal/mol. Overall, the results reflect a systematic improvement in accuracy along the functional development series from LH to RSLH to scRSLH. However, our analysis also indicates that the real-space nondynamical correlation model is not the primary driver of this improvement, which is instead dominated by other features of the functional design. Additionally, we studied (sc)(RS)LHs for the ENZYMES22 set of enzyme reactions comprising metal-free active site models as well as the ECR20 benchmark of net reaction energies of enzyme-catalyzed reactions. MAEs of 1.45 and 0.63 kcal/mol confirm ωLH23tdB-D4 as one of the top-performing functionals for these sets. Due to inaccuracies in the original CCSD(T)/aug-cc-pVDZ + [SCS-MP2/aug-cc-pVTZ - SCS-MP2/aug-cc-pVDZ] reference values for the ECR20 set, we devise new canonical CCSD(T)/CBS reference values using an aug-cc-pVXZ (X = D, T, Q)-based basis set extrapolation. Because scRSLHs have only very recently emerged, this benchmark study is accompanied by critical assessments of their use as a practical tool. We analyze the grid dependence of their seminumerical implementation in Turbomole and evaluate the computational cost and parallel scaling on a shared-memory OpenMP architecture using up to 192 physical CPU cores.