Efficient Computational Strategies of the Cluster-in-Molecule Local Correlation Approach for Interaction Energies of Large Host-Guest Systems

We propose a heterogeneously accelerated reduced cluster-in-molecule (CIM) local correlation approach for calculating host-guest interaction energies. The essence of this method is to compute only the clusters that make significant contributions to the interaction energies while approximately neglecting those clusters with smaller contributions. Benchmark calculations at the CIM resolution-of-identity second-order Moller-Plesset perturbation (CIM-RI-MP2) or CIM spin-component-scaled RI-MP2 (CIM-SCS-RI-MP2) levels, involving three medium-sized protein-ligand structures, demonstrate that the reduced CIM method achieves over 48% time savings without compromising accuracy, as the interaction energy error remains within 0.5 kcal/mol compared to the full CIM method. To further enhance cluster computation efficiency, we developed a heterogeneous parallel version of the CIM-(SCS-)RI-MP2 method. It achieves over 93% internode parallel efficiency and over 98% multi-GPU card parallel efficiency for the tested large complexes. Ultimately, the hardware-accelerated reduced CIM-(SCS-)RI-MP2 method is applied to calculate the interaction energies of six protein-ligand systems, ranging from 913 to 1425 atoms. Remarkably, the method requires only 4.3-22.8% of the clusters to achieve accurate results, and under the condition of using only a single node, the wall time is within 2 days. Additionally, the reduced CIM domain-based local pair natural orbital coupled cluster with singles, doubles, and perturbative triples [CIM-DLPNO-CCSD(T)] method is successfully applied to the calculation of a 1425-atom protein-ligand system. These computations demonstrate the capability of a specific electronic structure to accurately calculate interaction energies for large host-guest systems.