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 Mo̷ller–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.