A Comparison of Quantum and Molecular Mechanical Methods to Estimate Strain Energy in Druglike Fragments

Abstract

Reducing internal strain energy in small molecules is critical for designing potent drugs. Quantum mechanical (QM) and molecular mechanical (MM) methods are often used to estimate these energies. In an effort to determine which methods offer an optimal balance in accuracy and performance, we have carried out torsion scan analyses on 62 fragments. We compared nine QM and four MM methods to reference energies calculated at a higher level of theory: CCSD(T)/CBS single point energies (coupled cluster with single, double, and perturbative triple excitations at the complete basis set limit) calculated on optimized geometries using MP2/6-311+G**. The results show that both the more recent MP2.X perturbation method as well as MP2/CBS perform quite well. In addition, combining a Hartree-Fock geometry optimization with a MP2/CBS single point energy calculation offers a fast and accurate compromise when dispersion is not a key energy component. Among MM methods, the OPLS3 force field accurately reproduces CCSD(T)/CBS torsion energies on more test cases than the MMFF94s or Amber12:EHT force fields, which struggle with aryl-amide and aryl-aryl torsions. Using experimental conformations from the Cambridge Structural Database, we highlight three example structures for which OPLS3 significantly overestimates the strain. The energies and conformations presented should enable scientists to estimate the expected error for the methods described and we hope will spur further research into QM and MM methods.

Approaching coupled cluster accuracy with a general-purpose neural network potential through transfer learning

“…The dihedral scans used to compute ANI-1ccx's error can be found inSupplementary Figure 6. Each ANI-1ccx restrained optimization (averaged over the 36 angles for each of the 45 torsions) took approximately 0.58s on a single NVIDIA V100 GPU.A similar timing comparison was reported in Sellers et al57 for the QM and MM methods. Comparing to this literature result, the ANI model on a single GPU is (on average) as fast as OPLS3 on a CPU and 6200 times faster than B3LYP-D3 on a CPU.…”

Outsmarting Quantum Chemistry Through Transfer Learning

“…al. 64 This benchmark provides a measure of accuracy for a model at reproducing potential energy profiles from a diverse set of molecular torsions. Figure 4 provides a comparison of results for three highly accurate but computationally expensive QM methods, four moderately computationally expensive QM methods, and two commonly used small molecule force fields.…”

“…The grey dots represent the MAD of a given torsion scan vs.gold standard CCSD(T)/CBS. The box extends from the upper to lower quartile, the black horizontal line in the box is the median, and the "whiskers" guide the eye to identify outliers as in Sellers et al64 .the ANI-1x DFT dataset plus active learning-based dihedral corrections, obtains a median MAD of 0.47 kcal/mol on the benchmark. The ANI-1x potential performs similarly to MP2/6-311+G** and to the ANI-1ccx-R potential.…”

Schrodinger-ANI: An Eight-Element Neural Network Interaction Potential with Greatly Expanded Coverage of Druglike Chemical Space

“…First we examined the Genentech rotamer test set of Sellers et al, 4 which consists of rotamers from 62 small druglike molecules. The HCNO molecules from the Genentech set were used by Smith et al in evaluating ANI-1x, but for Schrödinger-ANI we were able to evaluate the full set, since its composition is entirely within H, C, N, O, S, F, Cl, P.…”