Non-covalent interactions in SQM methods: from PM6-DH to machine learning

Jan Rezac


More than a decade ago, we introduced the first corrections that allowed reasonably accurate descriptions of non-covalent interactions with semiempirical quantum mechanical (SQM) methods[1]. The later version of this approach, PM6-D3H4X[2,3], became the mainstay of applications of SQM methods to large supramolecular systems, including biomolecules. Recently, these corrections have been reparametrized using a new, larger set of benchmark data.

To illustrate the capabilities of current SQM methods, I will present examples from our application MOPAC in computer-aided drug design. Our SQM-based protocol for scoring protein-ligand interactions hits the sweet spot of high accuracy and fast computation time[4].

I will also share some results from the current work in progress – using machine learning to develop more general corrections that solve problems that could not be fixed in the SQM methods themselves or by simple standalone corrections.


(1) Řezáč, J.; Fanfrlík, J.; Salahub, D.; Hobza, P. J. Chem. Theory Comput. 2009, 5 (7), 1749–1760.

(2) Řezáč, J.; Hobza, P. J. Chem. Theory Comput. 2012, 8 (1), 141–151.

(3) Řezáč, J.; Hobza, P. Chemical Physics Letters 2011, 506 (4), 286–289.

(4) Pecina, A.; Eyrilmez, S. M.; Köprülüoğlu, C.; Miriyala, V. M.; Lepšík, M.; Fanfrlík, J.; Řezáč, J.; Hobza, P. ChemPlusChem 2020, 85 (11), 2362–2371.