“…Remarkable progress in the eld of arti cial intelligence and increasing availability of high-quality reference data have resulted in a rapid development of protein-ligand interaction scoring functions (Huang, Grinter, & Zou, 2010;Nguyen, Zhou, & Minh, 2018;Yadava, 2018) using machine learning algorithms (Chen, Engkvist, Wang, Olivecrona, & Blaschke, 2018) such as vector support machines or neural networks. Neural networks designed for the prediction of binding energies between receptors and ligands are typically based on the pattern recognition and computer vision ideas and have deep architecture utilizing 2D-or 3Dconvolution (Gomes, Ramsundar, Feinberg, & Pande, 2017;Gonczarek et al, 2018;Ragoza, Hochuli, Idrobo, Sunseri, & Koes, 2017;Stepniewska-Dziubinska, Zielenkiewicz, & Siedlecki, 2018;Sunseri, King, Francoeur, & Koes, 2019) or graph-convolution (Feinberg et al, 2018;Lim, Ryu, Park, Choe, & Ham, 2019;Torng & Altman, 2018) approaches.…”