Fig. 4: Evaluation of torsional energy descriptor (TED) in torsion energy predictions.
From: Assessing conformation validity and rationality of deep learning-generated 3D molecules
a Histogram distribution of torsion fragments in the DFT-5K dataset across Pearson correlation coefficients between density functional theory (DFT) level energy values and predictions from our TED-Model and GFN2-xTB13. For each torsion fragment in the DFT-5K dataset, torsion energies were computed for 24 conformations derived by enumerating torsion angles with increments of 15°. Pearson correlation coefficients between both methods and DFT were calculated per torsion fragment. b Case illustration of our model’s superior performance to GFN2-xTB’s in the scenario of sigma-hole interactions. There is a sigma-hole associated interaction between the sulfur of the thiazole group and the oxygen of the carbonyl group when the dihedral angle of S-C-C-N is ±180°. c Case illustration of our model’s superior performance to GFN2-xTB’s in the scenario of lone-pair repulsion. There is a lone-pair repulsion between the nitrogen of the thienopyridine group and the oxygen of the carboxy group when the dihedral angle of C-O-C-C is ±180°. For panels (b) and (c), the atoms defining the dihedral angle are highlighted in red.
