Fig. 4: Comparison of interaction energy predictions using high level and approximate methods.
From: Extending quantum-mechanical benchmark accuracy to biological ligand-pocket interactions
Specifically, Density Functional Theory (DFT), semiempirical, and empirical methods are compared to Local Natural Orbitals - Coupled Cluster with Singles, Doubles, and perturbative triplets (LNO-CCSD(T)59,60,61,62) and to each other. a Distributions of interaction energy Eint predictions w.r.t. LNO-CCSD(T), ΔEint, showed via box plots, for a selection of computational methods - DFT methods: PBE017+Many-Body Dispersion (MBD)18,19, ωB97X-V70, PBE017+D423,24, PBE87+MBD18,19, ωB97X21+D389,90, PBE017+ eXchange-hole Dipole Moment (XDM)95, PBE-QIDH88+D389,90, PBE017+MBD-NL20, B3LYP91,92+D389,90, PBE017+TS25; semiempirical methods: DFTB315+MBD18,19, GFN2-xTB16; and classical force fields: AMBER-GAFF211 and CHARMM-CGenFF12. The negative ΔEint values signify underbinding, while the positive ones overbinding. Each boxplot shows the median of the error distribution with a vertical median, a box covering the interquartile range from the 25th to the 75th percentile is shown in color, with whiskers spanning horizontally 1.5 times the interquartile range and data points outside that range plotted individually as outliers. b A heatmap of mean absolute error (MAE) values of predicted Eint w.r.t. LNO-CCSD(T) for the 42 QUID equilibrium dimers in the first column, and the MAE of all methods w.r.t. each other in subsequent columns. The computational methods to predict Eint were the same methods as in (a). *For the LNO-CCSD(T) method, the value shown with asterisk is the mean absolute of the uncertainty estimates for Eint.
