Conclusions
The current work extends previous efforts to consider the accuracy of modern computational chemistry methods to rank the energies of drug-like conformers. Since such energy differences are small, this poses a challenging benchmark even for density functional methods. Use of dispersion-corrections for density functionals are required – the slim time required is offset with dramatically increased accuracy. While triple-zeta and larger basis sets also provide higher accuracy, likely because of better treatment of non-covalent interactions, the large number of possible conformers forces trade-offs in accuracy and computational time required.
Current ML methods show great promise, particularly the ANI-1ccx method trained in part on coupled-cluster energies,\cite{S_Smith_2019} since they provide accuracy comparable to the semiempirical GFN2 method and can be performed in batch and accelerated on GPUs. Despite claims of reaching and exceeding DFT accuracy, we do not find these methods yet meet the accuracy of modern dispersion-corrected methods. Nevertheless, we expect these methods will provide increased accuracy in the future. An important caveat is the need to train on accurate data, such as dispersion-corrected density functional, MP2, or coupled-cluster calculations.
We expect continued improvement from other methods, particularly multiple efforts to improve classical force fields,\cite{Wahl_2019,van_der_Spoel_2018,Roos_2019,Harder_2015} inclusion of polarizable atomic charges,\cite{Lin2019,Inakollu_2020,Warshel_2007,Jing_2019,Zhang_2018,Rackers_2017,Ponder_2010,Liu_2019} novel force fields from experimental data, density functional and other quantum methods,\cite{parsley,Beauchamp_2015,Zanette_2018,Waldher_2010,Zahariev_2017,Grimme_2014} and continued development of approximate semiempirical quantum methods.\cite{Bannwarth_2018}
At present, we can highly recommend methods at each tier of the accuracy-time tradeoff, particularly the recent GFN2 semiempirical method, the B97-3c density functional approximation, and RI-MP2 for accurate conformer energies. Previous efforts to use a hierarchy of methods are still useful, for example, the use of GFN2 methods to refine initial conformer ensembles, followed by refinement of a smaller set of low-energy geometries with more accurate methods. Batch evaluation with ANI methods are also efficient, although they do not yet span the range of elements supported by semiempirical methods such as GFN2 or density functional methods.
The current benchmark reflects conformational preferences in a vacuum as judged by enthalpy differences alone. Since free energy differences drive experimental conformers, introducing entropic considerations will be needed for further work.\cite{Johansson_2008} Moreover, much chemistry is performed in solution, thus work on understanding conformer energies in solvation is also critical.\cite{Basdogan_2019,Basdogan_2018}
Acknowledgments
GRH and DLF acknowledge the National Science Foundation (CHE-1800435) for support and the University of Pittsburgh Center for Research Computing through the computational resources provided. The authors thank Olexandr Isayev and Justin Smith for access to the ANI-2x model.
Supporting Information