Fig. 5: Energy landscape and structure of four representative drug molecules.

Predicted polymorphic energy landscapes for (a) cocaine, (b) MK-2022 (c) olanzapine and (d) GSK269984B using r²SCAN-D3 energy computed on PBE-D3 optimized crystal structures. The corresponding energy landscape dependence on density functional theory (DFT) functionals are shown in (e–h). For cocaine (a), the only known experimental structure is predicted to have much lower energy than all other predicted candidate structures with an energy gap of about 1.0 kcal/mol by PBE-D3 and PBE0-MBD functionals (e). For MK-2022 all functionals predict experimentally known form I to be most stable (f). For olanzapine, the three resolved experimental structures are predicted among the top five candidate structures with three different DFT functionals (g). For the unresolved form III, two predicted candidate structures have the spectra characteristics of form III from experiment, one corresponding to a previously identified structure (III*) in the Pbca space group (j), and one novel structure in the P2₁/c space group with an identical 2D layer (i). For GSK269984B, candidate structures with favorable intermolecular hydrogen bonds (k) and (l), including the typical bidentate hydrogen bonds for carboxylic acids, as highlighted in (k) and (l), are predicted to have lower energy than the experimental structure with only intramolecular hydrogen bonds (m), indicating the possibility of other polymorphs with competitive stability yet to be discovered by experiment. The grey frames in (k), (l), and (m) represent the unit cell of the crystals. Source data are provided as a Source Data file.