Extended Data Fig. 6: The initial structural models used for benchmarking.

a, ARES score vs RMSD of the dynamic fitting trajectory for RNase P RNA catalytic domain (see Methods). ARES scores were calculated using geometric deep learning²⁸. Models are colored as a function of their total energy: dark red for lower energies and dark blue for higher energies. b, Ribbon diagrams of the trajectory model with the lowest ARES score, k158597 (orange), used as the initial structure for BM0, with an RMSD of 21.4 Å for 3DHS (black) crystal struture. c, Crystal structure of the RNase P RNA catalytic domain (PDB 3DHS)⁴⁰ with missing residues modeled, which was used as the GT structure. The models from FARFAR2 structure calculation using the secondary structure information (d) S257, (e) S1076 (BM2), and (f) S142 (BM1) were selected from the pool of 10,000 models based on, respectively, the best ARES score (S257), best combined ARES and FARFAR2 scores with RMSD > 20 Å from the GT structure(S1076), or best combined ARES and FARFAR2 scores and RMDs relative to the GT structure of > 10 Å and <15 Å (S142). The ARES scores and the RMSDs to the GT structure are indicated. The missing residues in the crystal structure were modeled and energy-minimized using PHENIX.