Fig 1.
Typical stages of molecular docking applied for interaction predictions for RNA-ligand complexes and methods tested within this publication.
The first stage is ligand conformer generation (either using the native conformation extracted from experimentally solved structure or generation of 3D structure), which is followed by molecular docking, and scoring of generated poses.
Fig 2.
Comparison of the performance of nine scoring functions, expressed as the RMSD between the best-scored pose and the reference pose (left panel), best among the top three scored poses (S(3), middle panel), and best among the top five scored poses (S(5), right panel).
Additional rows represent the median and minimal values of RMSD obtained during docking. Each dot represents one complex from the testing set. Docking was performed using rDock with dock desolvation potential with the native conformation of a ligand as an input.
Fig 3.
Comparison of scores of the experimentally determined ligand structures (black dots) with the distribution of scores of docked poses (gray violin plots) for the testing set.
Additional row [RMSD] represents the distribution of the RMSD values for docked poses and experimentally determined structures. Docking scores and RMSD values were normalized independently for each complex to values [0, 1]. The quartiles of the distribution are marked by the inner vertical dashed lines. Docking was performed using rDock with dock desolvation potential with the native conformation of a ligand as an input.
Table 1.
Percentages of the experimentally determined poses determined as the top scoring pose, in the top three, five, 25%, and 50% of scoring poses, among poses resulting from molecular docking.
Docking was performed using rDock with the dock desolvation potential with the native conformation of a ligand as an input.
Fig 4.
Comparison of the performance of three scoring functions (rDock dock_solv, LigandRNA, and AnnapuRNA) expressed as RMSD to the reference pose of best in top three scored poses (S(3)), calculated for four docking programs.
The fourth column represents the best (lowest) RMSD obtained during docking for each program. Each dot represents one complex from the testing set. Docking was performed with the native conformation of a ligand as an input.
Table 2.
Average and median values of the lowest RMSD reported for rDock (dock and dock_solv potentials), AutoDock Vina, and iDock.
Values calculated for all structures in the testing set (A), only for structures for which all programs completed docking (B), and only for structures for which RMSD values of poses found by all programs are below or equal 10 Å (C). Docking was performed using a native conformation of a ligand as an input.
Fig 5.
Comparison of the performance of three scoring functions, expressed as a RMSD of the best-scored pose to the reference pose of best in top three scored poses (S(3)), calculated for three conformer generation methods.
The fourth column represents the best (lowest) RMSD obtained during docking for each program. Each dot represents one complex from the testing set. Docking was performed using rDock with dock desolvation potential.
Table 3.
P-values obtained in a Wilcoxon signed-rank test comparing the distribution of S(1), S(3), and S(5) values of AnnapuRNA and other docking programs.
Docking was performed using native conformation of a ligand as an input. Two-tailed p-values ≤ 0.05 are in bold. Data referenced in the main text are underlined.
Fig 6.
Selected docking solutions for structures from the testing set.
Best pose found by rDock (with dock desolvation potential, left) and AnnapuRNA (DL 2013, right), together with scatter plots of RMSD and normalized score. Scatter plots show a regression line with a 95% confidence, and the top scoring pose according to each method is indicated by a red arrow. (A) 3SUX (Crystal structure of THF riboswitch, bound with THF fragment), (B) 1FYP (Decoding region A-site in complex with Paromomycin) and (C) 1AJU (HIV-2 TAR-argininamide complex). RNA molecules are presented as a gray cartoon, ligands as sticks; reference structure—green, solution found by rDock—red, solution found by AnnapuRNA—cyan. Heteroatoms are colored: O—red, N—blue.
Fig 7.
Prediction of the complex of FMN Riboswitch with 4-{benzyl[2-(7,8-dimethyl-2,4-dioxo-3,4-dihydrobenzo[g]pteridin-10(2H)-yl)ethyl]amino}butanoic acid.
(A). Redocking experiment—ligand extracted from the crystal structure 6DN2 was redocked to this structure; (B). Ligand extracted from the crystal structure 6DN2 docked to FMN structure solved with a different ligand (2YIE); (C). Ligand extracted from the crystal structure 6DN2 docked to the low-resolution APO FMN structure (2YIF). All RNA structures are superimposed. The reference ligands structures are shown in green. Structures predicted by rDock (dock_solv) are shown in magenta (left), predicted by AnnapuRNA (DL 2013) in light blue (right).
Fig 8.
The left column represents the method development pipeline, while the right one displays the method usage. Optional post-processing steps are placed in boxes with a dotted border.
Table 4.
Summary of the 2013 and 2016 datasets.
Fig 9.
Atoms and pseudoatoms used in the coarse-grained representation of RNA (upper left pane) and ligand molecules (upper right pane), an example of a ribonucleotide (guanosine monophosphate) in SimRNA representation (bottom left) and a small molecule—S-adenosylmethionine (bottom right) in the pharmacophore representation.
Fig 10.
Descriptors collected for RNA-ligands complexes.
The distance d and the angle α between the two selected pseudoatoms of the RNA and the pharmacophore (pane A), the angle β between the two selected pseudoatoms of the RNA and the pharmacophore vector (pane B), the angle γ between the nucleotide base plane and the pharmacophore (pane C), and the angle δ between the base plane and the pharmacophore vector (pane D).
Fig 11.
AnnapuRNA scoring function calculates the probabilities p1…pn of interactions between ligand pharmacophores (white circles) with RNA pseudoatoms (black circles) within the distance of 10 Å.
The total score is calculated as a negative sum of all probability values p calculated for the given ligand.
Fig 12.
Docking volume definitions for docking programs (left) and input conformer sources used for docking (right).
For AutoDock Vina and iDock a box of sizes 20 Å ✕ 20 Å ✕ 20 Å was set, with a center in the center of mass of the ligand (pane A), while for the rDock it is defined as a volume around the known ligand at the maximum distance 10 Å from the ligand atoms (pane B). Three sources of ligand conformers were tested in docking: native pose extracted from the experimentally determined structure (pane C) and conformations generated de novo by two programs: OpenBabel and Balloon (pane D).