Fig 1.
A schematic indicating the hypothetical mechanism by which the Transform algorithm exhibits improved performance compared to the TransRot algorithm.
A) When the TransRot algorithm is used, a Cartesian starting coordinate is specified as the starting position for the ligand. B) This starting point is then translated to a random location which does not overlap with the protein backbone. C) The ligand is centered at the new random location within a user specified starting radius, and a set of diverse, minimally- clashing rotational binding positions are selected. D) A single random binding pose is selected for refinement. E) When the Transform algorithm is used, the starting Cartesian coordinate is specified as the starting position for the ligand. F) The simultaneous translations and rotations within a user specified radius is sampled using a MCM algorithm. G) The best scoring model is selected from step (F) for refinement.
Fig 2.
A schematic of the Transform (left) and TransRot (right) docking protocols described in this paper.
Because the initial placement and refinement steps are independent, the two initial placement algorithms can be alternatively selected to produce a total of four ligand docking algorithms.
Fig 3.
Kernel Density Estimate curves showing the time necessary to generate a single model using the four RosettaLigand protocols.
TransRot/MCM is the protocol previously published by Davis et al. [15].
Fig 4.
The fraction of protein systems in which the lowest scoring model has an RMSD of less than 2.0 Å to the native structure as a function of CPU time using the four RosettaLigand docking algorithms and three starting protein models.
A) Experimental structures, B) models in which only the sidechains are repacked, and C) models in which all atoms have been minimized using the Rosetta energy function. A large pool of models were generated, and random subsamples were taken corresponding to time points at 5 minute intervals. The number of structures included in each time point was based on the average time to generate a model for each algorithm. 20 random samples were taken for each time point, and the means are plotted, with the error bars representing the standard deviation. Docking protocols which make use of the Transform algorithm are reliably converged after approximately 15 minutes (dotted line).
Fig 5.
The fraction of protein systems in which the lowest scoring model has a RMSD of less than 2.0 Å to the native structure as function of the total number of structures generated using the four RosettaLigand docking algorithms and three starting protein models.
A) Experimental structures, B) models in which only the sidechains are repacked, and C) models in which all atoms have been minimized using the Rosetta energy function. A large pool of models was generated, 20 random subsamples were taken for each point, and the means and standard deviation are plotted. Docking protocols which make use of the Transform algorithm are reliably converged after approximately 150 models (dotted line).
Fig 6.
Plots comparing the performance of various docking protocols when docking ligands into relaxed structures.
A) For RMSD versus RMSD plots 20 samples of 150 models were collected, and the average of the RMSD of the lowest scoring model is plotted for each protein/ligand system. The standard deviation of these 20 samples is shown with error bars. Dotted lines indicate the 2.0 Å RMSD cutoff used to classify correct vs incorrect binding positions. B) For score versus score plots the change in average all-atom Rosetta score of the lowest scoring model generated by several pairs of docking algorithms.
Fig 7.
Comparison of specific successes and failures between the RosettaLigand protocols.
Native structures are in grey, lowest scoring models generated by the Transform/MCM protocol in blue, and lowest scoring models generated by TransRot/MCM in pink. A) A case in which the TransRot/MCM protocol was unsuccessful but the Transform/MCM protocol was successful (PDB ID: 1fhd). B) A case in which the Transform/MCM protocol was unsuccessful but the TransRot/MCM protocol was successful (PDB ID: 2otz). C) A case in which both methods were successful (PDB ID: 1bky). D) A case in which neither method was successful (PDB ID: 1q4w).
Fig 8.
Scatter plots showing the weak correlation between experimental-log(Kd) and predicted Rosetta energy score for models in the 43 protein benchmark.
Scores from models generated using the Transform/MCM protocol are in red while scores from models generated using the Transform/MCM protocol are in black.