Evaluation of Predicted Protein–Protein Complexes by Binding Free Energy Simulations

Abstract

The accurate prediction of protein−protein complex geometries is of major importance to ultimately model the complete interactome of interacting proteins in a cell. A major bottleneck is the realistic free energy evaluation of predicted docked structures. Typically, simple scoring functions applied to single-complex structures are employed that neglect conformational entropy and often solvent effects completely. The binding free energy of a predicted protein− protein complex can, however, be calculated using umbrella sampling (US) along a predefined dissociation/association coordinate of a complex. We employed atomistic USmolecular dynamics simulations including appropriate conformational and axial restraints and an implicit generalized Born solvent model to calculate binding free energies of a large set of docked decoys for 20 different complexes. Free energies associated with the restraints were calculated separately. In principle, the approach includes all energetic and entropic contributions to the binding process. The evaluation of docked complexes based on binding free energy calculation was in better agreement with experiment compared to a simple scoring based on energy minimization or MD refinement using exactly the same force field description. Even calculated absolute binding free energies of structures close to the native binding geometry showed a reasonable correlation to experiment. However, still for a number of complexes docked decoys of lower free energy than near-native geometries were found indicating inaccuracies in the force field or the implicit solvent model. Although time consuming the approach may open up a new route for realistic ranking of predicted geometries based on calculated free energy of binding.

Computational prediction of protein–protein binding affinities

“…Due to the instantaneous solvent response (at every time point in equilibrium with the solute structure) and the possibility to use a low viscosity, it also allows for faster convergence than explicit solvent simulations (in each U.S. window). In a recent study, it has been demonstrated that it is possible to directly employ such binding free energy simulations to score a reasonable set of 50 decoy complexes for 20 test complexes within about a day on a GPU cluster . The calculated binding free energies for the near‐native complex geometries were in reasonable agreement with experiment.…”

“…The calculated binding free energies for the near‐native complex geometries were in reasonable agreement with experiment. Also, improved ranking of near‐native docking solutions compared to simple single complex structure evaluation after energy minimization or short MD simulations was observed . The study demonstrated the feasibility for systematic evaluation of predicted complexes based on calculated binding free energies instead of single point interaction, knowledge‐based scores, or mean interaction energies.…”

Prediction of protein–protein complexes using replica exchange with repulsive scaling

“…In addition to simulations starting far away from the native binding geometry, H‐REMD and regular MD simulations were also performed for arrangements in the vicinity of the native complex structure obtained by an initial protein–protein docking run using the program ATTRACT . The same set of structures and docking procedure as used in a previous study were employed (see Supporting Information Table S2). Since the H‐REMD method for refinement of docked complexes is computationally demanding the number of test complexes was limited to 20 complexes from the docking benchmark 3.0 .…”

“…Intramolecular pairwise harmonic distance restraints between the C α atoms of each individual protein were applied (force constant 0.5 kcal mol −1 Å −2 ) together with a COM distance restraint of interfacial C α atoms between the ligand and receptor (atoms with distances between 10 and 15 Å were considered) with a half parabolic shape (force constant 1.5 kcal mol −1 Å −2 ) that prevents full dissociation in the high replicas and shrinks the possible sampling space for these short simulations. The same simulation conditions and restraints were applied for regular MD simulations of each pose (no replica exchange and bias involved) of the same simulation time (4 ns) following a standard refinement protocol developed previously . For evaluating the interaction energy a short MD simulation (30 ps) was applied on the reference replica of the REMD simulations followed by a minimization (500 steps of steepest descent, 2000 steps of conjugate gradient), which was also applied to evaluate the final structures from regular MD simulations.…”

“…For evaluating the interaction energy a short MD simulation (30 ps) was applied on the reference replica of the REMD simulations followed by a minimization (500 steps of steepest descent, 2000 steps of conjugate gradient), which was also applied to evaluate the final structures from regular MD simulations. Finally, the minimized structures were scored by subtracting the potential energy of the partners from the energy of the complex . To access the deviation of the refined structures from the native binding site the Rmsd ligand was calculated, the root mean square deviation of the ligand to the native ligand after superpositioning the receptor on the native receptor (only heavy atoms were considered).…”