BDOCK : An Implementation of the FFT Protein-Protein Docking Method Using the BALL Library Bingding Huang Center of Bioinformatics Saarland University
Bingding Huang 2 Overview Introduction of BALL(Biochemical Algorithms Library) Protein-Protein docking FFT (Fast Fourier Transform) method Our implementation Result Conclusion
Bingding Huang 3 BALL(Biochemical Algorithms Library) by Dr. Oliver Kohlbacher and Prof. Hans-Peter Lenhof
Bingding Huang 4 BALL What is BALL BALL is a C++ application framework for rapid software prototyping in the area of Molecular Modeling and computational bioiolgy What can BALL do It provides an extensive set of data structures as well as classes for molecular mechanics,advanced solvation methods,comparison and analysis of protein structure,file import/export and visualization of molecule,etc
Bingding Huang 5 Molecular Docking The molecular docking problem Given two molecules with 3D conformation at atomics level Do the molecules bind to each other?if yes How strong is the binding affinity How does the molecule-complex look like? Docking problem in biochemistry Protein-Ligand docking Protein-Protein docking Protein-DNA docking DNA-Ligand docking
Bingding Huang 6 Overview of a typical docking procedure Coordinates of two molecules to be docked Perform a rigid- body search for favorable complexes Generate a number of possible docked complexes Re-rank complexes based on a scoring function Introduce flexibility to refine and re- rank complexes List a few complexes for experimental design and test Generator Scoring function
Bingding Huang 7 Protein-Protein Docking Problem features Stable conformation(rigid) Large contact surface Good geometric shape complementarity Applications Understanding Protein-Protein interaction Prediction of Protein-Protein interaction Predicting protein complex structures classification Unbound docking Bound docking
Bingding Huang 8 FFT method for protein-protein docking First proposed by Katchalski-Katzir (1992) Further developed by Sternberg and Gabb (1997) Features Proteins are projected into 3D grids to measure geometric shape complementarity Assign interior and surface grid cell values Use Fast Fourier Transform to decrease the computational time
Bingding Huang 9 FFT method Surface:+1 Interior: -15 Interior: +1 Blank: 0 Protein A Protein B
Bingding Huang 10 FFT method Good shape complementarity complex
Bingding Huang 11 FFT method Protein A a l,m,n = 1, surface cell p, interior cell 0, elsewhere Protein B b l,m,n = q, interior cell 0, elsewhere Here we use p = -15 Here we use q = +1 Correlation: Find the grid step that maximise the correlation Overcost O(N 6 ) – and have to rotate protein B and repeat..
Bingding Huang 12 FFT method DFT A p,q,r = Forward FFT A = DFT (a) Forward FFT B= DFT (b) Computer C=A*B Inversed FFT c=IFFT(C) Totally, FFT can reduce O(N 6 ) to O(N 3 lnN 3 )
Bingding Huang 13 The strategy for FFT Protein-Protein Docking Protein A Protein B static grid mobile grid discretise Stack FFT Inverse transform Multiply loop rotate protein B discretise Score complexes finish loop Filter local refinement Predicted complexes
Bingding Huang 14 Rotational conformations An uniformly distributed Euler angle set is used to ensure minimal orientations are required to cover the whole rotational space A schematic diagram of rotational search 15 o o o o8o o6o 68760
Bingding Huang 15 Our implementation of FFT using BALL Realize all the functionalities into a class:geometricfit initGridSize(Atomcontainer &pro_a,&pro_b) makeGrid(Atomcontainer &pro) FindInsidePoints() FindSurfacePoints() RotateProtein(Atomcontainer &pro_b) doFFT() FFTGridMulti() doIFFT() getPeakValue() ……….
Bingding Huang 16 The main function 2int main(int agrc,char ** agrv){ 2PDBFile pdb_a(“recetor.pdb”) ; 3PDBFile pdb_b(“ligand.pdb”) ; 4System pro_a; 5System pro_b; 6pdb_a>>pro_a; 7pdb_b>>pro_b; 8GeometricFit geofit(pro_a,pro_b); 9geofit.initGridSize(pro_a,pro_b); 1geofit.makeFFTGrid(pro_a); 1geofit.doFFT(pro_a); 1RotationAngles rotAngle; // the main docking program loop 1for ( int i=0;i< rotAngle.getNumber();++i ){ 1float phi = rotAngle.getXAng(i); 1float theta = rotAngle.getYAng(i); 1float psi = rotAngle.getZAng(i); 1System sys_b = pro_b; 1geofit. RotateProtein(sys_b,phi,theta,psi); 2geofit.makeFFTGrid(sys_b); 2geofit.doFFT(sys_b); 2geofit.FFTMutil(); 2geofit.doIFFT(); 2float peak_value = geofit.getGlobalPeak(); 2Vector3 trans = geofit.getTranslation(); 2} // finish docking loop 2}
Bingding Huang 17 Distribute the rotational conformations Each rotational conformation is independent,so we can distribute the total rotational conformations to a number of different processors to perform docking together using MPI (Message Passing Interface)
Bingding Huang 18 A good scoring function should be able to eliminate the false positives to screen the docked complexes Initial stage of docking Geometric shape complementarity alone – very fast to compute Re-ranking stage Empirical residue-residue pair potentials Binding free energy: Scoring function
Bingding Huang 19 Evaluate the docked Complex Ideally, the prediction complexes having higher score should be near-native complex Evaluation RMSD (Root mean square Deviation) of all C atoms between prediction complex and native complex RMSD below 3 Angstrom is acceptable
Bingding Huang 20 We apply our implementation to an unbound/unbound protein-protein data set Enzyme / Inhibitor Antibody / Antigen Parameters 1 Angstrom grid spacing 2 Angstrom surface thickness 15 degree interval We obtain The number of hits (RMSD below 3.0 Angstrom )at top 2000 The rank of best hit The best RMSD value Result
Bingding Huang 21 Result Complex IDHitsRankBest rmsdComplex IDHitsRankBest rmsd 1ACB FSS ATN MAH AVW PPE AY PPF BRC TGS BRS UDI CGI UGH CHO KAI CLV PCC CSE PTC Table 1. The docking results at 15 degree based on shape complementarity
Bingding Huang 22 Re-ranking Table 2. The re-ranking results using residue-residue pair potential and binding free energy scoring functions Complex IDshaperpscoreEnergyComplex IDShaperpscoreEnergy 1ACB23131FSS ATN271591MAH AVW5211PPE112 1AY711661PPF723 1BRC TGS1347 1BRS3121UDI212 1CGI UGH634 1CHO KAI CLV PCC411 1CSE PTC
Bingding Huang 23 1PPE. Shape complementarity vs. RMSD Fig1.1PPE shape complementarity vs. RMSD (unit Angstrom)
Bingding Huang 24 1PPE. Pair potential vs. RMSD Fig 2.1PPE. residue-residue pair potential vs. RMSD (unit Angstrom)
Bingding Huang 25 1PPE. binding free energy vs. RMSD Fig 3.1PPE. binding free energy (unit KJ/mol) vs. RMSD (unit Angstrom)
Bingding Huang 26 Prediction Complex structures Fig 4. 1PPE. RMSD 0.42 Fig 5. 1UGH. RMSD 1.67
Bingding Huang 27 Only half an hour to two hours is needed at 15 o on a single Xeon 2.8 G processor. Using 8 processors (connected by giganet) the general running time is about 5 to 15 minutes When docking at 6 o, the running time is 8 hours to two days. Using 8 processors, it is one hour to 6 hours Running time
Bingding Huang 28 We implemented the FFT protein-protein docking method using the BALL library Our program can predict a number of near-native complex structures based on shape complementarity alone Pair potential and binding free energy can improve the ranking Our program is more faster than FTDOCK The rapid prototyping capabilities of BALL saves us a lot of time in implementation source codes and we only need to focus on the new technologies, algorithms and methods BDOCK is freely available to academic users Conclusion
Bingding Huang 29 Prof. Volkhard Helms and Prof. Hans-Peter Lenhof Dr. Julie Mitchell Mr. Andreas Hildebrandt Mr. Hongbo Zhu Aknowledgement
Bingding Huang 30 Thank you! Questions?