1. A new protein-protein docking scoring function based on interface residue properties Reporter: Yu Lun Kuo (D95922037) E-mail: sscc6991@gmail.comsscc6991@gmail.com Date: November 10, 2008 Bioinformatics Vol. 23 no. 5 2007(Pages 555-562)

2. Introduction A protein-protein docking procedure traditionally consist of two successive tasks –A search algorithm generates a large number of candidate solutions Over the rotational/translational degrees of freedom Two partners contact each other in many different orientations –A scoring function is used to rank them The best solutions are selected by evaluating a score 2

3. Introduction Scoring function –Express the geometric complementarity of the two molecular surfaces in contact –The strength of the interaction Based on the physico-chemical characteristics of the amino acids in contact with each other 3

4. Introduction The formation of a complex –Concern only the side chain conformation of amino acid residues at the interface –Imply motions of the protein backbone The nature and amplitude of which remains very difficult to predict Unbound docking predictions much more difficult than bound ones 4

5. Introduction Protein-protein complexes in a complete genome with thousands of genes –Very reliable The score of the best solution is high The best solution is close to the native complex –Very fast The inspection of a whole genome requires the modeling of many hundreds of thousands potential complexes 5

6. Methods The Voronoi diagrams –Includes all points of space that are closer to the cell centroid than to any other centroid –Smallest polyhedron defined by bisecting planes between its centroid and all others The Dalaunay tessellation is obtained by tracing the vertices joining centroids 6

7. Voronoi diagrams 7

8. Tessellation Definitions (1/2) –Two residues are neighbors If their Voronoi cells share a common face –A residue belongs to the protein interior If all its neighbors are residues of the same protein –A residue belongs to the protein surface If one ore more of its neighbors is solvent 8

9. Tessellation Definitions (2/2) –A residue belongs to the protein–protein interface If one or more of its neighbors belongs to the other protein –An interface residue belongs to the core of the interface If none of its neighbors is solvent –The cell facets shared by residues of both proteins constitute the interface 9

10. Voronoi Description (1p2k) 10 Protein chains Solvent By Voronoi polyhedra of residues of two proteins, representating the interface

11. Training Set The training set consists in two subsets –Positive examples Complexes of known 3D structure –The 2004 #1 of the Protein Data Bank (PDB) –Negative examples Generated from the positive examples using a docking procedure Decoys 11

12. Decoys An imperfect scoring function can –Mislead by predicting incorrect ligand geometries by selecting nonbinding molecules over true ligands –These false-positive hits may be considered decoys 12

13. Training Attributes (1/2) Number of parameters that may be used in –training is limited by the size of the training set To define pair attributes, we grouped residue types in six categories –Hydrophobic H –Aromatic Φ –Positive charged + –Negatively charged – –Polar P –Small S 13

14. Training Attributes (2/2) –Attributes includes 84 parameters in sixe classes P1 The Voronoi interface area (1 parameter) P2 Total number of core interface residues (1) P3 Number fraction of each type of core interface residues (20) P4 The mean volume of the Voronoi cells for the core interface residues of each type (20) P5 number fraction of pairs of each category (21) P6 The mean centroid-centroid distance in pairs of each category (21) 14

15. 15

16. Learning Methods The values of the 84 parameters were measured on –The 102 native complexes (positive) –The decoys of training set (negative) Logistic function SVM (Support Vector Machines) ROGER (a ROc based GEnetic learner) 16

17. Learning Methods Logistic function –Linear combination of the parameters with weighted optimized to yield a value Close to 1 on the native models 0 on the decoys –Using the general linear model (GLM) of the R software 17

18. Learning Methods SVM (Support Vector Machine) –Divide a high-dimensional space into regions containing only positive examples or only decoys –Using SVMTorch Efficiently solve large-scale regression problems 18

19. Learning Methods ROGER (Roc based genetic learner) –The receiver operating characteristics (ROC) procedure is often used evaluate learning By cross-validation on examples –Uses genetic algorithm to find a family of functions that optimized the ROC criterion 19 Central value ci and weight wi

20. Results and Discussion Performance of the learning procedures –The ROC curve was evaluated on the training set for four different scores Sum of the mean square deviation Logistic function SVMs ROGER scoring function –A perfect selection (100% true positive &no false positive) should make the area under the ROC curve AUC equal to 1 20

21. ROGER and SVMs did much better –AUC of 0.98 and 0.99, respectively –Very few false positives among their best scoring solutions Retained the ROGER score for further studies as the SVMs only give a binary classification Ill-suited to our problem of “finding a needle in a hay stack” 21

22. Results on the targets of CAPRI rounds 3-6 We tested the scoring functions on models of the targets of CAPRI rounds 3-6 –By two docking programs DOCK (1991) HADDOCK (2003) 22

23. –Fnat: fraction of native contacts present in the solution –Fint: fraction of interface residues correctly predicted Class 1: Fnat > 0.75 Class 2: 0.5 < Fnat < 0.75 Class 3: 0.25 < Fnat < 0.5 Class4: 0 < Fnat < 0.25 Class5: Fint > 0 and Fnat =0 Class6: Fint = 0 and Fnat = 0 23

24. 24 target target 18: the best solution in the set has 31% of native contacts. The best solution is the set is thus class 3. The best rank given by our scoring function to a solution of the best class in the set he first solution of the second best class (class 4 in this case) is ranked 1 Original rank of the rank 1 solution Number of models with ROGER ranks < 50 that belong to the best class Rank given by our function is better than given by HADDOCK Best class (class 2), re-ranked 4 Next best class (class 3) was1st Top 50 ROGER scores included 4 models of class 2, and 44 models of class 3 Thus Very few if any of top50 were false positive Best class (class 2), re-ranked 4 Next best class (class 3) was1st Top 50 ROGER scores included 4 models of class 2, and 44 models of class 3 Thus Very few if any of top50 were false positive

25. Conclusion For most targets, a best or second best class solution was found in the top 10 ranking solutions –More than half of the cases the top ranking solution belonged to the best or second best class –For all targets but one Rank given to the first best class solution by our scoring function is better than the rank given by the original method (DOCK or HADDOCK) 25

26. 2016/7/726 Thanks for your attention