Protein docking prediction using predicted protein-protein interface

Abstract

Background: Many important cellular processes are carried out by protein complexes. To provide physical pictures of interacting proteins, many computational protein-protein prediction methods have been developed in the past. However, it is still difficult to identify the correct docking complex structure within top ranks among alternative conformations.

Results: We present a novel protein docking algorithm that utilizes imperfect protein-protein binding interface prediction for guiding protein docking. Since the accuracy of protein binding site prediction varies depending on cases, the challenge is to develop a method which does not deteriorate but improves docking results by using a binding site prediction which may not be 100% accurate. The algorithm, named PI-LZerD (using Predicted Interface with Local 3D Zernike descriptor-based Docking algorithm), is based on a pair wise protein docking prediction algorithm, LZerD, which we have developed earlier. PI-LZerD starts from performing docking prediction using the provided protein-protein binding interface prediction as constraints, which is followed by the second round of docking with updated docking interface information to further improve docking conformation. Benchmark results on bound and unbound cases show that PI-LZerD consistently improves the docking prediction accuracy as compared with docking without using binding site prediction or using the binding site prediction as post-filtering.

Conclusion: We have developed PI-LZerD, a pairwise docking algorithm, which uses imperfect protein-protein binding interface prediction to improve docking accuracy. PI-LZerD consistently showed better prediction accuracy over alternative methods in the series of benchmark experiments including docking using actual docking interface site predictions as well as unbound docking cases.

Figure 1

Surface points used in geometric hashing. A, Example of surface points. Surface points of 1A2K chain B (left, purple, 248 residues) are colored in blue for non-PPI surface (1652 points) and yellow for the PPI region (25 residues, 248 points). Critical points of 1A2K chain C (right, yellow, 196 residues) are colored in green for non-PPI surface (1332 points) and red for the PPI region (16 residues, 165 points). B, Two schemes for restricting docking interface using predicted PPI regions in geometric hashing. Two base points are selected from points that locate within the predicted PPI regions (gray ellipsoid). Given the two base points (crosses), points in their neighborhood (within 15Å; showed in two circles) are transformed based on the coordinate system defined by the base points, which are then stored in a hash table. Fitness of local regions around the base points of the ligand and the receptor proteins are evaluated by the number of equivalent matching points (the voting stage). In a permissive search, points from outside of the PPI region (squares) as well as points within the PPI region (triangles) are considered. For a restrictive search, only the points in the PPI region (triangles) are considered.

Figure 2

Overview of the PI-LZerD algorithm. See the text for the explanation. "LZerD" is the original version of the LZerD docking program. Modified LZerD (p) and (r) stands for the permissive search space and the restrictive search space employed in the geometric hashing stage, respectively.

Figure 3

Accuracy of the shifted PPI regions. Distributions of A, the sensitivity (= specificity); B, the fnat value; of the shifted PPI regions by 5, 10, 12, and 15 residues.

Figure 4

The prediction accuracy of the naive post-filtering method. The percentage of the cases among the tested complexes is shown where the naive post-filtering method obtained a near native structure of below A, 2.5 Å iRMSD; B, 4.0 Å iRMSD; within specified ranks at the x-axis. PPI site information of five different accuracy levels are used: correct PPI (pentagons); 5 residue shifted (downward triangles); 10 residue-shifted (upward triangles), 12 residue-shifted (diamonds); and 15 residue-shifted PPI regions (squares). For comparison, results of the base LZerD which do not use PPI information are also shown (filled triangles).

Figure 5

Docking prediction with simulated protein interface predictions. Four other methods are listed to compare with PI-LZerD-2: the base LZerD (LZerD), LZerD with clustering using ciRMSD (LZerD+Clustering), LZerD with one interaction of modified LZerD (PI-LZerD-1), and Simple residue filtering method (post-filtering). The x-axis indicates the ranks in logarithmic scale and the y-axis shows the percentage of cases where correct predictions are ranked equal or better than the corresponding ranks. Left panels, A, C, E, G, I, use the 2.5 Å as threshold for correct predictions, while right panels, B, D, F, H, J, use 4.0 Å as the cutoff for near hit predictions. A &B use the correct protein interface information; C/D use the simulated predictions with 5 residue shifts, E/F, G/H, and I/J use the simulated predictions with 10, 12, and 15 residue shifts, respectively.

Figure 6

Docking prediction results using meta-PPISP binding site prediction. The meta-PPISP server predicted PPI regions of 127 complexes selected from the iPFAM database. Distribution of A, the sensitivity and B, the specificity of the meta-PPISP prediction. The docking prediction results using C, 2.5 Å iRMSD cutoff; D, 4.0 Å iRMSD to define correct predictions.

Figure 7

Docking results on the unbound dataset using meta-PPISP prediction. 118 unbound-unbound protein complexes were taken from the docking benchmark dataset version 3.0. The distribution of A, the sensitivity; and B, the specificity. C and D use 2.5 Å and 4.0 Å, respectively, for the iRMSD cutoff value to define correct prediction.

Figure 8

Docking results on the unbound dataset using meta-PPISP prediction are classified by the docking difficulty. The same results shown in Figure 7 are classified into three docking difficulty classes, rigid-body (87 cases), medium (16 cases), and, difficult (15 cases). A, B, results using 2.5 Å and 4.0 Å, respectively, for the iRMSD cutoff value to define correct prediction. Black/red circles, the rigid-body category results by PI-LZerD/LZerD; green/yellow triangles, the medium category by PI-LZerD/LZerD; Black/red squares, the difficult category by PI-LZerD/LZerD.

Figure 9

Examples of docking prediction. The best prediction within 50 ranks by PI-LZerD-2, the base LZerD, and naive post-filtering is shown in green, yellow, and red, respectively. The correct docking pose is shown in blue. Receptor proteins are shown in gray. First, predictions using 10 residue-shifted PPI site information are shown for A, 1BUH; B, 1MLC. For 1BUH, the iRMSD (Å) of prediction by PI-LZerD-2/post-filtering/base LZerD was 1.03/9.09/9.91. The rank of the decoys was 8/24/17. The sensitivity (= specificity, since the size of the shifted PPIs are same as the correct PPI) of the shifted PPI region was 0.33 for 1BUHA and 0.44 for 1BUHB. For 1MLC, iRMSD (Å) and the rank of the best prediction within 50 ranks by PI-LZerD-2/post-filtering/base LZerD were 0.89 Å (34)/8.37 Å (9)/14.35 Å (22). The ranks are shown in the parentheses. The sensitivity (= specificity) of the shifted PPI region was 0.55 for chain A and B, and 0.44 for chain E. C and D are predictions using actual PPI site predictions by meta-PPPISP for proteins from the iPFAM dataset. C, 1ADU. The iRMSD (Å) and the rank of the predictions by PI-LZerD-2/post-filtering/base LZerD were 1.04 Å (48)/14.90 Å (37)/10.85 Å (43). The sensitivity/specificity of PPI site predictions were 0.77/0.47 for 1ADUA, and 0.00/0.00 for 1ADUB. D, 1BMT. The iRMSD (Å) and the rank of the predictions by PI-LZerD-2/post-filtering/base LZerD were 2.31 Å (36)/14.44 Å (31)/13.02 Å (44). The sensitivity/specificity of PPI site predictions were 0.11/0.06 for 1BMTA, and 0.22/0.11 for 1BMTB. The last two examples are unbound docking cases with actual PPI predictions by meta-PPISP. E, 1OPH. The iRMSD (Å) and the rank of the predictions by PI-LZerD-2/post-filtering/base LZerD were 3.76 Å (42)/5.71 Å (39)/10.28 Å (16). The sensitivity/specificity of PPI site predictions were 1.00/0.32 for 1OPHA, and 0.70/0.18 for 1OPHB. F, 1IQD. The PI-LZerD-2/post-filtering/base LZerD: 2.91 Å (23)/6.97 Å (16)/12.80 Å (32). The sensitivity/specificity of PPI site predictions were 0.53/0.16 for 1IQDA, and 0.89/0.20 for 1IQDB.

Figure 10

Comparison of the docking prediction performance by PI-LZerD and CPORT. The dataset contains 57 unbound proteins. The predictions were downloaded from the CPORT website. Distribution of A, the sensitivity and B, the specificity of the PPI site predictions by CPORT. C, hits within 2.5 Å iRMSD; D, hits within 4.0 Å iRMSD.