Docking of protein models

Abstract

An adequate description of entire genomes has to include information on the three-dimensional (3D) structure of proteins. Most of these protein structures will be determined by high-throughput modeling procedures. Thus, a structure-based analysis of the network of protein-protein interactions in genomes requires docking methodologies that are capable of dealing with significant structural inaccuracies in the modeled structures of proteins. We present a systematic study of the applicability of our low-resolution docking method to protein models of different accuracies. A representative nonredundant set of 475 cocrystallized protein-protein complexes was used to build an array of models of each protein in the set. A sophisticated procedure was created to generate the models with RMS deviations of 1, 2, 3,., 10 A from the crystal structure. The docking was performed for all the models, and the predictions were compared with the configurations of the original cocrystallized complexes. Statistical analysis showed that the low-resolution docking can determine the gross structural features of protein-protein interactions for a significant percent of complexes of highly inaccurate protein models. Such predictions may serve as starting points for a more detailed structural analysis, as well as complement experimental and computational data on protein-protein interactions obtained by other techniques.

Fig. 1.

Model used to calculate the statistical significance of the docking results. Proteins are approximated by spheres. Matches, represented by ligand's center of mass, are placed around the receptor using the uniform random distribution for each match. The binding site is shown in gray. The size of the proteins, shown relative to the size of the binding site (<10 Å), approximately corresponds to the average R_A and R_B values in the database.

Fig. 2.

The concept of the database of protein models for validation of docking. Each of 475 protein–protein complexes is complemented by 10 models of both protein subunits. The accuracy of the models ranges from 1 to 10 Å, with a 1 Å interval.

Fig. 3.

The outline of the algorithm for generating protein structures with the predefined RMSD from the native structure. Areas with flexible ϕ and ψ angles (see magnified inset) are shown in green. Secondary structure elements, shown in gray, are treated as rigid bodies. See text for details.

Fig. 4.

Percent of original interface residues that remain on the surface in model structures of different accuracies. The surface residues were defined as residues with the side-chain accessible surface >7% of the total accessible surface of the side chain for the residue type (Mizuguchi et al. 1998). The accessible surface was calculated by PSA (Sali and Blundell 1993). The data is averaged over all 475 complexes and over 110 complexes with large interfaces (>4000 Ų).

Fig. 5.

Average change of distance between residues (d-RMSD) in model structures of different accuracies. The noninterface data is the average over the interface-size surface patches outside the interface.

Fig. 6.

Evolution of the interface in a set of models with increasing RMSD from the native structure. The structure shown is 1cyd, subunit D. The interface is with the subunit C. The definition of the interface is in the text. The interface residues are in black.

Fig. 7.

Results of the low-resolution docking of trypsin and its protein inhibitor BPTI. (a) Experimental structures. (b) Low-resolution models (RMS = 6 Å, both trypsin and BPTI). The red spheres are the BPTI center of mass in 100 lowest energy positions. The yellow sphere (indicated by an arrow) is the BPTI center of mass in the cocrystallized complex. For comparison, the experimental structure of trypsin, in green, is overlapped with the model. The docking of the models clearly preserves the cluster of correct predictions in the area of the binding site.

Fig. 8.

Percent of correctly predicted complexes for protein models of different accuracies. (a) All 475 complexes. (b) Complexes with small interfaces (1000–2000 Ų, 189 complexes). (c) Complexes with large interfaces (>4000 Ų, 110 complexes). The black horizontal line indicates the estimated upper limit on the percent of complexes where the correct prediction could occur by chance, due to the clustering of matches only. See text for the definition of the correct prediction and other details.