Figure 1 shows a snapshot of the input page. The user uploads two Protein Data Bank (PDB) format files (13), one ‘receptor’ and one ‘ligand’, along with their email address. The web server currently makes no distinction between chains; it simply reads in each line in the PDB file starting with an ‘ATOM’ field. The maximum number of residues is limited to 1500. The email address is where the output/results will be sent (as a file attachment). Hydrogen bonds and missing atoms are built and optimized on the uploaded structures using the molecular software CHARMm. Figure 1.

Snapshot of input page.

The results from a ‘FastContact’ server run are returned to the user via email as a file attachment (with a normal response time of ∼1 min). The attached file is a gzipped archive (.tar.gz) containing three results files: (i) the main results file (‘output.txt’); and, the processed (including H-bonds) and renumbered (ii) receptor PDB file (‘protein1’) and (iii) the ligand PDB file (‘protein2’). All of the files are prefixed with the user name (email prefix) and timestamp of the server run for easy reference.

The main source of errors in the output file relates with the format of the input PDB files. For instance, columns usage must strictly follow the PDB standards, and ATOM keyword must describe only protein amino acids. The server cannot minimize the backbone of sequences with gaps, and missing heavy atoms are sometimes not able to be reconstructed by the server. If the server detects an error, it will report a message with possible problems and suggestions.

The main results file (‘output.txt’) returns two components of a free energy function, electrostatic energy and desolvation free energy, and evaluates the solute vdW energy using CHARMm. The latter is sometimes useful to compare between different models (16), but here it is given only as a reference since it is not used in the analysis of contacts. Often vdW energies larger than about −500 kcal/mol suggest structural overlaps. Although ‘FastContact’ smoothes the potentials to tolerate some limited overlaps, these are, in general, detrimental to the quality of the computational estimates. Figure 2 shows the summary energy output and part of the contact analysis, for the barnase–barstar complex 1BRS. We should caution that, when submitting co-crystallized receptor and ligand structures, the automated minimization implemented in the server leads to an over optimization of the electrostatic contacts of ∼10–20% (5). The reason is because the direct electrostatic term used in the server does not have an angular dependence for Hydrogen bonds. Hence, these interactions tend to ‘double-dip’. This effect is compensated when scoring unbound models that always have some built in frustration due to the less optimal backbone and side chain conformations. Figure 2.

Output page. It includes a summary of electrostatics, desolvation and vdW energies, followed by a list of the 20 most attractive and 20 most repulsive residues and contacts for the electrostatic, desolvation, and the sum of these two components that correlates with the binding free energy. For modeling, the repulsive information is sometimes useful as an indication of a wrong structural motif.

The method implemented in ‘FastContact’ has been successfully applied in the CAPRI experiment both as a free energy filtering procedure of the ‘ClusPro’ server (14) that predicts protein complexes and in protein–protein refinement (4) (using a 9 and 6 Å desolvation range, respectively). ‘FastContact’ has been instrumental in the success of our group in blind predictions (17,18). In rounds 1 and 2 of CAPRI, Camacho and Gatchell (19) produced some of the best model structures, appropriately distinguishing between near-native and false positive structures for three targets. In rounds 3–5, the automated server ‘ClusPro’ (the only server participating in CAPRI) predicted good models for 5 targets (15), while our manual predictions resulted in good predictions for 6 targets (20) (missing the 3 targets that had a significant structural rearrangement upon binding).

The robustness of our method was further supported by the analysis of the full set of models submitted for CAPRI (rounds 3–5) for the 6 targets that did not undergo a large structural rearrangement upon binding (18). For these targets, we showed that ‘FastContact’ was able to discriminate near-native predictions from docked conformations far from the binding site for 5 of the targets (10), and for all but one of the manual predictions submitted to CAPRI. For instance, Figure 3 shows the re-scoring of models submitted for targets 8 and 12 by 13 different groups around the world. In all cases, ‘FastContact’ correctly identified the near-native conformation, even when the modeler failed to do so. Figure 3.

Examples of ‘FastContact’ scoring for a subset of high quality docked models from eight groups for targets 8 and 12 of CAPRI rounds 3–5, from http://capri.ebi.ac.uk/. For each of these targets, we run the models in our server and re-rank the models accordingly. In all cases, the server was able to correctly rank a low RMSD model as the one with the lowest free energy score. For comparison, we also marked with a diamond symbol the model ranked number 1 by the modeler (23–33). (A) target 8; (B) target 12. For a complete analysis of all models please see reference (10).