Virtual Ligand Screening Against Comparative Protein Structure Models

Abstract

Virtual ligand screening uses computation to discover new ligands of a protein by screening one or more of its structural models against a database of potential ligands. Comparative protein structure modeling extends the applicability of virtual screening beyond the atomic structures determined by X-ray crystallography or NMR spectroscopy. Here, we describe an integrated modeling and docking protocol, combining comparative modeling by MODELLER and virtual ligand screening by DOCK.

1. Introduction

Structure-based methods have been widely used in the design and discovery of protein ligands (1–4). Given the structure of a binding site on a receptor protein, its ligands can be predicted among a large library of small molecules by virtual screening (1, 5–11): Each library molecule is docked into the binding site, then scored and ranked by a scoring function. High-ranking molecules can be selected for testing in the laboratory. Virtual screening methods can significantly reduce the number of compounds to be tested, thus increasing the efficiency of ligand discovery (12–16).

Many protein structures are relatively flexible, and can adopt different conformations when binding to different ligands. Docking a ligand to a protein structure with current methods is most likely to be successful when the shape of the binding site resembles that found in the protein-ligand complex. Therefore, the protein structure for docking is best determined in complex with a ligand that is similar to the ligand being docked, by X-ray crystallography or NMR spectroscopy. Induced fit and differences between protein conformations bound to different ligands limit the utility of the unbound (apo) structure and even complex (holo) structures obtained for dissimilar ligands. The problem of the protein conformational heterogeneity is especially difficult to surmount in virtual screening, which involves docking of many different ligands, each one of which may in principle bind to a different protein conformation (17).

An even greater challenge is that many interesting receptors have no experimentally determined structures at all, especially in the early phases of ligand discovery. During the last seven years, the number of experimentally determined protein structures deposited in the Protein Data Bank (PDB) increased from 23,096 to 67,421 (November 2010) (18). In contrast, over the same period, the number of sequences in the Universal Protein Resource (UniProt) increased from 1.2 million to 12.8 million (19). This rapidly growing gap between the sequence and structure databases can be bridged by protein structure prediction (20), including comparative modeling, threading, and de novo methods. Comparative protein structure modeling constructs a three-dimensional model of a given target protein sequence based on its similarity to one or more known structures (templates). Despite progress in de novo prediction (21, 22), comparative modeling remains the most reliable method that can sometimes predict the structure of a protein with accuracy comparable to a low-resolution, experimentally determined structure (23).

Comparative modeling benefits from structural genomics (24). In particular, the Protein Structure Initiative (PSI) aims to determine representative atomic structures of most major protein families by X-ray crystallography or NMR spectroscopy, so that most of the remaining protein sequences can be characterized by comparative modeling (http://www.nigms.nih.gov/Initiatives/PSI/) (25, 26). Currently, the fraction of sequences in a genome for whose domains comparative models can be obtained varies from approximately 20% to 75%, increasing the number of structurally characterized protein sequences by two orders of magnitude relative to the entries in the PDB (27). Therefore, comparative models in principle greatly extend the applicability of virtual screening, compared to using only the experimentally determined structures (28).

Comparative models have in fact been used in virtual screening to detect novel ligands for many protein targets,(28) including G-protein coupled receptors (GPCR) (29–41), protein kinases (42–45), nuclear hormone receptors, and a number of different enzymes (14, 15, 46–57). The relative utility of comparative models versus experimentally determined structures has been assessed (17, 29, 42, 43, 58–60). Although the X-ray structure of a ligand-bound target often provides the highest enrichment for known ligands, comparative models yield better enrichment than random selection and sometimes performs comparably to the holo X-ray structure. Recently, we assessed our automated modeling and docking pipeline (17) based on MODELLER (61) for comparative modeling and DOCK (62, 63) for virtual screening. We demonstrated that when multiple target models are calculated, each one based on a different template, the “consensus” enrichment for multiple models is better or comparable to the enrichment for the apo and holo X-ray structures in 70% and 47% cases, respectively; the consensus enrichment is calculated by combining the docking results of multiple structures — for each docked compound, the best docking score across all structures was used for ranking the compound — thus, the ranking relied on optimizing the protein conformation as well as protein-ligand complementarity. Another similar criterion for ligand ranking was also described (64).

The modeling and docking protocol is carried out in 7 sequential steps (Figure 1). Steps 1–4 correspond to comparative modeling: (1) template search finds known structures (templates) related to the sequence to be modeled (target), (2) target-template alignment aligns the target sequence with the templates, (3) model construction computes multiple target models based on the input alignment, (4) model selection identifies the best-scoring model. Steps 5–7 correspond to virtual screening: (5) binding site preparation involves creating input files for generating spheres and scoring grids used in docking, (6) database screening docks database molecules into the binding site, and (7) database prioritization scores and ranks the docking poses of the database molecules. Comparative modeling is carried out by program MODELLER that implements comparative modeling by satisfaction of spatial restraints derived from the target-template alignment, atomic statistical potentials, and the CHARMM molecular mechanics force field (61). The spatial restraints are combined into an objective function that is optimized by a combination of conjugate gradients and molecular dynamics with simulated annealing; this model-building procedure is formally similar to structure determination by NMR spectroscopy. Virtual screening is performed by the DOCK suite of programs (63, 65, 66). DOCK uses a negative image of the receptor — spheres that fill the receptor site — to describe the space into which docked molecules should fit. Docking poses are generated by matching the atoms of a small molecule with the centers of the spheres. The generated poses are evaluated using a grid-based approach in which interactions between the docked molecules and the receptor are precomputed at each grid point.

Figure 1.

The automated modeling and docking pipeline. Numbers in parentheses indicate the corresponding section in the text.

2. Materials

2.1. Software for Comparative Modeling

2.1.1

The MODELLER 9v8 program can be downloaded from http://salilab.org/modeller/.

2.1.2

A typical operation in MODELLER consists of (1) preparing an input Python script, (2) ensuring that all required files (eg, files specifying sequences, structures, alignments) exist, (3) executing the input script by typing ’mod9v8 input-script-name’, and (4) analyzing the output and log files. A tutorial for the use of MODELLER 9v4 or newer is available at http://salilab.org/modeller/tutorial/.

2.2. Database for Comparative Modeling

2.2.1

Sequence database (UniProt90) contains all sequences from UniProt (clustered at 90% to remove redundancy), and can be downloaded from http://salilab.org/modeller/supplemental.html.

2.2.2

Template sequence database (pdball) contains the sequence for each protein structure in PDB, and can be downloaded from http://salilab.org/modeller/supplemental.html.

2.3. Software for Virtual Screening

2.3.1

DOCK 3.5.54 (62, 63) is available under the UCSF DOCK license http://dock.compbio.ucsf.edu/Online_Licensing/dock_license_application.html (Note 4.1). Documentation for DOCK 3.5 is provided at http://wiki.bkslab.org/index.php/Image:Dock3_5refman.pdf.

2.3.2

Third party applications.

2.3.2.1

DMS is a program that calculates the solvent-accessible molecular surface of the protein binding site (67), and can be downloaded at http://www.cgl.ucsf.edu/Overview/ftp/dms.shar.

2.3.2.2

SYBYL is a commercial molecular modeling program that can build and manipulate molecules.(68) In our study, SYBYL is used to add hydrogen atoms to polar atoms in a protein receptor (in the PDB format) that contains only non-hydrogen atoms; it can be downloaded from http://tripos.com/index.php?family=modules,General.DownloadPortal,Home.

2.3.2.3

Delphi is a program that computes numerical solutions of the Poisson-Boltzmann equation for molecules of arbitrary shape and charge distribution (69); a request for access to this program can be made at http://luna.bioc.columbia.edu/honiglab/software/cgi-bin/software.pl?input=DelPhi.

2.4. Docking Database of Small Molecules

2.4.1

The Directory of Universal Decoys (DUD) is a docking database designed to help test docking algorithms by providing challenging decoys (70). DUD contains a total of 2,950 compounds that bind to a total of 40 targets; in addition, for each ligand, it also contains 36 "decoys" with similar physical properties (eg, molecular weight, calculated LogP) but dissimilar chemical topology. DUD can be downloaded from http://dud.docking.org/r2/.

3. Method

The automated modeling and docking pipeline will be illustrated with one example taken from our benchmark study (17), adenosine deaminase (ADA, EC 3.5.4.4). ADA is a metalloenzyme in whose binding pocket one catalytic zinc ion is coordinated by three histidine residues and one aspartic acid residue (71, 72). The bovine ADA has been co-crystalized with a non-nucleoside inhibitor (PDB code 1NDW). The DUD database was screened against comparative models and the ligand-bound (holo) crystal structure of the bovine ADA, to compare the utility of comparative models and holo crystal structures for virtual screening.

3.1. Comparative Modeling of Protein Structures

3.1.1. Template search

First, a file with the bovine ADA sequence in the MODELLER “PIR” format is prepared (Figure 2; Note 4.2). Then the ADA sequence is scanned against all sequences in the PDB (stored in file “pdball”) to identify suitable templates, with the MODELLER “profile.build” routine (Figure 3; Note 4.3). In this example, one holo structure (PDB code 1UIO) (73) with 85% sequence identity to the target and one apo structure (PDB code 2AMX) (74) with 27% sequence identity are selected as templates (Note 4.4), to be used independently for calculating two models of ADA.

Figure 2.

File “ADA.ali” in the “PIR” format. This file specifies the target sequence. See the MODELLER manual for the detailed description of the format.

Figure 3.

File “search_templates.py”. This script searches for potential template structures in a database of non-redundant PDB sequences.

3.1.2. Target-template alignment

For each target-template pair (ie, ADA-1UIO and ADA-2AMX), the target and template sequences are scanned against all sequences in UniProt90 independently with the “profile.build” routine, resulting in the target profile and the template profile, respectively. Next, the target profile is aligned against the template profile with the “profile.scan” routine (a sample script is given at http://salilab.org/modeller/examples/commands/ppscan.py). The resulting alignment is presented in Figure 4, for the 2AMX template (Note 4.5; the ADA-1UIO alignment is not shown).

Figure 4.

File “align.ali” in the “PIR” format. The file specifies the alignment between the sequences of ADA and 2AMX (A chain).

3.1.3 Model construction

Once the target-template alignment is generated, MODELLER calculates 500 models of the target completely automatically, using its “automodel” routine (Figure 5; Note 4.6). The best model (defined in 3.1.4. Model selection) is then subjected to a refinement of binding site loops (Note 4.7) with the “loopmodel” routine (Figure 6). All three binding site loops were optimized simultaneously, resulting in 2500 conformations of ADA (Note 4.8).

Figure 5.

File “build_model.py”. The script generates 500 models of ADA based on 2AMX with “automodel” routine.

Figure 6.

File “loop_model.py”. Input script file that generates 2500 models with the “loopmodel” routine.

3.1.4. Model selection

When multiple models are calculated for the target based on a single template (by “automodel”, and “loopmodel”, if there are binding site loops), it is practical to select the model or a subset of models that are judged to be most suitable for subsequent docking calculations (Note 4.9). In this example, for each template, we select the model with optimized loops that has the lowest value of the MODELLER objective function (ada-loop.BL16340001.pdb for 2AMX), which is reported in the second line of the model file (Note 4.10). The most suitable model can also be selected by the Discrete Optimized Protein Energy (DOPE) (75), which is calculated using the “assess_dope” routine (Note 4.11).

3.2. Virtual Screening Against Comparative Models

As described in the previous section, a single comparative model of bovine ADA is selected from models calculated based on the 2AMX template. Another model is selected from models based on the 1UIO template. The DUD database is then screened against each of the two models independently. We will only describe the docking to the ADA model based on 2AMX.

3.2.1. Binding site preparation

3.2.1.1. Prepare input files for the automated docking pipeline

The file containing the ADA model based on 2AMX is renamed to “rec.pdb”, followed by (1) removing all lines that do not contain coordinates of non-hydrogen atoms; (2) replacing “HETATM” in the line containing the coordinates of the zinc ion by “ATOM”; and (3) removing all chain identifiers (Note 4.12). Next, the file “xtal-lig.pdb” is created, containing the binding site specification in the same format as that of “rec.pdb”. In this example, the ligand observed in the holo crystal structure of the target is given in “xtal-lig.pdb”; this ligand is transferred into the model by superposing the crystal structure on the model using the binding site residues (Note 4.13).

3.2.1.2. Automated spheres and scoring grids generation

First, the environment variable “DOCK_BASE” is defined to be the “dockenv” directory of the DOCK 3.5.54 installation. Second, file “Makefile” from “dockenv/scripts/” is copied to the current working directory, which also contains the “rec.pdb” and “xtal-lig.pdb” files. Third, file “.useligsph” is generated. Finally, command “make” is executed to generate the spheres and scoring grids (Note 4.14).

3.2.2. Database screening

The DUD database contains 2950 annotated ligands and 95,316 decoys for 40 diverse targets (70); the DUD database is stored in 801 DOCK 3.5 hierarchy database files (DUD 2006 version) (63). 801 sub-directories corresponding to the 801 hierarchy database files are created. In each sub-directory, two files are required for docking. One is file “INDOCK” that contains the input parameters for DOCK 3.5.54 (Figure 7) (Note 4.15). Another file, “split_database_index”, contains the location and name of the corresponding database file. In file “INDOCK”, “split_database_index” is given as the value for the parameter with the keyword “ligand_atom_file”. Docking is performed by running the DOCK executable “dockenv/bin/Linux/dock” in each sub-directory. Two output files are produced: (1) the compressed file “test.eel1.gz” contains the docking poses of database molecules in the extended PDB format and (2) the compressed file “OUTDOCK.gz” contains the docking scores for the database molecules as well as the input file names and parameter values.

Figure 7.

A section of file “INDOCK” containing some input parameters for DOCK 3.5.54.

3.2.3. Database prioritization

First, the conformations of database molecules are filtered for steric complementarity using the DOCK contact score. The conformations that do not clash with the receptor are then scored by the DOCK energy function (the DOCK contact score is not included):

$$E_{\mathit{score}}=E_{\mathit{vdw}}+E_{\mathit{elec}}+\mathrm{\Delta }{G}_{\mathit{desolv}}^{\mathit{lig}}$$ (1)

where E_vdW is the van der Waals component of the receptor-ligand interaction energy based on the AMBER united-atom force field, E_elec is the electrostatic potential calculated by DelPhi, $\mathrm{\Delta }{G}_{\mathit{desolv}}^{\mathit{lig}}$ is the ligand desolvation penalty computed by solvmap, as described in Section and 3.2.2. For each ligand conformation, the total energy and all the individual energy terms are written out to file “OUTDOCK” (Figure 8; Note 4.16). The single conformation with the best total energy is saved in file “test.eel1” as the docking pose of the database molecule. The docking pose of one ADA ligand – 1-deazaadenosine (PubChem ID: 159738, ZINC ID: C03814313) – is shown in Figure 11B. After the virtual screening, the best total energy of each database molecule and the corresponding molecule ID are extracted from the “OUTDOCK” files in all sub-directories. The molecules in the docking database are ranked by their total energies. The top 500 ranked molecules are then inspected visually. Molecules forming favorable interactions with the receptor (eg, a docking pose is similar to the binding mode found in crystal structures of proteins in the same family) can be chosen for subsequent experimental testing.

Figure 8.

A section of file “OUTDOCK” containing docking scores of two DUD molecules.

Figure 11.

(A) The matching spheres (dark grey) and DelPhi spheres (light grey) generated for the binding site of the ADA model (cartoon) based on 2AMX. (B) The docking pose (stick) and the 2D structure of one ADA ligand – 1-deazaadenosine (PubChem ID: 159738, ZINC ID: C03814313) – as well as the matching spheres (light grey)

In this benchmark example, we can quantify the accuracy of modeling and docking by computing the enrichment for the known ADA ligands among the top scoring ligands:

$${EF}_{\mathit{subset}}=\frac{\left({\mathit{ligand}}_{\mathit{selected}}/N_{\mathit{subset}}\right)}{\left({\mathit{ligand}}_{\mathit{total}}/N_{\mathit{total}}\right)}$$ (2)

where ligand_total is the number of known ligands in a database containing N_total compounds and ligand_selected is the number of ligands found in a given subset of N_subset compounds. EF_subset reflects the ability of virtual screening to find true positives among the decoys in the database compared to a random selection. An enrichment curve is obtained by plotting the percentage of actual ligands found (y-axis) within the top ranked subset of all database compounds (x-axis on logarithmic scale). To measure the enrichment independently of the arbitrary value of N_subset, we also calculated the area under the curve (logAUC) of the enrichment plot:

$$\mathit{logAUC}=\frac{1}{{\mathit{log}}_{10}100/\lambda }\sum _{\lambda }^{100}\left\{\frac{{\mathit{ligand}}_{\mathit{subset}}}{{\mathit{ligand}}_{\mathit{total}}}•\left(\lambda •{\mathit{log}}_{10}\frac{N_{\mathit{subset}}}{N_{\mathit{total}}}\right)\right\}$$ (3)

where λ is arbitrarily set to 0.1. A random selection (ligand_selected/ligand_total = N_subset/N_total) of compounds from the mixture of true positives and decoys yields a logAUC of 14.5. A mediocre selection that picks twice as many ligands at any N_subset as a random selection has logAUC of 24.5 (ligand_selected/ligand_total = 2 * N_subset/N_total; N_subset/N_total ≤ 0.5). A highly accurate enrichment that produces ten times as many ligands than the random selection has logAUC of 47.7 (ligand_selected/ligand_total = 10 * N_subset/N_total; N_subset/N_total ≤ 0.1). In this example, the ADA model based on 2AMX yielded the logAUC of 40.3 (Figure 9). When multiple structures are available (either models or experimental structures), consensus enrichment can be calculated (Introduction).

Figure 9.

The enrichment curve for virtual screening of the DUD database against the ADA model based on 2AMX. The ligand enrichment is quantified by the logAUC of 40.3.

Figure 10.

Schematic description of the automated preparation of receptor binding site, including sphere and scoring grids generation.