Multiple virtual screening approaches for finding new Hepatitis C virus RNA-dependent RNA polymerase inhibitors: Structure-based screens and molecular dynamics for the pursue of new poly pharmacological inhibitors

BMC Bioinformatics 2012, 13(Suppl 17):S5

Mahmoud ElHefnawi, Mohammad ElGamacy, Mohamed Fares.

Informatics and Systems Department, Biomedical Informatics and chemoinformatics group, Division of Engineering Research and Centre of Excellence for Advanced Sciences, National Research Centre, Tahrir Street, 12311 Cairo, Egypt


The RNA polymerase NS5B of Hepatitis C virus (HCV) is a well-characterised drug target with an active site and four allosteric binding sites. This work presents a workflow for virtual screening and its application to Drug Bank screening targeting the Hepatitis C Virus (HCV) RNA polymerase non-nucleoside binding sites. Potential poly-pharmacological drugs are sought with predicted active inhibition on viral replication, and with proven positive pharmaco-clinical profiles.The approach adopted was receptor-based. Docking screens, guided with contact pharmacophores and neural-network activity prediction models on all allosteric binding sites and MD simulations, constituted our analysis workflow for identification of potential hits. Steps included: 1) using a two-phase docking screen with Surflex and Glide Xp . 2) Ranking based on scores,and important H interactions. 3)  a machine-learning target-trained artificial neural network PIC prediction model used for ranking. This provided a better correlation of IC50 values of the training sets for each site with different docking scores and sub-scores.4) interaction pharmacophores–through retrospective analysis of protein-inhibitor complex X-ray structures for the interaction pharmacophore  (common interaction modes) of inhibitors for the five non-nucleoside binding sites were constructed. These were used for filtering the hits according to the critical binding feature of formerly reported inhibitors. This filtration process resulted in identification of potential new inhibitors as well as formerly reported ones forthe thumb II and Palm I sites(HCV-81) NS5B binding sites. Eventually molecular dynamics simulations were carried out, confirming the binding hypothesis and resulting in 4 hits.


Molecular modelling, docking screens, pharmacophore modelling, target-trained models, molecular dynamics, HCV NS5B polymerase inhibitors, polypharmacology, DrugBank, Interaction pharmacophore, target-trained models.


It takes too long and costs too much to develop a new drug. Therefore, drug repositioning efforts are gathering more attention (i.e.: to screen available drugs for new uses). Currently, fifty plus drugs have been repositioned ( ). Off-label uses of drugs are widespread and legal in the USA. Also, multi-targeting compounds have been used in various diseases (e.g: Receptor-Thyrasine Kinase inhibitors for various cancers such as GleeVec and Nexavir).

This study presents a workflow for virtual screening and its application to Drug Bank screening targeting the Hepatitis C Virus (HCV) RNA polymerase non-nucleoside binding sites. Potential poly-pharmacological drugs are sought with predicted active inhibition on viral replication.   Hepatitis C virus (HCV) infects over 3% of the world population and is one of the leading causes of chronic liver diseases. About 80% of HCV-infected patients develop chronic hepatitis, 20% progress to cirrhosis and eventually develop Hepatocellular carcinoma . Currently there is no vaccine available for HCV . Current standard care of treatment for chronic hepatitis C is based on the combination of subcutaneous pegylatedinterferon-ɑ and oral nucleoside drug ribavirin. However, serious side effects and poor response rates render the development of novel anti-HCV therapy an urgent need . Several clinical trials are currently progressing for specifically targeted antiviral therapies (STAT-C) inhibitors that target specific protein pockets to inhibit HCV functions, while additional trials proceed on compounds which target host cell proteins that the virus utilizes for its survival/replication .

Currently, different targets for therapeutic intervention include structural as well as non-structural proteins and RNA structures in addition to post-transcriptional silencing. Non-structural targets include the NS3 Protease Covalent and non-covalent Inhibitors, NS3-NS4A Complex Inhibitors, NS3 Helicase Inhibitors, NS4B Inhibitors, NS5A Inhibitors, Nucleoside inhibitors and NS5B Polymerase Non-Nucleoside Inhibitors. The RNA-dependent RNA polymerase NS5B in particular has been subject of intense research in the past decade because of its essential role in viral replication, its distinct features as compared to human enzymes, and ultimately due to its highly druggable nature.

Although NS5B has the right-handed fingers, thumb and palm domains typical of polymerases, extensions of the fingers and thumb lead to a more fully-enclosed active site  (Figure 1). The inhibitors of HCV NS5B polymerase consist of two main classes: nucleoside inhibitors (NI) and non-nucleoside inhibitors (NNI). The NIs bind to the active site of the polymerase such as GS-7997, RGB7128, TMC649128 , PSI-7977 and PSI-938 .They currently offer the best candidates for cross-genotypic coverage and low resistant mutants. NNIs are a structurally and chemically heterogeneous class and do not induce premature termination of the RNAs synthesis. Moreover, NNIs are almost invariably allosteric inhibitors believed to block the enzyme, preventing a conformational transition needed for initiation of RNA synthesis ; the fact that assumed a solvated, and essentially flexible receptor . These NNI classes bind to one of the four allosteric binding sites within the NS5B polymerase  ( Figure 1) including: Site I (Thumb I) for JTK-109,benzimidazoles and Indoles , Site II (Thumb II) for  dihydropyrols, phenylalanine analogs and thiophenes (PF-868554, VCH-759, VCH-916 and VCH-222), Site III (Palm I) for chemically heterogeneous leads such as ANA-598, A-848837 and ABT-333, Site IV (Palm II) for benzofurans as HCV-796  and Site V (palm III) as phenylpropanylbenzamides. For details, refer to the methods and results sections below and Figures 1,2, 3 for a schematic of the NS5B polymerase and is important residues for each NNI site in addition to the minimum interaction pharmacophore for some major classes of NNI inhibitors.

Virtual Screening is the computational analogue of High Throughput Screening. It is defined as the in silico evaluation of properties, such as activity , or physiochemical properties like drug-likeness of different molecular scaffolds. Different applications of machine learning to virtual screening have been presented in the literature including both ligand-based similarity searching and structure-based docking. The main purpose of such applications is to prioritize databases of molecules as active against a particular protein target. In silico approaches such as virtual screening and structure-based design have emerged as a reliable, cost and time-saving technique complementary to in vitro screening for the discovery and optimization of leads and hits. VS can be divided into ligand-based, structure-based and mixed approaches such as the approach implemented here ( Figure 2). Activity-prediction /ranking models could be based on the set of ligands only which would be a purely ligand-based approach. Or, it could be based on the 3d structure of the ligand-receptor complex (interaction pharmacophore) (such as those shown in Figure3). The same holds true for screening models which could be based on the ligand pharmacophore / 3D quantitative structure-activity relationship ( 3D QSAR); or it could be based on the score of binding to the receptor docking-based screen (as employed here).

Attempts to perform focused screening on specialized databases have been implemented before. These databases include some whose compounds have acceptable pharmacokinetic/pharmacodynamic properties (absorption, distribution, metabolism, excretion and toxicology) (ADMET) properties. Virtual screening of physiochemical properties as a first filter before activity-based screening has also been highly recommended in virtual screening protocols. It has been highlighted as a means of preserving time, and money. This has been triggered after the incompletion of a high percent of drug discovery projects with good activity due to problems with ADMET properties. Several studies have since indicated the importance of such prefilters and considerations of ADMET properties from the beginning.

Here, a structure-based docking approach is used to find promiscuous drugs/compounds from the drug bank that could target the HCV polymerase allosteric sites. This shortcut approach has yielded candidate hits that can immediately enter into clinical trials for dosage determination with minimal cost, pharmacokinetic, and toxicological profiling studies, which could offer a potential of improving treatment outcome with HCV chronic patients .

Both Ligand-based and receptor-based drug design approaches have been heavily implemented in finding new candidate inhibitors for HCV polymerase .Yet to date, a comprehensive docking-based virtual screening of the Drug Bank for finding novel multivalent compounds has not been performed, although several studies reported the use of high throughput docking for lead identification and optimization . Furthermore, the combinations of docking tools that are based on higher accuracy scoring functions such as the XP (extra precision) and constrained docking approach in Glide were used to filter off potential false positives from the initial screening.This was followed by molecular dynamics-based investigation of the binding profiles of the resulting hits as detailed below.

Validation of the docking and selection was performed in multiple steps. These included reproducing the original interactions of the reference enzyme-ligand complexes, comparing the root-mean square distance of the experimentally determined pose with the docked pose, and correlation of the enzymologic inhibition concentration (IC) 50/90  with the docking scores and sub scores.These validations for choosing the higher accuracy score for filtration were performed on datasets of known NNI binding sites inhibitors obtained from the Binding DB  in order to use the highest correlating score in the filtration of initial hits. Extending this idea further, here, we describe aneural-network artificial intelligence model that was constructed to provide better correlations of docking scores with experimental data through a target-trained model. The model is based on a multitude of scores and sub-scores from different scoring functions for each binding site. These are combined non-linearly via an artificial neural network classifier, that was used here for ranking the hits obtained from first-phase docking with Surflex. It could be later used after some statistical validations for screening massive compound databases. These multiple validation approaches were necessary in order to build confidence in the final predicted  compounds to have novel inhibitory potential against the HCV polymerase. We are currently working on the experimental validation of these hits and extending this protocol.

Thus, through this work, novel inhibitors for the RdRp of HCV are sought. Combinations of ligand-based, receptor-based and incorporation of machine-learning classifiers were introduced along with molecular dynamics experiments to investigate the prospective inhibitors. Also, a meta-retrospective analysis to generate common contact pharmacophores that represent features required for efficient binding to NNI-sites for the HCV RdRp was performed by collecting all PDB files for each site, and finding common physical interaction moieties that are shared across all inhibitors of the same class that target that site.

For justification of our receptor-based approach, a ligand-based pharmacophore was built on a promising lead that is currently in clinical trials, Filibuvir (PF-00868554), targeting the thumb II site. This was used to screen different chemical databases with a few hits retrieved that need more activity and ADMET profile characterisation. These resulting hits were then short-listed using the docking approach. In the structure-based screening, a refined docking, ranking, and validation approach that employed machine-learning classification during the ranking process was performed for all sites of the RdRp on the drug bank database. The structure-based approach relied initially on virtual high-throughput docking of the drug bank on the four allosteric sites yielding tens of potential hits. This was followed by a second stage of more rigorous docking for the top candidates resulting from the former stage. Between them, ranking using the ANN model was applied. Also, validation using IC correlation for first-stage and rmsd for second-stage was done. Hit binding analysis selecting top poses and use of the interaction pharmacophores generated for each site followed. Further testing through molecular dynamics simulations culminated in potential hits acting on the palm I and thumb II were concluded that scored higher than threshold reference drugs, hadlow predicted IC values, and stable binding poses with molecular dynamics.

We have generated the common trends of receptor-ligand interactions pharmacophores, calculated from 37 PDB coordinates compiled according to the corresponding sites. Categorizing the palm region into three sub-sites, it seemed that hydrogen bonding with the backbone amide of TYR448 and hydrophobic interaction with the deep pocket were found to be a common binding feature among all chemical subclasses of palm I inhibitors, also polar interaction with ASN316 was common between palm II and palm III inhibitors. While SER476, TYR477 and ARG501 polar contacts constituted the major features for thumb II inhibitors, ARG503 formed the only common polar interaction for thumb I inhibitors. These “essential” interactions that have been defined here constitute a method of selection that could be used in various virtual screening exercises on the NS5B protein. This concept of polypharmacology could be utilised for various new drug discovery endeavours.  Reuse of already available safe compounds should thus shorten and lower the expenses of the long drug discovery cycle.

The novel idea of basing a machine-learning activity prediction model on interaction simulation scores was used here for ranking the retrieved hits from first-stage docking with  Surflex. It could be further improved as a screening tool, or including essential pharmacophoric interaction data as features in the future.

Conflict of interest

Authors report no conflict of interest.


We are thankful for funding support through the STDF program from Ministry of scientific research and technology, grant # 1169, Cairo, Egypt.



Ahmed Yahia-1

Fig.1 A diagram showing the 3D structure of NS5B, showing key interaction residues (polar interactions; blue, hydrophobic; green). The three different domains are shown (thumb; blue, palm; yellow, fingers; red).

Ahmed Yahia-2

Fig.2 A workflow describing the steps taken in both ligand-based and structure-based approaches to find novel inhibitors for HCV polymerase NS5B. A) the ligand-based search consisted of pharmacophore generation and screening, followed by docking and selection.B)  Structure-based NNI work flow consists of identifying the target binding sites and their interaction pharmacophore, a two-stage docking screen, combined with a neural-network ranking model for the hits, and finally, molecular dynamics simulations for the promising hits.

Ahmed Yahia-3

Fig.3 Interaction pharmacophore for the Palm I subdomain.  Overlay of PDB coordinates of the three major chemical classes of NS5B palm I inhibitors A) class I (Benzothiazoles), B) class II (Benzothiadiazines), C) class III (Benzodiazepines)). D) Overlaid complexes at the thumb II site (green dotted lines show polar interactions, different coordinates are colored differently, partial receptor surface is colored according to the interpolated charge–showing the whitish neutral regions)(polar hydrogens displayed in section C) all classes for Palm I shown in A ,B, and C fill the deep hydrophobic pocket and shared TYR448 as a backbone HB acceptor. A) ASP318 also had polar interactions at the backbone. SER556 and ASN291 had hydrogen bond (HB) interactions through the terminal hydroxyl and amide groups with all members of class I.B) ASP318, GLY449 and ASN 291 were also involved in the same manner in addition to the terminal polar groups of SER 556 and CYS 366.C) shows  that replacing the ketone on the hexene ring by a sulphone group expands the hydrogen bonding from TYR448 to TYR448 and GLY449, additionally, SER 367 HB seemed a common feature and to a lesser extent SER368. D) The main residues forming common polar interactions were SER476, TYR477 –as backbone HB acceptors and ARG 501 through the guanidinium group. A well defined π stacking was noticed between the histidine’s imidazole ring and the filibuvir’s (3FRZ ligand) triazolo pyrimidine group

Ahmed Yahia-4

Fig.4 A) Overlay of docked and PDB coordinates of 2JC1 ligand for the Palm I  site (original ligand pose coloured blue, docked ligand pose coloured pink); B) docked pose of DB05039 it shows a very good placement of the diethyl-indanyl group into the deep pocket. A salt bridge strengthens the binding with ARG394. Similarly,  hydrogen bonding with ARG386 and TYR415, and Pi stacking with TYR415 ; C) docked pose of DB01940 forming a HB inside the deep hydrophobic pocket (as with one of benzodiazepines) in addition to a HB with GLN446 (just four atoms away along the backbone’s TYR448 amidic nitrogen) ; D) docked pose of DB04142.(red dotted lines represent salt bridges, green dotted lines represent hydrogen bonds, orange solid lines represent π-interactions, backbone shown as curved purple lines, protein transparent surface with interpolated charges; bluish(+ve), reddish(-ve)). The furan ring show similar interaction with  TYR448 . These three hits preserved these binding modalities after molecular dynamics simulations.

Multiselect Ultimate Query Plugin by InoPlugs Web Design Vienna | Webdesign Wien and Juwelier SchönmannMultiselect Ultimate Query Plugin by InoPlugs Web Design Vienna | Webdesign Wien and Juwelier Schönmann