The virtual screening problems associated with acetylcholinesterase (AChE) inhibitors were explored using multiple shape, and structure-based modeling strategies. The employed strategies include molecular docking, similarity search, and pharmacophore modeling. A subset from directory of useful decoys (DUD) related to AChE inhibitors was considered, which consists of 105 known inhibitors and 3732 decoys. Statistical quality of the models was evaluated by enrichment factor (EF) metrics and receiver operating curve (ROC) analysis. The results revealed that electrostatic similarity search protocol using EON (ET_combo) outperformed all other protocols with outstanding enrichment of
Acetylcholinesterase (AChE; EC 3.1.1.7) terminates signaling at cholinergic synapses by rapid hydrolysis of the neurotransmitter acetylcholine [
AChE inhibitors are chemically diverse; the active site of AChE is multifaceted and complex in architecture allowing numerous structurally diverse ligands to bind to different subsites [
In this study, we have explored both ligand-based and structure based approaches for virtual screening of AChE. Ligand based approaches such as similarity search and pharmacophore mapping were used whereas molecular docking was used as a structure based approach. The following virtual screening tools were used for this study: (a) molecular docking using AutoDock and Glide [
Known ligands and decoys set for AChE as reported in the directory of useful decoys (DUD) [
PDB code 1b41 was downloaded (
For docking studies all the ligands were energy minimized in the Macromodel minimization panel using the OPLS-2005 force field and GB/SA water model with a constant dielectric of 1.0. Polak-Ribiere first derivative, conjugate gradient minimization was employed with maximum iterations of 1000 and convergence threshold of a gradient to <0.05 kJ/Å-mol. LigPrep2.0 module of Schrödinger was used to generate possible ionization states at target pH
All the docking experiments were performed with AutoDock4.0 and Glide. A grid size of 110 × 110 × 110 centered on the ligand was used. For Auto Dock, Lamarckian Genetic Algorithm was employed as the docking algorithm. For making the virtual screening protocol automated a separate script was written and validated [
ROCS shape-based virtual screening: Multiconformer files, which were generated by OMEGA, were saved in oeb.gz format. These generated multi-conformational files were used as input database for performing Rapid Overlay of Chemical Structures (ROCS) [
EON electrostatic similarity-based virtual screening: ROCS output structures (oeb.gz) were used as input for analysis using EON. EON [
PHASE-Shape based similarity: similar to the ROCS program PHASE-Shape program [
PHASE version 3.0 was used for pharmacophore elucidation. For this dataset, we performed the PHASE procedure with six built-in types of pharmacophoric features: hydrogen bond acceptor (A), hydrogen bond donor (D), hydrophobe (H), negative ionizable (N), positive ionizable (P), and aromatic ring (R). The graphical user interface of maestro was used. Ligands were processed with the LigPrep program to assign protonation states appropriate for pH 7.0.
Conformer generation was carried out with the Macromodel. Potentials were computed using the OPLS2005 force field. The default pharmacophore feature definitions were used in site generation. After the sites were generated hypotheses were generated by a systematic variation of the number of sites. The number of matching active compounds was kept default, that is, entire training set. The process started with five sites, but the set of five-point hypothesis did not survive the scoring process. Gradually the number of sites was reduced to four. Ten hypotheses were generated by the program, out of which only two survived. The scoring was done using the default parameters. The top hypotheses were then used to build pharmacophore-based model. All the molecules were considered active; no thresholds (active or inactive) were applied to the training set for developing the hypothesis. Before considering the model for virtual screening, the hypothesis was overlayed on the enzyme active site and then visually analyzed. After this overlay study, the generated hypothesis was then used to screen the entire DUD dataset. Receptor excluded volume was subsequently considered.
The virtual screening protocols were validated by their enrichment factors [
ROC curve analysis is considered as one of the best approaches for the performance characterization of virtual screening protocols so far. The ROC is represented equivalently by plotting the fraction of true positives (TPR: true positive rate) versus the fraction of false positives (FPR: false positive rate).
The active site of the enzyme is primarily hydrophobic, subdivided into several subsites: esteratic subsite also called the catalytic triad, acyl binding pocket, peripheral anionic subsite, and ligand recognition site (Figure
Active site of AChE. (a) Surface view of the active site gorge and donepezil in the active site of PDB ID 1eve. Surface view generated using AutoDock 4.0. (b) Active site of the different PDBs was aligned. Depicting the amino acid residues of different subsites of the active site. Different ligand engages different water molecules. PDBs: 1eve (donepezil): grey; 1gpk (huperzine): blue; 1odc (tacrine): red; 1gqr (rivastigmine): green; and 1ax9 (edrophonium): yellow. Representation: amino acid residues in wire, ligands in tube, and water molecules in ball and stick.
Thus, it is not practically feasible to choose intrinsic water molecule(s) for structure based virtual screening studies. Thus, the complexity associated with active site of AChE needs a valid model and working algorithms to overcome the virtual screening difficulty. Therefore, exploration of significant search protocol that reliably identifies the subsite where a ligand is most likely to bind is warranted.
When screening a large dataset, a reasonable model encompassing the entire chemical space is essential. If a model of this category is developed, it can easily serve as query tool for screening any dataset. The virtual screening problem for AChE inhibitors was addressed. We have compared different VS tools and discussed the advantages and disadvantages of them. The molecular docking approach was considered for pose analysis and scoring; in addition ligand based strategies were employed for pharmacophore mapping and similarity search.
The active site of AChE has a large volume and is divided into many subsites. Therefore, we have used a larger grid (110 × 110 × 110 grid units) that encompasses the entire active site (for both protocols AutoDock and Glide). The results of structure based virtual screening are as follows: for Glide-XP enrichment of 19%, 20% and 24% was observed for top 1%, 2%, and 5% of the dataset, respectively. Whereas, for AutoDock the enrichment was 30%, 34%, and 32%, respectively, for top 1%, 2%, and 5% of the dataset (Figures
Enrichment factors (EF) calculated at different stages (percent of database screened) of virtual screening. Shape based screening methods such as ROCS and EON outperformed all the other protocols employed.
(a) ROC curves for the various protocols employed in the current study. (b) Area under the curve (AUC) values of different protocols used for the virtual screening of AChE. These analyses clearly accentuate the excellent performance of ROCS and EON.
During our redocking experiment, we observed that increasing the grid size consistently increases the RMSD for both AutoDock and Glide docking.
Taking the example of donepezil, when we redocked with a 60 × 60 × 60 grid size, it yielded an outstanding RMSD of 0.69 Å. Whereas when the same molecule is redocked with 110 × 110 × 110 grid size, it yielded a very poor RMSD of 4.34 Å. The redocked poses are shown in Figure
Redocking experiment of donepezil using pdb 1b41. (a) When redocked with a 60 × 60 × 60 grid size, RMSD = 0.69 Å. (b) When redocked with 110 × 110 × 110 grid, RMSD = 4.34 Å.
We have also studied the effect of docking studies on known mutant of human AChE E202Q. We have used the known ligands and cross-docked them with the mutant form of AChE. As discussed by Krygar and coworkers [
Aligned structures of the wild type human AChE and E202Q mutant AChE. (a) With donepezil (b) With rivastigmine.
GLU202 is near to catalytic triad; the presence of this anionic residue is important in molecular recognition for rivastigmine or ligands specific for binding to this triad. Rivastigmine with the protonated nitrogen is naturally attracted towards the anionic site. However, due to mutation in GLU202, we observed a significantly lower affinity for rivastigmine. The affinity of other ligands was unaffected. However, separate studies considering the protein flexibility and intrinsic water molecule(s) are warranted.
The pharmacophore model was created with the same queries that were used for shape based screening. The model was obtained through an automated mode in PHASE. Before choosing the final model, it was visually inspected through superimposition study on the enzyme active site along with ligand pharmacophore map. Top scoring, four-point AHHR hypothesis was considered further. This generated hypothesis was overlaid on the enzyme active site. The pharmacophoric features were compared with crystallographic information; these interactions were compared and substantiated (Figure
Chosen pharmacophoric hypothesis. (a) Overlaid on the binding site of donepezil bound to PDB ID 1eve. (b) The pharmacophore model depicting the inter-pharmacophoric distances. Ligand structure aligned on the pharmacophore model is donepezil. Hydrogen bond acceptor (A), hydrophobe (H), and aromatic ring (R).
A1 (H-bond acceptor): the ligands should have an acceptor group that accepts a hydrogen bond from protonated W279 and H3 and H5 (Hydrophobes); H3 seems to be accommodated in the hydrophobic pocket created by the side chains of W279 and Q74; H5 occupies the hydrophobic pocket created by the four aromatic amino acids F288, 330, 331 and Y334; R9 (ring aromatic feature) is stacked with W84. The interpharmacophoric distances are presented in Table
Inter-pharmacophoric distances for the newly developed pharmacophore model.
Site 1 | Site 2 | Distance |
---|---|---|
A1 | H3 | 3.133 |
A1 | H5 | 9.420 |
A1 | R9 | 14.103 |
H3 | H5 | 9.626 |
H3 | R9 | 13.585 |
H5 | R9 | 5.757 |
Three computational strategies were employed: PHASE shape-based similarity, ROCS shape based similarity search, and EON electrostatic search. An outstanding enrichment factor (EF) of >90% for analysis by a shape based screening using ROCS was observed in all the cases (1%, 2%, and 5%). However, rescoring of this dataset by EON (ET_combo) improved the results to the best possible outcome. It outperformed all other protocols with outstanding enrichment of >95% in top 1% and 2% of the dataset, with an AUC of 0.958. ROCS (AUC = 0.944) and PHASE-Shape based (AUC = 0.898) protocol performed well, but not as well as EON (ET_combo) (Figures
The shape based screening protocol adopted in the current study also shows the importance of using multiple queries (Figure
Percentage of different ligands in the top 1% of the dataset during the virtual screening process. (a) Using ROCS. (b) Using EON. It clearly shows the importance of using multiple queries. Using a single query during virtual screening may omit the chemical space encompassed by the others.
We have used five known ligands donepezil, edrophonium, huperzine, rivastigmine, and tacrine. Selectivity towards AChE and BuChE is presented in Table
Comparison of AChE and BuChE selectivity of the selected ligands for similarity search analysis.
S. No. | Compound | AChE |
BuChE |
---|---|---|---|
(1) | Donepezil | 2.9 | 640 |
(2) | Huperzine | 0.026 | 120 |
(3) | Tacrine | 7 | 6.9 |
(4) | Edrophonium | 1.6 | 340000 |
(5) | Rivastigmine | 37 | 37 |
Donepezil identified around 60% of the HITs through similarity search. It is a selective AChE inhibitor, so the identified HITs are more likely to be AChE selective inhibitors. On the other hand, tacrine and rivastigmine together identified around 20% of the HITs. These identified HITs are more likely to be nonselective. However, there is no way to prove this hypothesis until we perform biological screening.
After an exhaustive validation of various virtual screening methodologies, we have found that shape based search using multiple queries performed better as compared to standard docking studies. The study provided a plausible solution to the virtual screening problem on AChE. We found shape based similarity search methods (ROCS and EON) performed significantly better than structure based methods. The study also revealed the importance of using multiple queries for exploring a larger chemical space and encompassing the majority of the pharmacophoric features.
This work was financially supported by University Grants Commission, India. The authors thank all the developers of freely available software, which greatly facilitated their work (cited in Material and Methods section). The authors are grateful to Dr. Ashoke Sharon for his contributions in the paper preparation.