DAKB-GPCRs: An Integrated Computational Platform for Drug Abuse Related GPCRs

Abstract

Drug abuse (DA) or drug addiction is a complicated brain disorder which is commonly considered as neurobiological impairments caused by both genetic factors and environmental effects. Among DA-related targets, G protein-coupled receptors (GPCRs) play an important role in DA therapy. However, only 52 GPCRs have been published with crystal structures in the recent two decades. In the effort to overcome the limitations of crystal structure and conformational diversity of GPCRs, we built homology models and performed conformational searches by molecular dynamics (MD) simulation. To accelerate and facilitate the drug abuse research, we construct a DA-related GPCR-specific chemogenomics knowledgebase (KB) (DAKB-GPCRs) for its research that can be implemented with our established and novel chemogenomics tools as well as algorithms for data analysis and visualization. Our established TargetHunter and HTDocking tools, as well as our novel tools that include target classification and Spider Plot, are compiled into the platform. Our DAKB-GPCRs provides the following results for a query compound: (1) blood−brain barrier (BBB) plot via our BBB predictor, (2) docking scores via HTDocking, (3) similarity score via TargetHunter, (4) target classification via machine learning methods that utilize both docking scores and similarity scores, and (5) a drug−target interaction network via Spider Plot.

Graphical Abstract

INTRODUCTION

Drug abuse (DA) is a complicated neurological disorder. The main symptoms of DA and addiction involve compulsive behaviors such as drug craving, seeking, persistent use, and relapses despite serious adverse consequences. In 2017, there were approximate 30.5 million people of age 12 or older in the U.S. who had used illicit drugs or abused psychotherapeutic medications (e.g., pain relievers, stimulants, or tranquilizers), representing 11.2% of the population.¹ According to Drug Facts,² the total overall costs of substance abuse in the U.S. exceed $740 billion annually, including approximately $193 billion for illicit drugs,^3,4 $300 billion for tobacco,^5,6 and $249 billion for alcohol⁷ (Figure S1). Thus, research for DA prevention and treatment is a high priority.

DA is a chronic disease with strong genetic and environ-mental influences that can often lead to patient relapse. The basic mechanisms of DA are well accepted that abused substances can disrupt activity of normal nerves cells by interacting with receptors in the brain and activating the downstream signaling pathways, which comprise a host of kinases and transcription factors. Specifically, there is a broad range of G-protein coupled receptors (GPCRs) related to the DA and central nervous system (CNS) side effects, including opioid receptors, cannabinoid receptors, and serotonin receptors.⁸ Further, it is estimated that genetic factors play a significant role, approximately 40−60% of the total risk.^8,9

To accelerate and facilitate the drug abuse research, we constructed a self-service online library named GPCR-specific chemogenomics knowledgebase (KB) for DA research (DAKB-GPCRs) that contains information about drug abuse-related protein targets, small molecules, as well as tools and algorithms for computational data analyses and visualization on those data. In our established drug DA-KB^8,10 regarding protein targets of abused drugs, we found that 86 out of 258 proteins (33.3%) are GPCRs including serotonin receptors (5HT 1A-5HT 7),¹¹ adenosine receptors (A1, A2a, A2b, and A3),¹² cannabinoid receptors (CB1 and CB2),¹³ acetylcholine receptors (M1-M5),¹⁴ C-X-C motif chemokine receptors (CXCR1−CXCR4), dopamine receptors (D1−D5),¹⁵ histamine receptors (H1−H4), melanocortin receptors (MC1−MC5),¹⁶ opioid receptors (Mu, Delta, and Kappa),¹⁷ glutamate receptors (metabotropic glutamate receptor 1−8),¹⁸ etc.⁸ More details can be found in our previous publication,⁸ our published DAKB website (https://www.cbligand.org/CDAR/) and Table S1. Moreover, in the past 18 years, only 52 GPCRs have been published with crystal structures or cryo-EM structures,¹⁹ in which only 29 GPCRs are related to DA (34% out of 86 GPCRs) (Figure S1). Recently, homology modeling with a sequence identity of 30% or greater,²⁰ and/or multiple conformations-based docking, is becoming a powerful tool and strategy for structural study and drug discovery.²¹

On the basis of our previously established molecular information databases⁸ and developed computing algorithms,²²⁻²⁴ we constructed a GPCR domain-specific drug abuse knowledgebase (DAKB-GPCRs, Figure 1). To achieve this, we have carried out a literature search to collect DA-related GPCR structures and corresponding drugs.⁸ Specifically, as discussed in the beginning, there were only a small percentage of GPCR crystal structures available, but we built the homology models for those GPCRs without crystal structure(s) based on published GPCR crystal structures for the use of molecular modeling study. To further increase the diversity of the protein conformation as well as optimize the accuracy of the homology models, molecular dynamics (MD) simulation was used to sample conformations of each GPCR, followed by computational prescreening validation for multiple conformation-based docking. Last but not least, we integrated established tools and algorithms (such as TargetHunter, HTDocking, NGL, ANN) and developed new tools (such as Spider Plot) for data visualization and analyses.

Figure 1.

Overall workflow of DAKB-GPCRs.

MATERIALS AND METHODS

Genes/Proteins.

Genes/proteins that related to drug abuse were collected from public databases such as Ensembl,²⁵ UniProt,²⁶ KEGG,²⁷ GPCRdb,²⁸ and the NCBI Protein Database.²⁹ Available crystal structures or cryo-EM structures of GPCRs were retrieved from the Protein Data Bank (PDB) (https://www.rcsb.org/).

Drugs and Chemicals.

The ChEMBL database (version 23)³⁰ was used in our work. The experimental data for each small molecule against its respective target proteins were collected using text mining techniques and cleaned up by manual inspection. Bioactivity data from different resources were normalized using the same standard. The number of small molecule agents for each GPCR is listed in Table S1. Especially, small molecules with IC₅₀ lower than 1 μM toward a GPCR target were regarded as the active compounds, while those larger than 10 μM were considered as inactive compounds. A training data set that consisted of both active and inactive compounds of each GPCR will be used for prescreening and similarity search using TargetHunter.³¹

Homology Models.

The sequences of the human GPCRs were collected from the UniProtKB/Swiss-Prot (https://www.uniprot.org/uniprot/) website. We will first truncate some residues from the N terminal and some residues from the C terminal. Moreover, the third intracellular loop (ICL3, between helix 5 and helix 6) has a long-flexible sequence, we only keep ∼15 residues. Then, Modeller 9.18³² was used to construct the homology models: (a) searching and selecting template(s) for the target protein, (b) conducting sequence alignment between target and template(s), (c) adjusting the sequence alignment using the residue tables from GPCRdb (http://gpcrdb.org/residue/residuetable), and (d) building and evaluating the homology models. After generating the 3D models of protein targets, SYBYL-X 1.3³³ was adopted to carry out the energy minimization. Then the best three models after energy minimization with the root-mean-square deviation (RMSD) of 2−3 Å was selected for further molecular dynamics simulation. This established protocol can be found in our previous publications.^34–36

Molecular Dynamics (MD) Simulation for Conformational Sampling.

A pioneer paper entitled “The Biological Functions of Low-Frequency Phonons”³⁷ and a series of following research works on biomacromolecules from dynamic point of view³⁸ suggest that low-frequency collective motions do exist in proteins and DNA.^39–42 By studying the low-frequency internal motions,⁴⁰ many important biological functions in proteins and DNA and their dynamic mechanisms, such as shifting between active and inactive states,⁴³ cooperative effects,⁴⁴ and allosteric transition,³⁸ can be revealed. Therefore, to further understand the action mechanisms of biomacromolecules, we should consider not only the static structural information (such as crystal structure or cryo-EM structure) but also the dynamical information acquired from molecular dynamic (MD) simulation.⁴⁵

In order to obtain the diverse conformations of each GPCR for multiple conformations-based docking, we carried out the MD simulations without ligand for either crystal structures or homology models. First, special caution was applied to His residues, because His was ionized at pH 7.40. VEGA ZZ 2.4.0⁴⁶ and PROPKA 3.1⁴⁷ were applied to predict the pK values of His and other residues. Sequentially, the VMD program was used to embed the receptor into a periodic and pre-equilibrated structure of 1-palmytoyl-2-oleoyl-sn-glycero-3-phosphatidylcholine (POPC). Lipid molecules within 3 Å of the receptor were eliminated. We then inserted these into a water box (TIP3P⁴⁸ water model, 67 Å × 67 Å × 74 Å) with eliminating the waters molecules within 3 Å of the receptor. Two minimizations were carried out. Each minimization was carried out in 50 000 steps. The first minimization was performed with the fixed protein, and the second minimization was carried out with the flexible protein. Then 1.0 ns of MD for heating from 0 to 300 K and equilibration was performed. Finally, the NAMD package (version 2.9b1)⁴⁹ using the CHARMM27⁵⁰ force field was applied to the MD simulation. The particle mesh Ewald (PME)⁵¹ method (with a 12 Å nonbonded cutoff and a grid spacing of 1 Å per grid point in each dimension) was used to calculate the electrostatics. A smooth cutoff (switching radius 10 Å, cutoff radius 12 Å) was used to calculate the van der Waals energies. The temperature and pressure were kept constant using a Langevin thermostat (310 K) and Langevin barostat (1 atm), respectively. The time step of the MD simulation was set to 1 fs. In order to select reasonable conformations of each homology model or crystal structure, we first performed 5 ns MD simulation for the protein with explicit water and lipid. During 5 ns MD simulation, we fixed the Cα atoms of seven transmembrane domains of the protein and made the side chains flexible. Then, we made the whole protein flexible and performed another 5 ns MD simulation. We selected ten conformations with the lowest energy and the RMSD over 2−3 Å to perform secondary energy minimization by SYBYL. This established protocol can be found in our previous publications.³⁴⁻³⁶

Prescreening.

Ten models of each GPCR that obtained after MD simulation were utilized to perform the prescreening against its training data set (see Drugs and Chemicals). The training data set of each GPCR included both the active and inactive compounds (Table S1). The docking program HTDocking (see below) was applied to constructed receptor−ligand complexes.

Database Infrastructure.

A query compound can be submitted using JSME Molecular Editor v2017–03-01.⁵² DAKB-GPCRs was implemented on the basis of our established molecular database prototype CBID (https://www.cbligand.org/cbid/) using the SQLite database management system (https://sqlite.org/) and Kestrel HTTP server (https://github.com/aspnet/KestrelHttpServer) with the Apache HTTP server (https://httpd.apache.org/) as its reverse proxy server. The overview of our design for DAKB-GPCRs is depicted in Figure 1.

HTDocking.

DAKB-GPCRs adopts our online high-throughput molecular docking technique—HTDocking,^22,23,53 for identifying possible interactions between protein targets and small molecules. There are three different conformations for each GPCRs that selected from MD sampling and validated by prescreening. For each query compound, HTDocking will automatically dock it into three different conformations and generate docking scores. A higher docking score indicates that the protein is more likely to be the candidate target of the queried small molecule. idock⁵⁴ can provide up to 9 predicted binding affinity values (ΔG values) from different docking poses for each compound in a binding pocket of a protein. In our HTDocking program, we only consider the best binding affinity value which is further transformed as a docking score. The docking score = −log₁₀(e^ΔG*^{4184/8.314/310.15}).

TargetHunter.

DAKB-GPCRs integrates our online target-identification service—TargetHunter,³¹ for predicting the potential off-targets for submitted compounds. TargetHunter exploits an important principle of medicinal chemistry: compounds with structural similarities often have similar physicochemical properties and biological profiles. For each query compound, TargetHunter calculates the similarity (from 0.0 to 1.0, totally different to 100% similar) with its known active compound’s data set that was collected from Drugs and Chemicals (see above).

Machine Learning-Based Target Classification.

For each GPCR, a data set consisting of three docking scores and one similarity score for each known compound were trained to build the target classification models using established machine learning (ML) algorithms. The compound data set collected for each target (as discussed in Drugs and Chemicals) was used for training and testing the classification models. The size of the data sets can be found in Table S1. Molecular docking scores and molecular fingerprint similarity scores were computed using the protocols discussed in HTDocking and TargetHunter, respectively. Four ML algorithms were adopted by our classification models: logistic regression,⁵⁵ support vector machine (SVM),⁵⁶ random forest (RF),⁵⁷ and artificial neural networks (ANN).⁵⁸ The training and the classification were implemented in Python using scikit-learn (https://scikit-learn.org/).

Spider Plot.

Based on the target classification, our online tool Spider Plot visualizes the molecule−protein interaction network. The average docking scores are displayed as connection labels and the protein targets on which the query compound is active are displayed as circular nodes. By default, a green node denotes a target with high similarity score (>0.7) comparing the best matched known compound and the query compound while a pink one denotes otherwise. With Spider Plot, the colors, font sizes, node sizes, border widths, and node shapes as well as the layout can be completely customized, and the entire network graph can be exported as an image file on particular browsers.

Blood−Brain Barrier (BBB) Predictor.

The BBB predictor^22,23 was integrated into DAKB-GPCRs. This predictor was built by applying the support vector machine (SVM) and LiCABEDS^59,60 algorithms on four types of fingerprints of 1593 reported compounds.⁶¹ It predicts whether or not a query compound can move across the BBB to the central nervous system (CNS).⁶² The BBB predictor is also available for access from https://www.cbligand.org/BBB/.

Software Requirements.

The DAKB-GPCRs website is compatible with modern web browsers (such as Chrome, Firefox, Microsoft Edge, and Safari) provided JavaScript and cookies are enabled. We recommend the latest release version of these web browsers for better rendering.

RESULTS AND DISCUSSION

Overview of Our Platform.

Our platform intrinsically works on a set of conformations of GPCR homology models/ crystal structures and their relevant compounds. The compounds serve as the training data set where both active and inactive compounds are included for docking and similarity calculation. The platform makes use of three conformations for each GPCR so as to obtain multiple docking scores for training the machine learning models. The predicted values of BBB penetration²² and the potential GPCR targets on which the submitted query compound is active are provided in the output. As shown in Figure 1, The platform first queues a task for the query compound and afterward computes three docking scores via HTDocking²² and a molecular similarity score via TargetHunter³¹ for each GPCR target. Subsequently a target classification is performed by our ensembled machine learning algorithms featuring docking scores and similarity score. The classification results are then passed to Spider Plot for visualization of the drug-targets interaction network.

Statistical Analysis of Multiple Conformations-Based Docking.

Ten models of each GPCR with the lowest energy and the RMSD over 2−3 Å (compared to the original homology model/crystal structure) obtained after MD simulation were utilized to perform the prescreening against its training data set (see Drugs and Chemicals). Three conformations of each GPCR with the best ROC curve were selected and integrated into our platform. Here, we take adrenoceptor alpha 1d (ADRA1D) as an example, which was constructed using beta-2 adrenergic receptor (PDB ID 3SN6, sequence identity between helixes 39%), beta-1 adrenergic receptor (PDB ID 2Y00, sequence identity between helixes 39%) and dopamine 3 receptor (PDB ID 3PBL, sequence identity between helixes 40%). Figure S2 shows the curves of its best three conformations. Figure S2A shows the statistical results of Model 1 of ADRA1D: the docking score of 6.15 was chosen as the best threshold, because the docking scores of 71% (1−0.29) inactive compounds were lower than 6.15, while the docking scores of 78% (0.78) active compounds were higher than 6.15. The threshold of docking score of other two models of ADRA1D were 6.59 (Figure S2B) and 6.71 (Figure S2C), respectively. We can find that the true positive rate (TPR) and false positive rate (FPR) of model 2 are 71% and 63%, while those of model 3 are 79% and 63%. The similar results can be found for other GPCRs in our platform, indicating the protocol of multiple conformations-based docking is reasonable.

Evaluation of the Machine Learning Classification Models.

A 10-fold cross validation was used to assess the predictive capability of all the models. The performance and evaluation for those classifications are listed in Figure S3. The performance of the ANN models was not robust in this case; thus, those models were eliminated. Therefore, for each protein target, the final prediction was determined by the classifications from the three selected models and their confidence levels.

Case Study.

To elaborate how the DAKB-GPCRs website can be used to facilitate research, we use a published compound (CHEMBL1779871, a positive allosteric modulator of human mGluR5) to showcase the functionalities step by step.

Home Page.

The DAKB-GPCRs can be accessed from https://www.cbligand.org/dakb-gpcrs/. On the top of the home page (Figure S4A) is the navigation bar which contains the HOME button, the ALL TASKS button and the HELP button. Clicking on the HOME button brings the user back to the home page while the ALL TASKS button directs the user to the task list page. The HELP button navigates the user to the User Guide page which provides a step-by-step guide for using DAKB-GPCRs.

Task List Page.

The task list page displays all the user-submitted tasks in a table layout (Figure S4B). Completed tasks are shown as Finished in the status column, and ongoing tasks are shown as Running. The user can click on the task name to access the detailed information on the input compound as well as the output of the computation task. The user can initiate a new task by clicking on the Create a new task button on the upper right corner of the page.

Start a New Task Using Structure Query.

The user can submit a query compound either by drawing its 2D structure with JSME Molecular Editor or by uploading a chemical file with the upload button.⁶³ A meaningful name for the task is also required. Clicking on the Create Task button will initiate a new task. Afterward, a background job worker which periodically monitors the task queue will start to allocate computation resources and dispatch the computation task. In Figure S4C, the structure of CHEMBL1779871 in SMILES format was uploaded.

Detailed Task Page.

After submitting the request, the website will automatically direct users to the detailed task page (Figure S4D). On this page, you can see the task information as well as the progress bar on the top showing the real time progress of computation for target prediction. Following the task information is the section of ligand information which includes 2D/3D structures, multiple structure files of the query compound, and various computed molecular fingerprints.⁶⁴

BBB Page.

The visualized results from the BBB predictor can be found on the BBB page (Figure S4E).

Output Page for Target Prediction.

The output page can be accessed by clicking on the See Detailed Output button from the detailed task page. On the output page, each block presents three docking scores by HTDocking computed against three protein target models and a similarity score by TargetHunter (Figure S4F and S4G). The classification result is presented suggesting whether the query compound is an active or inactive compound for this protein target, and the confidence level of the prediction is also provided. When clicking on the download button, the user can download the docking results of the submitted compound. The most similar compound within the compound library of a protein target can be seen from the Best Match field. Its comparison with the query compound is shown in a popup window (Figure S4H) after clicking the compound ID. When clicking on the target name, a window that consists of additional resource for the protein target (Figure S4I) will be shown, which includes a 3D interactive visualization,^65,66 multiple links to other database websites, and models for direct downloading.

The SPIDER PLOT button on the top of the page leads user to the drug−target network plotting tool for data visualization and analysis (Figure S4J). The most promising predictions (with similarity score > 0.70) for the query compound are shown as green nodes connected with a solid line while other predictions are shown as purple nodes connected with a dashed line. The name and structure of the query compound are placed in the center of the plot and the average docking scores are displayed as connection labels. By clicking on a node, a toolbox will appear for users to customize the colors, font sizes, node sizes, and border widths. Also, the entire network graph can be exported as an image file on particular browsers. As illustrated in Figure S4J, GRM5 was predicted as the most promising target for CHEMBL1779871, which was consistent with the bioactivity data where CHEMBL1779871 is an active compound for GRM5 (EC₅₀ = 7.5 nM).

Prediction Accuracy.

As the statistics show in Table S1, the overall accuracy of our machine learning based prediction was 90% on the testing set containing 178 812 compounds collected in the method discussed in the Drugs and Chemicals section. As a comparison, the overall accuracy using just docking scores with an optimized cutoff value on the same testing set was ranged between 61% and 62%, while the overall accuracy using just similarity score was 93%. In the final output of DAKB-GPCRs all the four scores and the machine learning prediction are provided for the users’ reference.

Future Perspective.

In the past decade, more and more prediction methods and computational tools were developed and released as publicly accessible websites with user-friendly interfaces.^67–70 Many of which are practically useful and have increasing impacts on medical science.⁷¹ We are aware of the importance of user experience and as such in our future work, to further improve the user-friendliness, we shall make efforts to add more flexibility for users to manipulate the findings according to their needs and more customizations to our visualizing tool Spider Plot.

CONCLUSIONS

To facilitate the DA research, we constructed an integrated online computing platform for drug abuse researchers and scientists who may have no experience in programming, statistics or in silico drug design. This is an optimal approach for tailoring medicinal chemistry drug design and customizing combination therapy in a precision systems pharmacology manner. To our knowledge, no such domain-specific database is available for the proposed computational applications. Our platform is the first web-based service that integrates DA-related genes, proteins, and drugs for DA research. State-of-the-art computational chemistry/chemoinformatics and machine learning algorithms established in our lab have been implemented for this chemogenomics database, which will help characterize the features of genes, proteins, and drugs in DA study. It will also facilitate new information exchange and data sharing of knowledge among relevant scientific communities.