Computational prediction of potential inhibitors for SARS-COV-2 main protease based on machine learning, docking, MM-PBSA calculations, and metadynamics

Isabela de Souza Gomes; Charles Abreu Santana; Leandro Soriano Marcolino; Leonardo Henrique França de Lima; Raquel Cardoso de Melo-Minardi; Roberto Sousa Dias; Sérgio Oliveira de Paula; Sabrina de Azevedo Silveira

doi:10.1371/journal.pone.0267471

Abstract

The development of new drugs is a very complex and time-consuming process, and for this reason, researchers have been resorting heavily to drug repurposing techniques as an alternative for the treatment of various diseases. This approach is especially interesting when it comes to emerging diseases with high rates of infection, because the lack of a quickly cure brings many human losses until the mitigation of the epidemic, as is the case of COVID-19. In this work, we combine an in-house developed machine learning strategy with docking, MM-PBSA calculations, and metadynamics to detect potential inhibitors for SARS-COV-2 main protease among FDA approved compounds. To assess the ability of our machine learning strategy to retrieve potential compounds we calculated the Enrichment Factor of compound datasets for three well known protein targets: HIV-1 reverse transcriptase (PDB 4B3P), 5-HT2A serotonin receptor (PDB 6A94), and H1 histamine receptor (PDB 3RZE). The Enrichment Factor for each target was, respectively, 102.5, 12.4, 10.6, which are considered significant values. Regarding the identification of molecules that can potentially inhibit the main protease of SARS-COV-2, compounds output by the machine learning step went through a docking experiment against SARS-COV-2 M^pro. The best scored poses were the input for MM-PBSA calculations and metadynamics using CHARMM and AMBER force fields to predict the binding energy for each complex. Our work points out six molecules, highlighting the strong interaction obtained for M^pro-mirabegron complex. Among these six, to the best of our knowledge, ambenonium has not yet been described in the literature as a candidate inhibitor for the SARS-COV-2 main protease in its active pocket.

Citation: Gomes IdS, Santana CA, Marcolino LS, Lima LHFd, Melo-Minardi RCd, Dias RS, et al. (2022) Computational prediction of potential inhibitors for SARS-COV-2 main protease based on machine learning, docking, MM-PBSA calculations, and metadynamics. PLoS ONE 17(4): e0267471. https://doi.org/10.1371/journal.pone.0267471

Editor: Alessio Lodola, University of Parma, ITALY

Received: August 9, 2021; Accepted: April 6, 2022; Published: April 22, 2022

Copyright: © 2022 Gomes et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: The source code and datasets are available at: https://github.com/IsabelaGomes/Prediction_SARSCOV2_inhibitors.

Funding: This study was financed in part by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior - Brasil (CAPES) - Finance Code 23038.004007/2014-82 - Grant 051/2013; Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq); Fundação de Amparo à Pesquisa do Estado de Minas Gerais (FAPEMIG). The funding agencies had no role in study design, data collection, analysis and interpretation, decision to publish, or preparation of the manuscript.

Competing interests: The authors have declared that no competing interests exist.

Introduction

During an outbreak, it is necessary to quickly respond to the unknown pathogen to avoid the uncontrolled spread of the disease. Just like what happened with the novel COVID-19 disease, caused by the Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-COV-2), that appeared for the first time in Wuhan, China at the end of 2019, spreading quickly all over the globe. In approximately one year, the infection reached more than 170 million cumulative confirmed cases and caused over than 3.5 million deaths (https://covid19.who.int/; as of Jun 2021). Considering that the development of new drugs is expensive and time-consuming, in this scenario computational strategies can potentially speed up the process of drug discovery. The repurposing of existing drugs to treat new diseases can accelerate the approval process, giving a quick response against unknown pathogens [1]. The use of the antiviral Remdesivir is an example, as this drug was initially indicated for the treatment of Ebola, and in October 2020, FDA approved it for use against COVID-19 [2]. However, the mortality rate for patients treated with Remdesivir is still quite high and does not differ significantly from placebo treatment in clinical trials [3, 4]. This shows that treatment with this antiviral alone is still not enough and further research to identify other promising drugs should continue.

Drug development often starts with the identification of key molecules, generally proteins, also called targets, which are crucial for the treatment of a specific disease. High-throughput screening (HTS) experiments are performed aiming to identify compounds that interact with a target protein to achieve a biological purpose, and these compounds can be used as drug candidates. However, designing HTS experiments is expensive, which requires time and resources such as advanced laboratories having chemical and biological libraries [5, 6]. The problem increases when considering the drug development as a whole since the cost of all process is upwards of US$2.8 billion [7]. Moreover, there is a lack of correlation between in vivo and in vitro assays, since HTS is not able to predict clinical failures, such as side effects and toxicity problems. To address these challenges, computational methods, such as virtual screening (VS), have been developed to mitigate costs and improve productivity in drug development [8–11].

VS is a computational approach in drug discovery, which aims to predict drug-like small molecules that can bind a drug target, generally, a protein [12]. The VS pipeline has the potential to highly reduce the cost and time required for HTS experiments, discarding unlikely compound-target pairs and, as result, potential active combinations are selected [5]. The binding affinity between a protein target and a ligand candidate is assessed by scoring functions associated with the method, in which potential ligands are ranked according to their binding capability. There are two main approaches of VS, structure-based and ligand-based. These can also be combined, generating a hybrid approach [13].

The structure-based virtual screening (SBVS) uses structural information of a protein target, such as biding pockets, to dock a ligand candidate to obtain energy predictions. It is highly dependent on the three-dimensional structure of the protein, which can be considered as a limitation because there is a considerable gap between the availability of protein sequences and protein structures. According to The UniProt Consortium, there are 189 million sequence records in Uniprot, a catalogue of all known protein sequences. On the other hand, in Protein Data Bank (PDB), the catalogue of all known structures, there are 175,000 macromolecules [14]. Therefore, depending on the target, it is necessary to create a protein structure model first, which can potentially impact on the quality of the final result [15, 16]. Furthermore, docking techniques may disregard dynamic features of the protein and, thus, the three-dimensional structure may not represent the most favorable conformation for binding a ligand that would be highly favorable in another thermodynamic state. For these reasons, structure-based methods are more commonly used for optimization, along with some other techniques, such as molecular dynamics [17]. The primary characteristic that distinguishes SBVS tools is how spatial conformations are ranked, according to the score function. There is a current trend to replace classical score functions with more sophisticated methods involving machine learning, to increase the accuracy of the tools [18]. Thus, recent methods look for improvements in the calculation of score functions, such as AtomNet [19], which uses deep convolutional neural network, and SIEVE [20], which uses features of intermolecular binding energy.

Another popular class of methods to perform virtual screening is the ligand-based, which explores chemical and structural features among small molecules to identify compounds having a similar pharmacological activity to the protein. Therefore, it requires a large volume of known data for the specific target, opening space for the employment of a variety of machine learning methods [21]. The main algorithms used for this purpose are support vector machines, k-nearest neighbors, random forests, genetic algorithms, and artificial neural networks, which can be combined [6, 22, 23]. There are some data structures usually employed to represent molecules in ligand-based virtual screening (LBVS) algorithms: graphs, based on the molecular structure [24, 25], fingerprint vectors, which characterize the ligand in terms of its physicochemical and structural properties [23, 26, 27], and ACFs (Atom-Centered Fragments) [28, 29]. We can highlight some examples involving each of the structures. LiSiCA [24] is a software that applies the click algorithm to find and rank similarities between pairs of two or three-dimensional molecules, represented in the shape of graphs. The LBVS [29] is a web server that uses BindingDB [30] and ChEMBL [31] as data sources for the Naive Bayes classifier, whose features are ACFs (Atom-Centered Fragments). Another web server is HybridSim-VS which performs virtual screening ligand prediction based on 2D and 3D fingerprints to calculate the similarity between molecules [27].

SARS-COV-2 is an enveloped RNA positive genome virus with about 30kb that encodes two polyproteins, which are cleaved to constitute the 16 non-structural proteins and four structural proteins, essential to virus cycle [8, 32]. The polyproteins cleaveage is mostly done by the main-protease (M^pro) which makes this protein a potential drug target to inhibit virus replication [10].

In this work, we combine an in-house developed machine learning strategy, docking, MM-PBSA calculations, and metadynamics to identify molecules approved by the Food and Drug Administration (FDA) that can potentially act as inhibitors of the enzyme activity of the main protease of SARS-COV-2, which is relevant for interrupting the viral cycle. Our computational strategy is hybrid as it combines LBVS and SBVS: (i) We devised a machine learning strategy that couples different molecule fingerprints to perform a first step of LBVS. (ii) Next, the resulting molecules went through SBVS steps. These molecules are docked against the target protein (SARS-COV-2 M^pro) using Autodock Vina. The poses with lowest vina scores show the preliminary favorable fit of the ligand inside the M^pro active site. Finally, we selected the best-scored poses to perform Molecular Mechanics Poisson-Boltzmann Surface Area (MM-PBSA) and metadynamic simulations using two distinct force fields, CHARMM and AMBER. These simulations intended to predict the binding energy for each complex and inspect the impact of the force field on this kind of simulation.

This work aims to identify, through computational techniques, promising molecules to interact with the target protein. The results obtained here serve as input for subsequent in vitro assays to validate the inhibition potential suggested by in silico experiments. All the source code and input datasets are available at https://github.com/IsabelaGomes/Prediction_SARSCOV2_inhibitors.

Materials and methods

This section details our computational strategy to predict candidate inhibitors for SARS-COV-2 M^pro. (Fig 1) presents the workflow that outlines the process. We explain data collection and preprocessing, prediction of molecules through the in-house developed supervised learning strategy, and refinement of these molecules through docking and molecular dynamics.

Download:

Fig 1. The workflow of our strategy.

It is composed of four blocks: Data preparation; Preprocessing data; Supervised learning; and Refining simulations. Rectangles indicate processing steps and ellipsoids denote datasets.

https://doi.org/10.1371/journal.pone.0267471.g001

Data preparation

Our strategy intends to predict small molecules to favorable interact with a protein in an active pocket. As a first step, it is necessary to collect robust data to feed the machine learning strategy.

Only positive samples are insufficient to build a robust predictive model and, to make a model that correctly filters non-active compounds, is important to have inactive molecules in the dataset. One challenge in molecule screening modeling is how to label compounds as inactive since the range of inactive molecules is very wide, so we use decoys to fulfill this task. We generated 50 molecular decoys for each active ligand using the “Database of Useful Decoys: Enhanced” server (DUD-E) [33].

An important step in machine learning is feature engineering, where the raw data is converted into a more informative value, called a feature. In the drug repurposing context, we aim to convert the chemical data into an informative numerical feature understandable for predictive models. This is achieved by computational processing for molecular fingerprint generation, which are numerical features extracted from chemical structures that can be used to assess the similarity of two compounds.

The structural similarity of two molecules is commonly evaluated by computing the Tanimoto coefficient on their chemical fingerprints [26]. High values indicate two compounds are similar but do not provide information about specific chemical groups they share. This is because the compounds are compared one by one, resulting in a local point of view, unlike when compounds are analyzed together, achieving a global view. In this scenario, we decide to use techniques to increase the overall quality of the data and to help the machine learning strategy better capture the relationships between features, beyond the Tanimoto coefficient.

The active ligands and the decoys composes our dataset and we described the molecule features through a combination of two molecular fingerprints, implemented by RDKit [34]. The first of them is a topological-based fingerprint inspired by Daylight fingerprint that extracts chemical patterns from the chemical graph, and the other one is a structure keys-based fingerprint implementation of the public MACCS keys, indicating the presence or absence in the compounds of particular substructures or features from a given list of structural keys [26]. Thus, we built a data matrix containing numerical information about active molecules and decoys on the rows and fingerprint features on the columns. The resulting matrix is composed of 2214 columns and used as input to build the supervised machine learning model.

Preprocessing data

The next step concerns processing the data in order to improve the predictive performance. Usually, the matrix data generated by fingerprints is sparse, because molecules have different sizes and functional groups. This results in many cells filled with zeros. Moreover, the assumption that similar molecules share similar activities might not always be valid, since minor modification on functional groups causes an abrupt change in activity [35]. Small changes of descriptor values can lead to considerable changes in molecular properties. Thus, before using raw fingerprints as input for a machine learning strategy, we perform singular value decomposition (SVD) that allows us to obtain lower-dimensional projections of the data [36]. The SVD reduces the dimensionality while preserving essential properties of the raw data matrix, detecting meaningful patterns while discarding weak signals and noises [37].

As 50 decoys were generated for each active molecule, there is an imbalance between actives and decoys, that is, positive and negative classes. Imbalanced data refers to a dataset in which one or more classes have much grater number of examples than others [38]. In drug discovery, the prevalent class, called majority class, comprises inactive molecules, while the rarest class, or minority class, is composed by active compounds. Generally, imbalanced data significantly challenges traditional machine learning models. To address this issue, a rebalance is performed before building the learning model, aiming to attain a more balanced input data and improve the model prediction capacity. We use a resampling technique to rebalance the sample space in order to relieve the effect of the skewed class distribution in the learning process. Given an imbalanced dataset, we split the majority class set into subsets similar in size to positive class. We combine each negative subset with the positive one, and create many smaller balanced subdatasets that will be used as input to an ensemble of classifiers, similar to the strategy implemented in [39].

Supervised learning

The matrix of active compounds and their respective decoys for each target after preprocessing data served as training data to build a prediction model. To select an appropriated learning model and optimize its parameters for each dataset, we used TPOT [40], an automated machine learning system written in Python. The pipelines were trained and evaluated under 5-fold cross validation using scikit-learn library in Python [41].

Nearly 2,000 approved molecules were obtained from Drugbank [42] and screened against each target using the best pipeline according TPOT. This set of approved drugs was used as a validation set, aiming to verify if the best machine learning pipeline is able to retrieve active compounds from Drugbank FDA approved drugs. After screening, Drugbank approved compounds were ranked according to the class probabilities returned by the respective pipeline. Best ranked compounds can be selected as candidates for future analysis in more robust experiments, such as molecular docking and molecular dynamics. The metric used to assess the performance of the VS strategy is the Enrichment Factor (EF), which is a measure of how much the compound dataset is enriched with actives after applying the VS strategy [6]. Usually, a sample of one percent (top-1%) of the original data, after labeling probabilities, containing the best classified compounds is used to calculate the EF. This metric is computed as the ratio between the proportion of actives after and before applying the VS strategy. (1) in which a1 = actives before the VS step, a2 = actives after the VS step, d1 = decoys before the VS step, and d2 = decoys after the VS step.

For our search for potential SARS-COV-2 M^pro inhibitors, we collected a benchmark on April 24, 2020 in PDB [43] consisting of 74 SARS-COV-2 main protease structures in complex with diverse ligands. The best machine learning algorithm calculated by TPOT under a 5-fold cross validation was the MLP Classifier, with parameters: alpha = 0.0001, learning_rate_init = 0.1, max_iter = 500, hidden_layer_sizes = (100,75,50,25,).

Docking

After using our VS strategy on Drugbank FDA approved drugs, we selected 17 compounds candidates to interact against SARS-COV-2 M^pro, excluding those that act on the central nervous system or are illegal—allowed to medical use under specific conditions. We executed the local docking focusing on the binding site: H41, S46, M49, Y54, F140, L141, N142, G143, C145, H164, M165, E166, L167, P168, H172, A173, F185, D187, Q189, T190, A191, and Q192 [10]. The grid box was elaborated using Autogrid from AutoDock Tool [44]. We added polar hydrogen to the macromolecule, intended to correct any deprotonations of the structure, and rotation was allowed to all rotatable bonds of the ligands. Every molecule was saved in pdbqt format so they could be used as input for AutoDock Vina [45]. We executed three simulations replica and we set the tool to generate seven poses for each drug in each replica. The best scored complexes were analyzed by Biovia Discovery Studio 2020 [46] to describe the contacts profile.

Docking indicates preferable areas where the ligands interact with the protein. However, this is a static result that can be improved with molecular dynamics approximating the system to real conditions. Moreover, the docking score function does not reflect the the binding energy, which can be estimated by molecular dynamics methods such as MM-PBSA and metadynamics.

MM-PBSA simulations

We selected the six best-scored poses from docking for molecular dynamics simulations, MM-PBSA and metadynamic, using two distinct force fields: CHARMM [47] and AMBER [48]. These simulations intended to predict the binding energy for each complex and inspect the impact of the force field on this kind of simulation. Using two different ways to estimate the interaction power of the two molecules allows us to view the complex binding from more than one perspective. The use of distinct force fields allows us to see if even with different parameters governing the molecular trajectories, the tendency of the interaction force is maintained.

We performed the MM-PBSA on GROMACS 5.1.4 [49]. We followed the SwissParam [50] protocol to generate drugs topologies with the CHARMM force field [47] and antechamber [51] to generate drugs topologies with the AMBER force field [48]. Each system was centered on a dodecahedral box with a distance of 14 Å between the complex and the edges, solvated with TIP3P water, and neutralized with 0.15 mol.L^-1 sodium chloride.

For energy minimization, we applied the steepest descent [52] and conjugate gradient [53] algorithms with convergence criteria of 0.24 kcal.mol^-1 in both cases. Subsequently, we equilibrated the system in 300 K and 1.0 atm using seven simulations stages with restriction forces being gradually removed. The first stage was in canonical coupling and, the others, in isotherm-isobaric coupling. The simulation protocol consisted of the Verlet Leap-frog integration algorithm [54], with an integration step of 2.0 fs. Also, v-rescale algorithms [55, 56] with a tau_t = 0.1 ps and Parrinello-Rahman [57] with a tau_p = 2.0 ps were used for the temperature and pressure couplings, respectively. For the CHARMM force field, we applied the settings suggested by GROMACS for van der Waals parameters on the molecular dynamics parameters file.

Next, we submitted the complexes to 100 ns of molecular dynamics each, under the same parameters. From the trajectories obtained, we compressed it into 500 snapshots due to MM-PBSA calculation using g_mmpbsa tool [58] and default parameters.

Since the sample data set is small to compare the energies calculated with both force fields, only six points, we used Spearman’s rank correlation to assess the concordance of the results.

Metadynamics

Metadynamics is an appropriate method to determine the binding energy for protein-ligand complexes, mainly when the active site is large and the ligands very small [59, 60] and it is a quicker simulation than MM-PBSA. This simulation also allows us to establish different sets of collective variables, which govern the simulation in specific ways. The combination of several of these sets of collective variables describes more clearly the unbinding behavior of the ligand from the active site of the protein.

We used the same six complexes to perform metadynamics simulations with the same force fields used previously, CHARMM, and AMBER. First, each system was solvated using the Solvate v1.7 plugin for Visual Molecular Dynamics software [61] with TIP3P water in a cubic box with 14 Å padding. Next, we neutralized the systems with 0.15 mol.L^-1 sodium chloride using Autoionize v1.7 plugin for VMD software [61].

We used the NAMD 2.14 software [62] to perform the simulations with a time-step of 2.0 fs, in an NPT ensemble with Langevin thermostat and barostat devices set at 300K and 1 atm. We set periodic boundary conditions and a cutoff of 12 Å for the nonbonded interactions, and we calculated the long-range electrostatic interactions by the Particle mesh Ewald (PME) method. The systems were minimized for 1000 steps by minimization with conjugate gradient, then a protocol of relaxation with harmonic restrains was performed. This protocol consists of:

500 ps MD with harmonic restrains in the M^pro and the ligand;
500 ps MD with harmonic restrains in the backbone of the M^pro and ligand;
500 ps MD with harmonic restrains in the M^pro;
500 ps MD without harmonic restrains;
8 ns MD without harmonic restrains, restarting the velocities;
500 ps MD with harmonic restrains in the backbone of the M^pro;
500 ps MD with harmonic restrains in the M^pro;
1 ns MD with harmonic restrains in the M^pro, restarting the velocities.

For the actual metadynamics, we restrained the M^pro harmonically, and we kept the ligands free for all the systems during the 7 ns. We executed three simulation replica for each ligand with each force field. We set the height of the Gaussians to 0.02 kcal.mol^-1, added every 1.0 ps with a width of 1.77. We chose two groups of collective variables (CV):

Group 1: The first CV were the distance between the center of mass of M^pro C145 and the center of mass of the closest atoms of the ligand to M^pro C145, varying between 0 Å and 30 Å with an amplitude fluctuation of 2 Å. For the second CV, we established the angle between the center of mass of M^pro C145, the center of mass of the closest atoms of the ligand to M^pro C145 and the center of mass of the entire ligand, varying from 0º to 180º with an amplitude fluctuation of 10º.
Group 2: The first CV were the same distance for group 1. For the second CV, we established specific angles for each ligand, considering their internal coordinate variations, varying from 0º to 180º with an amplitude fluctuation of 10º. The atoms involved in each angle are described in Table 1. The label of each ligand is available on Supporting Information.

Download:

Table 1. Set of atoms for each angle component for the selected ligands.

https://doi.org/10.1371/journal.pone.0267471.t001

We chose this pair of collective variables based on the principal movements of the ligands inside the protein on the equilibrium molecular dynamics stage. These CVs showed better sampling and convergence about the configurations on each minimum before the recrossing.

Despite operating the metadynamics by varying two collective variables, distance (CV_dist), and a specific angle (CV_ang), the unbinding process is mainly described by pull off the ligand from the protein. For this reason, we can estimate the metadynamics free energy onto the CV_dist according to the (Eq 2) [60]. (2) in which β = 1/k_bT, where k_b is the Boltzmann constant 1.9858 x 10^-3 kcal.mol^-1·K^-1, T = 300K, and G(CV_dist, CV_ang) is the free energy value for each distance and angle pair described in the PMF (Potentials of Mean Force) map.

Since Mpro is an enzyme with a shallow and exposed active site, the 7.0 ns of simulation maps conformations of the ligand onto the active site, the bulk, and even allows the molecule to interact with the protein again in other regions. Thus, it is critical to analyze the distance profiles and potential and electrostatic energies over time to select the most appropriate PMF files, since they are constructed by overlapping Gaussians during the simulation. These files describe the energies within the catalytic site and on the bulk to reconstruct the free energy landscape of the system before the ligand explores states beyond the two intended to avoid overestimating the energies in these two states. We calculated the distance in each frame using CPPTRAJ [63] and nonbonded energies using NamdEnergy. We analyzed the results with VMD software [61] and Python in-house scripts.

The binding free energy from each ligand with both force fields was determined by the subtraction of the minimum energy of the ligand inside the protein binding site (Gⁱⁿ) and the minimum energy of the ligand in the water (G^out), without the protein influence according to (Eq 3). (3)

This influence could be noticed from the distance profile, in which we established a cutoff of 20 Å to consider the ligand inside the protein, concomitantly with energy profiles. When the complex energy is null, it indicates no influence of the M^pro on the ligand.

For each ligand, we then performed six simulations with each force field. We calculated the final free energy by weighting the free energy of each simulation using the Boltzmann average, which highlights more favorable states, which are the more negative ones. The Boltzmann average applied for our systems can be written as the summation in (Eq 4). Thus with it was done for the MM-PBSA essay, we measured the concordance of the two force fields with Spearman rank correlation. (4) in which R_i CV_j is the i^th replica with the j CV set.

Results and discussion

This section discusses detailed results for each step of our method. First, we show that our supervised learning strategy is able to enrich compound datasets for three well known targets. Then our strategy is applied to point out potential molecules to inhibit SARS-COV-2 M^pro that will be refined through docking and molecular dynamics. Next, molecules that went through docking simulation and their respective scores are presented. Finally, the binding energies calculated through MM-PBSA and metadynamics simulations for two force fiels are shown.