Discovery of new Zika protease and polymerase inhibitors through the open science collaboration project OpenZika

Abstract

Zika virus (ZIKV) is a neurotropic arbovirus considered a global threat to public health. Although there have been several efforts in drug discovery projects for ZIKV in recent years, there are still no antiviral drugs approved to date. Here, we describe the results of a global collaborative crowdsourced open science project, the OpenZika project, from IBM’s Word Community Grid (WCG), which integrates different computational and experimental strategies for advancing a drug candidate for ZIKV. Initially, molecular docking protocols were developed to identify potential inhibitors of ZIKV NS5 RNA-dependent RNA-polymerase (NS5 RdRp), NS3 protease (NS2B-NS3pro), and NS3 helicase (NS3hel). Then, a machine learning (ML) model was built to distinguish active vs. inactive compounds for the cytoprotective effect against ZIKV infection. We performed three independent target-based virtual screening campaigns (NS5 RdRp, NS2B-NS3pro and NS3hel), followed by predictions by the ML model and other filters, and prioritized a total of 61 compounds for further testing in enzymatic and phenotypic assays. This yielded five non-nucleoside compounds which showed inhibitory activity against ZIKV NS5 RdRp in enzymatic assays (IC₅₀ range from 0.61μM to 17 μM). Two compounds thermally destabilized NS3hel and showed binding affinity in the micromolar range (K_d range from 9 μM to 35 μM). Moreover, the compounds LabMol-301 inhibited both NS5 RdRp and the NS2B-NS3pro (IC₅₀ of 0.8 and 7.4 μM, respectively), and LabMol-212 thermally destabilized the ZIKV NS3hel (K_d of 35 μM), and both also protected cells from death induced by ZIKV infection in in vitro cell-based assays. However, while eight compounds (including LabMol-301 and LabMol-212) showed a cytoprotective effect and prevented ZIKV-induced cell death, agreeing with our ML model for prediction of this cytoprotective effect, no compound showed a direct antiviral effect against ZIKV. Thus, the new scaffolds discovered here are promising hits for future structural optimization and for advancing the discovery of further drug candidates for ZIKV. Furthermore, this work has demonstrated the importance of the integration of computational and experimental approaches, as well as the potential of large-scale collaborative networks to advance drug discovery projects for neglected diseases and emerging viruses, despite the lack of available direct antiviral activity and cytoprotective effect data, that reflects on the assertiveness of the computational predictions. The importance of these efforts rests with the need to be prepared for future viral epidemic and pandemic outbreaks.

Graphical Abstract

INTRODUCTION

Zika virus (ZIKV) is a flavivirus that belongs to the Flaviviridae family, transmitted mainly by mosquito bite^1,2 but also through perinatal,³ sexual,⁴ and blood transfusion.⁵ Since the last ZIKV outbreaks in 2014 in French Polynesia and in 2015 in Brazil, and the increasing reports of severe neurological disorders, such as microcephaly^6-8 and Guillain-Barré syndrome,^9,10 contributed to the inclusion of ZIKV as a world public health concern by the World Health Organization (WHO).¹¹ Moreover, in 2018, Zika was added to the WHO Blueprint List, a list of priority diseases for research and development in the emergency context. In a recent study (2022), researchers mimicked the natural transmission cycle of ZIKV from mosquito to mice and shed light on the possibility of new ZIKV outbreaks, due to mutations in the ZIKV NS2 protein that could enhance transmissibility and pathogenicity¹². Besides the efforts and research on vaccines and drugs,^13-16 there is still no approved treatments for ZIKV.¹⁷

Like other flaviviruses, ZIKV is enveloped and the viral genome consists of a single positive-stranded RNA and an Open Reading Frame (ORF).^18,19 The translated polyprotein is proteolytically processed into three structural proteins (capsid, envelope and membrane) and seven nonstructural (NS) proteins (NS1, NS2A, NS2B, NS3, NS4A, NS4B, and NS5), that regulate viral replication and can contribute to pathogenicity and virulence.^20,21 Given the lack of progress there is still an urgency to discover antivirals against ZIKV, where viral replication represents an attractive strategy to halt the infection.²² The nonstructural proteins involved in the replication process are promising targets for the drug design of antivirals due to their essentiality in the viral life cycle.²³

The ZIKV NS3 protein is a key component for the processing of viral polypeptide and genomic replication.²⁴ The NS3 N-terminus is a serine-protease.²⁵ The association with NS2B, a trans-membrane cofactor, stimulates NS3 protease to the active conformation.²⁶ NS2B-NS3 protease complex (NS2B-NS3pro) cleaves the viral polyprotein at NS2A/NS2B, NS2B/NS3, NS3/NS4A, and NS4B/NS5 junctions.^27,28 The NS3 C-terminus is the NS3 helicase (NS3hel) responsible for the unwinding of the RNA double-stranded after hydrolyzing nucleoside triphosphate (NTP) and works in coordination with NS5 polymerase during the de novo viral RNA synthesis.^29,30 The ZIKV NS5 RNA-dependent RNA-polymerase (NS5 RdRp) performs RNA synthesis, as well as antagonizes human host interferon response.^31,32 NS5 RdRp has an essential role in replication and host immune response modulation, and has been considered a promising target for ZIKV drug discovery^33,34.

In 2016 we built a team of researchers in Brazil and the USA in an open drug discovery effort for the Zika virus.^35,36 This became the OpenZika Project^37,38 which utilized the IBM’s World Community Grid (WCG) crowdsourced computational network for virtually screening millions of compounds against all ZIKV protein structures (and related flavivirus protein crystal structures) and included various computational and experimental strategies to discover novel candidate compounds for ZIKV. Here, we describe our further efforts from this project focused on the ZIKV NS3hel, NS2B-NS3pro and NS5 RdRp proteins, through docking-based virtual screening campaigns of a commercial database, followed by prioritization of compounds with machine learning models to predict the cytoprotective effect of compounds against ZIKV infection, molecular dynamics (MD) simulations and experimental validation with biophysical and enzymatic assays as well as cell-based assays. We identified seven potent inhibitors for ZIKV replication proteins as promising scaffolds for hit-to-lead optimization as well as eight compounds that were able to protect cells from ZIKV infection-induced cytopathic effect (CPE), which is the host-cell death caused by the virus infection. Our findings support the efficiency and capability of integrated in silico approaches and the community efforts to accelerate the discovery of promising drug candidates against ZIKV in the absence of direct funding for our work.

METHODS

Computational

IBM’s World Community Grid is an internet-distributed network of millions of computers (Mac, Windows and Linux) and Android-based tablets or smartphones in over 80 countries. Over 715,000 volunteers donate their dormant computer time (that would otherwise be wasted) towards different projects that are both (a) run by an academic or nonprofit research institute, and (b) are devoted to benefiting humanity. OpenZika harness the World Community Grid to dock millions of commercially available compounds against multiple ZIKV homology models and crystal structures (and targets from related viruses) using AutoDock Vina. This ultimately produced candidates (virtual hits that produced the best docking scores and displayed the best interactions with the target during visual inspection) against individual proteins. All computational data against ZIKV targets are open to the public on our website and OpenZika results are also available upon request. Figure 1 shows the general workflow used to identify inhibitors of ZIKV NS5 polymerase, NS3 helicase and NS2B-NS3 protease proteins in this study.

Figure 1.

General workflow applied in this work to identify ZIKV NS5 polymerase, NS3 helicase and NS2B-NS3 protease inhibitors. First the ZINC database and ZIKV targets were prepared for molecular docking. In parallel, were developed a machine learning model for prediction of cytoprotective effect against ZIKV. Then, were performed VS campaigns using the computational filters: (i) molecular docking at OpenZika, (ii) ML model for cytoprotective effect against ZIKV, (iii) Bayesian models for BBB permeability, (iv) filters for LogP, PAINS and aggregators, (v) MedChem-based inspection. The prioritized virtual hits of each ZIKV target were validated through enzymatic assays, and for the most promising compounds we performed molecular dynamics simulations to characterize the molecular mechanism of action. Finally, cell-based assays to estimate the cytoprotection and potential cytotoxicity of the hits were performed.

Molecular Docking

Docking input files of the targets and ligands were prepared, and positive control docking studies were performed using AutodockVina 1.1.2 program³⁹ that uses knowledge-based potentials and empirical scoring functions. The pdbqt files of the libraries of compounds we screened are also openly accessible (see https://web.archive.org/web/20170829223457/http://zinc.docking.org/pdbqt/). We have prepared the docking input files for ~36 million compounds from ZINC (i.e., the libraries that Perryman and coworkers, previously used in the GO Fight Against Malaria project on World Community Grid)^40,41.

Initially, the Gasteiger-Marsili charges were added to all ligands atoms using AutoDock Tools 4.2⁴². Then, the 3D structures of NS5 RdRp (PDB ID: 5TFR⁴³ and 6LD3⁴⁴); NS2B-NS3pro open conformation (PDB ID: 5GXJ⁴⁵ and 7M1V⁴⁶) and closed conformation (PDB ID: 5YOD⁴⁷ 6L50⁴⁸ and 6L4Z⁴⁸); and NS3hel (PDB ID: 5MFX⁴⁹) were downloaded from Protein Data Bank⁵⁰ and prepared as follows: hydrogen atoms were added using the MolProbity server,⁵¹ whereas Gasteiger partial charges were computed, and non-polar hydrogens and their charges were merged onto their heavy atoms. For NS5 RdRp, the grid was built on the active, NTP and RNA binding sites, respectively. The grid coordinates for docking against NS2B-NS3pro was built on the catalytic and on the allosteric binding sites^52,53. For NS3hel, the grid was built on the ATP and RNA binding sites²⁹. All grid coordinates are specified in Supp. Info. 1. The “exhaustiveness” setting used was 20, for all systems. Positive-control redocking experiments were performed, using the co-crystallized inhibitors: compound 13 that was docked to the 3D structure 6LD3 of ZIKV NS5 RdRp; small fragment inhibitors that were docked to the 6L50 and 6L4Z of ZIKV NS2B-NS3pro; and compound NSC86314 that was docked to the 7M1V of ZIKV NS2B-NS3pro structure.

Details of each docking protocol are described below.

• NS5 RdRp: The docking was performed at three small binding sites, very close to each other (Supp. Info 1 Fig. S1): (i) the active site (composed by Asp535, Asp665 and Asp666)^54,55 where nucleoside triphosphates (NTPs) are incorporated; (ii) the RNA binding site⁵⁴ (Trp797, Lys403, Asn494, Tyr609, Thr608, Val607, Ser603, Gln605, Phe487, Arg483) and (iii) the allosteric site or N-pocket,³⁴ that includes a highly conserved loop known as priming loop (Thr796, Trp797, Ile799, Lys802, Glu804 and Trp805).^55,56 The priming loop stabilizes the de novo initiation complex and releases the new synthesized double stranded RNA.⁵⁷

• NS2B-NS3pro: The docking was performed on both catalytic and allosteric binding sites. The ZIKV NS2B-NS3pro active site is composed of the catalytic triad His51, Asp75, and Ser135, and the residues Ser81, Asp129 and Tyr161 (Supp. Info. 1 Fig. S2).⁵⁸ While the allosteric site is composed of Trp69, Lys73, Leu76, Trp83, Leu149, Asn152 and Val154 residues (Supp. Info. 1 Fig. S2).⁵⁹ ZIKV NS2B-NS3pro presents two main conformations⁶⁰. These include the closed or active conformation, that presents the NS2B cofactor wrapped around the NS3 active site, (and is the preferred conformation to investigate the catalytic site) and the open conformation, were the NS2B cofactor is dissociated from NS3pro and is preferred to investigate the allosteric site, which is described in the literature as the region of binding of flavivirus noncompetitive inhibitors⁶¹.

• ZIKV NS3hel: The ZIKV NS3hel presents two known binding sites: (i) the RNA binding site (Arg226, Thr265, Asp291, Val366, Arg388, Asp410, Lys431, Lys537, Asp540) and (ii) the ATP binding site (Glu231, Val228, Gly415, Thr201, Gly199, Lys200, Pro196, Gln455, Met414, Asp285, Glu286, Arg462, Gln455, Gly197)²⁹ (Supp. Info. 1 Fig. S3).

Machine Learning Models

A dataset of 5,414 molecules tested to measure caspase-3 activity induced by ZIKV infection in human progenitor cells (hNPCs) was downloaded from the PubChem Bioassay database (AID 1224857)^62,63 and was used to develop machine learning models. Briefly, all chemical structures and corresponding EC₅₀ data were carefully curated according to the protocols proposed by Fourches and colleagues.^64-66 Although there is no activity threshold stablished in literature for compounds with anti-ZIKV activity, compounds with activity <10 μM are considered good starting points for the subsequent hit-to-lead optimization, without solubility issues. Tropical infectious diseases such as tuberculosis, leishmaniasis and Chaga’s disease established EC₅₀ of 10 μM for a hit⁶⁷. Therefore, 43 compounds with EC₅₀ ≤ 10 μM were categorized as actives whereas 3478 compounds with EC₅₀ >10 μM were categorized as inactives. Then, 12 classification models were developed in Python v.3.6⁶⁷ through the combination of Support Vector Machine (SVM)⁶⁸ and Random Forest (RF)⁶⁹ algorithms along with three fingerprint sets and hybrid descriptors: Molecular ACCess System (MACCS) keys, Functional-Class Fingerprints (diameter 4: FCFP4), Extended Connectivity Fingerprints (diameter 4: ECFP4), Mordred⁷⁰ descriptors + MACCS (Mordred_MACCS), Mordred + FCFP4 (Mordred_ FCFP4), and Mordred + ECFP4 (Mordred_ ECFP4).^{71, 70, 68, 69} Details of model building and validation are available in the Supp. Info.1.

Virtual screening (VS)

We performed three independent VS campaigns of the ZINC15 library^72,73 aiming to identify anti-ZIKV compounds for experimental validation. Initially, molecular docking protocols were employed to screen inhibitors of ZIKV proteins (NS5 RdRp, NS2B-NS3pro, NS3hel) through the Autodock Vina 1.1.2³⁹ on the WCG platform, for large-scale virtual screening. Then, top-scored compounds were filtered by our ML model and the Bayesian ML models in the Assay Central software^®74 to predict the cytoprotective potential against ZIKV and blood-brain barrier (BBB) permeability⁷⁵, respectively. The BBB ML model was developed by Urbina and coworkers, 2021⁷⁵. The dataset used to build the model contain 1,777 BBB actives and 519 BBB inactive compounds. This dataset was modeled using extended connectivity fingerprint descriptors and Bayesian algorithm from Assay Central^®, with high accuracy metrics obtained in the validation ⁷⁴. The remaining compounds were then filtered to remove pan-assay interference compounds (PAINS)⁷⁶ and potential aggregators⁷⁷ using PAINS-Remover (https://www.cbligand.org/PAINS/) and Aggregator advisor⁷⁷ (http://advisor.bkslab.org/), respectively. Finally, a medicinal chemistry-based inspection was performed to select 61 putative virtual hits. Those were purchased, resuspended in DMSO, and prepared for experimental validation. The purity and chemical structures of the acquired compounds were confirmed by proton nuclear magnetic resonance (1H-NMR) spectra (Sup. Info. 1) and liquid chromatography–mass spectrometry (LC-MS/ELSD) analysis (Sup. Info. 1), at minimum purity ranging from 85% to 90% for all compounds⁷⁸.

Molecular dynamic simulations

All simulations were performed in the CUDA version of the AMBER19 software with the CUDA-version PMEMD, using FF14SB⁷⁹ and gaff2 force fields. Each molecule’s formal charge at physiological pH of 7.4 were predicted with MarvinSketch v.16.9.26, 2016, ChemAxon⁸⁰. Subsequently, the RESP method implemented on the R.E.D Server interface for PyRED⁸¹ was used to calculate the substrate charges applying HF/6-31G* level of theory using the Gaussian16 package ⁸². The Antechamber package from AmberTools18 generated the gaff2 (general amber force field) library and topology file for all ligands, including the 2,4-dimethoxy-5-(thiophen-2-yl) benzoic acid (PDB ID: 6LD2⁴⁴).

NS5 RdRp and NS2B-NS3pro docking structures had their hydrogen atoms removed and protonatable residues renamed according to their protonation state. C and N terminal extremities were capped as needed with NME (methylamine) and ACE (acetyl) groups. NS2B-NS3pro and NS5 RdRp systems were solvated in a truncated octahedral box extending its volume of at least 15 Å of any protein atom, with the TIP3P⁸³ and the SPC/E⁸⁴water models, respectively. The latter enabled applying the 12-6-4 Lennard-Jones-type nonbonded model⁸⁵ to account for the charged-induced dipole interactions of Zn²⁺. Charges were neutralized and salt concentrations were adjusted to 0.15 M with Na⁺ and Cl⁻ counterions using the LEaP module implemented on AmberTools18.

Systems’ equilibration consisted of three minimization steps, followed by NVT heating to 310 K and water equilibration in NPT ensemble, both with constraints of 10 kcal.mol⁻¹ on protein and ligand atoms, and a final 5 ns unconstrained MD using NVT ensemble (Supp. Inf. 1). The production simulations ran in the NTP ensemble, with coordinates saved every 50 ps. MD simulations used a 10 Å cut-off for nonbonded interactions. SHAKE algorithm⁸⁶ constrained all bonds with hydrogen atoms, and Hydrogen Mass Repartitioning (HMR) allowed a 4 fs time step^87,88. Langevin dynamics-controlled systems temperatures with the collision frequency set to 2 ps⁻¹, and the Monte Carlo barostat was applied during the two final equilibration steps and production simulations, with a pressure relaxation time of 1 ps.

Each of the 20 docking poses for both NS5 and NS2B-NS3 was considered a starting point for MD simulations. The simulations ran in duplicates with different initial velocities, generating 80 distinct trajectories. The resulting trajectories were analyzed with cpptraj⁸⁹ software. Two-dimensional interaction diagrams of the final MD frames were produced with Discovery Studio Visualizer v. 21.1⁹⁰.

Experimental

Proteins cloning, expression and purification

NS5 polymerase, NS2B-NS3 protease, and NS3 helicase were cloned at pETTRX- or pETSUMO by LIC method, expressed and purified according to the protocols described by Silva and coworkers (2019)⁹¹, Lima and coworkers (2021)⁹² and Godoy and coworkers (2017)⁹³. Briefly, the proteins were expressed in ZYM 5052 auto-induction medium or Luria Bertani (LB) medium and purified in four steps: (i) a HisTrap HP 5.0 mL with a Ni Sepharose resin (GE Healthcare); (ii) a buffer exchanged by dialysis and a concomitantly TEV protease cleavage from 6His-TRX-tag or 6His-SUMO-tag; (iii) an inverse HisTrap HP 5.0 mL to separate protein from 6His-TRX-tag and 6His-SUMO-tag; and (iv) a size-exclusion chromatography at an XK 16/60 Superdex 75 column (GE Healthcare).

NS5 RdRp and NS2B-NS3 protease activity assays

NS5 RdRp and NS2B-NS3pro activity assays were performed as described by Fernandes and coworkers⁹⁴. To evaluate the inhibitory activity of the compounds against ZIKV NS5 RdRp, we performed enzymatic assays, previously described by Sáez-Álvarez and coworkers⁹⁵. The concentration-response experiments for NS5 RdRp were performed with serial dilution of each compound ranging from 80 to 0.039 μM. The results were analyzed and plotted using the GraphPad Prism 5.0 program⁹⁶.

To investigate the mechanism of inhibition of LabMol-301 against ZIKV NS2B-NS3 protease, we used the fluorescence of the synthetic fluorogenic peptide Benzoyl-Nle-Lys-Arg-Arg-4-methylcoumarin-7-amide (Bz-nKRR-AMC) to determine the initial velocity of the enzymatic reactions in the presence of varying concentrations of the inhibitor. Kinetic assays were performed for LabMol-301 at 37°C, using the protocol described by Fernandes and coworkers⁹⁴, 4 nM of protein, 50 μM to 3.1 μM of substrate Bz-nKRR-AMC and 3.75 μM and 7.5 μM of LabMol-301. Enzymatic assays were performed in duplicate on a 96-well flat plate and in a SpectraMax Gemini EM Microplate Reader (Molecular Devices Co.). Data were plotted and analyzed on GraphPad Prism 5.0 program. Concentration-response data were adjusted using a Hill fitting to obtain IC₅₀ values, and kinetic plots were adjusted using a Non-Competitive Mixed Mechanism Model⁹⁶.

NS3 helicase Thermal Stability assay

NS3 helicase thermal stability was investigated using 200 μM of compound, 20 μM of protein in 20 mM Bis-Tris (Sigma), pH7, 500 mM NaCl (Sigma), 10% glycerol supplemented with 5x Sypro^® Orange (Sigma Aldrich). The assays were performed as described by Silva and coworkers.⁹¹ The results were analyzed and plotted using the GraphPad Prism 5.0 program⁹⁷.

NS3 helicase ATPase activity assay

NS3 helicase ATPase activity assay was performed using the commercial QuantiChrom^™ ATPase/GTPase Assay Kit (BioAssay Systems), 2mM of each compound and proteins and buffer as described by Silva and coworkers.⁹¹ The results were analyzed and plotted using the GraphPad Prism 5.0 program⁹⁷.

NS3 helicase binding assay

Experiments were performed on a Monolith^® NT.115 (Nanotemper technologies). NS3 helicase was labeled on histidine-tail using Monolith His-Tag Labeling Kit RED-tris-NTA 2^nd Generation as per the manufacturer’s instructions. The concentration of protein indicated for MicroScale Thermophoresis experiments⁹⁸ was 50 nM and with a serial dilution of each compound from 5 mM to 150 nM. The dissociation constant K_d was obtained by fitting the binding curve with the Hill function.

ZIKV Cell-based assays

We measured the indirect antiviral activity based on a cytopathic effect on the host cell. Glioblastoma stem cells (GSC387) (5000 cells/well) were seeded into 1536-well black, clear-bottom plates containing 20-point, 2-fold serial dilutions of a compound or DMSO vehicle controls and were either mock-infected or infected with ZIKV H/PAN/2016/BEI-259634 at a multiplicity of infection (MOI) of 10. Cell viability to determine virus-induced cytopathic effect was measured using the CellTiter-Glo assay (Promega, Madison, WI, USA) 72 h post-infection. An Envision plate reader (PerkinElmer) was used for readouts of luminescence intensity. All data were normalized to DMSO vehicle controls and were expressed as % relative luminescence intensity. Normalized activity and toxicity data were plotted against compound concentration and fit to a sigmoidal dose-response curve with variable slope using CDD Vault to obtain EC₅₀ and CC₅₀ values (Collaborative Drug Discovery Inc., Burlingame, CA).

To test if the hits had direct antiviral activity, plaque assays on Vero cells were performed to measure the amount of new virus released by treated cells. Serial dilutions of infected cell supernatants were incubated on confluent monolayers of Vero cells for 2h. Agarose overlays were applied and plates were incubated at 37° C in 5% CO₂ for 3 days. Plates were fixed with formaldehyde, overlays were removed, and monolayers were stained with crystal violet (0.025% in 2% EtOH). Plaque assays were performed and counted by a blinded experimenter to avoid quantification bias.

Cytotoxic assessment against human hepatocellular carcinoma cells

A culture of human hepatocellular carcinoma (HepG2) cells was kept in a flask in a humidified atmosphere of 5% CO₂ at 37°C. The culture medium used was RPMI 1640 supplemented with 25 mM HEPES (pH 7.4), 24 mM sodium bicarbonate, 11 mM D-glucose, 40 μg/mL penicillin-streptomycin, and 10% (v/v) bovine fetal serum. Treatment with a 0.25% trypsin solution was used to release cells from the flask walls every three to four days, and a 1:4 proportion of the cells were maintained in culture.

An adaptation of the MTT assay described by Denizot and Lang (1986)⁹⁹ was employed to determine cytotoxic activity. The HepG2 cells were counted and distributed in a 96-well plate, in a proportion of 5x10⁶ cells per well. The plate was incubated overnight at 37°C, in a humidified atmosphere of 5% CO₂. Serial dilutions of the compounds were added in a 1:9 proportion to each well and the plates were incubated for another 24 hours. Positive (no compound) controls were added to each plate for normalization of results. The supernatant was then removed, and a solution of MTT salt (3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide) was added to each well. After two to four hours of incubation, the formazan crystals formed by MTT reduction were solubilized in DMSO. The absorbance of the plate at 570 nm was measured, and the intensity values obtained were converted to viability values using the equation below. Concentration-response curves were built for each compound using the OriginPro 9.0 software (OriginLab), and the minimum inhibitory concentration for 50% of cells (IC₅₀^HepG2) was determined for each compound.

RESULTS AND DISCUSSION

Machine Learning Models

We built ML models for classifying the compounds as cytoprotective or not against ZIKV infection, based on a dataset of compounds tested to measure caspase-3 activity induced by ZIKV infection in human progenitor cells (hNPCs). The increase in caspase-3 activity indicates an activation of the apoptosis cell death pathway caused by ZIKV infection in hNPCs. A decrease in the luminescence signal in this assay by small molecule treatment indicates an inhibitory effect of that small molecule on ZIKV induced caspase-3 activity that may correlate with its effect on reduction of Zika virus caused cell death. The literature highlights that there are complex relationships between caspase activity, viral replication and cell survival. Some viruses appear to exploit the apoptosis response, by utilizing caspase activity to cleave viral proteins and facilitate replication¹⁰⁰. Flaviviruses utilize various strategies to activate or inhibit cell apoptosis, in a dual regulation apoptosis mode¹⁰¹. In this way, compounds that induce caspase-3 activity could inhibit the virus to exploit apoptosis, reduce viral replication and reduce cell death caused by viral infection.

The statistical results of the ML models are summarized in Table 1.

Table 1.

Statistical characteristics for the ML models obtained by 5-fold cross validation.

Fingerprint	Method	PT	ACC	SE	SP	MCC	AUC
FCFP4	SVM	0.01	0.70	0.72	0.70	0.10	0.71
FCFP4	RF	0.01	0.72	0.77	0.72	0.12	0.75
ECFP4	SVM	0.01	0.74	0.65	0.74	0.1	0.7
ECFP4	RF	0.02	0.73	0.77	0.73	0.12	0.75
MACCS	SVM	0.03	0.88	0.56	0.89	0.15	0.72
MACCS	RF	0.02	0.76	0.77	0.76	0.14	0.76
Mordred_FCFP4	SVM	0.01	0.77	0.77	0.77	0.14	0.77
Mordred_FCFP4	RF	0.02	0.85	0.74	0.85	0.18	0.80
Mordred_ECFP4	SVM	0.01	0.8	0.72	0.81	0.14	0.76
Mordred_ECFP4	RF	0.02	0.84	0.74	0.84	0.17	0.79
Mordred_MACCS	SVM	0.03	0.81	0.70	0.81	0.14	0.75
Mordred_MACCS	RF	0.03	0.86	0.72	0.86	0.18	0.79

FCFP4, functional-class fingerprints with diameter 4; ECFP4, extended-connectivity fingerprints with diameter 4; MACCS, Molecular ACCess System keys; SVM, Support Vector Machine; RF, Random Forest; PT, probability threshold; ACC, accuracy; SE, sensitivity; SP, specificity; MCC, Matthews correlation coefficient; AUC, area under ROC curve. The best model is highlighted in bold.

The model was built using hybrid descriptors (Mordred_FCFP4) and RF (ACC = 0.85; SE = 0.74; SP = 0.85; and AUC = 0.80) demonstrated the best performance among all other models for cytoprotective effect, indicating that this model may be useful for VS campaigns.

Virtual Screening

Three independent VS campaigns of part of the ZINC15 library (~36 million compounds)^73,102 were performed against ZIKV (i) NS5 RdRp (ii) NS2B-NS3pro and (iii) NS3hel, respectively (Figure 2). These nonstructural proteins are part of the viral replication complex, that regulate viral replication and can contribute to pathogenicity and virulence. Thus, inhibitors of these proteins are essential agents to prevent viral replication. The viral replication reduction would reduce the infection caused by ZIKV and consequently reduce the cell death. Moreover, inhibiting a viral protease and/or polymerase is a well-established therapeutic strategy for viruses e.g., HIV¹⁰³ and hepatitis C virus¹⁰⁴. Initially, compounds were filtered using molecular docking, as part of the OpenZika project,^35,38 an open science collaboration with WGC³⁹. The OpenZika project lasted approximately four years (2016 to 2019) and in total 9.29 billion docking jobs were submitted, involving 427 different target sites. The screens used approximately 36 million compounds from the ZINC15 library. 80,000 WCG volunteers donated their spare computing power to OpenZika with 92,696 CPU years’ worth of docking calculations.

Figure 2. Virtual screening workflow used to identify inhibitors of ZIKV NS5 polymerase, NS2B-NS3 protease and NS3 helicase.

The independent virtual screenings of ZINC15 database were filtered through: (i) molecular docking, (ii) machine learning model for cytoprotective effect against ZIKV, (iii) Bayesian models for BBB permeability, (iv) LogP, PAINS and aggregators, (v) MedChem based inspection. At the end, we selected 61 compounds for experimental validation.

After docking, the compounds were filtered using the best ML model developed for cytoprotection against ZIKV infection. The compounds predicted as active by the ML model were prioritized. Then, a Bayesian ML model available in the Assay Central^® software¹⁰⁵ was applied to predict BBB permeability¹⁰⁶. Compounds that presented BBBpred score ≥ 0.5 were predicted as promising BBB permeant agents. The BBB filter was applied to select compounds that possible could act also on the central nervous system and treat neurologic issues induced by ZIKV infection. The LogP filter was then applied and only compounds with LogP ≤ 5 were prioritized, to avoid compounds with poor permeability. Compounds predicted having PAINS alerts or aggregator alerts were discarded. A detailed medicinal chemistry-based inspection of all selected compounds structures and binding mode was performed using Pymol program¹⁰⁷. This inspection considered favorable scores for higher numbers of hydrogen bonds between ligand and important protein residues, pi-stacking and pi-cation interactions, salt bridges; and unfavorable scores for nonpolar regions of the ligand exposed to solvent⁴⁰.

At the end of the VS, we virtually prioritized 15 hits for ZIKV NS5 RdRp (Supp. Info. 2 excel file) (four for the active site, five for the RNA site and six for the allosteric site), 13 hits for ZIKV NS2B-NS3pro (Supp. Info. 2 excel file) (five for the active site and eight for the allosteric site) and 33 compounds for ZIKV NS3hel (13 for the ATP binding site and 20 for the RNA binding site) (Supp. Info. 2 excel file), totaling 61 virtual hits for experimental validation through enzymatic activity assays, kinetic assays and cell-based assays. Molecular dynamics simulations were also performed for selected compounds to investigate protein-ligand stability, binding mode and mechanism of action.

ZIKV NS5 RdRp enzymatic assays

Five compounds inhibited NS5 RdRp activity greater than 80% threshold at a single compound concentration of 20 μM. At this concentration, compounds LabMol-301, LabMol-309, LabMol-319, LabMol-204 and LabMol-202 showed 99%, 99%, 98%, 85% and 80% inhibition, respectively. The IC₅₀s and concentration-response curves of compounds LabMol-202, LabMol-204, LabMol-319, LabMol-301 and LabMol-309 are shown in Fig. 3A-E.

Figure 3. ZIKV NS5 RdRp assays.

Concentration-response curves adjusted with Hill to determine IC₅₀ ± Δ IC₅₀ values for compounds A) LabMol-202; B) LabMol-204; C) LabMol-319; D) LabMol-301 and E) LabMol-309.

ZIKV NS5 RdRp molecular dynamics simulations

To further assess the conformational dynamics and the main interactions involving ligand recognition by ZIKV NS5 RdRp, we performed MD simulations of NS5 RdRp in complex with the five enzymatic hits for NS5 RdRp, LabMol-309, LabMol-301, LabMol-319, LabMol-204 and LabMol-202, in their several docking poses. The Root Mean Square Deviation (RMSD) concerning the original docking pose and the RMSD regarding each replicate's final structure were used to determine all poses’ stability and convergence. We also performed the MD of a recently co-crystallized ligand 2,4-dimethoxy-5-(thiophen-2-yl)benzoic acid bounded to the ZIKV NS5 RdRp allosteric site (PDB ID 6LD2)⁴⁴. This ligand was used as a reference for the standard motion dynamics of the small molecules during MD. The co-crystallized ligand’s RMSD, regarding the initial crystallographic structure, presented values ranging from 1.5 Å to 5.5 Å (Supp. Info. 1 Fig. S4), showing that the NS5 RdRp allosteric site is large and allows a wide movement of ligands.

The three most potent compounds in the enzymatic assays, LabMol-309, LabMol-301 and LabMol-319, showed a diverse interaction repertoire. These three ligands are located at the interface between RdRp's thumb and palm subdomains and interact with the priming loop (residues 785-810) which connects two α helices from the thumb region, hence obstructing the RNA binding tunnel^43,108. The priming loop takes part in both the RNA de novo initiation process and the elongation step, where it undergoes an extensive conformational change to accommodate the double-stranded RNA. Moreover, it is also responsible for recognizing the stem-loop A (SLA) promoter for RNA synthesis¹⁰⁹. The convergence RMSD of poses suggests that LabMol-301, LabMol-309, and LabMol-319 target this region, leading to the disruption of the polymerase activity confirmed by the submicromolar IC₅₀ of those compounds. We highlighted the MD results for LabMol-301, whereas LabMol-309 and LabMol-319 MD results are shown in the Supp. Info. 1 – Supporting Results and Fig. S5-S8.

LabMol-301 presented two possible binding modes, from replicates of two different docking poses. Convergence RMSD towards the last frame of the first replicate from one pose showed RMSD ≤ 5 Å (Fig. 4A,B). The main anchoring residue for this binding mode was Cys711, forming a recurrent interaction with LabMol-301 in all three trajectories (Fig. 4C, D). Additionally, interactions between the aminopyridine and Ser603 or Tyr609 were also detected (Supp. Info. 1 Fig. S7). Nevertheless, an alternative binding mode converging from simulations starting from three different docking poses, was stabilized by Asp534 and Pro709 (Fig. 4E).

Figure 4. Binding modes of LabMol-301 during MD simulations with ZIKV NS5 RdRp.

A) Binding mode 1 convergence RMSD of the two replicates from docking pose 8 (regular and light orange) and the first one from pose 19 (blue) towards the last frame of pose 8 replicate 1 MD. B) Binding mode 2 convergence RMSD of the first replicates from poses 6 (green), 17 (dark blue), and 59 (brown) regarding the last frame of pose 6 replicate 1 MD. C) LabMol-301 (sticks and surface) interacting with multiple RdRp subdomains (color-coded ribbons: palm in grey, thumb in blue, fingers in light yellow, and N- terminal extension in pink) in the two binding modes identified. Detailed representation of final MD frames from the converged simulations of LabMol-301 D) Binding mode 1 and E) Binding mode 2 and its respective poses and replicates.

Some ZIKV NS5 RdRp inhibitors have been reported in the literature so far, classified in two categories: nucleoside inhibitors (NI) that present structural similarity to glycosylamines and non-nucleoside inhibitors (NNI). Adenosine triphosphate analogs^110,111, 2′ NTP analogs¹¹², ribonucleotide 5′-triphosphate analogs¹¹³, sofosbuvir¹¹⁴, remdesivir^111,115, 7-deazaadenosine nucleosides^111,116 have already been reported as NI of ZIKV NS5 RdRp. Among the NNIs, there are few natural compounds such as the chalcone xanthoangelol (IC₅₀ of 6.9 μM) ¹¹⁷ and the alkaloid lycorine (60% inhibition at 50 μM)^111,118, as well as the synthetic compounds 10-undecenoic acid zinc salt^111,119 (IC₅₀ of 1.13 μM) and 3-chloro-N-[({4-[4-(2-thienylcarbonyl)-1-piperazinyl]phenyl}amino)carbonothioyl]-1-benzothiophene-2-carboxamide(TBP)^111,120 (IC₅₀ of 94 nM), emetine^121,122 (IC₅₀ of 121 nM) and 2-morpholino-5-((pentyloxy)methyl)-N-(2-phenoxyethyl)-6-((4- sulfamoylphenyl)amino)pyrimidine-4-carboxamide^122,123 (IC₅₀ of 2.4 μM) that act by inhibiting ZIKV NS5 RdRp. NNIs act at the ZIKV NS5 RdRp allosteric site and, in general, display fewer side effects since they are more selective for viral over host polymerase targets^56,124. Our results therefore contribute to the ZIKV NS5 RdRp NNIs discovery, promoting the identification of distinct structures when compared with previous reported ZIKV NS5 RdRp NNIs. We analyzed the similarity of our best hit for ZIKV NS5 RdRp (LabMol-301) with the reported ZIKV NS5 RdRp NNIs: TBP; lycorine; 10-undecenoic acid zinc salt; 2-morpholino-5-((pentyloxy)methyl)-N-(2-phenoxyethyl)-6-((4-sulfamoylphenyl) amino)pyrimidine-4-carboxamide; xanthoangelol and emetine, using MACCS descriptors and Tanimoto coefficient (Tc) (Fig. 5). As we can see, our best hit is very dissimilar to all known inhibitors of ZIKV NS5 RdRp, with Tc < 0.5, showing that LabMol-301 represents a new scaffold of ZIKV NS5 RdRp NNIs.

Figure 5. Radial plot of LabMol-301 and the NS5 RdRp NNIs reported in the literature.

This analysis was performed with MACCS descriptors and the Tanimoto coefficient (Tc) between LabMol-301 and (A) TBP (Tc = 0.5); (B) lycorine, (Tc = 0.4); (C) 10-undecenoic acid zinc salt, (Tc = 0.37); (D) 2-morpholino-5-((pentyloxy)methyl)-N-(2-phenoxyethyl)-6-((4-sulfamoylphenyl)amino)pyrimidine-4-carboxamide, (Tc = 0.36); (E) xanthoangelol, (Tc = 0.16); and (F) emetine, (Tc = 0.03).

ZIKV NS2B-NS3pro enzymatic activity assays

For the investigation of NS2B-NS3pro enzymatic activity we used the same activity and enzymatic assays described by Li and collaborators¹²⁵. First, we evaluated anti-protease activity at a single time point using a single concentration of each selected compound (200 μM). The positive control aprotinin was used and had an IC₅₀ of 0.13 ± 0.02 μM¹²⁶. We observed that the compounds LabMol-301 and LabMol-314 inhibited 86 % and 83 % of the protease activity, respectively. Therefore, these compounds were selected for concentration-response experiments to determine the IC₅₀ values. The concentration-response assays showed that LabMol-301 is a low micromolar inhibitor (IC₅₀ = 7.4 ± 0.3 μM) (Fig. 6A) of the protease activity, whereas LabMol-314 showed a 7-fold decreased activity (IC₅₀ = 53.5 ± 4.1 μM) compared to LabMol-301. Thus, compound LabMol-301 was the most promising protease inhibitor.

Figure 6. ZIKV NS2B-NS3pro enzymatic assays.

A) Concentration-response curve adjusted with Hill to determine IC₅₀ ± Δ IC₅₀ for compound LabMol-301. B) Kinetic curves adjusted with Michaelis-Menten model, showing a non-competitive mechanism of action. C) Predicted pose by molecular docking of LabMol-301 at allosteric site of ZIKV NS2B-NS3pro.

To investigate the mechanism of inhibition of LabMol-301 against ZIKV NS2B-NS3pro, we used the fluorescence of the synthetic fluorogenic peptide Benzoyl-Nle-Lys-Arg-Arg-4-methylcoumarin-7-amide (Bz-nKRR-AMC) to determine the initial velocity of the enzymatic reactions in the presence of varying concentration of the inhibitor. These values were adjusted to the Michaelis-Menten and Mixed model of inhibition (Figure 6B). The K_m and V_max values are described in the Table S1 (Supp. Info. 1).

Our results showed that compound LabMol-301 is a non-competitive inhibitor with respect to the substrate, thereby acting as protease allosteric inhibitor. Both the K_m^app and V_max^app values decreased as increasing inhibitor concentration was used. These findings indicated that LabMol-301 binds to both the free enzyme and the enzyme-substrate complex with different binding constants (K_i and αK_i, respectively). LabMol-301 has increased affinity to the free form of the enzyme (K_i = 7 ± 2 μM) compared to the enzyme-substrate complex (K_i = 111 ± 34 μM). In Fig 6C, we showed the predicted docking pose of LabMol-301 at the allosteric site of ZIKV NS2B-NS3pro.

ZIKV NS2B-NS3pro molecular dynamics simulations

MD simulations were used to determine the feasibility of the interactions between LabMol-301 and NS2B-NS3pro at allosteric and catalytic sites. We evaluated the ligand in the protonated and neutral states since both forms could be present at the pH 8.5 of the enzymatic assays. Briefly, we recovered one stable conformation of the ligand's neutral form into the allosteric pocket and one of the ligand's protonated form into the catalytic pocket (Supp. Info. 1 Fig. S11-12). MD simulations did not preclude interactions at the catalytic site at pH 8.5 with the protonated LabMol-301. However, since the neutral molecule should be predominant at neutral pH, we argue that LabMol-301 interacts more competently with the allosteric site in physiological conditions.

LabMol-301 interactions with the ZIKV NS2B-NS3pro at the allosteric site relied on a highly hydrophobic complementarity between the molecule’s surfaces and the protein (Fig. 7A). The binding site's hydrophobic core residues that interact with the compound were Leu76, Trp83, Leu85, Ala87, Ile147, Gly148, Leu149, and Val155 (Fig. 7B). Interactions with residues Leu76 and Trp83 were also observed in the ZIKV NS2B-NS3pro- NSC86314 complex crystal structure (PDBID 7M1V). The hydrophobic nature of the allosteric pocket led to a tight coupling with the ligand's aromatic groups, leaving the polar groups exposed to the solvent. Such coupling also explains the observed convergence between MD replicates (Fig. 7C).

Figure 7. MD convergence and interactions of LabMol-301 at ZIKV NS2B-NS3pro allosteric pocket.

A) Tight coupling of LabMol-301 (green sticks and transparent surface) at the interaction site of NS2B-NS3pro (grey and yellow ribbons, transparent surface). The arrow highlights the subpocket B present exclusively on NS2B-NS3pro allosteric conformation. B) Ligand’s binding mode (green sticks) with NS2B-NS3pro (white ribbons and sticks) after simulation convergence. Blue sticks represent the interacting residues of the hydrophobic core. C) LabMol-301 converged into a stable conformation with low RMSD variation compared to the final replicate 1 structure.

Considering their essential function, proteases are promising targets for antiviral drug design leading to the successful HIV protease inhibitors saquinavir¹²⁷, boceprevir an HCV NS3-4A inhibitor¹²⁸, and more recently the Pfizer^® compound PF-07321332, an oral available main protease inhibitor for SARS-CoV-2¹²⁹. In contrast, since the ZIKV outbreak, there has been no drug candidate in clinical trials. Despite innumerous studies looking for competitive inhibitors for ZIKV NS2B-NS3pro, such as peptides¹³⁰, fragments¹³¹ and small molecules¹³² that bind at the active site, the recent discovery of allosteric inhibitors which interfere with the interaction of NS2B with NS3pro, provides a starting point for finding additional inhibitors¹³³. Among the allosteric inhibitors of NS2B-NS3pro reported in literature, compound NSC135618 (Fig. 8B) presented activity against ZIKV NS2B-NS3pro (IC₅₀ of 0.38 μM), acting as an allosteric inhibitor in kinetic assays⁵². Currently, there is only one crystal structure with a ZIKV NS2B-NS3pro allosteric inhibitor co-crystallized available (PDBID 7M1V)⁴⁶, NSC86314 (Fig. 8C) that presented an IC₅₀ of 1.12 ± 0.11 μM¹³⁴.

Figure 8. ZIKV NS2B-NS3pro inhibitors and their respective IC₅₀s at NS2B-NS3pro.

Superimposing the ZIKV NS2B-NS3pro-NSC86314 complex with the best docking pose of LabMol-301 (Supp. Info. 1 Fig. S14), there is an overlap between the structures, and both shared hydrogen bonds and π-stacking interactions with Trp83, and hydrophobic interactions with Leu76 residue. Moreover, we have calculated the 2D-similarity between LabMol-301 and the two known allosteric inhibitors of ZIKV NS2B-NS3pro, NSC135618 and NSC86314, using MACCs fingerprints and Tanimoto coefficient, and the results showed that our hit is very dissimilar from these compounds, with a Tc of 0.24 and 0.40, respectively. Our results agree with experimental evidence, and suggest that LabMol-301 prevents the formation of the active form of NS2B-NS3pro.

ZIKV NS3hel enzymatic activity assays

Dynamic Scattering Fluorescence assays at a single compound concentration (200 μM) to investigate NS3hel thermal stability. From the 33 compounds, 14 decreased NS3hel thermal stability, indicating that they could bind to the protein (Fig. 9A). Moreover, four compounds LabMol-212, LabMol-220, LabMol-298 and LabMol-307 showed a typical binding curve. LabMol-220 and LabMol-298 had poor binding to the protein, with K_d > 5 mM (the highest concentration on the assay). LabMol-212 and LabMol-307 showed K_d values that were low micromolar (Fig. 9B-C), with K_d ± Δ K_d values of 35 ± 6 μM and 9 ± 2 μM, respectively.

Figure 9. ZIKV NS3hel activity.

(A) Thermal stability assessment of NS3hel in the presence of the ligand candidates. Kd values determination by microscale thermophoresis (MST) (B) of LabMol-212 and (C) LabMol-307. Docking poses of (D) LabMol-307 and LabMol-212 at the RNA biding site of ZIKV NS3hel. LabMol-307 (carbon atoms in yellow sticks representation) interacting with the adenine nucleobase of viral RNA and with the protein residues of NS3hel and (E) LabMol-212 (carbon atoms in cyan sticks representation) interacting with the guanine of RNA and NS3hel residues. Hydrogen bonds are represented in green doted lines and hydrophobic interactions are in transparent surface.

Additionally, the compounds were submitted to the NS3hel ATPase activity assay at 40 μM, but none of them inhibited the enzyme activity. We then increased the compounds’ concentration and assessed the endpoint helicase ATPase activity assay at 2 mM. At this concentration, only compound LabMol-221 showed some inhibitory activity that met the selection criteria (inhibition > than 80%) to concentration-response assays. LabMol-221 showed IC₅₀ ± Δ IC₅₀ value of 2.14 ± 0.01 mM.

Since LabMol-212 and LabMol-307 did not inhibit NS3hel ATPase activity, we suggest that they could possibly bind to the RNA binding pocket, agreeing with the computational docking results (Fig. 9D-E). Although there are some DENV NS3hel inhibitors described^135-137, so far there are few inhibitors for ZIKV NS3hel described such as epigallocatechin-3-gallate^37,138. Our results add to the discovery of new ZIKV NS3hel inhibitors that possibly bind to the RNA binding pocket.

In vitro ZIKV cell-based assays and cytotoxicity assays

All 61 purchased compounds were tested in two independent phenotypic assays to evaluate (i) the cytoprotective effect on GSC387 cells (Supp. Info. 3 excel file); and (ii) direct antiviral activity in Vero cells and cytotoxicity assays. We plotted cell viability versus antiviral efficacy for the promising compounds (Supp. Info. 1 Fig. S15). LabMol-297, LabMol-201, LabMol-212, LabMol-298, LabMol-194, LabMol-300, LabMol-301 and LabMol-305 were able to protect from cell death induced by ZIKV, with an EC₅₀ ranging from 0.00339 to 31.9 μM. These compounds presented low cytotoxicity in Vero cells with CC₅₀s > 100 μM. Table 2 summarizes the predictions and biological evaluation for the best selected hits, with enzymatic and cytoprotective effect on glioblastoma stem stem cells.

Table 2.

Computational predictions with the probabilities of the selected hits from the virtual screening and respective in vitro activities at enzymatic assays and GSC387 cells.

ID	Chemical structure	ML probability active		NS5 RdRp IC50 (μM)	NS2B- NS3pro IC50 (μM)	NS3hel Kd (μM)	EC50 GSC387 (μM)
		ZIKV	BBB
LabMol-301		0.74	0.87	0.8 ± 0.1	7.4 ± 0.3	-	6.68
LabMol-309		0.84	0.28	0.61 ± 0.02	-	-	≥100
LabMol-319		0.63	0.99	1.6 ± 0.2	-	-	≥100
LabMol-204		0.66	0.64	6 ± 2	-	-	≥100
LabMol-202		0.71	0.47	17 ± 3	-	-	≥100
LabMol-307		0.58	0.53	-	-	9 ± 2	50.6
LabMol-212		0.80	0.99	-	-	35 ± 6	0.13
LabMol-297		0.50	0.99	-	-	-	0.00339
LabMol-201		0.78	0.93	-	-	-	0.0603
LabMol-298		0.69	0.99	-	-	-	0.212
LabMol-194		0.77	0.96	-	-	-	0.219
LabMol-300		0.99	0.99	-	-	-	0.658
LabMol-305		0.75	0.96	-	-	-	31.9

ML, Machine Learning probabilities predicted by ZIKV cytoprotective and BBB models; NS5 RdRp enzymatic activities; NS2B-NS3pro enzymatic activities; NS3hel enzymatic activities; cytoprotective effect on GSC387 cell. Dashed values mean the activity was not observed even in the highest concentration during the assay.

We also performed cytotoxicity assays against the hepatocellular carcinoma cell line HepG2 to verify potential effects of toxicity on human liver, an important cause of drug failure in the clinic¹³⁹. Overall, the compounds demonstrated low cytotoxicity against HepG2 cells. LabMol-201 had an IC₅₀ of 80 μM and was the only compound that showed an IC₅₀ value < 100 μM, thereby suggesting a low propensity to cause toxicity in liver cells.

Unfortunately, plaque assays on Vero cells revealed the compounds did not show direct antiviral activity. We also performed cell-based antiviral assays with Huh7 cells (Supp. Info. 1 Methods) and none of the compounds were able to inhibit viral activity in this other experiment. Cell permeability, solubility, efflux, metabolism, degradation and lack of sufficient concentration to reach protein targets could all be possible issues involved in the lack of direct antiviral activity to be explored in future, despite the inhibitory activity of some compounds against specific ZIKV proteins.

Moreover, the cytoprotection effect observed for LabMol-297, LabMol-201, LabMol-212, LabMol-298, LabMol-194, LabMol-300, LabMol-301and LabMol-305, could be related to the blockage of human apoptotic pathways that ZIKV commonly induces, such as caspase-3-mediated pathways; pro-apoptotic B-cell lymphoma protein 2 (Bcl-2) mediated family pathway, through NS proteins; C/EBP Homologous Protein (CHOP); double-stranded RNA and toll-like receptor (DsRNA/TLR3), through NS2B-NS3pro; or fibroblast growth factor 2 (FGF2).¹⁴⁰ We can hypothesize that our compounds could act by inhibiting caspase-3, which would agree with our GSC387 cell assays. The literature describes caspase-3 inhibitors that protect cells from cytopathic effects induced by viral infection.¹⁴¹ As an example emricasan is a pan-caspase inhibitor which inhibited the ZIKV-induced increase of caspase-3 activity and promoted neuroprotection in human cortical neural progenitors assays.¹⁴²

CONCLUSIONS

Despite the global health impact that ZIKV has caused, there are still no antivirals available for the treatment ZIKV infection. Therefore, continuous efforts are needed for the discovery and development of anti-ZIKV agents. The OpenZika project aimed to fill this gap in the search for ZIKV proteins inhibitors, leveraging computational approaches and the World Community Grid volunteer’s computer network. We successfully developed and integrated molecular docking, machine learning models, physicochemical and ADME properties filters applied for virtual screening of the ZINC database. The 61 virtual hits were tested against ZIKV replication proteins NS5 polymerase, NS2B-NS3 protease and NS3 helicase. Five of them, LabMol-309, LabMol-301, LabMol-319, LabMol-204 and LabMol-202, showed considerable inhibitory activity against ZIKV NS5 RdRp. LabMol-301 was also able to inhibit ZIKV NS2B-NS3pro, as a non-competitive inhibitor. Two compounds LabMol-212 and LabMol-307 did not inhibit NS3hel ATPase activity, but thermally destabilized the ZIKV NS3hel and could possibly bind to NS3RNA binding site. Moreover, molecular dynamics simulations revealed important dynamic features regarding these compounds binding mode on these proteins. Cell-based assays showed that eight compounds protected cells from ZIKV-induced death, agreeing with our developed ML model, and demonstrated low cytotoxicity. From these eight compounds, two LabMol-301 and LabMol-212 also presented activity against ZIKV NS proteins. We have therefore developed a robust and externally predictive machine learning model for prediction of cytoprotective effect of compounds, suggesting prevention of ZIKV-induced cell death, that could be applied to other viruses. Future structure-activity relationship studies of these compounds would generate useful information for hit-to-lead optimization and to discover compounds with direct antiviral activity. Furthermore, this work has also highlighted the importance of the integration of computational and experimental approaches, as well as the benefits of a collaborative network leveraging crowdsourcing resources and open science to advance drug discovery projects for these neglected and emerging viruses. The lack of available data regarding direct antiviral activity and cytoprotection effect of compounds tested against ZIKV in public databases such as ChEMBL and PubChem can decrease the accuracy of the computational predictions. These issues contribute to the differences observed in the computational predictions and experimental assays, and make the development of a computational approaches to ZIKV drug discovery a challenge requiring more data. Increasing ZIKV drug discovery research and making the data available in public databases would contribute to improve computational tools in future.

In an era where we are still struggling with COVID-19 we would do well to remember that Zika, and before that Ebola and Middle East Respiratory Syndrome (MERS) were earlier virus outbreaks that should have prepared us for the next epidemic and pandemic. OpenZika represents the first crowdsourced drug discovery approach for a newly discovered virus undertaken during an emerging pandemic, preceding the efforts during COVID-19, a strategy which can be applied to others to come.

Highlights.

• The OpenZika project, through the World Community Grid computational network, enabled massive docking-based virtual screening campaigns for ZIKV;

• Machine learning models were developed to predict the cytoprotective effect of compounds over ZIKV infection;

• Five non-nucleoside compounds were identified as potent ZIKV NS5 polymerase inhibitors;

• One compound was able to inhibit both ZIKV NS2B-NS3 protease and NS5 polymerase and to protect against ZIKV-induced cell death;

• Two compounds bound and thermally destabilized the ZIKV NS3 helicase, one of them was also able to protect ZIKV-induced cell death;

• Molecular dynamics simulations shed light on the binding mode of the experimental hits;

• Eight compounds were able to protect glioblastoma stem cells (GSC387) from ZIKV infection.

Abbreviations

AD

Applicability Domain

AUC

Area Under the ROC curve

BBB

Blood Brain Barrier

CC₅₀

Half-Maximal Cytotoxic Concentration

DENV

Dengue virus

EC₅₀

Half-Maximal Concentration

FN

False Negatives

FP

False Positives

IC₅₀

Half-Maximal Inhibitory Concentration

hNPCs

Human Progenitor Cells

GSC

glioblastoma stem cells

MD

Molecular Dynamics

MERS

Middle East Respiratory Syndrome

ML

Machine Learning

NS

Nonstructural Protein

NS2B-NS3pro

NS2B-NS3 protease

NS3hel

NS3 helicase

NS5 RdRp

NS5 polymerase

PAINS

Pan-assay Interference Compounds

TN

True Negatives

TP

True Positives

RMSD

Root Mean Square Deviation

SI

Selective Index

VS

Virtual Screening

ZIKV

Zika virus