Modern approaches to accelerate discovery of new anti-schistosomal drugs

Abstract

Introduction:

The almost exclusive use of only praziquantel for the treatment of schistosomiasis has raised concerns about the possible emergence of drug-resistant schistosomes. Consequently, there is an urgent need for new anti-schistosomal drugs. The identification of leads and the generation of high quality data are crucial steps in the early stages of schistosome drug discovery projects.

Areas covered:

Herein, the authors focus on the current developments in anti-schistosomal lead discovery, specifically referring to the use of automated in vitro target-based and whole-organism screens and virtual screening of chemical databases. They highlight the strengths and pitfalls of each of the above-mentioned approaches, and suggest possible roadmaps towards the integration of several strategies, which may contribute for optimizing research outputs and led to more successful and cost-effective drug discovery endeavors.

Expert opinion:

Increasing partnerships and access to funding for drug discovery have strengthened the battle against schistosomiasis in recent years. However, the authors believe this battle also includes innovative strategies to overcome scientific challenges. In this context, significant advances in in vitro screening as well as computer-aided drug discovery have contributed to increase the success rate and reduce the costs of drug discovery campaigns. Although some of these approaches were already used in current anti-schistosomal lead discovery pipelines, the integration of these strategies in a solid workflow should allow the production of new treatments for schistosomiasis in the near future.

1. Introduction

Schistosomiasis is a neglected tropical disease (NTD) caused by S. mansoni, S. haematobium, and S. japonicum, three species of Schistosoma genus parasitic flatworms, accounting for the majority of human infections. Schistosomal infections cause chronic and often debilitating disease that ends up impairing development and productivity of affected individuals, and is strongly linked to extreme poverty [1]. Recent estimates of the World Health Organization suggest that around 258 million people are infected resulting up to 200,000 deaths annually. Furthermore, the disease is endemic in 78 countries worldwide, mainly in sub-Saharan Africa, Middle East, Caribbean, and South America, where the number of cases is positively correlated to poor knowledge about the disease, poor sanitation, and a lack of effective health policies [2].

Currently, the control of schistosomiasis relies on presumptive treatment or case management with a single drug, praziquantel (PZQ), which has been used in mass drug administration programs for almost four decades [3]. However, the disseminated and repeated use of this drug in endemic areas as well as high rates of reinfection raise concerns about the emergence and evolution of drug-resistant parasites [4,5]. This problem may be further aggravated by the lack of efficacy of PZQ against schistosomula and juvenile worms [6], often a potential cause of treatment failure in endemic areas. Hence, there is an urgent need for discovering new anti-schistosomal drugs. This paper focuses on current developments in anti-schistosomal lead discovery, with particular emphasis on virtual and automated in vitro target-based and whole-organism screenings. In addition, we highlight recent progress in each area and suggesting possible solutions to existing pitfalls.

2. Challenges to discovering new anti-schistosomal drugs

The long-voiced concerns associated with PZQ argue for increased efforts to identify new anti-schistosomal candidates in drug research and development (R&D) programs. However, the decades-long availability of PZQ as a well-tolerated, affordable (or donated), oral and single dose drug, associated with low financial viability of new anti-schistosomal drugs in poor resource countries does not offer an incentive to enable the high and risk-associated investments in R&D required for the discovery of new treatments for schistosomiasis. Nonetheless, assuming that ‘it is impossible develop resistance to PZQ’ is both reckless and risky, as it may minimize the interest in R&D of alternative anti-schistosomal drugs [7]. Furthermore, while public-private partnerships have been formed for some of the NTDs, e.g., the Drugs for Neglected Diseases Initiative (DNDi) focusing mainly on human African trypanosomiasis, Chagas disease, leishmaniasis, filariasis, pediatric HIV, and mycetoma, corresponding drug discovery and development programs do not yet exist for Schistosomiasis. Consequently, nowadays the number of anti-schistosomal drug candidates undergoing clinical trials is very small.

3. Schistosome drug screening strategies

The majority of anthelminthic drugs approved for human use have been derived from veterinary medicine and discovered through in vivo screening of selected compounds in animal models [8]. Since these screens are labor-intensive and expensive, attention has shifted to developing primary in vitro screens. Typically, worms are cultured for a period of days and morphological changes (e.g., shrinkage, curling, tegumental disruption, worm disintegration) and motility (e.g., increased activity, sluggishness, or paralysis) can be determined using a predetermined scale [9]. However, manual visualization of drug efficacy in vitro is subjective, laborious, and unsuitable for high-throughput screening (HTS) [10]. This limitation can be overcome by the application of automated readouts incorporated into HTS platforms, allowing very large compound collections to be screened against relevant schistosome targets or in the whole-organism, invigorating the lead discovery pipeline. Below, we will discuss the studies highlighting the progression of screening technologies developed to accelerate discovery of new lead candidates for schistosomiasis.

3.1. Target-based screens

Target-based HTS campaigns have been emphasized in last decade as a way of harvesting the investment made in parasite genomics consortiums [11]. Such efforts eventually led to translational research-based groups and platforms taking up genome sequencing of S. mansoni [12,13], S. japonicum [14] and S. haematobium [15], resulting in new information on the parasites’ biological pathways, facilitating identification of relevant targets for therapeutic intervention, and opening new possibilities to HTS on recombinant schistosome proteins. Several changes have encouraged this evolution, including price decreases of the automation and instrumentation sector, availability of commercial compound library datasets, simplification of systems and software, and recruitment of industry-experienced personnel into research groups [16–18].

Literature examples of HTS campaigns on anti-schistosomal lead discovery field are scarce, but some HTS-adapted assays for some S. mansoni targets have been published [19,20]. For instance, Simeonov and colleagues [19] developed a HTS assay for identification of thioredoxin glutathione reductase (TGR) inhibitors, i.e., a key enzyme in redox cascade pathway and a validated target for schistosomiasis [21]. As a result, oxadiazole 2-oxides were identified as TGR inhibitors and nitric oxide donors, with inhibition activities in the low micromolar to low nanomolar range. Incubation of parasites with these compounds led to rapid parasite death. Further, the treatment of schistosome-infected mice with 4-phenyl-1,2,5-oxadiazole-3-carbonitrile-2-oxide, also known as furoxan, killing all relevant stages of S. mansoni (schistosomula, juveniles, and adults), and egg-associated pathologies. The results of these studies verify the utility of oxadiazole-2-oxides as new leads in control of schistosomiasis [22].

3.2. Automated whole-organism screens

The whole-organism schistosome screen approach (phenotypic screening) is an indispensable method to discover new anti-schistosomal agents. Compared to target-based screens, whole-organism screens have, at best, a medium throughput, but validated lead compounds identified from these methods must have been able to reach the designated target within the assayed organism by crossing several biological membranes and resisting to degradation by detoxification enzymes. Conversely, compounds obtained through target-based screening are frequently found to lack activity against the whole-organism [8,23]. In view of these advantages, improvements in robotics and automated acquisition technologies over the last decade have meant that the state-of-the-art for whole-organism screening is now fast, quantifiable, and rigorous [23]. In the following subsections, we discuss advantages and limitations of state-of-the-art medium- and high-throughput methods.

3.2.1. Isothermal micro-calorimetric assay

The isothermal micro-calorimetric assay effectively measures heat flow of biological processes (endo- or exothermic reactions) generated by worms using a 48-channel isothermal microcalorimeter to determine schistosome viability over the time. Sharp peaks in the heat flow occur as the organism moves, allowing the measurement of worm motility. This method reliably distinguishes real-time changes in metabolism and motility of S. mansoni schistosomula and adult worms after treatment with anti-schistosomal compounds. Although isothermal micro-calorimetry method seems to have many advantages, some drawbacks have to be noted. For instance, the heat-flow curves obtained reflect the overall activity of the worms. It is not possible to distinguish between worm contractions and increased movement, because both physiological processes result in increased metabolic activity with heat production. Other disadvantages includes the high cost of the microcalorimeter, the number of replicates (at least 400) required to achieve analyzable signals, and reduced throughput caused by extensive time for each analysis done in a real time fashion [24].

3.2.2. Impedance-based assay

Measurement of electrical impedance has been applied to assess motility of S. mansoni cercariae, adult worms, and hatching of eggs. The system named xCELLigence™ worm real-time motility assay [25], also known as xWORM assay [26], comprises a 96-well E-plate covered in gold microelectrodes which measure changes in conductivity due to contact of the worms with the electrodes. This technique has been favorably received in the field of lead discovery because of its sensitivity, broad applicability and adaptability, ability to measure motility under high-throughput, and minimal effort and training required [25,26].

3.2.3. Image-based assays

With advances of high-resolution cameras and computer-based image analysis programs, imaging also has become a powerful tool for demonstrating anti-schistosomal activity. Image-based assays have the advantage over the use of fluorescent labels in requiring fewer manipulation steps and avoiding use of potentially toxic fluorophores that may affect the outcome of the assay. In this context, high-content screening (HCS) microscopes/cameras are now able to capture high resolution images of schistosomula and adult worms in order to demonstrate phenotypic and/or motility changes [23].

HCS machines are so termed due to the large amounts of data produced requiring use of advanced automated image analysis software. Regarding HCS analysis for schistosomula, there has been recent progress in development of automated imaging algorithms utilizing video microscopy, so that time-dependent phenotypic changes following drug treatment could be monitored [27,28]. Of special interest, is the well validated high-throughput system developed by Paveley and colleagues [29]. This system includes compound stamping, robotic integration, incubation, parasite resuspension, and automated image analysis. Test compounds are cultured with schistosomula for 3 days in 384-well plates; then well content is first re-suspended by repeat aspiration to redistribute individual larvae across the field of view and to disrupt clumping of schistosomula. The HCS then collects 4-tiled objective images for phenotype analysis and five time-lapse images for motility analysis (with 4x objective). The image analysis algorithms integral to the HCS are then applied in three distinct stages. The first step in this algorithm is to identify sufficient schistosomula or to segment them from the background image using basic segmentation algorithms to detect individual larvae. Following successful segmentation, a phenotype score for individual larvae is assigned by a Bayesian classification model built using a dataset of 20,000 manually phenotyped schistosomula treated with a range of known anti-schistosomal compounds. Finally, to assess larvae motility, a score is produced for each larva based on the change in area between successive time-lapse images which is averaged across the well [23,29].

In contrast, WormAssay was developed to quantify motility of adult worms. This system is based on dark-field microscopy imaging, where plates are illuminated from the side. This enhances the contrast between the background and the worms, allowing a more efficient detection of the parasites, and consequently generates ideal images for recording and analysis. Images comprising the entirety of the plate are collected, and a further analysis assesses movement of all worms within each well by two separate algorithms based on contour velocity and area occupation. As result, WormAssay quantifies each worm’s movement simultaneously on the entire plate, with each plate taking approximately 30 seconds to 1 minute to read. The system is cost-effective and capable of screening a large number of compounds. The application supports 6-, 12-, 24-, 48-, and 96-well plates, but the assay throughput is currently limited by number of worms produced and delivered [30].

3.3. Luminescence- and fluorescence-based assays

During the last years, notable advances were obtained for detection of schistosomula viability through quantitation of luminescence and fluorescence emitted by cellular markers and metabolites. These assays can be read by an automatic plate reader, making a cheaper, simpler, more practical and more trainable read-out alternative, requiring little extra software or equipment [31]. Currently, four fluorescent assays (i.e., (i) Alamar blue [32] and (ii) resazurin [33] reagents; (iii) the fluorescein diacetate/propidium iodide multiplex assay [34]; and (iv) L-lactate commercial kit [35]) and luminescent assays based on quantification of adenosine triphosphate [31,36] have been studied. Although these assays are viable tools in anti-schistosomal lead discovery, they have some limitations. For instance, the Alamar blue assay [32] could achieve signal to concentration tests in schistosomula only after 24 hours of incubation with the marker, while the assay developed by Marxer and colleagues [33] using resazurin failed to generate dose-dependent viability curves and to distinguish between live and dead schistosomula after 72 hours of drug exposure. Fluorescein diacetate/propidium iodide assay [34] requires high number of schistosomula to assess viability in a drug sensitivity assay, which reduces its throughput. The assay developed by Howe and colleagues [35] also has a reduced throughput, since screening involves measurement of L-lactate concentrations by removal the supernatant of the well plate (without aspirating the schistosomula) and then diluting it to an acceptable fluorescence range.

4. The open data era

The current strategy to discover lead compounds consists on a data-driven process that is reliant on producing bioactivity data for large chemical libraries in high-throughput campaigns [37]. However, the creation of automatized schistosome screening platforms is beyond the reach of most researchers due to its high cost and infrastructure requirements. To make in vitro tests less expensive, researchers have used data retrieved from assays to explore computer-aided drug design (CADD) approaches. Currently, several datasets exclusively dedicated to lead discovery are freely available in publicly databases, such as PubChem Bioassay [38], BindingDB [39], and ChEMBL [40]. The data related to schistosomiasis could also be found in these databases. Deposited data include information regarding compounds, their bioactivities tested against a specific target or schistosome life stage, and the description of the biological assay extracted from the reference paper. Full exploitation of this rich source of data is of high interest to researchers in the field of CADD to explore the anti-schistosomal chemical space [41,42].

5. Virtual screening

Virtual Screening (VS) has emerged as a powerful and straightforward computational method to guide the identification of new hits from large chemical libraries. In principle, this method is often compared to a funnel, where a large number of compounds in chemical libraries (i.e., 10⁵ to 10⁷ compounds) are reduced by a computational tool to a smaller number that will then be tested experimentally (i.e., 10¹ to 10³ compounds) [43,44]. In general, VS approaches do not hold the drawbacks that are characteristic of experimental approaches, such as accessible chemical space, time, automatization level, or number of worms produced. Nevertheless, the expectation that VS could completely replace in vitro or in vivo approaches is over-optimistic. In addition, typical hit rates from in vitro approaches can range between 0.01% and 0.14%, while hit rates for prospective VS typically range between 1% and 40% [45]. Thus, compounds from the subset that pass the initial VS are found to be biologically actives at a higher rate and at a lower cost. Because of their obvious advantages, VS approaches are widely employed in pharmaceutical industry and academic organizations. Nevertheless, in anti-schistosomal research, the application of VS approaches is limited by a small number of recent studies [46–50]. The main VS approaches that can be implemented in a hit identification pipeline and examples of successful applications of VS workflows leading to the identification of new anti-schistosomal hits are discussed in next sections.

5.1. Structure-based virtual screening (SBVS)

SBVS is becoming an essential tool in assisting fast and cost-efficient lead discovery. The application of SBVS strategies in schistosome lead discovery facilitates understanding the molecular basis of parasite biology and development and utilizes the knowledge of the 3D structure of the biological target in the process to select ligands with acceptable affinity and complementarity with the binding site. In this context, genomic and proteomic studies have provided basis into the identification of validated schistosome targets [51,52]. Currently, 3D structures of 35 different schistosome targets have been solved by X-ray crystallography and nuclear magnetic resonance (NMR) and stored in public databases such as the RCSB Protein Data Bank (PDB) [53], leading to attractive opportunities for the application of SBVS strategies. The main SBVS approaches are discussed below.

5.1.1. Protein-ligand docking

The most extensively used SBVS method is docking, a tool that involves fitting the ligands into the binding site of a 3D structure in order to predict their binding affinities [54]. Generally, docking calculations can be accomplished using two types of algorithms: the search algorithm and the scoring function. The search algorithm generates the various possible poses to fit the ligand into the binding pocket of the receptor. The most known search methods are: (i) Monte Carlo and genetic algorithms; (ii) molecular dynamics (MD) and energy minimization methods; and (iii) systematic methods. Scoring function ranks the different poses and locations of the ligand that are generated by the search algorithm, and orders them by a score. Commonly used scoring functions can be categorized as follows: (i) empirical scoring functions; (ii) force field-based functions; and (iii) knowledge-based functions. The discussion about search algorithms and scoring functions is beyond the scope of this review, so detailed information can be obtained elsewhere [52,54,55].

5.1.2. Structure-based pharmacophores

Structure-based pharmacophores complement docking procedures, including the same level of information, but are less computing-demanding. In structure-based pharmacophore approach, pharmacophore models are derived from the 3D structure of the protein. This method can work both with a free (apo) structure or a ligand-target complex (holo) structure. The generation of models using a holo structure allow complete exploration of interactions with binding site, and inclusion of shape and volume information derived directly from the structural data. In absence of co-crystalized ligand in binding site, the complex can also be obtained by a docking approach, where a small ligands’ binding orientation in the protein is predicted [56,57]. For apo structures, pharmacophores can also be obtained using functional groups or small fragments, also referred as molecular probes, to map possible interaction sites or hot spots within a binding site. Selected hot spots can then be converted into pharmacophoric features in order to generate a structure-based pharmacophore model [57].

5.1.3. Piggy-back

The ‘piggy-back’ is a useful strategy for exploring targets studied in other diseases (human or parasite either) once it facilitates the development of new drugs against homologous targets in other organisms. The efforts in finding validated drug targets and identification of their homologs in schistosome provide evidence of target druggability and may offer new chemical scaffolds, which can be optimized for the schistosomiasis [58]. In this context, gene ontology informations available for schistosomes can be used in order to compare targets, and consequently, identify lead candidates, based on their sequence motifs and binding sites similarities or using 3D structural information. In many cases, these approaches focus on those residues that are known from experimental molecular recognition studies or based on prediction of functional residues. For instance, we use sequence alignment to identify several S. mansoni macromolecules with homologous human and microbial targets [59]. As a result, several drugs with good bioavailability were predicted as new drugs for treatment of schistosomiasis, opening new opportunities for experimental validation [59].

5.2. Ligand-based virtual screening (LBVS)

LBVS is the approach of choice when the biological target is not known or its 3D structure is not available. In this approach, prior knowledge of a reference set of compounds with biological activity against a target is used to identify structural rules to identify new lead candidates for experimental evaluation [60]. There are four main LBVS methods: (i) similarity search; (ii) ligand-based pharmacophores; (iii) 3D shape matching; and (iv) QSAR.

5.2.1. Similarity search

Similarity search is used to identify new compounds by measuring the level of their structural similarity to the known active compounds. This approach is based on the assumption that structurally similar compounds have similar biological properties. Here the compounds are typically represented by molecular fingerprints [61], a high-dimensional vector of bits that accounts for the presence (1) or absence (0) of a fragment in a compound, turning it into a sequence of bits that can then be easily compared with other bit sequences. This comparison must then be expressed in a way that can be quantified [61]. There are many ways to assess the similarity between two vectors; the most common is Euclidean distance [62]. Whereas for molecular fingerprints, the standard estimation of similarity method is the Tanimoto coefficient, which is equal to the number of common bits set to 1 in both fingerprints divided by the total number of bits set to 1 between both fingerprints [62].

5.2.2. Ligand-based pharmacophores

Ligand-based pharmacophores are commonly used when the 3D structure of the biological target is absent. They may be explored using a set of distance constraints of known ligands to locate the relative positions of common chemical features, such as hydrogen-bond donors or acceptors; aromatic rings; partial charges; and hydrophobicity, in order to identify key common features and the relative orientations of known active ligands not shared by inactive. This method delivers good results if enough ligand information is available and if the dataset compounds are known to bind to a target in the same way. One the other hand, conformational flexibility represents one of the main difficulties in pharmacophore generation, because the same ligand could have several productive conformations, which are usually unknown [63].

5.2.3. Shape-based method

Shape-based approach is an established and effective method for identifying hits that are similar in shape and bioactivity to a reference ligand [64]. The shape of a compound refers to the part of space occupied by the chemical structure as determined by its external boundary, abstracting from other aspects the object can have such as its color, atom type, and size [65]. In biological context, shape is a fundamentally important molecular feature that often determines the fate of a compound in terms of molecular interactions with biological targets. Most ligands bind to their target through van der Waals interactions, and hence desolvation effects dominate the binding free energy. Unlike hydrogen bonds and electrostatic interactions, the strength of these hydrophobic interactions is mainly governed by the proximity of the respective electron densities. It is assumed that compounds sharing similar 3D surface shapes are likely to share similar biological activities [66]. Thus, 3D LBVS shape matching tools could offer a ‘scaffold hopping’ way to find new anti-schistosomal leads. Nonetheless, conformational flexibility remains the same obstacle as for ligand-based pharmacophores.

5.2.4. QSAR

Quantitative Structure-Activity Relationship (QSAR) analyses have been increasingly used in VS because of its high speed of screening and good hit rate. Once developed on a set of compounds with known activity, QSAR models are applied to untested chemical compounds for numerical prediction of biological activity (continuous models) or the discrimination between active and inactive compounds (binary models) [67]. In QSAR molecular fingerprints or descriptors are calculated for chemical compounds and correlated with a determined endpoint (experimental data) using a machine learning method. Currently, several machine learning methods, such as Random Forest [68] and Support Vector Machine [69], among others, are available to build QSAR models. The influence of various factors on QSAR performance is decreasing in the following row: data quality > molecular descriptors > machine learning approach [70].

5.3. Successful VS stories in anti-schistosomal hit discovery field

In the current scenario, VS remains a poorly explored strategy in anti-schistosomal R&D field. Despite the lower rate of examples available in literature, most of VS works led the discovery of hits with potency or affinity values in the low micromolar range. Below, we described the successful identification of potent anti-schistosomal hits by use of two different VS workflows.

Among these studies, the identification of histone deacetylase 8 (HDAC8) inhibitors, a target that catalyze posttranslational modifications in schistosomes could be cited [46]. Recent studies [71,72] showed that exposure to generic histone deacetylase inhibitors led to protein acetylation and dose-dependent mortality of schistosomula and adult worms. Initially, Kannan and colleagues [46] built a homology model for S. mansoni HDAC8 based on the available X-ray structures of the human orthologue. Next, MD simulations were employed to optimize structure and evaluate the structural flexibility of the residues at the binding site of the build model. The optimized homology model was then used in a docking-based VS the drug-like subset of the ZINC database. At this stage, authors also refined chemical space of the chemical library using sub-structure filters (i.e., hydroxamate, anilinobenzamides, or thiazole-sulfonamide groups), which are known as zinc binding groups, based on information derived from human HDAC8 inhibitor complexes. Next, 75 predicted compounds were selected based on their intermolecular interactions with the zinc ion located at the binding site and docking scores. Identified hits was then tested for estimate its inhibition activity on human and S. mansoni HDAC8, resulting in the identification of eight hydroxamate derivatives as potent and lead-like inhibitors of the parasitic enzyme. Solving of the X-ray structure of S. mansoni HDAC8 with two of the VS hits, authors also confirmed the predicted binding mode.

In another study, Jacques and colleagues [50] describe a VS approach to identify inhibitors of S. mansoni NAD+ catabolizing enzyme (NACE), a target suspected to be involved in immune evasion by the parasite at the adult stage. An initial filtering biased of the chemical library toward compounds was performed to select small molecular weight compounds, in agreement with the small size of the binding site of NACE. Docking of filtered compounds into a homology model of the enzyme has led to the discovery of 1,701 virtual hits. Then, best poses were filtered using a machine learning model developed using a training set of positive and negative reference binding modes described with protein-ligand interaction fingerprints. The automated selection protocol returned 188 compounds. A visual inspection rejected 90 compounds because of high similarity between structures. Predicted compounds were then tested for estimate its inhibition activity on S. mansoni NACE, resulting in the identification of six hits active in nanomolar concentrations, accompanied by a largely enhanced selectivity for the parasitic enzyme over the human homologue CD38.

5.4. Best practices in virtual screening

Prior to performing any VS workflow, it is important to consider many pitfalls that can influence the erroneous selection of inactive compounds, also known as false positives [73,74]. It is well known that the success of any VS campaign largely depends on development of integrated approaches, combining different search methods as well as structure- and ligand-based methods, along with use of additional filters and statistical approaches to improve the selection of hits. In the following subsections, we will give some important recommendations to increase the reliability of a VS campaign.

5.4.1. Quality of dataset

Publicly available datasets contain a fraction of erroneous records, which presence is caused by lack of or incomplete data curation, measurement variations, and insufficient quality check. In short, these ‘bad’ data encompass entries containing unclear stereoisomer annotations, wrong chemical structures, experimental values, and activity units, and duplicated records [75]. The data compiled from different laboratories under possibly different experimental conditions also add certain discordance [76]. Therefore, the concordance analysis between original and compiled bioactivities as well as structure standardization procedures should be applied for the data to ensure consistency and quality of the data [77].

Pan-assay interference compounds (PAINS) could also lead the VS campaign astray. The apparent activity of PAINS is typically caused by their reactivity rather than non-covalent binding. Such compounds typically interact nonspecifically with proteins of buffer in a high percentage of bioassays [78]. Time and money are consequently wasted in attempts to optimize the biological activity of these compounds. Therefore, the use of computational filters for identifying and removing PAINS in a VS workflow is highly recommended [78].

5.4.2. Balancing the dataset

Unbalanced datasets have been widely discussed in QSAR-based VS field [79,80]. Most classifiers assume equal weighting of the classes in terms of both number of instances and the level of importance.. However, prediction a minority class of an unbalanced dataset results in a large number of erroneous predictions. Two following strategies can be used for unbalanced datasets: dataset balancing and cost-sensitive classifiers. Dataset balancing could be employed by either removing some compounds from the majority class (i.e., under-sampling) or adding some artificially-generated compounds to the minority class (i.e., over-sampling) [79,80]. The cost-sensitive classifiers were used to reduce the level of importance of majority class, a strategy in which samples are predicted to have the class that has the lowest expected cost [81].

Other VS approaches, such as structure- and ligand-based pharmacophore models, docking-based and shape-based models are highly sensitive to balanced datasets or datasets with lower rates of inactive compounds. In these specific cases, it is highly recommended the use of unbalanced datasets at a standard ratio of 36 inactive compounds for each active compound [82]. This data distribution corresponds to what would be found in a completely unbiased approach, i.e., the HTS of a large library of compounds, which could be termed as the ‘natural’ distribution.

5.4.3. Protonation and tautomerism

Ionization and tautomerism may significantly change properties in the ligand, and therefore will have an impact in descriptor calculation, definition of pharmacophore interaction points, and interaction possibilities between ligand and a target protein in docking simulations. Therefore, those phenomena could have an important effect on CADD. Consequently, accurate and relevant assignment of protonation and tautomeric states is crucial in order to match the biological system targeted in the screening process [74,83].

5.4.4. Protein flexibility

Flexibility of the target binding site is an essential but frequently overlooked aspect in structure-based studies. Proteins are flexible, which may be stabilized by ligand binding in one conformer out of an ensemble of conformations of similar energy in the unbound state. Taking into account the flexibility of the protein by docking programs is still an area of active development [84]. The ideal approach to increase protein flexibility would be screening of all possible conformations of a ligand against the full degrees of freedom of the protein structure, using MD simulations. Unfortunately, such an approach is infeasible because of the high cost of computing resources [85,86]. Currently, two streamlined strategies have been developed to solve this problem: ensemble methods and induced-fit docking. To incorporate protein flexibility, ensemble methods make use of multiple input conformations of a target using a set of different X-ray and NMR data available on PDB or extracted from multiple time steps produced by MD or Monte Carlo simulations [87,88]. The induced-fit docking allows for conformational search of the protein and the ligand is performed in parallel (i.e., both conformations are altered at the same time). To limit the combinatorial explosion of conformations that need to be generated and energetically evaluated, a small number of degrees of freedom are considered. Variables that characterize the magnitude of the conformational change along the selected degrees of protein flexibility are optimized in parallel to the ligand degrees of freedom, such as translational, rotational, and torsion changes [89].

5.4.5. Statistical validation of VS models

The quality of a VS model is mostly measured in terms of their ability to discriminate between known active and inactive compounds experimentally evaluated against a target or a whole-organism. Irrespective to the approach chosen, the VS model should be rigorously externally validated before applying it to prospective VS. The most useful statistical metrics to measure the quality of developed models for VS are: sensitivity (SE), specificity (SP), positive and negative prediction values (PPV and NPV), correct classification rate or balanced accuracy (CCR or BA), Cohen’s kappa (κ), receiver operating characteristic (ROC) curve, area under the ROC curve (AUC), determination coefficient (Q²), enrichment factor (EF), and Boltzmann-enhanced discrimination of ROC (BEDROC). The choice of these metrics depends on the approach used; for instance, CCR, SE, PPV, etc., are commonly used for QSAR, while BEDROC, AUC, EF, etc., are usually used for docking and pharmacophore modeling. We also want to remind that these metrics are just an indicator that the model could be useful for VS, but its real quality will be determined only by experimental testing of selected VS hits. In a previous paper, we have extensively discussed those and several other metrics for VS validation [90].

5.4.6. Consensus modeling

The main reason for errors for selection of inactive hits in VS lies in the usage of a single classifier or a single scoring function that very often fail to predict the activity class or binding affinity of screened compounds. This could lead to the selection of many false positives that will be erroneously placed in the top scorers of a ranked list. In an effort to optimize VS models, the consensus strategy could be applied by averaging the predicted values from the individual models, providing better statistical fit and predictive ability as compared to the individual models. This can be very useful, as it combines the advantages of different approaches and simultaneously attenuates the prediction errors of each classifier [91].

5.4.7. ADME/Tox filters

It is well established that poor pharmacokinetics and toxicity (ADME/Tox) properties represent one of the main reasons drug candidates failing in clinical trials. However, during the last decade, combinatorial chemistry and HTS have significantly increased the number of compounds for which early data on ADME/Tox are needed, and, therefore, has increased efforts and costs in the lead discovery process [92]. Obviously, the use of computational tools cannot replace the experimental screens, but they reveal an early ADME/Tox profile in VS campaigns. In this context, our group has been working to overcome or reduce these failures by developing computational tools and web services for the early prediction of some ADME/Tox properties [74,93,94]. Our knowledge points to the importance of these filters in the early stages of hit-to-lead discovery campaigns, which helps in the selection of virtual hits before acquisition and biological evaluation. However, usually the application of all possible ADMETox and other possible structural filters could result in elimination of all VS hits [95]. At first glance, this statement seems to be paradoxical, however, all computational approaches are just an instrument developed to help the researcher but not to substitute his expertise and knowledge.

5.4.8. Integrated VS workflows

The development of integrated workflows, i.e., combining several VS methods and computational tools, is also a useful approach to increase hit rates during experimental validation. Compilation of these workflows could be either in sequential or parallel forms. In the sequential approach, different methods are used in a pipeline to sequentially filter the number of hits retrieved, until the number is small enough for extensive biological testing. Often, methods like pharmacophore screening are used in the beginning of the multi-step screening process. As the number of hits decreases, computationally more expensive methods can be applied to further filter the retrieved compounds, such as docking. In the parallel approach, different methods are run independently and the predicted hits in both screens are combined in a consensus approach for biological testing. The methods used should be complementary and ideally should include ADME/Tox filters, pharmacophore models, QSAR models, similarity-based models, as well as docking-based models [96,97].

6. Conclusions

In conclusion, we emphasize the importance of integrating modern experimental and computational approaches to accelerate schistosomiasis drug discovery. We propose a workflow, presented in Figure 1, showing the possible road map to integrate such approaches. Combinatorial chemistry and organic synthesis can provide a large number of compounds to be screened using automated assays, for the whole organism in different life stages or for selected schistosome targets. The large amount of data generated from these HTS/HCS assays and public availability of this information started the era of big and open data. These data boost the use of computer-assisted technologies for VS, allowing generation of predictive models for prioritization of VS hits or structural design of novel compounds with desired properties. Further, these computational hit are assayed in the target-based, whole-organism, luminescence or fluorescence-based platforms. Successful integration of all these technologies can accelerate the discovery of new anti-schistosomal leads that could ultimately enter clinical trials.

Figure 1.

Suggested antischistosomal lead discovery workflow in open data era.

7. Expert opinion

Despite the urgent need in finding a new treatment for schistosomiasis and the relative success that has been recently achieved, defeating this NTD may still be far reaching. There are multiple potential ways of battling schistosomiasis. The first one is the development a vaccine. At present, there are ongoing clinical trials regarding several vaccines against various Schistosoma strains [98]. To this purpose, a vaccine against Schistosoma haematobium is under Phase III of clinical trials in Senegal and Niger and two vaccines against Schistosoma mansoni are under Phase I of clinical trials in US and Brazil [98]. However, despite this, there are numerous challenges in vaccine development related to protective antigen discovery, product development, and the preclinical and clinical testing stages. For instance, reverse vaccinology approaches that work well for small bacterial and viral genomes have not yet been successful for eukaryotic parasites because of the greater complexity of the parasite genome and the requirement to employ eukaryotic expression systems to reliably produce soluble and properly folded antigens. In addition to scientific challenges, socioeconomic issues also present a bottleneck for the development and implementation of new vaccines that should be available to those who need them most, especially in developing countries in South America, Africa, and Asia.

Another way of tackling schistosomiasis is the development of new drug(s) which, ideally, should be active against all the different species of Schistosoma. Thus, such drug(s) need(s) to be promiscuous enough to cover a variety of species and selective enough not to hit other targets in human cells, because the interaction with some of them could lead to strong undesirable side effects. Although this is likely to be achievable, as in the case of vaccines, similar socioeconomic challenges could limit the development and distribution of such drug in poor resource countries. Another interesting possibility is the use of combination therapy, which could help overcome scientific bottlenecks such as drug resistance, toxicity, etc. Despite the fact that combinatorial therapy may eventually lead to undesired drug-drug interactions, careful planning of drug combinations will easily circumvent this issue

One of the solutions to decrease the potential economic burden related to drug R&D for schistosomiasis is the use of computational techniques for targeted design of new chemical entities that will decrease both the time and costs associated. Expected future development of combined therapy will in turn boost the development and/or optimization of reliable computational tools for handling and analyzing mixtures [99]. In addition, these approaches should take into account both synergistic and antagonistic effects. We strongly believe that socioeconomic battle against schistosomiasis emerged with London Declaration of 2012 as evidence of willingness to advance through partnerships and provision of funding to R&D [100], but the future battle may include novel approaches that will help to overcome scientific challenges. In conclusion, the era of big and open data has just started and data is increasingly accumulating. Integration and curation of these data will greatly improve the success of drug discovery efforts. Furthermore, development of novel computational approaches capable of handling millions of compounds is essential, in order to avoid that the amount of data generated exceeds the amount of data that can be processed. In turn, these improved tools may also contribute to reducing the number of animals that are required for in vivo testing of new compounds. Finally, we note that even the best computational and VS approaches are just the tools and the overall success of the project still depends only on the ability of a scientist to think critically, to understand the advantages and disadvantages of computational and other tools, to perceive the final goals of his research, and to interpret the consequences of his discoveries.

Article highlights box.

• The possible evolution of praziquantel-resistant schistosomes and the lack of new effective drugs call for a paradigm shift from contemporary to modern anti-schistosomal lead discovery strategies;

• The main advances in high- and medium-throughput lead screening technologies have opened up new possibilities for anti-schistosomal drug discovery pipelines.

• Data produced on extract bioactivity from medium- and high-throughput assays, i.e., open data available on publicly databases, can now be explored in a computational data-driven process.

• Both experimental (i.e., target-based assays, automated whole-organism assays, and luminescence- and fluorescence-based assays) and computational (i.e., structure-based and ligand-based strategies) drug discovery approaches have strengths and pitfalls, but carefully planned integration of these strategies could lead to new treatments for schistosomiasis in the near future.

Abbreviations

NTD

neglected tropical disease

PZQ

praziquantel

R&D

research and development

DNDi

Drugs for Neglected Diseases Initiative

HTS

high-throughput screening

TGR

thioredoxin glutathione reductase

HCS

high-content screening

CADD

computer-aided drug design

SBVS

structure-based virtual screening

LBVS

ligand-based virtual screening

HDAC8

histone deacetylase 8

NACE

NAD+ catabolizing enzyme

QSAR

quantitative structure-activity relationships

PAINS

pan-interference assay compounds

MD

molecular dynamics

CCR

correct classification rate

ROC

receiver operating characteristic

AUC

area under the ROC curve

EF

enrichment factor

BEDROC

Boltzmann-enhanced discrimination of ROC

ADME/Tox

pharmacokinetics and toxicity