Repurposing of Anti-Infectives for the Management of Onchocerciasis Using Machine Learning and Protein Docking Studies

Abstract

There is a need to improve the discovery of new drugs for neglected tropical diseases (NTDs), as the lack of financial incentives has slowed their development. Currently, ivermectin and moxidectin are used in the management of onchocerciasis. We present a proof-of-concept study based on computational methods to find anti-infectives that can be repurposed or serve as lead compounds for onchocerciasis. A combination of exploratory data analysis, machine learning (ML), and molecular docking studies was used to evaluate 58 anti-infective agents. Out of the 58 test drugs, 14 were predicted by at least 5 ML models to be potentially useful in managing onchocerciasis. Molecular docking studies with the 14 predicted drugs using glutamate-gated chloride channel, a known target of ivermectin, an onchocerciasis drug, yielded good results. Cridanimod, diminazene, and vandetanib were the top 3 agents showing the highest binding affinities of −7.8, −7.2, and 7.1 kcal/mol, respectively, higher than the native ligand glutamate, which has a value of −4.5 kcal/mol. The binding interactions of these agents also showed overlaps with that of doramectin and pyrvinium agents that have demonstrated activity against onchocerciasis and ivermectin, the gold standard for onchocerciasis management. This study highlights the potential of cridanimod, diminazene, and vandetanib as promising candidates for developing new treatments for onchocerciasis.

Introduction

Neglected tropical diseases (NTDs) affect low-income populations in the developing countries of Africa, Asia, and the Americas.^1-3 Approximately 1 billion people, predominantly in developing countries living in remote rural areas, urban slums, or conflict zones, are at risk of NTDs.^3,4 Onchocerciasis (river blindness or Robles’ disease) is a skin and eye infection caused by Onchocer meca volvulus, a parasitic worm.^5,6 The disease is transmitted through the bites of infected blackflies, and symptoms of the disease include bumps under the skin, extreme itching, psoriasis, and loss of skin elasticity, among others. The most severe manifestations are ocular lesions that can progress to visual impairment and blindness. ⁷

Ivermectin remains the main drug used to treat the disease.^8-10 Side effects such as neurotoxicity, central nervous system depression, and liver toxicity have been reported.^11,12 Studies conducted in Ghana revealed that subjects repeatedly treated with ivermectin exhibited sub-optimal responses to the drug. ¹³ Even though several approaches are used to control the spread and the management of onchocerciasis, research on finding new therapeutic agents is not actively pursued. ¹⁴ Other drugs used in the past for managing onchocerciasis included diethylcarbamazine, suramin, and moxidectin.^15-23 Suramin use had to be discouraged due to the risk of optic atrophy, ¹⁶ while diethylcarbamazine is less effective even at high doses with frequent and severe side effects. ²⁴ Pharmacokinetic studies on moxidectin show that it is an attractive long-acting therapeutic option for treating human onchocerciasis ²⁵ and was approved by the United States Food and Drug Administration (FDA) in 2018.^22,23

New drug discovery is confronted with many challenges such as the increasing stringency of regulatory agencies in evaluating the efficacy of new drug candidates.^26-29 Research into finding new chemotherapeutic agents for NTDs is highly unattractive because endemic areas are largely poverty-stricken. Due to the onerous nature, associated risk, and low return on research and development investments, there is limited interest in developing medications for treating NTDs, including onchocerciasis. ²⁸

Drug repurposing/repositioning is finding new therapeutic indications for existing drugs.^25,26,30,31 This reduces the time of drug discovery, as information on doses and toxicities is already established. Repurposing of existing drugs implies that research associated with all pre-clinical work is also exempted.^25,31 As an alternative, it offers scientists and pharmaceutical companies an efficient strategy to identify novel targets and uses for already approved pharmaceutical agents. This provides the advantage of reducing the effort, time, and cost of every drug development stage for an area plagued by a need for affordable and effective therapies. ³¹ In fact, the latest drug moxidectin is a veterinary anthelmintic which has been repurposed to manage onchocerciasis. Even though high-throughput screening allows for the screening of several thousands of agents in a relatively short time, techniques that can weed out agents with low probability of being hits will be beneficial.

Machine learning (ML) has emerged as a transformative tool in drug discovery and development. The ML leverages computational power to analyze complex data sets and identify patterns that would be challenging to discern using traditional methods. Recently, the use of ML methods to screen compounds in databases to find new indications has become popular. ²⁷ ,^32-35 Among others, concepts such as artificial intelligence, deep learning, and re-enforcement learning have been used.^32,33,^36-39 Halicin, an effective antibiotic, was discovered using deep learning approaches. ³⁷ The ML methods have been used to identify drug candidates, which have been repurposed for managing diseases such as Alzheimer disease, cancers, and COVID-19.^32,33,36,^40-44 Unfortunately, there is a dearth of literature on the use of ML methods to identify potential drug candidates that can be repurposed for NTDs such as onchocerciasis.

The interaction between a drug and its biological target is fundamental to its pharmacological and therapeutic effect. ⁴⁵ In silico studies with molecular docking predict the preferred binding orientation and affinity of a ligand (drug) with its target, thereby aiding drug discovery and design.^41,46 It involves pose prediction between ligands and the target and scoring to estimate binding affinity. It has been utilized to predict potential therapeutic agents for the management of NTDs, including leishmaniasis. ⁴⁷

Antionchocerciasis drugs like ivermectin specifically bind to glutamate-gated chloride channels (GluCl) in Onchocerca volvulus.^48-50 This causes an increase in chloride ion permeability and the subsequent hyperpolarization of nerve and muscle cells, leading to paralysis and death of the parasite. ⁵¹ These channels are absent in humans, ensuring selective toxicity. ⁵² Hence, chemical agents that selectively bind to GluCl have a strong potential of having activity against Onchocerca volvulus.

In this proof-of-concept study, we used exploratory data analysis (EDA) and ML techniques to screen several anti-infective agents to identify potential candidates that may have activity against onchocerciasis. Subsequently, molecular docking studies were done on potential candidates to determine their binding interactions with GluCl, a known target of doramectin, a drug used to manage onchocerciasis.

Methods

Data set generation and pre-processing

Three-dimensional (3D) structures of 602 drugs, mostly anti-infective drugs, were obtained from online repositories, including DrugBank, ⁵³ PubChem, ⁵⁴ and ChemSpider. ⁵⁵ Where the 3D conformers of the compounds were unavailable, their 2-dimensional (2D) conformers were converted to their optimized 3D structures using WebMO (version 21.0). ⁵⁶ Structural geometry was optimized using geometry using Restricted Hartree-Fork (RHF) method and 6-31G(d) basis set. The web-based cheminformatics platform, online chemical modeling environment (OCHEM), and AlvaDesc (version 2.0.4) were used to generate molecular descriptors.^57-59 The descriptors are broadly two main categories: 2D descriptors, comprising over 4000 parameters such as constitutional descriptors, topological indices, and functional group counts, among others; and 3D descriptors, which include autocorrelation descriptors, weighted holistic invariant molecular (WHIM) descriptors, and geometry, topology, and atom-weights assembly (GETAWAY) descriptors. The complete list of calculated descriptors is provided at www.alvascience.com/alvadesc-descriptors. This yielded a data set matrix of 602 rows × 5666 columns (rows × columns). The data set was thoroughly examined and confirmed to be free of duplicate entries. Data analysis, including ML models, was performed using in-house tools in Python.

Descriptors with recorded values in less than 30% of the samples were excluded, reducing the total number of descriptors from 5666 to 4474. Of the 602 samples, 16 are/were used (or have demonstrated activity) in the management of onchocerciasis, while 588 were not. The data thus comprised 2 classes, which are onchocerciasis and non-onchocerciasis drugs, which were labeled oncho⁺ and oncho⁻, respectively. The data were mean-centered and auto-scaled to unit variance.

Data analysis

Linear discriminant analysis and cluster analysis (k-means)

Using the scaled data, linear discriminant analysis (LDA) was performed as a dimensionality reduction step using the singular value decomposition solver. The indication of the drugs was used as the target for the LDA model. These indications were obtained from the drugbank.com database. The indications of these drugs were antiviral, antibacterial, antiseptic, antiprotozoal, antifungal, anthelmintic, pesticides, statin, antineoplastics, antiemetics, antipsychotic, and anti-inflammatory agents (as shown in Supplemental Figure S1). This was followed by cluster analysis using the k-means cluster algorithm to identify drugs like the oncho⁺. The pairwise distance between the oncho⁺ drugs and all other anti-infectives was determined. Five anti-infectives closest to each onchocerciasis drug were selected and put into a test set. In total, 70 samples were selected, but after the removal of duplicates, the 56 remaining were captured and put into a test set. This reduced the training set data to 546 (ie, 16 oncho⁺, 530 oncho⁻, 56 test drugs). Doramectin and pyrvinium agents demonstrating activity against onchocerciasis were added to the test drugs to evaluate our proposed routine. Hence, the distribution of the data set for subsequent analysis became 14 oncho⁺, 530 oncho⁻, and 58 test drugs (including doramectin and pyrvinium).

Class imbalance and machine learning

The imbalance in the data, (14 oncho⁺, 530 oncho⁻) was corrected using the synthetic minority oversampling technique^60,61 (SMOTE). The SMOTE was implemented using the imblearn python package version 0.12.1. The SMOTE operates by creating synthetic samples along the line segments joining k-nearest neighbors of the minority class in the feature space. The SMOTE was applied on the latent value (LV) obtained after LDA. A total of 516 artificial compounds were generated to balance the data. Thus, the final data set for subsequent ML modeling was 1060 (530 oncho⁺ and 530 oncho⁻).

The data set was split into 2/3 (707 drugs) and 1/3 (353 drugs) for training and validation sets, respectively. Ten supervised learning algorithms, namely, support vector machines (SVMs), random forest (RF), logistic regression (LR), k-nearest neighbor (KNN) classification, stochastic gradient boosting (SGB), boosted decision trees (BDT), naïve Bayes classifiers (NBC), AdaBoost (ABC), multi-layer perceptron (MLP), and Gaussian Naïve Bayes (GNB), were used. These ML models were generated with the training set data using scikit-learn. Hyperparameter tuning was performed automatically to achieve the optimum model. The models’ performance was evaluated using their prediction accuracy as defined by Long et al ⁶² on the external validation set. The models with prediction accuracies of at least 80% were used to evaluate the 58 test compounds.

Protein target retrieval and virtual screening (molecular docking)

The binding affinity of test drugs predicted to have oncho⁺ activities was evaluated using molecular docking studies. Glutamate-gated chloride channel (PDB ID: 3RHW, UniProtKB Entry: Q25634), ⁶³ a known target of doramectin, was used. The PDB files for the crystal structures of GluCl were downloaded from the RCSB Database (https://www.rcsb.org/). The retrieved structure was cleaned up with Discovery Studio version 21.1.0 (BIOVIA, San Diego, California) ⁶⁴ to correct issues associated with incomplete structures due to missing atoms or water and the presence of multimers and interaction partners of the receptor molecule.

Virtual screening was performed using the Python prescription virtual screening tool (PyRx 0.8) ⁶⁵ containing AutoDock Vina module. ⁶⁶ The prepared protein structure was fed into the PyRx tool along with the 3D conformers of the test drugs predicted to be oncho⁺. The drug and protein molecule was converted to a PDBQT file using the AutoDock module of the PyRx tool. A grid box was then constructed to define docking spaces. The dimensions of the grid box were set along the x-, y-, and z-axes at: 154.1, 75.0, 122.7 Å. These parameters were set to encompass the entire 3D structure of the protein so that the ligand could freely move and rotate in the docking space. The 2D and 3D interactions between the protein-ligand were analyzed using Discovery Studio version 21.1.0 (BIOVIA). ⁶⁴ Using a similar approach, binding studies using the native ligand of GluCl, glutamate, as well as the gold standard drug for onchocerciasis, ivermectin were also determined.

Further details of molecular docking studies including software, docking parameters, and protein preparation can be found in the Supporting Information (SI).

Results and Discussion

The use of EDA, ML, molecular docking, and dynamics studies to identify potential anti-infectives to be repurposed for the management of onchocerciasis yielded very promising results. The incorporation of a known onchocerciasis drug, doramectin, in the test samples further demonstrated the robustness of our finding as well as our multi-step computational pipeline.

The ML algorithms benefit from the availability of large data sets. Unfortunately, fewer agents are used in the management of onchocerciasis, leading to an imbalance in the dataset. With only 14 samples assigned to the oncho⁺ class, the high number of descriptors relative to the sample size could lead to overfitting. Using a typical feature selection algorithm would not be ideal; hence, LDA was performed as a dimensionality reduction step.

Figure 1A and B shows the LDA results using the major indications of the drugs as the output. Subsequently, 11 LVs were determined to be ideal since they capture 99.77% of the variability in the data (as shown in Supplemental Figure S2).

Figure 1.

Visualizing the LDA model of anti-infective agents using the major indications/drugs: (A) LD1 vs LD2 and (B) LD3 vs LD2.

To find agents that could be repurposed for managing onchocerciasis from the data, an unsupervised learning approach, k-means cluster analysis was employed. In cluster analysis, drug samples that have similar properties/indications are expected to be clustered closer to each other. The k-means clustering algorithm was used due to its fast computation and general ability to produce tighter clusters.^67,68 Using the folk mallows and rand scores, an optimal k value of 11 was identified. The 11 clusters were just one less than the total number of indications of the drugs in the data set. Subsequently, the 14 oncho⁺ samples were identified, and the pairwise Euclidean distance from these 14 drugs and the other samples was computed. The 5 drugs that clustered closest to the oncho⁺ drugs were selected. It is expected that their closer proximity to the oncho⁺ drugs implied the likelihood to have similar therapeutic properties. A total of 70 drugs would be found in the test set; however, there were 56 drugs, as some agents were found to be close to more than 1 oncho⁺ drug. Supplemental Figure S3 in SI is the LD2 vs LD3 plot of the LDA model showing the location of oncho⁺ (red), oncho⁻ (blue), and test drugs (orange). It must be highlighted here that doramectin and pyrvinium with known oncho⁺ were added to investigate whether this approach would have identified them as potentially useful drugs.^69,70

Generation of ML models with imbalanced data is not ideal, as it leads to high model accuracies even if the models perform woefully in the minority class. The imbalanced-learn SMOTE algorithm was implemented to generate samples to balance the data in the minority class. The SMOTE helps address the algorithmic challenge of class imbalance; it does not generate new chemically validated compounds. Rather, the synthetic instances are interpolations in the descriptor space and should be interpreted as computational aids rather than biologically verified molecules.

The SMOTE algorithm risks overfitting, synthetic sample bias, and class overlap. These risks are mitigated by the use of dimensionality reduction and feature selection methods; hence, LDA was used. In total, 516 artificial oncho⁺ samples were generated to balance the data set. A plot of LD1 vs LD2 and LD2 vs LD3 for the data set with and without the artificial samples can be found in Supplemental Figure S4 in the SI.

The ML models were generated using the balanced data as described earlier. Two thirds of the data was used for the training set, while a third was used for external validation. The following ML algorithms were used as follows: SVM, RF, GNB, NBC, KNN, LR, SGB, BDT, MLP, and ABC. All these algorithms were employed because it was not possible to determine in advance which would perform best for this specific data set. Hyperparameter tuning was performed automatically. The model prediction accuracies on the external validation sets were used to evaluate the ML models. Figure 2 shows the model prediction accuracy of the 10 ML algorithms on the external validation set of data (the data for Figure 2 is shown in Supplemental Table S1 in the SI). It is evident that using the same data set and pre-processing steps, different algorithms yield varying prediction accuracies. Hence, valuable information may be lost if a single model is relied on. Consequently, the test drugs in this study were evaluated using an ensemble of the models. Given the variation in prediction accuracies across the models, a threshold of at least 80% was chosen. Seven models met this criterion: SVM, RF, KNN, BDT, SGB, MLP, and ADB. Subsequently, the 7 models were used to predict the onchocercidal activity of the 58 test drugs. The threshold prediction probability was set to a default value of 0.5. Thus, for a model, any drug with a y-predicted ⩾0.5 is deemed to have the potential to have onchocercidal activity. The choice of a 0.5 threshold is a standard approach in binary classification tasks, representing an equal probability of classification into either class. Supplemental Figure 3A to G shows the y-predicted probabilities of the 58 test drugs. The results indicate varying predicted probabilities for the test drugs by different ML algorithms. These variations in predicted probabilities across different algorithms are expected due to the differences in the approximation methods and objective functions employed in solving ML problems. Thus, employing multiple algorithms enables a more robust identification of drugs with true potential. The number of test drugs predicted by each algorithm is shown in Supplemental Figure S5 in the SI. Supplemental Figure S5 shows KNN predicted 38 of the test drugs as having the potential to be used for onchocerciasis with SGB predicting the least, ie, 13. To account for variations in the number of predicted samples across the ML models, drugs identified by all 7 ML models were prioritized for molecular docking studies.

Figure 2.

The model prediction accuracies of the 10 machine learning algorithms using the external validation set data.

Figure 3.

The y-predicted probability plots of the 16 test samples (including pyrvinium and doramectin) for the top 7 best-performing ML algorithms based on the model prediction accuracy on the external validation set (A) SVM, (B) RF, (C) KNN, (D) BDT, (E) SGB, (F) MLP, and (G) ADB. The red dashed line represents the 0.5 predicted probability value above which a drug is classified as oncho⁺.

Supplemental Table S2 in SI shows the list of the test drugs and their predictions by each of the 7 selected models. Sixteen drugs representing 27.59% of the test data were predicted to be oncho⁺. It is noteworthy that onchocercidal drugs doramectin and pyrvinium were predicted as oncho⁺ by all 7 models. This underscores the potential of combining EDA and ML as a powerful approach for repurposing anti-infectives to address NTDs.

Molecular Docking Studies

The drug of choice in the management of onchocerciasis is ivermectin. Its antiparasitic activity is attributed to its binding to ivermectin-sensitive GluCl.^48,71 Thus, for any of the predicted test drugs to exhibit activity against onchocerciasis, it must demonstrate binding affinity to GluCl. Therefore, the binding interaction between predicted drugs and GluCl can provide useful information that can reduce the attrition rate of any attempt to repurpose the predicted drugs. The binding affinity, Kd (kcal/mol) measures the interaction between 2 molecules, ie, a macromolecule (eg, protein/target) and a ligand (drug). ⁴⁶ Low Kd values indicate the higher affinity, while higher Kd indicates the lower affinity.

The results between GluCl and the 16 predicted test drugs (including doramectin and pyrvinium) are shown in Figure 4. It is not surprising that our evaluation drugs, doramectin and pyrvinium, rank first and second with binding affinities of −8.4 kcal/mol and −8.1 kcal/mol, respectively, values close to ivermectin (red dashed line), the gold standard drug, which has a binding affinity of −8.5 kcal/mol (indicated by red dashed lines). Thus, agents with Kd values closer to ivermectin could be potential candidates for further investigation as GluCl modulators. The binding affinities for the top 3 drugs namely, cridanimod, diminazene, and vandetanib, were −7.8, −7.2, and −7.1 kcal/mol in that order. These values are higher than glutamate, the human ligand for GluCl with a binding affinity of −4.5 kcal/mol.

Figure 4.

The binding affinity (Kd) of the 14 test samples (and 2 external validation drugs, pyrvinium and doramectin) for the GluCl receptor.

Figure 5 shows the binding interaction between GluCl and cridanimod, diminazene, and vandetanib (3D figures of the binding of vandetanib and diminazene with GluCl are shown in Figures S7 and S8 in SI). These interactions were compared with the known onchcocercidial agents, doramectin, ivermectin, and pyrvinium (3D figures of their binding with GluCl are shown in Supplemental Figures S9 to S11). This comparison provided insights into why the test drugs may share similar activity. Table 1 shows binding residues and interaction types of predicted drugs and known onchocercidal agents. The similarities in binding domains and amino residues of all the agents are shown in Supplemental Figure S6, which shows overlap in binding residues that is suggestive of similarities in mechanism of action.

Figure 5.

The binding interaction between GluCl and cridanimod, diminazene, and vandetanib.

Table 1.

Amino acid residues interacting with various drugs in GluCl and their corresponding interaction types.

Drug	Interacting residues	Types of interaction
Cridanimod a	ASN D:45	H-bond
	ASN D:46	H-bond
	GLU E:25	H-bond, Pi-anion bond
	GLN E:103	H-bond
	ARG D:41	Pi-alkyl bond
	VAL D:40	Pi-alkyl bond
Diminazene a	PRO A: 216	H-bond
	ARG A:251	Pi-alkyl bond
	ILE A:70	H-bond
	ASP A:71	Attractive charge
	ASP A:72	Pi-anion, attractive charge
	TYR A:255	H-bond
	TYR A:254	Pi-Pi stacked
Vandetanib a	GLN A:258	C-H bond, Halogen bond
	LYS A:305	Pi-cation bond
	TYR A:255	Alkyl bond
	LEU A:259	Alkyl bond
	ILE A:302	Alkyl bond, Pi-alkyl bond
	PRO A:262	C-H bond, Alkyl
	LEU A:266	Alkyl bond
Doramectin b	TYR A:190	H-bond
	PHE A:128	Pi-alkyl bond
Pyrvinium c	LYS A:305	Pi-alkyl bond
	TYR A:254	Pi-Pi stacked
	ARG A:251	Pi-alkyl bond
Ivermectin c	GLY A:320	C-H bond
	PRO A:308	Alkyl bond
	ASP A:316	C-H bond
	ASN: A:303	C-H bond

^aTest drug predicted to have onchocercidal activity.

^bGold standard drug clinically used for onchocerciasis.

^cAgents with known/documented onchocercidal activity.

Cridanimod binds to ASN D:45, ASN D:46, GLU E:25, GLN E:103, ARG D:41, VAL D:40 via H-bond, pi-alkyl, and pi-anion interactions. Similar hydrogen bonding residues and pi interactions are observed in doramectin and pyrvinium (eg, pi-alkyl interactions in both cases). The presence of interactions involving polar residues (eg, ASN and GLU) may contribute to the drug’s ability to stabilize specific binding sites involved in the onchocercidal mechanism.

Diminazene, on the contrary, binds to PRO A:216, ARG A:251, ILE A:70, ASP A:71, ASP A:72, TYR A:255, TYR A:254 based on H-bonds, attractive charge, pi-anion, and pi-pi stacking. TYR A:254 and ARG A:251 are also found in pyrvinium, suggesting a shared binding region. The involvement of charged residues (eg, ARG and ASP) indicates potential interactions with the active site critical for onchocercidal activity.

Vandetanib binds to GLN A:258, LYS A:305, TYR A:255, LEU A:259, ILE A:302, PRO A:262, and LEU A:266 in GluCl via pi-cation, alkyl, halogen, and C-H bonds. Its binding to LYS A:305 and TYR A:255 is shared with ivermectin and pyrvinium, pointing to a potential overlap in their interaction sites. The diverse interactions, hydrophobic, pi-cation, and halogen bonds suggest that vandetanib has a strong and specific binding affinity for its target site. This is consistent with its potential as a drug candidate with onchocercidal activity.

Conclusion

Using EDA, ML, and molecular docking studies, 14 anti-infectives have been identified to have the potential to be repurposed for the management of onchocerciasis. Three agents, namely, cridanimod, diminazene, and vandetanib, showed binding interactions comparable to the treatment of choice (ivermectin) and agents with known activity (doramectin and pyrvinium) against onchocerciasis. Cridanimod, diminazene, and vandetanib showed binding affinities of −7.8, −7.2, and 7.1 kcal/mol. This high affinity suggests that these may be a promising candidate for further studies aimed at repurposing them for use in treating onchocerciasis. These compounds could also serve as lead compounds in the discovery of new and improved therapies for onchocerciasis. The future direction of these studies includes but is not limited to performing molecular dynamic simulations and experimental studies (in vitro and in vivo) to evaluate the efficacy of these agents. These techniques and routines implemented here can be applied to other diseases to repurpose existing drugs.