Predicting Lead Compounds against Mycobacterium tuberculosis using Machine Learning and Molecular Docking
This project aims at building an in silico pipeline to identify novel antimicrobial compounds against Mycobacterium tuberculosis, in the fight against Tuberculosis, using high-throughput screening (HTS) data sourced from PubChem; and possibly evaluate the in vitro effects of identified lead compounds. We will incorporate Machine Learning (ML), using the best-performing model to predict active substances from curated databases, and Molecular Modeling techniques into our designed pipeline. This method facilitates the identification of the most promising drug candidates amongst predicted compounds while allowing for the visualization of intermolecular interactions between critical residues in the active site and the compounds.
Tuberculosis (TB), caused by Mycobacterium tuberculosis (Mtb), remains one of the most significant global health challenges. It is the leading cause of death from a single infectious agent, surpassed only by coronavirus disease (COVID-19), and the most significant cause of mortality among individuals living with HIV. According to WHO, approximately 10.8 million people were affected by TB in 2023, including 6.0 million men, 3.6 million women, and 1.3 million children, resulting in 1.25 million deaths, of which 161,000 were co-infected with HIV. Drug resistance poses a significant challenge to TB management. Resistance to first-line drugs necessitates using second-line treatments, such as para-aminosalicylic acid (PAS), ethionamides, cycloserine, viomycin, and ciprofloxacin. However, these therapies are associated with higher toxicity, extended treatment durations, and reduced patient compliance, resulting in poorer outcomes and increased drug resistance. Machine learning (ML) provides a promising avenue for the enhancement of the drug-discovery process, with different models like Random Forest, Naive Bayesian Classification, Logistic Regression, Linear Discriminant Analysis, Probabilistic Neural Networks, Multi-Layer Perceptron, and Support Vector Machine (SVM) being considered in this context. The main antimycobacterial agents used (isoniazid and rifampicin) were discovered via various experimental compound screening methods. The development of computational tools for virtual screening has great potential in effectively identifying potent compounds, with a reduction in time and cost, and enhancing the quality of compounds tested in the laboratory.
Our approach seeks to train various Machine Learning models using the Anti-Mtb H37Rv dataset from PubChem to discriminate potential antimycobacterial compounds from within prospective antimicrobial compounds. Subsequently, we will screen the predicted compounds against a Mtb protein target for downstream analysis. Details of the study pipeline can be found in the description section below.
- Study Goal
- Overview
- Objectives
- Description
- Manuscript
- Results
- Data Acquisition and Processing
- Model Development and Evaluation
- Prediction of Inhibitors and Compounds from curated databases
- Target Selection and Molecular Docking of Predicted Compounds
- Mechanism of Binding Characterization of Selected Compounds
- ADMET Screening of Selected Compounds
- Molecular Dynamics Simulations
- MMPBSA Computations
- How to use
- Data Availability
- Reproducibility Prerequisites
- Credits
- Identify the Mycobacterium tuberculosis protein target for our assessment.
- Identify Mycobacterium tuberculosis ligand database for ML training and molecular modeling method validation.
- Determine ML
pythonalgorithm to be utilized in the project. - Process ligand database and train ML model.
- Evaluate ML performance and perform EDA.
- Validate molecular modeling method using prepared ligand database (Actives vs Non-actives).
- Virtual screening of predicted actives into identified protein crystal structures.
- Assess and identify hits using criterion: docking score, interactions with important residues.
- Assess hits ADMET properties.
- Conduct MD simulations to determine compounds' binding mode stability and binding free energy, and MMPBSA computations.
- Compile results and observations.
- Finalize write-up.
- Tentative in vitro testing of predicted leads on whister rats.
This figure illustrates a schematic representation of the process pipeline to be utilized to identify potent antimycobacterial agents against Mycobacterium tuberculosis.
Figure 1. Tuberculosis Drug Identification Pipeline using Machine Learning and Molecular Docking.
The ligand database was obtained from PubChem BioAssay AID: 1626. The ligand database was experimentally generated using a Mtb "in vivo" High Throughput Screen assay to Identify Inhibitors of Mycobacterium tuberculosis H37Rv, adapted from the microdilution Alamar blue assay method by Collins and Franzblau, 1997, to suit a 384-well microtiter plate format and 7H12 broth. A total of 216,163 compounds were initially screened, but overall, 215,098 compounds were considered and analyzed for their antimycobacterial inhibitory effect with 2,044 actives and 209,564 inactives identified. Data were analyzed using the IDBS Activity Base software. Results for each concentration were expressed as percent inhibition (% Inhibition) and were calculated as:
Columns 1 and 2 in the assay plates contained media + 0.4% DMSO for negative control, and columns 23 and 24 contained media in 0.4% DMSO and 100 μM hyamine as a positive control. A stacked-plate dose-response method was used, and the final test concentrations for the compounds ranged from 40 μg/ml to 0.078 μg/ml in 2-fold dilutions with a final DMSO concentration of 0.4%. An active is represented as a compound that can exhibit a percent inhibition above 80% at these concentrations.
-
The unprocessed database can be found here.
-
The molecular descriptors of the actives and inactive were calculated using PaDEL-Descriptors. The descriptors of the actives and inactives were calculated using the
Descriptor CalculatorPython script. -
The actives and inactives databases were combined and all missing descriptors were filled with the value 0. Next, dimensionality reduction was conducted using a variance filter (scikit-learn VarianceThreshold library).
-
The data was then standardized using the mean and standard deviation metrics of various assessment parameters.
-
The data was split into training , test, and external datasets. The training dataset was equivalent to 70% (14875 compounds) of the data set and the test and external data sets were equivalent to 15% (~3188) each. The training dataset contained 3105 actives vs 11770 inactives.
-
The ML models were constructed using lazy predict python package. The models that exhibited the greatest Accuracy, F1-score, Balanced Accuracy, and ROC AUC metrics were selected for validation.
-
The models chosen for further validation were K-Nearest Neighbours, Gaussian Naïve Bayes, Support Vector Machine, Random Forest and Logistic regression. Using K-fold splitting of the training data, the models were cross-validated and the model's suitability was evaluated using the Accuracy, F1-score, Precision, Recall, and Specificity, and false and true positive and negative rate as selection metrics.
-
The models' prediction ability was assessed using the test data. The model's prediction accuracy was determined using Accuracy, F1-score, Precision, and Recall evaluation metrics.
-
The logistic regression (LR) model exhibited the greatest results on the test dataset and therefore was evaluated on the external dataset. The LR model obtained an 82% active and 98% inactive accuracy.
-
The LR model was employed to screen the Northern African Natural Products Database (NANPD), East African Natural Products Database (EANPD), AfroDB and Tradtional Chinese Medicine (TCM) database.
-
The natural compounds' chemical structures were prepared similarly to the training dataset and ~43,000 compounds were screened using the LR model.
-
7,722 compounds were predicted to be active and subsequently utilized for molecular docking.
-
AutoDock Vina was utilized to screen the 7,722 compounds into the Dengue 2 virus envelope protein.
-
The potential hits were selected using the criterion:
- AutoDock Vina binding score.
- Presence of binding interactions between important binding site residues and ligand (LigPlot + v1.4.5).
- The ADMET properties of the identified hits will be predicted using SwissADME, applying Veber's rule and Lipinski's Rule of Five (Ro5).
- The hits with potential pharmacokinetic and toxicity moieties will be removed.
- The hits binding mode stability will be assessed through 200-nanosecond (ns) MD simulations utilizing GROMACS.
- The stability will be assessed using metrics like root-mean-square deviation (RMSD) and fluctuation (RMSF), Radius of Gyration (Rg), etc., using Xmgrace.
- The compounds binding interactions retention with important residues throughout the MD simulations will be assessed with the ProLIF python library.
- The compounds' binding free energies throughout the MD simulation were calculated using Molecular Mechanics Poisson-Boltzmann Surface Area (MMPBSA).
Important
When using the pipeline or findings for research or commercial purposes, please cite our research and star the repository.
From the various analyses and computations performed throughout this in silico exploration, the following results were obtained and observations made.
The data obtained from PubChem was preprocessed after computing molecular descriptors and fingerprints for each compound using PaDEL. PaDEL is a versatile chemoinformatics tool that generates unique machine-readable dsecriptions of compunds, including key features such as the number of hydrogen bond donors (nHBd) and acceptors (nHBa), molecular weights (MW), water partition coefficient (xlogP), and other properties critical for drug characterization.
Due to insignificant imbalance in the dataset, with 2,044 active compounds and 209, 564 inactive compounds, random sampling was applied to the inactive compound library before computing descriptors and fingerprints. This approach was chosen over duplicating existing instances to avoid the risk of overfitting (Gnip et al., 2021; Wongvorachan et al., 2023) .As a result, descriptors were generated for 2,044 active compounds and 6,997 randomly sampled inactive compounds. To further enhance model performance, a low variance filter with a threshold ofvalue 0.1 was applied to remove irrelevant columns from the dataset. This was achieved using the variance threshold function from the feature selection module in scikit-learn(version 1.5.0).
Scikit-Learn was used to develop the following models for the prediction of inhibitors against _Mycobacteriums tuberculosis_: Logistic Regression, Support Vector Machine, Random Forest, K-nearest neighbours, and Gaussian Naive Bayes. Each dataset was partitioned into a train set comprising 70% (6328) of the total and and a test set of 30% (2712). However, before training, certain hyperparameters or settings of machine learning algorithms were configured to improve the performance of the models. The models were developed using Scikit-Learn, Numpy, Pandas, and other Python libraries. A 10-fold cross-validation was used to evaluate the models and further evaluation was carried on the test set to calculate various statistical parameters for analysis and evaluation. The parameters include precision, F1 score, accuracy, recall, Matthew's correlation coefficient(MCC) and Area Under the Receiver Operating Curve(AUROC). The performance of the SVM model gave the highest accuracy score of 96% followed by LR and RF with an equal accuracy score of 95% (Table 1) according to cross-validation. When tested on test data, all the five models had an accuracy score of 97% except for NB which had 83%. The SVM and LR had similar scores including MCC and F1 scores according to the CV demonstrating their robustness in accurately predicting active compounds. NB was ranked the lowest according to all the calculated metrics from cross-validation and eveluation of the test data as shown in (Table 1). The area under the curve is the measure of the ability to distinguish between positive and negative classes. As shown in (Figure 2, high AUC values for all models (>0.85) in both (a) and (b) indicate strong classification performance. From the overall evaluation, SVMand LR demonstrated the best perfomances and thus were selected for the prediction.
Table 1. Performance of models developed for prediction of inhibitors based on the 10-fold cross-validation (CV) and on external test data.
| Model | Process | Accuracy | Precision | Recall | F1 Score | MCC Score |
|---|---|---|---|---|---|---|
| KKN | CV | 0.93 | 0.93 | 0.77 | 0.84 | 0.80 |
| Test | 0.96 | 0.98 | 0.84 | 0.91 | 0.89 | |
| NB | CV | 0.81 | 0.58 | 0.62 | 0.60 | 0.47 |
| Test | 0.83 | 0.59 | 0.63 | 0.61 | 0.50 | |
| SVM | CV | 0.96 | 0.94 | 0.87 | 0.90 | 0.87 |
| Test | 0.97 | 0.99 | 0.86 | 0.92 | 0.90 | |
| RF | CV | 0.95 | 0.96 | 0.82 | 0.89 | 0.86 |
| Test | 0.96 | 0.98 | 0.84 | 0.90 | 0.88 | |
| LR | CV | 0.95 | 0.92 | 0.88 | 0.90 | 0.87 |
| Test | 0.97 | 0.96 | 0.89 | 0.92 | 0.90 |
 cross-validation and b) external data evaluation for the five developed models_page-0001 (1).jpg)
Figure 2. Receiver Operating Characteristic Curve for a) cross-validation and b) test data evaluation for the five developed models. The higher the AUC, the better the performance of the model.
The descriptor standardization approach indicates that a small fraction of compounds fall outside the Applicability Domain, with training data outliers being 0.98% and test data outliers being 0.59% (figure 3). The model's predictions for molecules that fall below this threshold are expected to be accurate since they are inside its applicability domain. Low outlier percentages for the test data suggest a well-trained model that generalizes well to new data ((Fry-Nartey et al., 2025).
Figure 3. Applicability domain of the five developed models using a threshold set to 2. The training data is colored blue while the test data is red.
The two best-performing models, LR and SVM, were applied to compound databases including AfroDB and FDA-approved drugs from the ZINC20 database, to classify active compounds against _M. tuberculosis_. Compounds classified as active by both models were selected for downstream analysis. Out of 3318 FDA-approved compounds submitted, 1,889 compounds were predicted to be active. In addition, 439 compounds from the Afrodb catalog were predicted to be active against _M. tuberculosis_ by both models.
Table 3. Top 15 compounds across the two databases predicted to be active against M. tuberculosis
| Number | Compounds | Logistic Regression Model | Support Vector Machine |
|---|---|---|---|
| 1 | ZINC000000001411 | Active | Active |
| 2 | ZINC000000001785 | Active | Active |
| 3 | ZINC000000001785 | Active | Active |
| 4 | ZINC000000001969 | Active | Active |
| 5 | ZINC000000013245 | Active | Active |
| 6 | ZINC000000013246 | Active | Active |
| 7 | ZINC000000035526 | Active | Active |
| 8 | ZINC000000039811 | Active | Active |
| 9 | ZINC000000057384 | Active | Active |
| 10 | ZINC000000057731 | Active | Active |
| 11 | ZINC000000057733 | Active | Active |
| 12 | ZINC000000113319 | Active | Active |
| 13 | ZINC000000119983 | Active | Active |
| 14 | ZINC000000120276 | Active | Active |
| 15 | ZINC000000120283 | Active | Active |
The crsytals structure of _Mycobacterium tuberculosis_ stain H37Rv was used in this study as a model for the molecular docking pipeline to identify potential inhibitors of M. tuberculosis. The three-dimensional structure of _M.tuberculosis_ DprE1 in complex with PBTZ169 (Macozinone) and its cofactor FAD was retrived from the protein Data Bank accessed on 11 February 2025. The structure was solved at a resolution of 1.88 A using the X-ray diffraction method, with the amino aci residue ranging from 6 to 461 (Makarov et al., 2014) a. In this study, the 4NCR structure of DprE1 was selected due to its low resolution and complex with a preclinical candidate compound, PBTZ169 (Ibrahim et al., 2024) , which served as a control for the molecular docking studies.The grid box used in this study was set on the active site of PBTZ169-DprE1 complex. Visualization of the chain A of the DprE1 structure complexed with PBTZ169 using Ligplot displayed hydrogen bonding with residue Cys387,Lys134, and Lys418 and hydrophobic interaction with residues such as Gly133, Ser228, His132, Phe369, Lys367, Val365, and Gly117 as shown in ( Figure 4). The sdf formats of the 2328 predicted active compounds were retrieved from the ZINC database specifically from the AfroDb and FDA-approved catalogs.
A total of 2328 predicted compounds and the known inhibitor PBTZ169 were docked into the binding site of the DprE1. Compounds with binding energies of -9.0 kcal/mol or less were selected for further analysis. This threshold was used to retain high-binding compounds since the two known inhibitors such as BTZ043 (Benzothiazinone) and PBTZ169 (Macozinone) added as a control had binding energies of -9.3 and -9.1 kcal/mol, respectively (Imran et al., 2023). Other computational methods reported binding energies of -9 and -9.2 kcal/cal for BTZ043 and PBTZ169, respectively. This cut-off resulted in 176 hit compounds possessing binding energies > -9.0 kcal/mol and thus were further tested. ZINC000003830925 showed the highest binding energy towards the DprE1 among all ligands with an energy of -11.0 kcal/mol followed by ZINC000003830767 and ZINC000013462588 with a binding energy of -10.9 kcal/mol.
Figure 4. The hit compounds firmly docked at the active site of the DprE1. A) A close-up view of the compounds docked. B) The ligands are shown as sticks, while the active site is rendered as a surface representation.
The pharmacokinetic properties of compounds play a crucial role in modulating their biological activities (Bowes et al., 2012; Watroly et al., 2021) . Pharmacokinetics describes the processes by which a drug is absorbed, distributed, metabolized, and excreted by the body as time and concentration of the drug change (Arnaut, 2021) . After molecular docking, 179 compounds with binding affinities below -9.0 kcal/mol were physicochemically profiled for drug-likeness.. Pharmacokinetic properties such as gastrointestinal (GI) absorption, solubility, TPSA, Lipinski, and Veber's violation were analyzed using SwissADME, as shown in Table 3. Data Warrior was then used to analyze the toxicity levels of the selected hit compounds. Compounds with poor solubility and low gastrointestinal absorption were eliminated as they impact how orally administered drugs are dissolved and absorbed, respectively (Suenderhauf et al., 2012; Gakpey et al., 2025). A total of 81 compounds exhibited good pharmacokinetics properties as they did not violate any of Lipinski's rule of five and Veber’s rule. 37 compounds were predicted to be poorly soluble while 3 were estimated to posses low absorption. DataWarrior classifies chemicals based on their predicted toxicity, categorizing them as high risk, low risk, or no toxicity(Azim et al., 2020; Kwofie et al., 2019). Out of the 58 compunds that passed the pharmacokinetics and druglikeness profiling, 54 were identified as having no tumorigenic properties and 4 demonstrated high tumorigenicity. A similar trend was observed for mutagenicity, however, two had low mutagenicity. Additionally, 21 were predicted to have significant reproductive impacts, 34 showed no reproductive effects, and 3 were associated with low reproductive effects. 28 compounds that exhibited good pharmacological properties and no toxicity properties were selected for downstream analysis.
Table 3. Pharmacokinetic evaluation of the hit compounds. .
| Compounds | MW (g/mol) | TPSA | ESOL Solubility Class | GI Absorption | No. of Lipinski’s Violation | No. of Veber’s Violation |
|---|---|---|---|---|---|---|
| ZINC000003830925 | 497.49 | 176.61 | Moderately soluble | Low | 1 | 1 |
| ZINC000003830767 | 376.49 | 46.53 | Moderately soluble | High | 0 | 0 |
| ZINC000003830923 | 497.49 | 176.61 | Moderately soluble | Low | 1 | 1 |
| ZINC000001550477 | 581.06 | 114.73 | Poorly soluble | low | 1 | 1 |
| ZINC000003797541 | 349.51 | 33.12 | Moderately soluble | High | 0 | 0 |
| ZINC000003830922 | 497.49 | 176.61 | Moderately soluble | Low | 1 | 1 |
| ZINC000003993855 | 389.4 | 74.87 | Moderately soluble | High | 0 | 0 |
| ZINC000001996117 | 426.55 | 55.56 | Moderately soluble | High | 0 | 0 |
| ZINC000003831201 | 600.53 | 186.22 | Soluble | Low | 2 | 3 |
| ZINC000030691797 | 349.38 | 58.68 | Moderately soluble | High | 0 | 0 |
Studies indicate covalent interaction with Cys387 amongst other favorable interactions between that adding suitable functionality off one or both of the phenyl rings to make additional energetically favorable interactions with residues like
Figure 6. PyMOL visualization of 4NCR structure (A) PBTZ169 shown as red stick docked in the active site (mesh); (B). Ligplot visualization of the interactions between the PBTZ169 (Purple) and the active site.
Table 4. Binding energies and protein-ligand interactions of top 15 selected hits (-9.0 kcal/mol threshold).
| Compounds | Binding Energy (kcal/mol | Hydrogen Bonding (Bond length/Å) | Hydrophobic Interactions |
|---|---|---|---|
| ZINC000003830925 | -11.0 | Asp389 (3.06), Leu363 (3.13), Lys134 (3.06), Tyr415 (2.79) | Trp230, Val365, Pro116, Gly117, His132, Ile131, Val121, Arg58, Ser59, Thr118, Tyr60, Cys387 |
| ZINC000003830767 | -10.9 | Ser59 (2.96), Lys418 (2.87) | Ile131, Val121, Arg58, Tyr60, Gly117, Trp230, His132, Gly331, Leu363, Asp389, Pro116, Val365, Cys387, Tyr415, Ala417 |
| ZINC000013462588 | -10.9 | His132 (3.12), Tyr60 (2.68) | Gly117, Lys418, Gly331, Asp389, Leu363, Tyr314, Trp230, Phe313, Asn385, Cys387, Pro116, Lys134, Val365, Gly133 |
| ZINC000003830923 | -10.8 | Gly331 (3.15), His132 (3.17) | Lys134, Pro116, Cys387, Gly133, Val365, Lys418, Ile131, Arg58, Tyr60, Val121, Thr118, Gly177, Tyr415 |
| ZINC000001550477 | -10.7 | His132 (3.01), Asn385 (2.81). Asp389 (3.22) | Ile131, Val121, Lys418, Tyr60, Gly117, Trp230, Tyr314, Phe313, Val365, Cys387, Lys367, Phe366, Phe369, Lys134, Gly133, Pro116, Arg58, Ala417 |
| ZINC000095486053 | -10.4 | Lys418 (3.03) | Ala417, Arg58, Ser59, Val121, Ile131, Tyr415, Pro116, Gly117, His132, Lys134, Val365, Trp230, Tyr60 |
| ZINC000095485954 | -10.4 | Asp389 (2.89), Lys418 (3.19), Tyr415 (2.93) | Arg58, Val121, Tyr60, Thr118, Val365, Cys387, Trp230, Tyr314, Gly117, Pro116, His132, Ile131 |
| ZINC000014768737 | -10.4 | - | Trp230, Asp389, Gly331, Gly117, Tyr60, Thr118, Arg58, Cys129, Ala417, Ser59, Val121, Pro116, Ile131, Lys134, Asn385, His132, Val365, Cys387 |
| ZINC000003797541 | -10.3 | Arg58 (2.96) | Val121, Ile131, Ala417, Tyr415, Pro116, Lys418, His132, Lys134, Gln336, Val365, Cys387, Thr118, Tyr60, Gly117, Ser59, |
| ZINC000003830922 | -10.3 | His132 (3.09, 3.07), Tyr415 (3.27), Tyr60 (2.74) | Arg58, Pro116, Ile131, Asn385, Cys387, Phe369, Ser228, Val365, Lys367, Phe366, Gly133, Lys134, Tyr314, Gly117, Val121, Thr118 |
| ZINC000003993855 | -10.3 | His132 (3.01) | Tyr60, Gly117, Tyr415, Pro116, Phe369, Asn385, Lys134, Val365, Cys387, Phe332, Gly331, Asp389, Lys418 |
| ZINC000001996117 | -10.2 | Gly117 (2.91) | Ile131, Ser59, Pro116, Val121, Tyr60, Gly331, Cys387, Leu363, Val365, Trp230, Asp389, Lys134, Phe313, Tyr314, Lys418, Arg58, Ala417, Tyr415 |
| ZINC000003831201 | -10.1 | Phe332 (3.03), Tyr60 (2.95), Ala417 (3.33), Lys418 (2.93), Gly117 (3.12) | Val121, Arg58, Lys134, Val345, Gly133, His132, Lys367, Pro116, Cys387, Phe313, Tyr314, Asp389, Trp230, Gly331, Ser59, Thr118 |
| ZINC000014952515 | -10.1 | Lys134 (3.21), His132 (2.95, 3.02), Tyr415 (2.90), Arg58 (3.02), Asp389 (3.03) | Gly331, Asn285, Val365, Cys387, Gly133, Pro116, Val121, Tyr60, Gly117, Lys418 |
| ZINC000014612849 | -10.1 | - | Cys387, His132, Gly117, Ile131, Arg58, Val121, Cys129, Pro116, Ser59, Tyr60, Ala417, Lys418 |
| PBTZ169 | - | Cys389 (2.91), Lys134 (3.16), Lys418 (3.03) | Gly133, Ser228, His132, Phe369, Val365, Gly117, Lys367 |
The data utilized for the project can be found here.
Note
The codes and scripts were run on Python 3., Anaconda3 2024.06.1 and Jupyter Notebook version .
R 4..0 was used for some of the data visualization to plot graphs from MMPBSA computations.
Please, cite and star the repository if you utilize the pipeline for research or commercial purposes.
Team members include:
| N° | First & Last Name | Email address | ORCID ID |
|---|---|---|---|
| 1 | George Hanson | george.hanson417@gmail.com | 0009-0007-2720-9102 |
| 2 | Joseph Adams | jkojoadams@gmail.com | 0009-0003-3871-1369 |
| 3 | Daveson Innocento Brank Kepgang | davesonbrank@gmail.com | 0009-0004-6104-184X |
| 4 | Maame Esi Annor-Apaflo | maaameesiannorapaflo@gmail.com | 0009-0005-8426-3071 |
| 5 | Richmond Kabutey Kudiabor | kabuteykudiabor98@gmail.com | 0009-0005-7167-5579 |
| 6 | Miriam Eyram Lawson Gakpey | lawsmimi40@gmail.com | 0009-0005-6595-5910 |
| 7 | Edgar Sungwacha | esungwacha@gmail.com | 0000-0002-8585-9337 |
| 8 | Clifford Seyram Yisufu | cyisufu@gmail.com | 0009-0001-8166-9552 |
| 9 | Olaitan Igbabo Awe | laitanawe@gmail.com | 0000-0002-4257-3611 |