INTRODUCTION:

Identifying and classifying compounds acting as either agonists or antagonists of the dopamine D2 receptor is crucial in the drug discovery and development. The dopamine D2 receptor, a member of the G-protein coupled receptor (GPCR) family, plays a vital role in various physiological and pathological processes, indicating it as an interesting target for therapeutic interventions^1–5.

In recent years, computational techniques in drug discovery have gathered substantial interest due to their ability to predict the characteristics and activities of new drug candidates efficiently^6,7.

Among these approaches, one notable method involves employing molecular descriptors generated from the computational software, which provides a comprehensive collection of chemical descriptors that characterize the molecular structure and properties of the compounds^2,7,8.

Machine learning has reformed drug classification, offering powerful tools for predicting compound properties and activities. Researchers can choose between structure-based and ligand-based approaches, leverage various datasets, and apply different algorithms to achieve accurate classifications. Despite challenges, the machine learning approach continues to drive innovation in drug discovery and development, ultimately benefiting the healthcare and pharmaceutical industries. Future research will likely focus on addressing challenges and further improving the accuracy and reliability of drug classification models^9-15.

This paper explores a systematic study of dopamine D2 receptor agonists and antagonists classification by utilizing machine learning algorithms, especially XGBoost, LGBM, ExtraTree, and AdaBoost Classifier. These algorithms have demonstrated promise across various drug discovery applications and hold the potential to categorize compounds accurately based on their molecular descriptors^16,18.

The proposed methodology involves the extraction of molecular descriptors from a dataset comprising dopamine D2 receptor ligands. These descriptors capture important chemical information such as atom type, bond type, and physicochemical properties. Subsequently, these extracted features serve as inputs for training and validating the performance of the four classifiers. The classifier performance evaluation includes several relevant metrics such as accuracy, precision, recall, and F1-score that comprehensively evaluate their predictive ability. Furthermore, the classification outcomes are compared and analyzed to find the most effective algorithm for the task^19,20.

The results of this study bring significant implications for a comprehensive design and refinement of dopamine D2 receptor modulators. The classification accuracy of the compounds into agonists or antagonists allows researchers to prioritize lead compounds, thereby reducing the time and cost associated with experimental screening. Additionally, the insights gathered from this research contribute to the comprehension of the structure-activity relationships governing interactions with the dopamine D2 receptor, thereby facilitating the development of more efficacious therapeutic agents. This paper illustrates the utility of RDKit molecular descriptors in combination with machine learning algorithms for the dopamine D2 receptor agonists and antagonists classification. The findings presented herein have the potential to boost drug discovery efforts directed at the dopamine system and enrich the broader field of computational chemistry and pharmacology^21,22.

MATERIALS AND METHODS:

Materials:

A dataset of dopamine D2 receptor ligands has been compiled, comprising known agonists and antagonists, along with their corresponding activity labels. The dataset was carefully curated, with a focus on ensuring the inclusion of diverse chemical structures and maintaining a balanced distribution between agonists and antagonists^7,11.

Molecular descriptor calculation:

Molecular descriptors for each compound in the dataset were calculated using the RDKit toolkit, a robust open-source software. An array of molecular descriptors, encompassing topological, physicochemical, and structural features, were computed through the utilization of RDKit functions.

Descriptors such as atom type, bond type, molecular weight, hydrogen bond acceptor, hydrogen bond donor, and lipophilicity indices were extracted as part of this process^8,22.

Classifier training and evaluation:

The dataset underwent division into training and testing sets through a random split, with careful consideration to maintain an appropriate balance of agonists and antagonists in both subsets. Subsequently, the training set served as the foundation for training the three classifiers: XGBoost, LGBM, ExtraTree, and AdaBoost Classifier. During this training phase, the classifiers modelled the relationships between the molecular descriptors and their corresponding agonist/antagonist labels.

The testing set was then employed to measure the performance of each classifier by comparing the predicted labels with the known labels. To assess the predictive ability of each classifier, various performance metrics, including accuracy, precision, recall, and F1-score, were computed²⁰.

Comparative analysis:

The classification results generated by each classifier were subjected to a comparative and analytical examination, aimed at identifying the most efficient algorithm for the task of classifying dopamine D2 receptor agonists and antagonists. In this evaluation, we examined the qualities and drawbacks of each classifier, giving due regard to factors like computational efficiency and the interpretability of outcomes.

RESULT:

Chemical diversity analysis:

In this study, a dataset comprising SMILES representations of 195 molecules was used to generate molecular descriptors. Within this dataset, there were 69 dopamine agonists and 126 dopamine antagonists. Figure 1 depicts the structures of 20 molecules selected randomly from the dataset.

The RDKit molecular descriptors were generated for a set of 195 molecules, resulting in a total of 208 descriptors. These descriptors encompassed 123 physicochemical properties and 85 molecular substructures. Physicochemical properties are generally represented as continuous or quantitative data, while molecular substructures are frequently presented as boolean or categorical data.

Generation and validation of machine learning models:

Multiple classification algorithms were utilized to build models with 75% of the designated training set, which included 195 molecule descriptors. The evaluation of these models was carried out using the remaining 25% as the test set, encompassing 49 molecule descriptors. It is worth noting that the AdaBoost, ExtraTree, LGBM, and XGB classifiers exhibited superior accuracy and F1 scores when compared to other algorithms, ultimately achieving an impressive accuracy rate of 92%, as outlined in Figure 2.

Figure 2. Performance evaluation of machine learning algorithms to classify dopamine D2 agonist-antagonist molecules.

Furthermore, the receiver operating characteristic curve (ROC) was computed and visualized for the various models. The ROC curve analysis involved the calculation of the area under the curve (AUC), a metric for evaluating the predictive capacity of the generated models. A higher AUC value indicates a better classification ability. Figure 3 presents the ROC-AUC values for the models under study. Notably, the AdaBoost and ExtraTree classifiers achieved AUC values of 0.92 and 0.90, respectively, while the LGBM and XGB classifiers attained AUC values of 0.87 and 0.89, respectively. These findings underscore the superior performance of the AdaBoost and ExtraTree algorithms in this study.

(a)

(b)

(c)

(d)

Figure 3. The receiver operator characteristic (ROC) curves for the machine learning model.

The assessment of classification model accuracy can also involve an examination of the true positives (TP) and true negatives (TN) within the entire dataset. Sensitivity, also referred to as the true positive rate (TPR), was calculated as TP/(TP+FN), representing the probability that a genuine positive instance will be correctly identified. Conversely, the true negative rate, known as specificity, indicates the probability that a genuine negative instance will be correctly classified.

The sensitivity values for the AdaBoost, ExtraTree, LGBM, and XGB classifiers were found to be 0.9118, 1.000, 0.9412, and 0.9706, respectively. Among these classifiers, the AdaBoost Classifier exhibited the lowest sensitivity. A chi-square analysis indicated that there were no significant differences in sensitivity among the four models, with a p-value of 0.9999.

Similarly, specificity, calculated as TN/(TN+FP), was assessed for each model. The specificity values for the AdaBoost, ExtraTree, LGBM, and XGB classifiers were 0.8235, 1.000, 0.8571, and 0.9231, respectively. The chi-square test for specificity also revealed no significant differences across all models studied, with a p-value of 0.9992.

To further analyze model performance, confusion matrices were employed to determine the counts of true positives, true negatives, false positives (FP), and false negatives (FN). Figure 4 presents a heatmap visually depicting the confusion matrices for the AdaBoost, Extra Tree, LGBM, and XGB classifiers.

Out of the 208 molecular descriptors employed as model estimators, the AdaBoost Classifier recognized 35 descriptors as important. In contrast, the ExtraTree Classifier identified a remarkable 167 descriptors as important, which was the highest among all the models. The LGBM classifier selected 90 essential features, while the XGBoost classifier chose 73 critical descriptors. The summary of these significant features chosen by each model is provided in Figure 5.

It's noteworthy that despite employing only 35 molecular descriptors, the AdaBoost Classifier achieve the highest level of accuracy compared to the other models.

The Extra Tree classifier, which identified 167 molecular descriptors as important, incorporated all the crucial features selected by the other classifier models. This relationship is illustrated in the Venn diagram presented in Figure 6. The intersection among the AdaBoost, LGBM, and XGBoost classifiers comprises 24 shared molecular descriptors. Specifically, these shared descriptors are as follows: BCUT2D_LOGPLOW, BCUT2D_MRLOW, BCUT2D_MWLOW, BalabanJ, Chi0v, Chi1v, Chi2v, Chi4v, Hall Kier Alpha, Max Partial Charge, Num Aliphatic Carbocycles, PEOE_VSA12, PEOE_VSA2, PEOE_VSA3, PEOE_VSA9, SMR_VSA1, SMR_VSA10, SlogP_VSA1, TPSA, VSA_EState2, VSA_EState4, fr_halogen, fr_piperzine, and QED. These descriptors collectively form the set of features shared by these classifiers.

(a) (b)

(c)

Figure 6. The intersection between AdaBoost, ExtraTree, LGBM, and XGBoost classifiers was selected as important features.

Despite all classification models selecting the same 24 molecular descriptors, the accuracy of classification predictions using only these 24 descriptors decreased for all models. The accuracy of AdaBoost decreased to 88%, ExtraTree achieved an accuracy of 90%, LGBM had an accuracy of 84%, and the XGBoost classifier achieved 86% accuracy.

Furthermore, the models were tested on molecules that were not included in the training and test compounds. Lisuride, ropinirole, and quinpirole were known dopamine D2 receptor agonists^23,24, while Sinedabet (Viozan™) is recognized as a D2 dopamine receptor and β2-adrenoceptor agonist²⁵. These four compounds were classified using the developed models, following the same procedure as the training datasets. The prediction results were presented in a heatmap plot shown in Figure 7.

Figure 7. Lisuride, ropinirole, quinpirole, and sinedabet prediction using the developed classification models.

The AdaBoost, ExtraTree, LGBM, and XGB classifiers collectively categorized Sinedabet as a dopamine D2 receptor antagonist, whereas they classified lisuride, ropinirole, and quinpirole as dopamine D2 receptor agonists.

DISCUSSION:

The research employed various classification algorithms, such as AdaBoost, ExtraTree, LGBM, and XGBoost, to classify compounds as dopamine D2 receptor agonists or antagonists. It was interesting to observe that despite using only 24 shared molecular descriptors, the accuracy of these models decreased compared to their performance with the full set of molecular descriptors. This indicates that the additional features beyond these 24 contribute significantly to the model's predictive power. It's essential for researchers to carefully consider feature selection strategies to optimize model performance in drug discovery. The 24 shared molecular descriptors were:

- BCUT2D_LOGPLOW: This term represents a molecular descriptor related to the partition coefficient (LogP) of a compound, which is a measure of its lipophilicity or hydrophobicity.

- BCUT2D_MRLOW: It refer to a molecular descriptor related to the molar refractivity (MR) of a compound, which is a property that characterizes its polarizability.

- BCUT2D_MWLOW: This descriptor is associated with the molecular weight (MW) of a compound, a fundamental property representing the sum of the atomic weights of its atoms.

- Balaban J: Balaban's J index is a topological descriptor used to measure the degree of molecular branching in a chemical compound. It helps assess molecular complexity.

- Chi0v, Chi1v, Chi2v, Chi4v: These terms represent various molecular topological indices (Chi indices) that provide insights into molecular structure, including atom connectivity and branching patterns.

- Hall Kier Alpha: This descriptor refers to a topological index used in quantitative structure-activity relationship (QSAR) studies, providing information about the structural complexity of a molecule.

- Max Partial Charge: This term represents the maximum partial charge within a molecule, which is a measure of the distribution of electrical charge within the molecule.

- Num Aliphatic Carbocycles: It indicates the number of aliphatic carbocycles within a molecule. Carbocycles are closed ring structures composed of carbon atoms.

- PEOE_VSA12, PEOE_VSA2, PEOE_VSA3, PEOE_VSA9: These descriptors related to Partial Equalization of Orbital Electronegativities (PEOE) and represent specific contributions of atoms or regions to the polarizability of a molecule.

- SMR_VSA1, SMR_VSA10, SlogP_VSA1: These terms are associated with different types of solvent-accessible surface area (SASAs) and provide insights into how molecular properties are distributed across a compound.

- TPSA: The Total Polar Surface Area (TPSA) is a descriptor that quantifies the polar surface area of a molecule, providing information about its hydrogen bonding capacity.

- VSA_EState2, VSA_EState4: These descriptors are related to the E-State indices, which capture the electronic properties of atoms within a molecule and are used for quantitative structure-activity relationship (QSAR) modeling.

- fr_halogen: This term refers to a descriptor that counts the number of halogen atoms within a molecule. Halogens include elements like fluorine, chlorine, bromine, and iodine.

- fr_piperzine: This descriptor counts the number of piperazine ring structures within a molecule. Piperazine is a common chemical moiety found in various compounds.

- QED (Quantitative Estimate of Drug-likeness): QED is a measure of a compound's drug-likeness, assessing its similarity to known drugs in terms of properties such as molecular weight, lipophilicity, and polar surface area²⁶.

These descriptors are important in computational chemistry and pharmacology for characterizing molecular properties and aiding in tasks like compound classification and drug design. They provide valuable information for understanding the structure-activity relationships of chemical compounds.

The study assessed the performance of these classifiers not only through standard metrics like accuracy, precision, recall, and F1-score but also by constructing receiver operating characteristic (ROC) curves and calculating the area under the curve (AUC) values. The results revealed that the AdaBoost and ExtraTree classifiers consistently outperformed the others in terms of accuracy, making them promising candidates for further research.

The models were tested on compounds that were not part of the training or test datasets, which included well-known dopamine D2 receptor agonists and an agonist with a broader target range. It's noteworthy that the models correctly classified these compounds, highlighting their practical utility in identifying potential drug candidates.

The accurate classification of compounds as dopamine D2 receptor agonists or antagonists is crucial in drug discovery and development. The ability to distinguish between these two types of compounds can significantly update the drug screening process, ultimately saving time and resources. The research provides a valuable tool for researchers to prioritize compounds and make informed decisions about which molecules to advance in the drug development pipeline.

The research also underscores the challenge of feature selection in machine learning for chemical compound classification. While the study showed that fewer features can lead to a decrease in model accuracy, finding the right balance between feature dimensionality and model performance remains an important task. Further research could explore advanced feature selection techniques to optimize model accuracy and reduce computational costs.

CONCLUSION:

This study investigated the classification of compounds as either dopamine D2 receptor agonists or antagonists using machine learning algorithms. It became evident that feature selection plays a crucial role in determining model performance. While employing a reduced set of 24 shared molecular descriptors resulted in diminished accuracy, it underscored the importance of utilizing a comprehensive set of descriptors for enhanced predictive capabilities.

Among the classifiers employed, AdaBoost and ExtraTree consistently outperformed others regarding accuracy and AUC values. This highlights their potential for accurate compound classification. The classification of dopamine D2 receptor agonists and antagonists using AdaBoost, ExtraTree, LGBM, and XGB classifiers produced impressive results, with the AdaBoost Classifier achieving the highest accuracy at 92%. Robustness was confirmed by robust AUC values ranging from 0.87 to 0.92 in ROC analysis. Notably, the ExtraTree classifier selected the most descriptors (167), emphasizing the task's complexity.

Classification results for sinedabet, lisuride, ropinirole, and quinpirole indicated sinedabet as a dopamine D2 receptor antagonist, while the other three compounds were classified as agonists. This suggests the influence of additional factors, potentially non-active dopamine D2 receptor compounds, on the classification. Accurate classification of dopamine D2 receptor agonists and antagonists was significant in drug discovery, reorganization lead compound prioritization, and minimizing the resources and time needed for experimental screening. The models underwent validation with compounds that were not included in the training datasets. They correctly classified well-known dopamine D2 receptor agonists and a broad-spectrum agonist, showcasing their practical applicability in real-world scenarios. This study highlighted the complicated balance between feature dimensionality and model performance. Future research suggests exploring feature selection techniques to optimize both accuracy and computational efficiency.

CONFLICT OF INTEREST:

The authors have no conflicts of interest regarding this investigation.

ACKNOWLEDGMENTS:

The authors gratefully acknowledge financial support from the Institut Teknologi Sepuluh Nopember for this work, under the project scheme of Publication Writing and IPR Incentive Program (PPHKI) 2024.

REFERENCES:

1. Im D. Inoue A. Fujiwara T. Nakane T. Yamanaka Y. Uemura T. Mori C. Shiimura Y. Kimura KT. Asada H. Nomura N. Tanaka T. Yamashita A. Nango E. Tono K. Kadji FMN. Aoki J. Iwata S. Shimamura T. Structure of the Dopamine D2 Receptor in Complex with the Antipsychotic Drug Spiperone. Nat Commun. 2020; 11(1): 6442. https://doi.org/10.1038/s41467-020-20221-0.

2. Bueschbell B. Barreto CAV. Preto AJ. Schiedel AC. Moreira IS. A Complete Assessment of Dopamine Receptor- Ligand Interactions through Computational Methods. Molecules. 2019; 24(7): 1196. https://doi.org/10.3390/molecules24071196.

3. Bhargava K. Nath R. Seth PK. Pant KK. Dixit RK. Molecular Docking Studies of D2 Dopamine Receptor with Risperidone Derivatives. Bioinformation. 2014; 10(1): 8–12. https://doi.org/10.6026/97320630010008.

4. Raghubabu K., Jagannadharao V., Ramu BK. Assay of Ropinirole hydrochloride in Pharmaceutical Preparations by Visible Spectrophotometry. Asian J. Pharm. Ana. 2012; 2(2): 41-45.

5. Choi J. Horner KA. 2022. Dopamine Agonists. In StatPearls StatPearls Publishing: Treasure Island FL.

6. Yamuna M. Elakkiya A. Mathematical Models in Drug Discovery, Development and Treatment of Various Diseases – A Case Study. Research J. Pharm. and Tech. 2017; 10(12): 4397-4401. doi: 10.5958/0974-360X.2017.00810.1

7. Goode-Romero G. Winnberg U. Domínguez L. Ibarra IA. Vargas R. Winnberg E. Martínez A. New Information of Dopaminergic Agents Based on Quantum Chemistry Calculations. Sci Rep. 2020; 10(1): 21581. https://doi.org/10.1038/s41598-020-78446-4.

8. Riniker S. Landrum GA. Similarity Maps - a Visualization Strategy for Molecular Fingerprints and Machine-Learning Methods. Journal of Cheminformatics. 2013; 5(1): 43. https://doi.org/10.1186/1758-2946-5-43.

9. Meenakshi K. Safa M. Karthick T. Sivaranjani NA. Novel Study of Machine Learning Algorithms for Classifying Health Care Data. Research J. Pharm. and Tech. 2017; 10(5): 1429-1432. doi: 10.5958/0974-360X.2017.00253.0

10. Chaurasia V. Pal S. Skin Diseases Prediction: Binary Classification Machine Learning and Multi Model Ensemble Techniques. Research J. Pharm. and Tech. 2019; 12(8): 3829-3832. doi: 10.5958/0974-360X.2019.00656.5

11. Ruiz-Moreno AJ. Reyes-Romero A. Dömling A. Velasco-Velázquez MA. In Silico Design and Selection of New Tetrahydroisoquinoline-Based CD44 Antagonist Candidates. Molecules. 2021; 26(7): 1877. https://doi.org/10.3390/molecules26071877.

12. Ali AMS. Pan Tompkins Algorithm based ECG Signal Classification. Research J. Pharm. and Tech. 2017; 10(12): 4365-4367. doi: 10.5958/0974-360X.2017.00802.2.

13. Indraja B. Annapurani K. Classification of Medicines using Naive bayes Classifier. Research J. Pharm. and Tech. 2018; 11(5): 1940-1944. doi: 10.5958/0974-360X.2018.00360.8.

14. Shyamala DM. Sruthi AN. Saranya JC. MRI Liver Tumor Classification Using Machine Learning Approach and Structure Analysis. Research J. Pharm. and Tech. 2018; 11(2):434-438. doi: 10.5958/0974-360X.2018.00080.X.

15. Kumar NK. Vigneswari D. Hepatitis- Infectious Disease Prediction using Classification Algorithms. Research J. Pharm. and Tech. 2019; 12(8): 3720-3725. doi: 10.5958/0974-360X.2019.00636.X.

16. Pillai N. Dasgupta A. Sudsakorn S. Fretland J. Mavroudis PD. Machine Learning Guided Early Drug Discovery of Small Molecules. Drug Discovery Today. 2022; 27(8): 2209–2215. https://doi.org/10.1016/j.drudis.2022.03.017.

17. Raju B. Narendra G. Verma H. Kumar M. Sapra B. Kaur G. jain SK. Silakari O. Machine Learning Enabled Structure-Based Drug Repurposing Approach to Identify Potential CYP1B1 Inhibitors. ACS Omega. 2017; 7(36): 31999–32013. https://doi.org/10.1021/acsomega.2c02983.

18. Kamath V. Pai A. Application of Molecular Descriptors in Modern Computational Drug Design –An Overview. Research Journal of Pharmacy and Technology. 2017; 10(9): 3237–3241. https://doi.org/10.5958/0974-360X.2017.00574.1.

19. Li J. Luo D. Wen T. Liu Q. Mo Z. Representative Feature Selection of Molecular Descriptors in QSAR Modeling. Journal of Molecular Structure. 2013; 1244: 131249. https://doi.org/10.1016/j.molstruc.2021.131249.

20. Pedregosa F. Varoquaux G. Gramfort A. Michel V. Thirion B. Grisel O. Blondel M. Prettenhofer P. Weiss R. Dubourg V. Vanderplas J. Passos A. Cournapeau D. Brucher M. Perrot M. Duchesnay É. Scikit-Learn: Machine Learning in Python. Journal of Machine Learning Research. 2011; 12(85): 2825–2830.

21. Dibia KT. Igbokwe PK. Ezemagu GI. Asadu CO. Exploration of the Quantitative Structure-Activity Relationships for Predicting Cyclooxygenase-2 Inhibition Bioactivity by Machine Learning Approaches. Results in Chemistry. 2022; 4: 100272. https://doi.org/10.1016/j.rechem.2021.100272.

22. Landrum GA. RDKit: Open-Source Cheminformatics Software. http://www.rdkit.org/ accessed 2022-03-15..

23. Schmidt LG. Kuhn S. Smolka M. Schmidt K. Rommelspacher H. Lisuride a Dopamine D2 Receptor Agonist and Anticraving Drug Expectancy as Modifiers of Relapse in Alcohol Dependence. Progress in Neuro-Psychopharmacology and Biological Psychiatry. 2022; 26(2): 209–217. https://doi.org/10.1016/S0278-584601.00214-7.

24. Wengrovitz A. Ivantsova E. Crespo N. Patel M. Souders CL. Martyniuk CJ. Differential Effects of Dopamine Receptor Agonists Ropinirole and Quinpirole on Locomotor and Anxiolytic Behaviors in Larval Zebrafish Danio Rerio.: A Role for the GABAergic and Glutamate System. Neurotoxicology and Teratology. 2023; 98: 107183. https://doi.org/10.1016/j.ntt.2023.107183.

25. Calverley P. Keating ET. Goldman M. Casty F. Conclusion Lessons from the Novel D2 Dopamine Receptor Β2-Adrenoceptor Agonist Viozan TM: Chronic Obstructive Pulmonary Disease and Drug Development Implications. Respiratory Medicine. 2003; 97: S71–S74. https://doi.org/10.1016/S0954-611103.80017-3.

26. Kamath V. Pai A. Application of Molecular Descriptors in Modern Computational Drug Design –An Overview. Research J. Pharm. and Tech. 2017; 10(9): 3237-3241. doi: 10.5958/0974-360X.2017.00574.1.

Received on 10.09.2023 Modified on 23.03.2024

Research J. Pharm. and Tech 2024; 17(9):4507-4514.

DOI: 10.52711/0974-360X.2024.00697