Evaluating and Improving In Silico Drug Candidate Prediction for Preventing Toxoplasmic Encephalitis

My name is Emily Huang and I'm a high school student from Waterloo, Ontario. I hope you enjoy learning more about my project!

Graphics obtained from Canva unless otherwise stated.

The Centers for Disease Control and Prevention (CDC) has designated toxoplasmosis as one of five neglected parasitic infections that require targeted public health action (Cantey et al., 2021). Currently, no therapeutics exist to target the latent (bradyzoite) stage of infection (Zwicker et al., 2020). Immunodeficient individuals face a high risk of parasite reactivation, resulting in a condition known as toxoplasmic encephalitis (Elsheikha et al., 2020) (Fig 1). The effects of toxoplasmic encephalitis are severe and can even result in death (Wang et al., 2017). Thus, the development of therapeutics to eliminate T. gondii bradyzoites before they can reactivate represents a high public health priority.

Traditional in vitro early-stage drug discovery is time-consuming and expensive. Machine learning models can be used to streamline this process, reducing time and cost (Blanco-González et al., 2023). In my previous project, I created a graph neural network-based drug screening model to target toxoplasmosis (Huang, 2025). However, deep learning models often act as black boxes and lack transparency (Patel & Shah, 2022) (Fig 2). As a result, promising in silico results do not always translate to promising in vitro results. In the context of drug discovery, understanding the “why” behind a model’s predictions is as crucial as the prediction itself.

This project seeks to improve upon current computational drug screening techniques by increasing clinical interpretability and model confidence, allowing for advancement towards the development of therapeutics to prevent toxoplasmic encephalitis.

T. gondii Cathepsin L (TgCPL) (Fig 3) was identified as a strong target for bradyzoite inhibition (Fig 4) based on prior work (Huang, 2025) involving a literature review and sequence/structure analysis. Alternative targets were considered, but deprioritized due to exhibiting a high level of conservation with their human homolog, reduced relevance to chronic infection, or limited open-source in vitro data (Fig 5) (Besteiro et al., 2011; Ortiz et al., 2025; Silva et al., 2021; Lourido et al., 2010).

Classical machine learning models were trained on molecular descriptors of pharmacological compounds downloaded from ChEMBL (Zdrazil et al., 2023). I applied SHAP (SHapley Additive exPlanations) (Lundberg & Lee, 2017) to quantify feature importance and validated those findings against existing scientific literature. Chemical features shown to be uninfluential were removed from the input data, and the models were retrained.

Next, a direct-message passing graph neural network (D-MPNN GNN) model was trained, and GNNExplainer (Ying et al., 2019) and RDKit were used to identify important substructures driving model predictions. The important substructures, in addition to commonly occurring substructures, were used to tokenize the input data. The tokenized data was used to train a transformer-based regression model.

The trained models were used to predict the binding affinity of compounds downloaded from ChEMBL (Zdrazil et al., 2023), BindingDB (Liu et al., 2006), and PubChem (Kim et al., 2025) with TgCPL based on their Simplified Molecular Input Line Entry Sequence (SMILES). Top-ranked candidates underwent molecular docking simulations and absorption, distribution, metabolism, excretion, and toxicity testing (Fu et al., 2024) (Fig 6).

Overall, this pipeline (Fig 7) enables the rapid identification of biologically relevant drug candidates, reducing time and cost compared to traditional in vitro screening, while mitigating the black box effect seen in in silico screening and integrating expert knowledge in the pipeline.

Based on the SHAP findings, heavy atom count and ring count had the largest influence on the linear regression and random forest regression models' predictions, respectively. For compounds in the training dataset, their heavy atom count and ring count were positively correlated with their predicted pKi value (Fig 8 & 9). A larger pKi value indicates a smaller Ki value, which indicates greater drug potency (Canadian Society of Pharmacology and Therapeutics (CSPT) - Inhibitory Constant (Ki), 2025).

The top ten substructures identified by GNNExplainer and RDKit exhibited common characteristics such as aromaticity, fluorination, and rigidity, consistent with existing literature showing that aromatic, hydrophobic, and conformationally constrained structures contribute to enhanced binding affinity and molecular stability (Lanzarotti et al., 2020; Monkovic et al., 2022; Nguyen et al., 2017) (Fig 10).

After incorporating SHAP-based insights, the test dataset R2 value of the random forest regression model improved from 0.68 to 0.71 (Fig 11). This demonstrated that removing less important features in the input dataset using explainable AI-driven insights improved the model’s ability to generalize to unseen data.

More significantly, after incorporating GNNExplainer-based insights to employ substructure-based tokenization of the SMILES data, the test dataset R2 value of the transformer model improved from 0.26 to 0.75 (Fig 12). The drastic improvement in the R2 value of the model showed that using XAI to enhance molecular structure encoding in the input data significantly increased the model’s “goodness of fit.” This finding showed that XAI-guided tokenization can be used to mitigate model overfitting in transformer models when a larger dataset is not available. This is an important finding for biomedical research applications, given the high cost of collecting curated biomedical data (Pentavere, 2024).

Overall, the SHAP-identified molecular feature findings align with existing literature.

Features like molecule size and rigidity, as indicated by ring count, are believed to influence binding affinity (Schmidt & Wittrup, 2009; Forrey et al., 2012)

These findings suggest that the classical models were able to successfully identify influential molecular features in drug binding

The top substructures identified by GNNExplainer and RDKit can be validated by existing literature and dataset analysis.

These identified substructures are important for binding affinity due to their aromaticity and structural rigidity and display structural similarity to previously studied chemical scaffolds for related protein targets (Deng et al., 2024; Borel et al., 2026; Theyagarajan et al., 2024)

Compounds containing the identified substructures displayed promising molecular docking-simulated binding values (Trott & Olson, 2009)

In the training dataset, compounds that contained one or more of the top substructures displayed a lower median Ki value compared to those without any of the top substructures.

These findings suggest that the deep learning model was able to successfully identify influential substructures in drug binding

Molecular docking simulations and absorption, distribution, metabolism, excretion, and toxicity (ADMET) testing effectively filtered out compounds exhibiting unfavourable properties like off-target binding, blood-brain barrier impermeability, poor oral bioavailability, and high neurotoxicity risk, and highlighted opportunities for further lead optimization in promising hit compounds (Fig 13).

The results demonstrate that:

Integrating explainable AI into drug screening not only improves model performance but also reveals biologically meaningful patterns that align with known pharmacological principles

Incorporating substructure-level insights significantly enhances model performance

Identifying chemically relevant substructures enables the model to prioritize compounds with stronger predicted binding affinity and greater stability

Together, these findings show that this approach produces more biologically interpretable predictions, improving the efficiency and reliability of early-stage drug discovery for preventing toxoplasmic encephalitis (Fig 14)

This project has several key contributions:

First integrated human-in-the-loop drug screening pipeline specifically targeting chronic toxoplasmosis

Demonstrated that integrating explainable artificial intelligence into in silico drug screening can enhance biological interpretability and also guide model performance improvement across both classical and advanced drug screening models, providing a versatile, scalable, and improved approach for drug discovery

First application, to my knowledge, to use GNN-scored substructures as a custom tokenization scheme for a transformer-based binding affinity model

Screened 86,000 compounds within 6 hours at virtually no cost (Fig 15)

Drug screening identified promising repurposing candidates showing strong binding affinity and favourable pharmacokinetic properties

With over 2 billion individuals infected worldwide, this project tackles one of the world’s “most successful human parasites” (Halonen & Weiss, 2013). The development of therapeutics to eliminate the dormant form of the parasite represents an urgent global health issue, and this improved pipeline serves as a solid starting step towards finding a solution to this issue.

Future steps to expand this work:

Increase the size of the training dataset through further data augmentation techniques and data simulations, allowing for less model overfitting (Understanding Overfitting: Strategies and Solutions, 2024) (Fig 16)

Further investigation of blood-brain-barrier-impermeable/neurotoxic candidates that displayed promising binding affinity results (Fig 17)

In vitro validation of predicted Ki values (Fig 18)

Apply this pipeline to other diseases, such as the four other neglected parasitic infections—Chagas disease (Fig 19), Cysticercosis (Fig 20), Toxocariasis (Fig 21), and Trichomoniasis (Fig 22)—highlighted by the Centers for Disease Control and Prevention (Cantey et al., 2021)

Special thanks to Dr. Brent Dixon of Health Canada, who provided valuable insights and literature regarding T. gondii for my background research.

I would also like to thank all of the volunteers at the Waterloo-Wellington Science and Engineering Fair who made the regional fair a great experience! Additional thanks to the Waterloo-Wellington delegates for their peer-editing work and helpful suggestions for ProjectBoard.

Al Salam, H. (2010). Acute necrotising toxoplasma encephalitis. Radiopaedia.Org. https://doi.org/10.53347/rid-10683

‌

Avva, A. (2025, May 3). High-Throughput Screening (HTS): a simple guide to faster drug and material discovery — ChemCoPilot: PLM + AI for chemical industry. ChemCopilot: PLM + AI for Chemical Industry. https://www.chemcopilot.com/blog/hts-material-discovery

Besteiro, S., Brooks, C. F., Striepen, B., & Dubremetz, J.-F. (2011). Autophagy Protein Atg3 is Essential for Maintaining Mitochondrial Integrity and for Normal Intracellular Development of Toxoplasma gondii Tachyzoites. PLoS Pathogens, 7(12), e1002416. https://doi.org/10.1371/journal.ppat.1002416

‌

Blanco-González, A., Cabezón, A., Seco-González, A., Conde-Torres, D., Antelo-Riveiro, P., Piñeiro, Á., & Garcia-Fandino, R. (2023). The Role of AI in Drug Discovery: Challenges, Opportunities, and Strategies. Pharmaceuticals (Basel, Switzerland), 16(6), 891. https://doi.org/10.3390/ph16060891

Borel, T. J., Hitchens, J. R., Kaur, H., Mueller, D., Ard, A., Carruthers, V. B., & Clasby, M. C. (2026). Improvements in potency and selectivity of triazine nitrile inhibitors targeting Toxoplasma gondii cathepsin L for the treatment of latent toxoplasmosis. Bioorganic & Medicinal Chemistry, 138, 118643. https://doi.org/10.1016/j.bmc.2026.118643

‌

Canadian Society of Pharmacology and Therapeutics (CSPT) - Inhibitory constant (Ki). (2025). Pharmacologycanada.Org. https://pharmacologycanada.org/Inhibitory-constant-ki

‌

Cantey, P. T., Montgomery, S. P., & Straily, A. (2021). Neglected Parasitic Infections: What Family Physicians Need to Know—A CDC Update. Cdc.Gov, 104. https://stacks.cdc.gov/view/cdc/117325

‌

CDC. (2024, September 4). About Cysticercosis. Cysticercosis. https://www.cdc.gov/cysticercosis/about/index.html

CDC. (2024, September 10). About Toxocariasis. Toxocariasis. https://www.cdc.gov/toxocariasis/about/index.html

‌‌

‌Deng, W., Hu, X., Tian, X., Zhang, Y., Shang, W., Zhang, L., & Shang, L. (2024). Peptidomimetic Analogues Act as Effective Inhibitors against SARS-CoV-2 by Blocking the Function of Cathepsin L. Journal of Medicinal Chemistry, 67(19), 17124–17143. https://doi.org/10.1021/acs.jmedchem.4c00656

Dubey, J. P., Lindsay, D. S., & Speer, C. A. (1998). Structures of Toxoplasma gondii Tachyzoites, Bradyzoites, and Sporozoites and Biology and Development of Tissue Cysts. Clinical Microbiology Reviews, 11(2), 267–299. https://doi.org/10.1128/cmr.11.2.267

Elsheikha, H. M., Marra, C. M., & Zhu, X.-Q. (2020). Epidemiology, Pathophysiology, Diagnosis, and Management of Cerebral Toxoplasmosis. Clinical Microbiology Reviews, 34(1). https://doi.org/10.1128/cmr.00115-19

Everlywell. (2023, June 14). How long can trichomoniasis be dormant? Everlywell. https://www.everlywell.com/blog/sti-testing/how-long-can-trichomoniasis-be-dormant/?srsltid=AfmBOoqHxBCyfnu1hRqFnuYN3CA6nVQIFs8Xlk4xA7n_Noc26zj8CyM8

Forrey, C., Douglas, J. F., & Gilson, M. K. (2012). The fundamental role of flexibility on the strength of molecular binding. Soft Matter, 8(23), 6385. https://doi.org/10.1039/c2sm25160d

‌

Fu, L., Shi, S., Yi, J., Wang, N., He, Y., Wu, Z., Peng, J., Deng, Y., Wang, W., Wu, C., Lyu, A., Zeng, X., Zhao, W., Hou, T., & Cao, D. (2024). ADMETlab 3.0: an updated comprehensive online ADMET prediction platform enhanced with broader coverage, improved performance, API functionality and decision support. Nucleic Acids Research, 52(W1). https://doi.org/10.1093/nar/gkae236

‌

Friedrich, N.-O., de Bruyn Kops, C., Flachsenberg, F., Rarey, M., & Kirchmair, J. (2019). Chemoinformatics in Drug Discovery: Methods and Applications. In J. Bajorath (Ed.), Applied Chemoinformatics: Achievements and Future Opportunities (Chapter 6). Wiley-VCH. https://doi.org/10.1002/9783527806539.ch6

Habtye, B., Habtu, D., Ebrahim, H., Mahmood, N., Alemu, G., Tilahun, M., Alemayehu, E., Mohammed, O., & Feleke, D. G. (2023). Global seroprevalence of Toxoplasma gondii infection among patients with mental and neurological disorders: A systematic review and meta‐analysis. Health Science Reports, 6(6). https://doi.org/10.1002/hsr2.1319

Hagberg, A., Swart, P., & S Chult, D. (2008). Exploring network structure, dynamics, and function using NetworkX.

Halonen, S. K., & Weiss, L. M. (2013). Toxoplasmosis. Handbook of Clinical Neurology, 114, 125–145. https://doi.org/10.1016/b978-0-444-53490-3.00008-x

Heid, E., Greenman, K. P., Chung, Y., Li, S.-C., Graff, D. E., Vermeire, F. H., Wu, H., Green, W. H., & McGill, C. J. (2023). Chemprop: A Machine Learning Package for Chemical Property Prediction. Journal of Chemical Information and Modeling, 64(1), 9–17. https://doi.org/10.1021/acs.jcim.3c01250

‌

Huang, E. (2025). QSAR-Based Virtual Screening for T. gondii Autophagy Inhibitors. Youth Science Canada. https://partner.projectboard.world/ysc/project/qsar-based-virtual-screening-for-t-gondii-autophagy-inhibitors

Huber, D. (2025, June 24). UGA researchers develop first test of cure for Chagas disease. UGA Today. https://news.uga.edu/uga-develops-first-test-of-cure-for-chagas/

‌

Hunter, J. D. (2007). Matplotlib: A 2D graphics environment. Computing in Science & Engineering, 9(3), 90–95.

Khurana, S. (2024, January 18). Explainability in Graph Neural networks with GNNExplainer. Medium. https://medium.com/@sambhav.khurana006/explainability-in-graph-neural-networks-with-gnnexplainer-6287ec82dc28

‌

Kim, S., Chen, J., Cheng, T., Gindulyte, A., He, J., He, S., Li, Q., Shoemaker, B. A., Thiessen, P. A., Yu, B., Zaslavsky, L., Zhang, J., & Bolton, E. E. (2025). PubChem 2025 update. Nucleic Acids Res., 53(D1), D1516–D1525. https://doi.org/10.1093/nar/gkae1059

Lanzarotti, E., Defelipe, L. A., Marti, M. A., & Turjanski, A. G. (2020). Aromatic clusters in protein–protein and protein–drug complexes. Journal of Cheminformatics, 12(1). https://doi.org/10.1186/s13321-020-00437-4

‌

Larson, E. T., Parussini, F., Huynh, M.-H., Giebel, J. D., Kelley, A. M., Zhang, L., Bogyo, M., Merritt, E. A., & Carruthers, V. B. (2009). Toxoplasma gondii Cathepsin L Is the Primary Target of the Invasion-inhibitory Compound Morpholinurea-leucyl-homophenyl-vinyl Sulfone Phenyl. Journal of Biological Chemistry, 284(39), 26839–26850. https://doi.org/10.1074/jbc.m109.003780

Liu, T., Lin, Y., Wen, X., Jorissen, R. N., & Gilson, M. K. (2006). BindingDB: a web-accessible database of experimentally determined protein-ligand binding affinities. Nucleic Acids Research, 35(Database), D198–D201. https://doi.org/10.1093/nar/gkl999

Lourido, S., Shuman, J., Zhang, C., Shokat, K. M., Hui, R., & Sibley, L. D. (2010). Calcium-dependent protein kinase 1 is an essential regulator of exocytosis in Toxoplasma. Nature, 465(7296), 359–362. https://doi.org/10.1038/nature09022

Lundberg, S., & Lee, S.-I. (2017). A Unified Approach to Interpreting Model Predictions. ArXiv.org. https://arxiv.org/abs/1705.07874

Lyons, R. E., McLeod, R., & Roberts, C. W. (2002). Toxoplasma gondii tachyzoite–bradyzoite interconversion. Trends in Parasitology, 18(5), 198–201. https://doi.org/10.1016/s1471-4922(02)02248-1

‌

McDonald, C., Smith, D., Di Cristina, M., Kannan, G., Dou, Z., & Carruthers, V. B. (2020). Toxoplasma Cathepsin Protease B and Aspartyl Protease 1 Are Dispensable for Endolysosomal Protein Digestion. mSphere, 5(1). https://doi.org/10.1128/msphere.00869-19

Monkovic, J. M., Gibson, H., Sun, J. W., & Montclare, J. K. (2022). Fluorinated Protein and Peptide Materials for Biomedical Applications. Pharmaceuticals, 15(10), 1201. https://doi.org/10.3390/ph15101201

‌

Müller, J., & Hemphill, A. (2023). Toxoplasma gondii infection: novel emerging therapeutic targets. Expert Opinion on Therapeutic Targets, 27(4-5), 293–304. https://doi.org/10.1080/14728222.2023.2217353

Nguyen, D. D., Xiao, T., Wang, M., & Wei, G.-W. (2017). Rigidity Strengthening: A Mechanism for Protein–Ligand Binding. Journal of Chemical Information and Modeling, 57(7), 1715–1721. https://doi.org/10.1021/acs.jcim.7b00226

‌

Offord, Catherine. (2025). No kittens required: Scientists find new way to study toxoplasmosis parasite in lab. Science | AAAS. https://www.science.org/content/article/no-kittens-required-scientists-find-new-way-study-toxoplasmosis-parasite-lab

Ortiz, J. O., Potter, A. K., & Imaan Benmerzouga. (2025). Novel Drug Targets for the Bradyzoite Form of Toxoplasma gondii. Research and Reports in Tropical Medicine, Volume 16, 25–30. https://doi.org/10.2147/rrtm.s431290

Pardridge, W. M. (2004). The blood-brain barrier: Bottleneck in brain drug development. NeuroRx, 2(1), 3–14. https://doi.org/10.1602/neurorx.2.1.3

Patel, V., & Shah, M. (2022). Artificial intelligence and machine learning in drug discovery and development. Intelligent Medicine, 2(3), 134–140. https://doi.org/10.1016/j.imed.2021.10.001

Pentavere. (2024, May 14). Real-World Data Curation to Transform Medical Investigation: Technology to reimagine the economics of evidence gathering and Support Regulatory Decision-Making - Pentavere. https://pentavere.ai/publications/real-world-data-curation-to-transform-medical-investigation-technology-to-reimagine-the-economics-of-evidence-gathering-and-support-regulatory-decision-making/

Ray, D. (2025, November 26). Understanding Model Predictions with SHAP. Medium; Women in Technology. https://medium.com/womenintechnology/understanding-model-predictions-with-shap-d7457f6a31c3

‌

RDKit: Open-source cheminformatics. https://www.rdkit.org

Schmidt, M. M., & Wittrup, K. D. (2009). A modeling analysis of the effects of molecular size and binding affinity on tumor targeting. Molecular Cancer Therapeutics, 8(10), 2861–2871. https://doi.org/10.1158/1535-7163.mct-09-0195

Services, Equipment & Pricing. (2024). Services, Equipment & Pricing. Carbone Cancer Center. https://cancer.wisc.edu/research/resources/ddc/smsf/equipment-services/

Silva, M. da, Teixeira, C., Gomes, P., & Borges, M. (2021). Promising Drug Targets and Compounds with Anti-Toxoplasma gondii Activity. Microorganisms, 9(9), 1960. https://doi.org/10.3390/microorganisms9091960

‌

Thaprawat, P., Chalasani, S., Schultz, T. L., Di Cristina, M., & Carruthers, V. B. (2024). Toxoplasma gondii PROP1 is critical for autophagy and parasite viability during chronic infection. bioRxiv (Cold Spring Harbor Laboratory). https://doi.org/10.1101/2024.10.02.616283

Theyagarajan, K., Sruthi, V. P., Mohanapriya, D., Thenmozhi, K., & Senthilkumar, S. (2024). Biosensor platform for testing active pharmaceutical ingredients. Health and Environmental Applications of Biosensing Technologies, 129–152. https://doi.org/10.1016/b978-0-443-19039-1.00007-9

‌

Trott, O., & Olson, A. J. (2009). AutoDock Vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. Journal of Computational Chemistry, 31(2), 455–461. https://doi.org/10.1002/jcc.21334

Understanding Overfitting: Strategies and Solutions. (2024, November 25). Lyzr. https://www.lyzr.ai/glossaries/overfitting/

‌

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30. https://doi.org/10.48550/arXiv.1706.03762

Wang, Z.-D., Wang, S.-C., Liu, H.-H., Ma, H.-Y., Li, Z.-Y., Wei, F., Zhu, X.-Q., & Liu, Q. (2017). Prevalence and burden of Toxoplasma gondii infection in HIV-infected people: a systematic review and meta-analysis. The Lancet HIV, 4(4), e177–e188. https://doi.org/10.1016/s2352-3018(17)30005-x

Wu, J., Luo, X., Huang, N., Li, Y., & Luo, Y. (2022). Misdiagnosis of HIV With Toxoplasmosis Encephalopathy With Progressive Memory Loss as the Initial Symptom: A Case Report. Frontiers in Neurology, 13. https://doi.org/10.3389/fneur.2022.809811

Ying, Z., Bourgeois, D., You, J., Zitnik, M., & Leskovec, J. (2019). GNNExplainer: Generating Explanations for Graph Neural Networks. Advances in Neural Information Processing Systems, 32. https://papers.nips.cc/paper_files/paper/2019/hash/d80b7040b773199015de6d3b4293c8ff-Abstract.html

Zwicker, J. D., Smith, D., Guerra, A. J., Hitchens, J. R., Haug, N., Roest, S. V., Lee, P., Wen, B., Sun, D., Wang, L., Keep, R. F., Xiang, J., Carruthers, V. B., & Larsen, S. D. (2020). Discovery and Optimization of Triazine Nitrile Inhibitors of Toxoplasma gondii Cathepsin L for the Potential Treatment of Chronic Toxoplasmosis in the CNS. ACS Chemical Neuroscience, 11(16), 2450–2463. https://doi.org/10.1021/acschemneuro.9b00674

Zdrazil, B., Felix, E., Hunter, F., Manners, E. J., Blackshaw, J., Corbett, S., de Veij, M., Ioannidis, H., Lopez, D. M., Mosquera, J., Magarinos, M., Bosc, N., Arcila, R., Kizilören, T., Gaulton, A., Bento, A., Adasme, M., Monecke, P., Landrum, G., & Leach, A. (2023). The ChEMBL Database in 2023: a drug discovery platform spanning multiple bioactivity data types and time periods. Nucleic Acids Research, 52(D1), D1180–D1192. https://doi.org/10.1093/nar/gkad1004