Predictive toxicology employs in silico methods to assess the potential toxicity of chemicals, drugs, and other products. Recent advances in computational methods, especially machine learning, are enhancing the ability to analyze large datasets and develop models that can predict the toxic effects of compounds with greater speed and accuracy.
Determining the toxicity of a new compound to assess whether it is harmful to humans is crucial but challenging. Conventionally, the prediction of toxicological effects relies on in vitro studies and, more importantly, animal testing, which is expensive, time-consuming, and raises ethical concerns.
The advent of machine learning can overcome these challenges. In predictive toxicology, machine learning models are built from the analysis of databases such as ToxCast, ChEMBL, and PubChem, or data in the public domain and published literature.
Image Credit: Gorodenkoff/Shutterstock.com
Machine Learning Basics
Machine learning involves the building of a model, followed by an iterative process of training and validation steps until acceptable performance is reached. Finally, the model is tested on data not previously exposed to the model itself.
Large datasets of chemical properties, biological interactions, and toxicological outcomes are used to build models that predict relationships between chemical structures and toxic effects.
This not only accelerates safety assessments but also reduces the reliance on animal testing, which is in line with the growing emphasis on ethical research practices.1
These models can be classified into supervised learning, unsupervised learning, and semi-supervised learning.
The most widely used approach in predictive toxicology is supervised learning, where models are trained using well-defined input parameters like molecular descriptors paired with corresponding output labels or toxicity values. The main methods used in predictive toxicology are support vector machines (SVMs), random forests (RF), and decision trees (DTs).2
Find out more about analytical chemistry
Applications in Predictive Toxicology
Predicting toxicity through machine learning can have a strong impact by enhancing risk assessments, determining clinical toxicities, and detecting harmful side effects of chemicals and drugs. The analysis of large datasets, which was impossible by manual approaches in the past, can rapidly provide insights into toxicity mechanisms.3
Risk assessments rely heavily on large datasets. Machine learning can improve the ability to evaluate and manage potential risks associated with chemical exposures, help unravel toxicity pathways, and assist policymakers in the decision-making process in line with the chemical next-generation risk assessment (NGRA), aiming to ensure human safety without animal testing.4
Machine learning is invaluable in drug discovery as it can enhance the optimization of candidate selection.
Quantitative structure-activity relationship (QSAR) models use data on known molecules to predict the toxicity of new candidates, helping to screen out potentially harmful chemicals before investing time and money in testing phases.
Supervised learning algorithms for ADMET (absorption, distribution, metabolism, excretion, toxicity) are frequently applied in predictive toxicology. Similarly, physiologically based pharmacokinetic (PBPK) models can efficiently predict toxicity for a large number of chemicals with accuracies comparable to those of in vivo animal experiments.5
Advantages
Traditional toxicological assessments can often take months or years to complete and involve extensive laboratory experiments and animal studies. Conversely, machine learning models enable faster and more accurate toxicology predictions.
Machine learning can rapidly analyze chemical data and predict potential toxic effects, allowing for quicker decision-making in regulatory and industrial settings.
In addition, the integration of machine learning into predictive toxicology reduces the need for large-scale animal testing, offering considerable cost savings and overcoming ethical issues.
Challenges
One of the main challenges in applying machine learning to predictive toxicology is the dependence on high-quality, diverse datasets.
The accuracy and reliability of machine learning models are directly related to the quality of the data they are trained on, and incomplete or biased data can lead to inaccurate predictions.
Another common issue is data imbalance. Some datasets like Tox21 and ToxCast have an overrepresentation of nontoxic compounds, which results in models being biased toward predicting compounds as nontoxic, with the risk of overlooking potentially harmful substances.6
There are also challenges related to the interpretability of machine learning models. Since many algorithms operate as "black boxes," it is difficult to understand how specific predictions are generated. Being able to understand how model predictions are made is critical for gaining insight into the toxicological mechanisms and can have an impact on regulatory decisions.
Ensuring ethical AI in research
Future Directions
Since machine learning is continuously evolving, it is expected that more models will be developed to assist predictive toxicology, especially thanks to technological advances that are lowering computational costs and the development of new data sources.
The integration of machine learning with genomics and other omics technologies is an important emerging trend. By incorporating genetic and molecular data, machine learning models can provide a more accurate prediction of how chemicals interact with biological systems and their related toxic effects.
Current efforts aim to develop interpretable models that can provide clear insights into the factors driving predictions.
In addition, emerging algorithmic techniques, such as transfer learning and reinforcement learning, are expected to further improve predictive models by enabling them to learn from related tasks and adapt to new data more effectively.
Conclusion
As a powerful resource in predictive toxicology, machine learning can enhance the analysis of complex toxicological data and predict the potential toxicity of new compounds. This approach is faster, more accurate, and less expensive than traditional methods.
Despite some challenges related to data quality and model interpretability, machine learning has the potential to transform predictive toxicology.
It offers a powerful tool for risk assessment and drug development, which can lead to safer and more effective chemical products.
References
- Guo, W., Liu, J., Dong, F., Song, M., Li, Z., Khan, M. K. H., Patterson, T. A. & Hong, H. (2023). Review of machine learning and deep learning models for toxicity prediction. Exp Biol Med (Maywood), 248, 1952-1973.10.1177/15353702231209421.
- Wang, M. W. H., Goodman, J. M. & Allen, T. E. H. (2021). Machine Learning in Predictive Toxicology: Recent Applications and Future Directions for Classification Models. Chemical Research in Toxicology, 34, 217-239.10.1021/acs.chemrestox.0c00316. Available: https://doi.org/10.1021/acs.chemrestox.0c00316
- Tonoyan, L. & Siraki, A. G. (2024). Machine learning in toxicological sciences: opportunities for assessing drug toxicity. Frontiers in Drug Discovery, 4.10.3389/fddsv.2024.1336025. Available: https://www.frontiersin.org/journals/drug-discovery/articles/10.3389/fddsv.2024.1336025
- Yang, C., Rathman, J. F., Bienfait, B., Burbank, M., Detroyer, A., Enoch, S. J., Firman, J. W., Gutsell, S., Hewitt, N. J., Hobocienski, B., Kenna, G., Madden, J. C., Magdziarz, T., Marusczyk, J., Mostrag-Szlichtyng, A., Krueger, C.-T., Lester, C., Mahoney, C., Najjar, A., Ouedraogo, G., Przybylak, K. R., Ribeiro, J. V. & Cronin, M. T. D. (2023). The role of a molecular informatics platform to support next generation risk assessment. Computational Toxicology, 26, 100272.https://doi.org/10.1016/j.comtox.2023.100272. Available: https://www.sciencedirect.com/science/article/pii/S2468111323000130
- Lin, Z. & Chou, W. C. (2022). Machine Learning and Artificial Intelligence in Toxicological Sciences. Toxicol Sci, 189, 7-19.10.1093/toxsci/kfac075. https://www.frontiersin.org/journals/drug-discovery/articles/10.3389/fddsv.2024.1336025/full
- Abou Hajal, A. & Al Meslamani, A. Z. (2024). Overcoming barriers to machine learning applications in toxicity prediction. Expert Opinion on Drug Metabolism & Toxicology, 20, 549-553.10.1080/17425255.2023.2294939. Available: https://doi.org/10.1080/17425255.2023.2294939
Further Reading
CHIEF Model Transforms Cancer Evaluation with Enhanced Generalizability Across Diverse Histopathology Samples