The use of Artificial Intelligence (AI) applications is quickly increasing and has the potential to revolutionize chemical risk assessment. For predicting the toxicity of nanomaterials, current models only work for relatively simple nanomaterials. There is a need to anticipate current developments and strategies to prepare for future AI implementation in chemical regulation.

The rise of AI: interest and concern

People are increasingly interested in developing and using Artificial Intelligence (AI) applications for a wide variety of purposes. However, there are also concerns about how the technology is being implemented. For example, regulators are frequently asked about their views on the use of AI for chemical risk assessment. 

Revolutionizing risk assessment with AI: predicting toxicity without animal testing 

AI has the potential to change how risk assessment is done. For example, it can predict toxicological effects without needing to use test animals or perform costly experiments. AI can even work for chemicals for which no experimental data are available, or for new chemicals. This would be particularly useful for predicting toxicity of nanomaterials. It can help prioritize which nanomaterials should be of most concern. Or it could help identify less toxic alternative materials to use in the process of innovation. These same principles apply to advanced material innovations. 

AI models show promise in predicting nanoparticle toxicity in aquatic environments 

A recent article reviewed how AI is being used to predict the aquatic toxicity of nanomaterials. The review included twenty-six relevant studies. According to the authors, most of the AI-based models demonstrated acceptable performance. However, most studies only focused on metal oxide and metallic nanoparticles. Virtually no studies were available for carbon-based nanomaterials like carbon nanotubes and graphene-based materials. The authors suggest that AI could be a useful tool to evaluate nanomaterial toxicity, as it could be fast, cheap and robust. Additionally, AI could be used to support risk assessment. It could help collect relevant literature and automatically extract and assess the quality of experimental data. 

Reflections by RIVM 

It is clear that AI has great potential for various applications in our (future) society. This requires us to reflect on how to deal with it at an early stage of development. Specifically, in the case of chemical regulation, AI is one of the New Approach Methodologies (NAMs) that can be used to replace costly and time-consuming animal testing. We need to keep track of AI developments to decide how to implement it in future chemical regulations. It is also important to be aware of the potential pitfalls. 

For nanomaterials, the cited review demonstrates that predictive models are available for prediction of nanomaterial toxicity, but they are limited to metallic nanomaterials. No predictive models exist for other important classes of nanomaterials like carbon nanotubes and modified graphenes. In part this is due to the lack of experimental data. Here AI can help select nanomaterials for testing and select tests that provide the most relevant information for model development. 

Another challenge is that the current models only work for relatively simple nanomaterials that consist of a single component. There are no predictive models for more advanced materials that are currently being innovated and are commonly made of multiple materials. 

To prepare for future AI implementation in chemical regulation, it is important to anticipate current developments and develop strategies. This is particularly important for advanced (nano)materials. In contrast with conventional chemicals, there is no system in place to distinguish unique nanomaterials. This lack of identification is a significant obstacle for the application of AI, but for instance within OECD work is ongoing to address this issue. 

RIVM is keeping a close eye on advancements in AI. It has initiated projects to gain practical experience developing and applying AI-based models in risk assessment.