A recent study applied machine learning to investigate which physicochemical and experimental factors were most involved in genotoxicity of titanium dioxide (TiO₂). The findings confirmed that exposure concentration, cell medium composition, and lysis temperature in the comet assay correlate with DNA damage. The identified correlations could provide valuable insights for standardizing this test. However, the study methods and findings are too limited to identify new parameters involved in genotoxicity. Also, the scope was not aimed at providing evidence on the genotoxicity of TiO₂, and therefore the study has no direct relevance for the discussion on the carcinogenicity classification of TiO₂ nanomaterials.

Using machine learning to investigate TiO₂-induced genotoxicity

Recent advancements in machine learning (ML), the most commonly used form of artificial intelligence, have brought exciting possibilities to the fields of toxicology, by performing pattern recognition in large datasets related to the environmental and human health effects of chemicals and materials. A recent article used ML to predict the genotoxicity of TiO₂, which is a current topic of debate due to contrasting results in both in vitro and in vivo studies.

Selection of the dataset

A dataset was created from a range of existing in vitro studies investigating the ability of TiO₂ nanoparticles to induce DNA strand breaks. By including physicochemical properties in the analysis such as particle size, exposure conditions, and other experimental details like dispersion protocol, the researchers sought to establish which factors were positively associated with genotoxicity. After excluding studies with missing values from the dataset —primarily related to particle characterisation— and removing outliers, the researchers used a dataset of 1032 datapoints. This dataset was divided into 722 datapoints for training the machine learning model, which was then used to predict DNA damage in the 310 datapoints for testing. The authors call for improved reporting of results, because much of the data could not be used due to missing information.

The comet assay to detect DNA damage

The comet assay is a test that detects DNA breaks. To measure the effect of TiO₂ nanoparticles on DNA integrity, the particles are suspended in cell medium using sonication and then added to cell cultures. After a specified treatment time, DNA is isolated from the cells while remaining in the form of tightly coiled balls called nucleoids. When an electric current is applied to these nucleoids, the negatively charged DNA moves toward the positive electrode. The more breaks there are in the DNA, the easier it is for the DNA to move out of the nucleoid. This process can be compared to pulling pieces of thread from a wool ball; the more threads you cut, the easier it is to pull wool from the ball.

Factors that impact DNA damage

The factor most strongly correlated with DNA damage was exposure concentration, although it was still relatively weak. Other factors included the cell medium composition and whether the lysis step was performed cold. To a lesser extent, particle diameter had a negative impact on DNA damage (i.e., larger NMs might cause less DNA damage), while longer exposure times and higher electrophoresis voltages increased DNA damage. The effects of sonication time and power were not apparent. Overall, these findings were mostly as expected.

Reflection by RIVM

The current ML approach identified experimental factors most strongly associated with a positive genotoxic result for various TiO₂ nanoparticle samples. This information could help standardise experimental procedures. However, limited information relevant for the genotoxic potential of TiO₂ nanoparticles was obtained. In fact, the model’s performance decreased when experimental factors were included in the analysis, compared to when only physicochemical characteristics were included. Additionally, the researchers’ method for handling outliers and validating the model warrants further assessment. The findings of this study are therefore not directly relevant to the discussion on the carcinogenicity classification of TiO₂ nanoparticles.

The choice of the comet assay for this approach is understandable, given its frequent use and the abundance of available data. However, the in-vitro comet assay is not an ideal test for predicting genotoxicity, as it primarily serves as an indicator test. It measures DNA strand breaks that are often repaired, leading to a high likelihood of false-positive results. Consequently, while the test has a high sensitivity (its ability to identify true positive substances), its specificity (the ability to identify true negative substances) is relatively low. Importantly, an OECD test guideline (TG) is not yet available for this assay, indicating a lower level of standardisation compared to other genotoxicity tests that do have a TG. This is especially relevant for nanomaterials, which can interfere with the test readout that relies on a fluorescent signal. Adaptation of in vitro genotoxicity tests for testing of nanomaterials is currently ongoing, with the in vitro micronucleus test (OECD TG 487) representing the first TG that will be adapted.

Overall, the paper presents an interesting approach that uses ML for hazard assessment. If developed further, these methods could be valuable for hazard and risk assessment. However, in its current set-up, the approach has several limitations that prevent it from identifying novel nanoparticle characteristics and experimental parameters that affect genotoxicity induced by TiO₂. While it is likely that additional data points are needed to draw stronger conclusions, it is also recommended to include other in vitro methods, such as the micronucleus test or chromosomal aberration test. These methods are more relevant for evaluating genotoxicity, especially when performed according to their respective OECD TG and the high standards of Good Laboratory Practice.