Engineered nanomaterials (ENMs) have found their applications in various technologies and consumer products. Manipulating chemicals at the nanoscale range introduces unique characteristics to these materials and makes them desirable for technological applications.

With the increasing production of ENMs, there have been adverse effects on the environment. Moreover, it is unfeasible to estimate the risks caused by ENMs each time via in vivo or in vitro experiments. To this end, in silico methods can come to the rescue to perform such evaluations.

In an article published in the journal Chemosphere, the performance of different machine learning algorithms was investigated for predicting well-defined in vivo toxicity endpoint and to identify the important features involved with in vivo nanotoxicity of Daphnia magna.

The results revealed comparable performances of all algorithms and the predictive performance exceeded approximately 0.7 for all metrices evaluated. Furthermore, artificial neural network, random forest, and k-nearest neighbor models showed a marginally better performance compared to the other algorithm models.

The variable importance analysis performed to understand the significance of input variables revealed that physicochemical properties and molecular descriptors were important within most models. On the other hand, properties related to exposure conditions gave conflicting results. Thus, the machine learning models helped generate in vivo endpoints, even with smaller datasets, demonstrating their reliability and robustness.

Role of Machine Learning in Nanotechnology

Nanotechnology has emerged as a key technology with implications agriculture, medicine, and food industries. Thus, ENMs are more appealing than their larger counterparts due to their outstanding features owing to their smaller size.

Despite their advantages, ENMs have also caused effects on the environment, impacting the health and safety of the environment, calling for environmental risk assessment associated with ENMs. However, this assessment via in vivo or in vitro testing for all fabricated nanoforms is impractical.

The challenge in risk assessment is not only due to extensive ENM production and applications but also due to the large diversity of materials. To this end, chemical modification at the nanoscale range may modulate the physicochemical properties and consequential toxicity profile of the materials.

Recent advances in machine learning offered new tools to extract new insights from large data sets and to acquire small data sets more effectively. Researchers in nanotechnology use machine learning tools to tackle challenges in many fields. Due to their compatibility with complex interactions, machine learning can help predict the toxicological effects of ENMs through large data sets.

The field of nanotoxicology lacks standardized procedures to depict common ontologies to measure ENM properties. However, the models from limited datasets can help generate the key nanotoxicological descriptors. The nanotoxicological models based on machine learning developed to date focused on endpoints like viability or cytotoxicity.

In Silico Machine Learning Tools for The Prediction of Daphnia Magna Nanotoxicity

Despite considerable efforts, various obstacles still exist for in silico modeling of nanotoxicological effects due to limited data availability and poor data curation. Hence, better agreement on data quality, experimental protocols, and availability are vital to acquiring homogenous data across different studies.

In the present work, the performance of machine learning algorithms for predicting in vivo nanotoxicity of metallic ENMs towards Daphnia magna was investigated. Various models were generated based on the sources obtained from immobilization data, which were in congruence with the principles of organization for economic co-operation and development (OECD). Furthermore, the limitations in obtaining consistent data for modeling were overcome by applying different methods of data curation.

Among the six machine learning models generated based on OECD, neural network, random forest, and k-nearest neighbor algorithms showed the highest performance, while the other models showed relatively similar performance. This indicates that machine learning is more suitable for in silico modeling of in vivo nanotoxicity than the actual algorithm. Additionally, key descriptors that modulated the toxicity of metallic ENMs towards Daphnia magna were also studied based on the generated machine learning models.

Conclusion

To summarize, machine learning algorithms were performed to predict the in vivo nanotoxicity of metallic ENMs. The collected Daphnia magna toxicity data for metallic ENMs were analyzed using six classification machine learning models based on the principles of OECD.

The results revealed that artificial neural networks, random forest, and k-nearest neighbor algorithms had the highest performances, which were in line with previous reports from the literature. On the other hand, the relative differences in other algorithm models were comparatively small. These results proved the compatibility of machine learning for in silico modeling of in vivo nanotoxicity.

Furthermore, feature importance analysis using machine learning algorithms revealed contradictory results in all the models, with physicochemical properties and molecular descriptors being significant features within models. The results demonstrated that the models with small datasets with few physicochemical properties and molecular descriptors result in machine learning models with good predictive performance.