Automated machine learning in nanotoxicity assessment: A comparative study of predictive model performance

Area of Science:

• Computational toxicology
• Materials science
• Machine learning applications

Background:

• Computational modeling is a promising alternative to animal testing for toxicity assessment.
• Developing optimized predictive models traditionally requires significant expertise, time, and intensive searching for algorithms and hyperparameters.

Purpose of the Study:

• To evaluate the efficacy of automated machine learning (autoML) platforms in developing nanotoxicity prediction models.
• To compare the performance of autoML-generated models against conventionally developed machine learning (ML) models.

Main Methods:

• Utilized seven publicly available datasets for oxides and metals to develop nanotoxicity prediction models.
• Employed three autoML platforms: Vertex AI, Azure, and Dataiku.
• Compared performance metrics (accuracy, F1 score, precision, recall) of autoML models with conventional ML models.

Main Results:

• AutoML platforms generated more reliable nanotoxicity prediction models, outperforming conventional ML models.
• No single autoML platform significantly outperformed the others, but differences in technical features offer user choice.
• Models built with higher quality datasets showed enhanced performance.

Conclusions:

• AutoML platforms streamline the development of effective nanotoxicity prediction models.
• The choice of autoML platform can be tailored to user expertise and technical feature preferences.
• Future improvements in nanotoxicity prediction models are expected with high-quality datasets.