Forecasting nanoparticle toxicity using nonlinear predictive regressor learning systems.

Area of Science:

• Nanotechnology
• Toxicology
• Computational Biology

Background:

• Nanoparticle (NP) toxicity assessment is complex due to numerous influencing factors.
• Accurate prediction of NP effects on biological tissues is crucial and urgent.
• Existing databases offer potential for data-driven toxicity prediction.

Purpose of the Study:

• To evaluate the predictive power of nanoparticle descriptors and assay characteristics for NP toxicity.
• To utilize machine learning models to mine the HORIZON 2020 MODENA COST NP cytotoxicity database.
• To predict toxicity endpoints such as EC25, EC50, and slope.

Main Methods:

• Employed nonlinear predictive regressor learning systems, including Support Vector Regressors (SVR) and Radial Basis Function (RBF) regressors.
• Utilized a nested-cross validation scheme for parameter optimization.
• Applied the ReliefF algorithm for feature selection from a comprehensive list of NP attributes.

Main Results:

• Machine learning models achieved high correlations (above 0.90) between predicted and real toxicity values after feature selection.
• Radial Basis Function (RBF) regressors demonstrated superior performance in predicting toxicity endpoints.
• Nanoparticle structural attributes were identified as the most informative features for toxicity prediction.

Conclusions:

• Data-driven, machine learning methods can accurately predict NP toxicity endpoints (EC25, EC50, slope) in Adenosine triphosphate (ATP)-based assays.
• Nanoparticle structural features are key predictors of toxicity.
• These findings support the use of computational approaches for NP safety assessment.