Ecotoxicity prediction by adaptive fuzzy partitioning: comparing descriptors computed on 2D and 3D structures

Area of Science:

• Environmental toxicology
• Computational chemistry
• Cheminformatics

Background:

• Accurate prediction of pesticide ecotoxicity is crucial for environmental risk assessment.
• Quantitative Structure-Activity Relationships (QSAR) models are widely used but require careful descriptor selection and validation.
• Comparing different molecular descriptors and computational methods is essential for optimizing model performance.

Purpose of the Study:

• To establish robust classification models for predicting ecotoxicity across different species (trout, daphnia, quail, bee).
• To compare the predictive power of various molecular descriptors, including 2D and 3D structural parameters.
• To evaluate the impact of different software packages on model performance and computational efficiency.

Main Methods:

• Development of classification models using Adaptive Fuzzy Partition (AFP) based on Fuzzy Logic concepts.
• Selection of relevant molecular descriptors using a genetic algorithm-based procedure.
• Comparison of descriptor sets computed by multiple software packages, focusing on 2D and 3D structural parameters.
• Validation of models using cross-validation and independent test sets.

Main Results:

• Satisfactory ecotoxicity prediction models were achieved for all four species, with best scores around 70-75%.
• Models utilizing 2D structural descriptors performed as well as or better than those using 3D descriptors.
• The computational time for 3D descriptors often did not improve predictive ability, suggesting potential for efficiency gains.
• Differences in model performance across various software packages were minimal, indicating robustness.

Conclusions:

• 2D molecular descriptors are often sufficient for developing accurate ecotoxicity models, challenging the necessity of computationally intensive 3D descriptors.
• The developed AFP models provide a reliable method for assessing pesticide ecotoxicity across diverse species.
• The findings support the use of different descriptor packages, enabling flexibility and potentially reducing computational costs in ecotoxicological modeling.