Factors influencing predictive models for toxicology.

Area of Science:

• Computational toxicology and cheminformatics
• Environmental risk assessment

Background:

• Model performance in predicting chemical toxicity is often limited by data variability.
• Understanding and quantifying this variability is crucial for reliable toxicological assessments.
• Pesticides represent a key area where accurate toxicity prediction is vital.

Purpose of the Study:

• To assess the impact of data variability on the comparison and evaluation of toxicity prediction models.
• To quantify the variability of molecular descriptors and compare it with experimental toxicological data variability.
• To investigate several toxicological endpoints for pesticides.

Main Methods:

• Calculation of hundreds of molecular descriptors (constitutional, electrostatic, geometrical, quantum-chemical, topological) using CODESSA, HyperChem, and Pallas.
• Evaluation of experimental toxicity data for multiple endpoints (e.g., Oncorhynchus mykiss, Daphnia magna, Acceptable Daily Intake, Anas Platyrhynchos, Colinus virginianus, Muridae) for pesticides.
• Assessment of descriptor variability influenced by molecular conformation and software, compared against experimental data variability.

Main Results:

• Molecular descriptor values exhibit variability dependent on molecular conformation and the software used for calculation.
• The extent of descriptor variability was quantified and compared directly with the variability observed in experimental toxicological data.
• Variability in toxicological data was considered across diverse endpoints and species.

Conclusions:

• Data variability, both in experimental results and calculated descriptors, significantly affects the reliability of toxicity prediction models.
• Quantifying descriptor variability is essential for accurate model evaluation and selection.
• This study provides insights into managing data variability for improved pesticide toxicity assessment.