Predictive Capability of QSAR Models Based on the CompTox Zebrafish Embryo Assays: An Imbalanced Classification...

Area of Science:

• Environmental Toxicology
• Computational Chemistry
• Developmental Biology

Background:

• The CompTox Chemistry Dashboard provides a large database of Zebrafish (Danio rerio) developmental toxicity data.
• This dataset includes 19 toxicological endpoints for 1018 compounds at low concentrations, focusing on effects in dechorionated embryos.

Purpose of the Study:

• To evaluate the predictive capability of quantitative structure-activity relationship (QSAR) models for zebrafish developmental toxicity.
• To assess the performance of machine learning algorithms (logistic regression, multi-layer perceptron, random forests) in predicting imbalanced toxicological endpoints.

Main Methods:

• Developed 209 QSAR models using machine learning with penalization techniques and diverse quality metrics.
• Utilized classical molecular descriptors, structural fingerprints, and their combinations as model inputs.
• Assessed model performance using the Matthew's correlation coefficient (MCC) threshold of 0.20.

Main Results:

• Only 8 out of 209 developed QSAR models met the predefined acceptable quality threshold (MCC > 0.20).
• The best predictive models were achieved for mortality (MORT), ActivityScore, and jaw deformation (JAW) endpoints.
• Overall, QSAR models exhibited low predictability for the zebrafish embryotoxicity data.

Conclusions:

• The limited predictability of QSAR models is primarily attributed to the sensitivity of the 19 measured endpoints in dechorionated zebrafish embryos exposed to low concentrations.
• Further research is needed to improve QSAR model accuracy for complex developmental toxicity endpoints.
• The findings highlight challenges in predicting toxicity from high-throughput screening data with imbalanced endpoints.