Systematically evaluating read-across prediction and performance using a local validity approach characterized by...

Area of Science:

• Toxicology
• Computational Chemistry
• Regulatory Science

Background:

• Read-across is a key method for filling data gaps in chemical safety assessments.
• Current read-across methods face challenges in acceptance and uncertainty quantification.
• Automated approaches are needed to improve the reliability and efficiency of read-across.

Purpose of the Study:

• To develop and evaluate an algorithmic, automated approach for read-across predictions.
• To assess the utility of in vitro bioactivity data and chemical descriptors for predicting in vivo toxicity.
• To establish a performance baseline for read-across predictions for repeated dose toxicity studies.

Main Methods:

• Generated over 3239 chemical structure descriptors for 1778 chemicals.
• Incorporated data from 821 in vitro assays (bioactivity descriptors) from EPA's ToxCast program.
• Developed a generalized read-across (GenRA) model using nearest neighbor analysis to predict toxicity for 600 chemicals with in vivo data.

Main Results:

• The automated approach successfully established a performance baseline for read-across predictions.
• In vitro bioactivity descriptors were frequently more predictive of in vivo toxicity than chemical descriptors alone or in combination.
• The GenRA model demonstrated the utility of bioactivity data in facilitating read-across for toxicity studies.

Conclusions:

• The developed generalized read-across (GenRA) approach offers a systematic method for toxicity predictions.
• This systemization of read-across is a valuable step towards reliable screening-level hazard assessments for new chemicals.
• The findings support the increased use of in vitro bioactivity data in regulatory toxicology.