Nonanimal Models for Acute Toxicity Evaluations: Applying Data-Driven Profiling and Read-Across.

Area of Science:

• Computational toxicology
• In vitro bioassays
• Predictive modeling

Background:

• In vitro bioassays offer a cost-effective alternative to animal testing for toxicity evaluation.
• Utilizing public in vitro bioassay data for chemical toxicity assessments like read-across is challenging due to data curation and analysis complexities.
• Massive amounts of unstructured public data are underutilized in toxicity testing.

Purpose of the Study:

• To develop a computational method for automated extraction of bioassay data from PubChem.
• To assess the prediction accuracy of animal toxicity using a novel bioprofile-based read-across approach.
• To identify relevant toxicity mechanisms through bioassay data analysis.

Main Methods:

• A training database of 7,385 compounds with rat acute oral toxicity data was used to create in vitro bioprofiles from PubChem.
• A subspace clustering algorithm identified bioassay groups relevant to acute oral toxicity mechanisms.
• Read-across methodology with cross-validation was employed to predict animal acute oral toxicity.
• An external test set of over 600 compounds validated the model's predictive performance.

Main Results:

• Several bioassay clusters demonstrated high predictivity for acute oral toxicity (62-100% positive prediction rate) via cross-validation.
• An ensemble model incorporating these clusters achieved a 76% positive prediction rate for acute toxicity in an external test set.
• Novel in vitro-in vivo relationships were identified, offering insights into animal toxicity mechanisms.

Conclusions:

• The developed in vitro bioassay data-driven profiling strategy addresses the needs of big data in computational toxicology.
• This approach can be extended to develop predictive models for other complex toxicity endpoints.
• The study highlights the potential of leveraging public bioassay data for robust toxicity prediction.