Comparing LD₅₀/LC₅₀ Machine Learning Models for Multiple Species.

Area of Science:

• Computational toxicology
• Machine learning in drug discovery
• Chemical safety assessment

Background:

• The lethal dose/concentration (LD50/LC50) quantifies chemical toxicity, crucial for safety decisions and personal protective equipment selection.
• Traditional LD50/LC50 assessments involve significant animal use, though methods are evolving to reduce this.
• Compounds with LD50 < 25 mg/kg are often classified as highly toxic, providing vital safety insights.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting in vivo LD50/LC50 values.
• To utilize public data for building classification and regression models across multiple species.
• To explore computational approaches for reducing animal testing in toxicity assessments.

Main Methods:

• Building machine learning models (classification and regression) using public in vivo LD50/LC50 data.
• Employing 5-fold cross-validation with various algorithms for statistical assessment.
• Utilizing an external curated test set for model validation, specifically for mouse LD50.
• Developing models for multiple species including rat, mouse, fish, and daphnia.

Main Results:

• Developed and validated machine learning models for predicting LD50/LC50 values across different species.
• Demonstrated the utility of these models in classifying toxicity and predicting quantitative values.
• Cross-validation statistics and external test set performance provide insights into model reliability.

Conclusions:

• Computational models offer a promising avenue to estimate chemical toxicity and reduce animal usage.
• Understanding model applicability domains is essential for reliable predictions of novel molecules.
• These models can bridge data gaps in toxicity datasets and aid in scoring large chemical libraries for potential hazards.