Two-Stage Machine Learning-Based Approach to Predict Points of Departure for Human Noncancer and...

Area of Science:

• Computational toxicology
• Environmental health
• Machine learning applications

Background:

• Chemical Points of Departure (PODs) are essential for human health risk assessment but are lacking for most commercial chemicals due to insufficient *in vivo* toxicity data.
• Existing methods for determining PODs are limited, hindering comprehensive risk evaluation and management for a vast number of chemicals.

Purpose of the Study:

• To develop and validate a machine learning (ML) framework for predicting human-equivalent PODs for organic chemicals based on their structure.
• To assess potential human health risks for a large set of environmental chemicals using the developed ML models.

Main Methods:

• A two-stage ML framework was created: Stage 1 used ML predictions of chemical properties (from OPERA 2.9) as features.
• Stage 2 involved training random forest regression models with human-equivalent PODs derived from *in vivo* data for general noncancer (n=1,791) and reproductive/developmental effects (n=2,228).
• Robust cross-validation was employed for feature selection and error estimation.

Main Results:

• The two-stage ML models accurately predicted PODs for both general noncancer and reproductive/developmental effects, with cross-validation errors under an order of magnitude.
• Application of the models to 34,046 environmental chemicals identified several thousand chemicals of moderate concern and several hundred of high concern.
• These findings are based on estimated median population exposure levels.

Conclusions:

• The developed ML framework effectively predicts human-equivalent PODs, significantly expanding the number of chemicals that can be evaluated for health risks.
• This approach provides a scalable method for prioritizing chemicals for further toxicological investigation and risk management.
• The study highlights a substantial number of environmental chemicals requiring attention due to potential health risks.