Identification of developmental and reproductive toxicity of biocides in consumer products using ToxCast bioassays data and machine learning models
Developmental and reproductive toxicity (DART) is a critical regulatory endpoint for protecting human health from chemical exposure. Biocides in consumer products are widely encountered in daily life, yet DART assessments often require substantial animal testing and are costly, leading to significan...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-08-01
|
| Series: | Environment International |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S0160412025003721 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Developmental and reproductive toxicity (DART) is a critical regulatory endpoint for protecting human health from chemical exposure. Biocides in consumer products are widely encountered in daily life, yet DART assessments often require substantial animal testing and are costly, leading to significant data gaps for these chemicals. This study aimed to identify ToxCast bioassays relevant to DART and develop machine learning models to screen biocides in consumer products for their DART potential. Initially, we compiled 201 bioassays linked to DART-related mechanisms using the Integrated Chemical Environment (ICE) database of the National Toxicology Program of (NTP). For these assays, we identified chemicals common to both ToxCast bioassays and in vivo ToxRefDB and conducted correlation analyses between bioactive concentrations (AC50) or hit-call data from in vitro assays and the lowest effect levels (LELs) from in vivo DART studies. This analysis revealed 25 bioassays with statistically significant correlations to in vivo DART data. Using the bioactivity data from these selected assays, we trained machine learning algorithms, including Random Forest, XGBoost, Gradient Boosting, Decision Tree, and Logistic Regression, on molecular fingerprints (MACCS, Morgan, Layered, RDKit). Based on the assumption that chemicals active in these ToxCast assays and their machine learning models may have DART potential, we prioritized active biocides in consumer products. For biocides with existing in vivo DART data, this approach achieved acceptable predictivity with F1 score of 76.2%. For biocides lacking in vivo DART data, we identified 100 active biocides as a high-priority group for further assessment. This study suggested the potential of ToxCast bioassays and machine learning models in predicting DART potential, offering a promising approach to address data-gap in consumer product safety. |
|---|---|
| ISSN: | 0160-4120 |