Machine Learning Models Based on Enlarged Chemical Spaces for Screening Carcinogenic Chemicals.

Area of Science:

• Toxicology
• Computational Chemistry
• Data Science

Background:

• Machine learning (ML) models are crucial for chemical safety and management.
• Previous models had limitations due to small datasets and lack of applicability domain (AD) characterization.
• Regulatory applications require reliable models with defined AD.

Purpose of the Study:

• To develop and validate improved ML models for screening carcinogenic chemicals.
• To define the applicability domain (AD) using a structure-activity landscape (SAL) approach (AD_SAL).
• To apply the validated model to large chemical datasets for risk assessment.

Main Methods:

• Curated an enlarged dataset of 1697 compounds (940 carcinogens, 757 non-carcinogens).
• Constructed screening models using 12 molecular fingerprints, 4 ML algorithms, and 2 graph neural networks.
• Defined the optimal model's AD using the AD_SAL methodology.

Main Results:

• An optimal random forest model with PubChem fingerprints achieved an 86.2% area under the receiver operating characteristic curve.
• The model demonstrated superior performance compared to previous efforts.
• The AD_SAL methodology provided robust characterization for regulatory use.

Conclusions:

• The developed ML model, coupled with AD_SAL, is a promising tool for identifying carcinogenic chemicals.
• Successfully screened 1282 chemicals from the IECSC and 841 plastic additives.
• Facilitates prioritization of chemicals for sound chemical management and regulatory oversight.