Data-Driven Quantitative Structure-Activity Relationship Modeling for Human Carcinogenicity by Chronic Oral Exposure

Area of Science:

• Computational toxicology
• cheminformatics
• predictive modeling

Background:

• Traditional chemical toxicity assessments are costly and time-consuming.
• Quantitative structure-activity relationship (QSAR) models offer a cost-effective alternative but often suffer from limited training data and low predictivity.
• Accurate prediction of carcinogenicity is crucial for public health and regulatory purposes.

Purpose of the Study:

• To develop and validate a data-driven QSAR modeling approach for predicting chemical carcinogenicity.
• To identify potential new human carcinogens using the developed models.
• To establish an automated technique for prioritizing toxicants based on validated QSAR models.

Main Methods:

• Utilized a US Environmental Protection Agency Integrated Risk Information System (IRIS) carcinogen dataset to identify relevant PubChem bioassays.
• Selected eight PubChem assays demonstrating significant carcinogenicity predictivity for QSAR model training.
• Developed 15 QSAR models using 5 machine learning algorithms and 3 chemical fingerprint types for each selected assay.

Main Results:

• The developed QSAR models demonstrated acceptable predictivity, with an average cross-validation classification accuracy of 0.71.
• The models accurately predicted and ranked the carcinogenic potentials of 342 IRIS compounds, achieving a positive predictive value of 0.72.
• Potential new carcinogens were identified and subsequently validated through a literature search.

Conclusions:

• The data-driven QSAR approach provides a reliable and efficient method for predicting chemical carcinogenicity.
• This methodology enables the identification and prioritization of potential new human carcinogens.
• The study highlights the potential for automated toxicant prioritization using validated QSAR models trained on extensive public data resources.