Classification of a Naïve Bayesian Fingerprint model to predict reproductive toxicity^$

Area of Science:

• Computational toxicology
• Cheminformatics
• Toxicology

Background:

• Predicting complex toxicological endpoints like reproductive toxicity is crucial for chemical safety assessment.
• Regulatory frameworks such as REACH and TSCA necessitate robust methods for evaluating chemical hazards.
• Existing models may require descriptors, posing limitations in certain applications.

Purpose of the Study:

• To develop and validate a novel classification model for predicting chemical reproductive toxicity.
• To explore the utility of Naïve Bayesian methods with extended connectivity fingerprints (ECFPs) as a descriptor-free approach.
• To assess the model's performance in discriminating between toxic and non-toxic compounds.

Main Methods:

• Utilized a curated dataset of 1172 compounds from the Leadscope database and Procter and Gamble researchers.
• Developed a classification model employing Naïve Bayesian algorithms.
• Employed extended connectivity fingerprint 2 (ECFP 2) bits as direct inputs for the model, avoiding traditional descriptor generation.
• Evaluated model performance using probability scores and the Matthews Correlation Coefficient (MCC).

Main Results:

• The Naïve Bayesian Fingerprint model demonstrated good predictive performance for reproductive toxicity.
• The model achieved a Matthews Correlation Coefficient (MCC) value of ≥0.4 during validation.
• The descriptor-free approach using ECFP 2 bits proved effective in discriminating toxicity classes.

Conclusions:

• The developed Naïve Bayesian Fingerprint model offers a promising computational tool for predicting reproductive toxicity.
• This descriptor-free methodology simplifies model development and may enhance regulatory assessments.
• The model's performance supports its potential application in chemical safety evaluations under global regulations.