Additive SMILES-based carcinogenicity models: Probabilistic principles in the search for robust predictions

Area of Science:

• * Cheminformatics
• * Computational Toxicology
• * Quantitative Structure-Activity Relationship (QSAR) modeling

Background:

• * Simplified molecular input line entry system (SMILES) descriptors are crucial for quantitative structure-activity relationship (QSAR) modeling.
• * Rare SMILES attributes can lead to overtraining in predictive models.
• * Identifying and excluding rare attributes is essential for robust model development.

Purpose of the Study:

• * To evaluate two systems for building carcinogenicity prediction models using SMILES descriptors.
• * To assess the impact of rare attribute exclusion on model performance.
• * To compare a classic training-test system with a balance of correlations approach.

Main Methods:

• * Calculation of optimal descriptors using SMILES attributes and Monte Carlo methods.
• * Implementation of a function (limS) to identify and exclude rare attributes.
• * Comparison of a classic training-test system with a balance of correlations approach using subtraining, calibration, and test sets.

Main Results:

• * The balance of correlations approach demonstrated more robust carcinogenicity predictions across three random data splits.
• * Predictive performance metrics (r-squared and standard error) were consistently better with the balance of correlations method.
• * Excluding rare attributes by fixing their correlation weights to zero mitigated overtraining.

Conclusions:

• * Balancing correlations in model building provides more reliable carcinogenicity predictions than traditional training-test splits.
• * The limS function effectively identifies and excludes rare SMILES attributes, enhancing model generalizability.
• * This approach offers a more robust method for developing predictive toxicology models.