A Comprehensive Study to Compare Different Compound Representations for Predicting Carcinogenicity In Vivo.

Area of Science:

• Computational toxicology
• Chemical risk assessment
• Machine learning in drug discovery

Background:

• In vivo carcinogenicity testing is resource-intensive and ethically complex.
• Machine learning (ML) models offer a promising alternative for carcinogenicity prediction.
• The impact of different molecular representations on ML model performance requires systematic investigation.

Purpose of the Study:

• To evaluate the influence of molecular embeddings, classical descriptors, and structural alerts on ML models for predicting in vivo carcinogenicity.
• To compare the predictive performance of various molecular representation strategies.
• To identify optimal approaches for regulatory-relevant carcinogenicity modeling.

Main Methods:

• Assembled a dataset of 2090 compounds with in vivo rodent carcinogenicity data from five toxicological databases.
• Represented compounds using classical molecular descriptors, structural alerts, SMILES-derived embeddings, and hybrid combinations.
• Benchmarked 24 ML classifiers using 10-fold cross-validation, assessing performance via accuracy, precision, recall, F1-score, and AUC-ROC.

Main Results:

• Hybrid representations combining molecular descriptors and structural alerts consistently yielded the best predictive performance across ML models.
• Molecular embeddings served as valuable complementary features but did not outperform classical representations.
• Chemically interpretable, expert-driven descriptors, especially those with genotoxic alerts, are vital for robust carcinogenicity modeling.

Conclusions:

• Combining molecular descriptors with structural alerts is a superior strategy for building accurate ML models for carcinogenicity prediction.
• Classical molecular representations remain essential, with embeddings offering supplementary predictive power.
• These findings support the use of interpretable descriptors in regulatory toxicology for chemical safety assessment.