Estimation of carcinogenicity using molecular fragments tree

Area of Science:

• Toxicology
• Medicinal Chemistry
• Computational Chemistry

Background:

• Carcinogenicity is a critical toxicological concern in drug discovery.
• Accurate prediction of carcinogenicity is essential for safe drug development.
• Existing methods for identifying carcinogenic structural alerts have limitations.

Purpose of the Study:

• To develop a novel, statistically driven method for extracting structural alerts (SAs) and modulating factors of carcinogens.
• To enhance the prediction accuracy of carcinogenicity in drug candidates.
• To provide an automated and unbiased approach for SA identification.

Main Methods:

• Utilized the Gaston algorithm for frequent subgraph mining to detect recurring substructures.
• Constructed and pruned a molecular fragments tree to identify high-quality SAs.
• Employed statistical significance (p-values) in binomial tests for SA selection.
• Developed three rules to extract modulating factors that mitigate SA toxicity.

Main Results:

• Developed a model with 77 SAs and 4 SA/modulating factor pairs.
• Achieved prediction accuracies of 0.70 for the training set and 0.65 for the test set.
• Demonstrated superior predictive ability compared to Benigni's model, particularly on the test set.

Conclusions:

• The developed method offers improved prediction accuracy for carcinogenicity.
• The identified SAs are valuable for both prediction and interpretation of toxicological data.
• This automated, unbiased approach facilitates SA extraction without requiring prior mechanistic knowledge.