Prediction of rodent carcinogenicity bioassays from molecular structure using inductive logic programming

Area of Science:

• Computational chemistry
• Toxicology
• Machine learning

Background:

• Structure-activity relationships (SARs) are crucial for predicting chemical toxicity.
• Existing SAR methods struggle with noncongeneric compounds and understanding underlying mechanisms.
• The National Toxicology Program (NTP) requires accurate prediction of carcinogenicity.

Purpose of the Study:

• To apply the inductive logic programming (ILP) program Progol for carcinogenicity SAR analysis.
• To evaluate Progol's predictive accuracy against established methods using NTP data.
• To generate and investigate novel structural alerts for carcinogenesis.

Main Methods:

• Utilized Progol, an ILP algorithm employing a fully relational chemical structure description (atoms and bond connectivities).
• Applied Progol to predict carcinogenicity for NTP-tested compounds, including those from the first round of predictions.
• Validated Progol's performance using 5-fold cross-validation on the full NTP database.

Main Results:

• Progol demonstrated high accuracy in predicting carcinogenicity, comparable to or exceeding human and other machine methods.
• Progol achieved 63% (+/- 3%) prediction accuracy on the NTP database, relying solely on chemical structure and Salmonella mutagenicity data.
• Generated statistically independent structural alerts for carcinogenesis, offering new insights into mechanisms.

Conclusions:

• Progol is a powerful tool for SAR analysis, particularly for complex datasets like carcinogenicity.
• The generated structural alerts provide valuable, understandable rules for risk assessment.
• Progol shows significant potential in understanding SARs for various cancer-related compounds.