Beyond QSARs: Quantitative Knowledge-Activity Relationships (QKARs) for enhanced drug toxicity prediction

Area of Science:

• Computational toxicology
• Drug safety assessment
• Artificial intelligence in pharmacology

Background:

• Quantitative Structure-Activity Relationships (QSARs) traditionally predict toxicity based on chemical structure.
• QSARs' structure-centric approach limits accuracy for drugs with minor structural changes but significant toxicity differences.
• Advances in AI offer new methods to integrate broader chemical knowledge for enhanced toxicity prediction.

Purpose of the Study:

• Introduce a novel framework, Quantitative Knowledge-Activity Relationships (QKARs), for predicting drug toxicity.
• Evaluate QKARs' performance against QSARs for drug-induced liver injury (DILI) and drug-induced cardiotoxicity (DICT).
• Explore the integration of knowledge-based and structure-based data for improved predictive accuracy.

Main Methods:

• Developed QKAR models using domain-specific knowledge representations for DILI and DICT.
• Applied five distinct machine learning algorithms to assess model complexity's impact on performance.
• Compared QKAR performance against QSARs using identical datasets and endpoints.
• Investigated combined knowledge and structure-based models (Q(K+S)ARs).

Main Results:

• QKAR models utilizing comprehensive drug knowledge showed superior prediction accuracy.
• Model complexity showed minimal association with performance across different machine learning algorithms.
• QKARs consistently outperformed QSARs for both DILI and DICT.
• QKARs effectively distinguished drugs with similar structures but varying liver toxicity profiles.

Conclusions:

• QKARs represent a robust alternative to traditional QSARs for drug toxicity assessment.
• Leveraging domain-specific knowledge significantly enhances toxicity prediction accuracy.
• Integrating knowledge-based and structure-based data offers further improvements in predictive modeling.