Local and global quantitative structure-activity relationship modeling and prediction for the baseline toxicity

Area of Science:

• Computational Chemistry
• Cheminformatics
• Toxicology

Background:

• Predictive accuracy is crucial for quantitative structure-activity relationship (QSAR) models in computational chemistry.
• Hypothesis: Models based on analogical chemicals outperform those using diverse compound sets.

Purpose of the Study:

• Develop a novel "clustering first, then modeling" scheme for QSAR.
• Build local QSAR models using subsets of structurally similar compounds.
• Enhance predictive performance for chemical properties and activities.

Main Methods:

• Hierarchical clustering for grouping training data by structural similarity.
• K-nearest neighbor algorithm for classifying validation and test sets.
• Partial least squares (PLS) for generating local QSAR models within subsets.

Main Results:

• Local QSAR models demonstrated significantly superior predictive performance compared to a global model.
• Validation on two independent datasets confirmed the effectiveness of the local modeling approach.
• Results align with the hypothesis that analogical chemical data improves model accuracy.

Conclusions:

• The "clustering first, then modeling" approach offers a promising strategy for enhancing QSAR predictive accuracy.
• This method is adaptable for modeling various physicochemical properties, biological activities, and toxicities.
• Local QSAR modeling based on structural similarity is a valuable advancement in computational chemistry.