Structure-activity relationship (SAR) modelling of mosquito larvicides

Area of Science:

• Computational chemistry
• Toxicology
• Medicinal chemistry

Background:

• Mosquito-borne diseases pose a significant global health threat.
• Developing effective and selective larvicides is crucial for vector control.
• Predictive models can accelerate the discovery of new larvicidal compounds.

Purpose of the Study:

• To derive quantitative structure-activity relationship (QSAR) models for predicting larvicidal activity.
• To identify novel chemical structures with potential larvicidal properties against Aedes aegypti.
• To evaluate various machine learning algorithms for predicting chemical activity.

Main Methods:

• Compilation of a database of 188 chemicals and their larvicidal activity (log 1/IC50) against Aedes aegypti.
• Chemical descriptor generation using CODESSA and autocorrelation methods.
• Application and comparison of multiple statistical and machine learning techniques including PLS, CART, Random Forest, ANN, and SVM.
• Development of a two-class classification model using a three-layer perceptron (TLP).

Main Results:

• Quantitative structure-activity models did not yield satisfactory results.
• The three-layer perceptron (TLP) model significantly outperformed other methods.
• The optimal TLP configuration utilized eight autocorrelation descriptors and four hidden neurons.
• The TLP model achieved over 96% prediction accuracy on both training and external test sets.

Conclusions:

• A highly accurate predictive model for Aedes aegypti larvicidal activity was successfully developed.
• The TLP model provides a robust tool for virtual screening and discovery of new larvicides.
• The study proposed novel candidate molecules for synthesis and experimental validation.