QSAR/QSPR studies using probabilistic neural networks and generalized regression neural networks

Area of Science:

• Computational Chemistry
• Cheminformatics
• Machine Learning in Drug Discovery

Background:

• Quantitative Structure-Activity Relationship (QSAR) and Quantitative Structure-Property Relationship (QSPR) studies are crucial for predicting molecular behavior.
• Traditional methods like k-nearest neighbor, multiple linear regression, and feed-forward neural networks have been employed for these predictions.
• There is a need for efficient and accurate predictive models that require minimal input features.

Purpose of the Study:

• To present Probabilistic Neural Network (PNN) and Generalized Regression Neural Network (GRNN) as powerful tools for QSAR and QSPR.
• To demonstrate the application of PNN for classification and GRNN for continuous function mapping in cheminformatics.
• To compare the performance of PNN/GRNN with traditional machine learning techniques.

Main Methods:

• Probabilistic Neural Network (PNN) methodology for binary classification tasks.
• Generalized Regression Neural Network (GRNN) methodology for continuous function mapping.
• Utilized molecular descriptors encoding structural features as network inputs.

Main Results:

• PNN successfully identified potential soluble epoxide hydrolase inhibitors.
• GRNN accurately predicted the aqueous solubility of small organic molecules.
• PNN and GRNN models demonstrated predictive power comparable to traditional methods.
• Significantly fewer input descriptors were required for PNN and GRNN models compared to traditional techniques.

Conclusions:

• PNN and GRNN are effective and efficient neural network techniques for QSAR and QSPR studies.
• These probability-based methods offer a simplified approach to molecular modeling.
• The reduced need for descriptors makes PNN/GRNN advantageous for predictive modeling in cheminformatics.