Non-linear QSAR modeling by using multilayer perceptron feedforward neural networks trained by back-propagation

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Pharmacology

Background:

• Quantitative structure-activity relationship (QSAR) studies are crucial for drug discovery.
• Non-linear relationships between chemical structures and biological activity often pose challenges for traditional modeling methods.
• Developing robust QSAR models requires advanced computational techniques.

Purpose of the Study:

• To present and detail the application of multilayer perceptrons (MLP) feedforward neural networks for non-linear QSAR model building.
• To compare the performance of MLP with conventional QSAR methods, including multiple linear regression (MLR), partial least squares (PLS), and quadratic PLS (QPLS).
• To evaluate the predictive accuracy of different QSAR models using a case study on alpha adrenoreceptor agonists.

Main Methods:

• A case study involving 18 alpha adrenoreceptor agonists across three pharmacological classes.
• Description of each compound using 15 chemical features (X block).
• Measurement of six pharmacological responses for each agonist (Y block).
• Training MLP neural networks using the back-propagation (BP) algorithm.
• Comparison of MLP performance against MLR, PLS, and QPLS models.

Main Results:

• MLP demonstrated slightly superior performance compared to MLR, PLS, and QPLS.
• The correlation coefficient of observed versus predicted response plots was used as the primary goodness-of-fit indicator.
• MLP models showed improved accuracy in predicting the pharmacological responses of alpha adrenoreceptor agonists.

Conclusions:

• Multilayer perceptrons (MLP) are a powerful tool for building non-linear QSAR models.
• MLP offers a slight performance advantage over traditional methods for this specific dataset.
• The findings support the utility of advanced neural network approaches in medicinal chemistry and drug design.