Polynomial neural network for linear and non-linear model selection in quantitative-structure activity relationship...

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Quantitative-structure activity relationship (QSAR) studies are crucial for drug discovery and chemical safety assessment.
• Developing accurate predictive models is essential for understanding the relationship between chemical structures and biological activity.
• Existing methods may lack flexibility in model complexity or predictive performance.

Purpose of the Study:

• To introduce a novel self-organising multilayered iterative algorithm for generating polynomial regression models.
• To enable user control over the number and power of terms in linear and non-linear models.
• To evaluate the algorithm's performance against established methods like partial least squares (PLS).

Main Methods:

• Development of a self-organising multilayered iterative algorithm.
• Application of the algorithm to generate polynomial regression models with user-defined complexity.
• Comparative analysis using fourteen diverse datasets in QSAR studies.
• Benchmarking against the partial least squares (PLS) algorithm.

Main Results:

• The proposed algorithm successfully generates both linear and non-linear polynomial regression models.
• Models generated by the algorithm are characterized by simplicity and high prediction ability.
• Demonstrated superior or comparable performance to PLS across fourteen QSAR datasets.
• The algorithm effectively selects parsimonious models relevant to QSAR.

Conclusions:

• The novel iterative algorithm offers a powerful and flexible tool for QSAR modeling.
• Its ability to create simple, highly predictive models makes it valuable for researchers.
• The algorithm's user-defined control over model complexity enhances its applicability.
• Accessible software implementation facilitates widespread adoption in QSAR research.