A novel method for building regression tree models for QSAR based on artificial ant colony systems

Area of Science:

• Computational chemistry
• Machine learning
• Cheminformatics

Background:

• Regression trees are valuable for quantitative structure-activity relationships (QSAR) due to their ability to handle large datasets, perform feature selection, and provide interpretable models.
• Recursive partitioning is a common, fast greedy algorithm for building regression trees, but it is not universally optimal.

Purpose of the Study:

• To introduce and evaluate a novel data partitioning method for regression trees using artificial ants.
• To compare the performance of the artificial ant-based method against conventional recursive partitioning in QSAR modeling.

Main Methods:

• Development of a new data partitioning algorithm inspired by artificial ant behavior.
• Application of the novel method and recursive partitioning to build regression tree models.
• Evaluation of model performance on three established QSAR datasets.

Main Results:

• The artificial ant-based partitioning method demonstrated superior performance compared to recursive partitioning.
• The novel method effectively handles large datasets and aids in feature selection for QSAR models.
• The developed models were readily interpretable, a key advantage of regression trees.

Conclusions:

• Artificial ant-based data partitioning offers a promising alternative to recursive partitioning for regression tree construction in QSAR.
• This novel approach enhances the accuracy and potentially the interpretability of quantitative structure-activity relationship models.
• The findings suggest broader applicability of swarm intelligence algorithms in cheminformatics and machine learning.