Simultaneous feature selection and parameter optimisation using an artificial ant colony: case study of melting point...

Area of Science:

• Computational Chemistry
• Cheminformatics
• Machine Learning

Background:

• Quantitative Structure-Property Relationship (QSPR) models are crucial for predicting chemical compound properties.
• Traditional QSPR model development often faces challenges with feature selection and parameter optimization, leading to overfitting.
• The Winnowing Artificial Ant Colony (WAAC) algorithm is introduced as an advancement over existing ant colony optimization methods.

Purpose of the Study:

• To develop a novel algorithm (WAAC) for simultaneous feature selection and model parameter optimization in QSPR.
• To evaluate the performance of WAAC in developing predictive models for melting point values.
• To assess WAAC's ability to improve model accuracy and reduce overfitting compared to other methods.

Main Methods:

• The Winnowing Artificial Ant Colony (WAAC) algorithm, an extension of modified ant colony algorithms, was employed.
• WAAC was tested for feature selection and parameter optimization in Partial Least Squares (PLS) and Support Vector Machine (SVM) models.
• The Karthikeyan dataset for melting point prediction was used for model development and validation.

Main Results:

• WAAC selected a PLS model with 68 descriptors (RMSE 46.6°C, R² 0.51) and an SVM model with 28 descriptors (RMSE 45.1°C, R² 0.54).
• The optimized SVM model demonstrated superior performance over kNN and comparable performance to Random Forest models.
• The WAAC-optimized SVM model exhibited reduced bias at the extremes of melting point values compared to Random Forest.

Conclusions:

• WAAC effectively optimizes regression models, mitigating overfitting through careful objective function selection.
• The algorithm simultaneously tunes model parameters for methods like SVM and PLS.
• WAAC's feature selection process, utilizing a winnowing procedure, efficiently removes irrelevant descriptors.