Implementing the Fisher's discriminant ratio in a k-means clustering algorithm for feature selection and data set...

Area of Science:

• Computational Chemistry
• Cheminformatics
• Bioinformatics

Background:

• Accurate feature selection and data trimming are crucial for understanding structure-activity relationships in drug discovery.
• HIV-1 protease inhibitors represent a significant class of antiviral drugs, necessitating efficient analysis methods.
• Existing methods may not optimally balance feature selection with outlier identification for complex datasets.

Purpose of the Study:

• To develop and validate a simultaneous feature selection and data trimming method for HIV-1 protease inhibitors.
• To identify class-sensitive molecular descriptors that correlate with inhibitory activity (pKi).
• To improve the quality of datasets for further quantitative structure-activity relationship (QSAR) analyses.

Main Methods:

• Implementation of Fisher's discriminant ratio as a class separability criterion within a k-means clustering algorithm.
• Application of feature evaluation indices including Shannon entropy, linear regression, and stepwise variable selection for descriptor filtering.
• Division of 221 HIV-1 protease inhibitors into five classes to optimize feature selection.

Main Results:

• The method successfully identified topological descriptors that are well-correlated with pKi values.
• K-means clustering effectively identified and removed outliers (trimmed inhibitors).
• Retaining 44% (98 inhibitors) with three selected descriptors resulted in significantly more class-sensitive data, validated by PLS regression.

Conclusions:

• The combined Fisher's discriminant ratio and k-means clustering approach offers an effective strategy for simultaneous feature selection and data trimming.
• Selected topological descriptors provide valuable insights into the activity of HIV-1 protease inhibitors.
• The refined dataset enhances statistical significance for subsequent QSAR modeling and drug design.