Structure-based classification of active and inactive estrogenic compounds by decision tree, LVQ and kNN methods

Area of Science:

• Computational chemistry
• Cheminformatics
• Toxicology

Background:

• Estrogenic compounds pose risks, necessitating accurate classification methods.
• Structure-activity relationship (SAR) studies are crucial for predicting compound behavior.
• Computational models can aid in identifying and classifying potentially harmful substances.

Purpose of the Study:

• To evaluate the performance of Decision Tree (DT), Learning Vector Quantization (LVQ), and k-nearest Neighbour (kNN) algorithms for classifying estrogenic compounds.
• To assess the predictive power of these machine learning models in a structure-activity relationship (SAR) context.
• To determine the most effective computational method for distinguishing active from inactive estrogenic compounds.

Main Methods:

• Utilized a dataset of 311 compounds for model development and validation.
• Employed principal components derived from DRAGON software-calculated molecular descriptors as structural variables.
• Applied and compared the classification performance of DT, LVQ, and kNN algorithms using separate training and testing datasets.

Main Results:

• The k-nearest Neighbour (kNN) algorithm exhibited the highest classification accuracy overall.
• Decision Tree (DT) showed the weakest performance, while LVQ performance varied.
• The best results were achieved with kNN on calf estrogen receptor data, with 98.3% correct classification in external tests.

Conclusions:

• All evaluated methods (DT, LVQ, kNN) are suitable for SAR-based classification of estrogenic compounds.
• The predictive power of the models ranged from adequate to excellent.
• kNN emerged as the most effective method for this specific classification task, highlighting its utility in cheminformatics and toxicology.