Comparison of methods for chemical-compound affinity prediction

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• Accurate prediction of chemical compound-protein binding affinities is crucial for effective virtual screening.
• Selecting optimal molecular descriptors and classification algorithms significantly impacts prediction accuracy.

Purpose of the Study:

• To compare the performance of various feature selection methods and classifiers for predicting compound-protein binding affinities.
• To identify the most effective strategies for enhancing virtual screening accuracy.

Main Methods:

• Evaluated multiple feature selection techniques, including the genetic algorithm.
• Assessed various classification algorithms such as Random Forests, Adaboosts, Baggings, and support-vector machines.
• Utilized six diverse compound series targeting different proteins (e.g., cytochrome P450 2C9, estrogen receptor).

Main Results:

• The genetic algorithm outperformed other feature selection methods.
• Combinations of the genetic algorithm with Random Forests, Adaboosts, or Baggings achieved performance comparable to support-vector machines and superior to other classifiers.
• Selected descriptors were plausible and aided in model interpretation.

Conclusions:

• The genetic algorithm is a highly effective feature selection method for predicting compound-protein binding.
• Ensemble methods and Random Forests, when combined with the genetic algorithm, offer robust classification performance for virtual screening.
• The study provides valuable insights into optimizing computational approaches for drug discovery.