Current Trends in Ligand-Based Virtual Screening: Molecular Representations, Data Mining Methods, New Application Areas, and Performance Evaluation

The increasing relevance and the use of data mining methods in virtual compound screening are studied. Virtual screening has evolved from traditional similarity searching, using single reference compounds, into an advanced application domain for data mining and machine learning methods that require compound reference sets of increasing size and of high-information content for training. Three approaches SVM learning, Bayesian methods and decision trees have been highlighted. Support vector machines represent a relatively new data mining methodology that has become popular during the 1990s based on the works of Vapnik and Cortes. Bayesian methods generally rely on the estimation of probability distributions of numerical representations of compounds based on property descriptors or fingerprints. Decision trees continue to be applied in chemoinfomatics, especially in the form of ensemble-based random forest models, which are utilized.