CHEMOMETRIC APPLICATIONS OF NAÏVE BAYESIAN MODELS IN DRUG DISCOVERY
Eugen Lounkine, Peter S. Kutchukian, Meir Glick · 2013
Naïve Bayesian models (NBMs) have proven a valuable and efficient approach to virtual screening of compounds and in silico target prediction. However, their application goes well beyond the mere scoring of compounds. A particularly useful quality of NBMs is their interpretability where weights of molecular features are associated with the class (e.g., activity). This opens up a variety of possibilities to project domain knowledge from distinct fields, such as chemistry, bioactivity, and adverse drug reactions (ADRs) into a common chemical reference space. We reflect on prerequisites for the practical utilization of NBMs for virtual screening including data quality and normalization and discuss their practical application in the context of phenotypic screening campaigns, whose focus shifts from a primary target to multiple (desired as well as unwanted) biological activities, providing a real-world example of mode of action (MOA) elucidation for phenotypic assay hits. We discuss mining for enriched features, both chemical and biological, and their interpretation. Going beyond these well-established applications, we show how decisions of medicinal chemists can be rationalized and expressed in terms of straightforward rules using a naïve Bayesian approach to modeling medicinal chemists' decisions. Furthermore, we illustrate how NBMs can be combined with molecular similarity principles to define biological context-sensitive compound similarity metrics and clustering.