Generative Topographic Mapping Approach to Chemical Space Analysis

Abstract

Generative Topographic Mapping (GTM) is a probabilistic, non-linear dimensionality reduction method, developed by C. Bishop et al. It essentially represents a fuzzy-logics-based enhancement of Kohonen Self-Organizing Maps (SOM). The probabilistic nature of this method is the source of the multivalent applications of GTM, which goes well beyond simple dimensionality reduction and visualization, but allows straightforward comparison of large compound libraries (in terms of diversity and coverage), supports regression or classification models with applicability domain control and herewith may serve as predictive tools of a large panel of properties (including polypharmacological profiles of bioactive compounds). A good predictive modeling implicitly validates a map and provides an objective criterion to select the best suited ones, out of the multitude of possibilities based on different initial molecular descriptors and user-defined mapping parameters. This multi-purpose “Swiss army knife” of dimensionality reduction may furthermore extract “privileged” structural patterns associated to bioactivities of interest, and hence contribute to an intuitive understanding of structure-activity relationships.