Data and Graph Mining in Chemical Space for ADME and Activity Data Sets

We present a classification method, which is based on a coordinate‐free chemical space. Thus, it does not depend on descriptor values commonly used by coordinate‐based chemical space methods. In our method the molecular similarity of chemical structures is evaluated by a generalized maximum common graph isomorphism, which supports the usage of numerical physicochemical atom property labels in addition to discrete‐atom‐type labels. The Maximum Common Substructure (MCS) algorithm applies the Highest Scoring Common Substructure (HSCS) ranking of Sheridan and co‐workers, which penalizes discontinuous fragments. For all compared classification algorithms used in this work we analyze their usefulness based on two objectives. First, we are interested in highly accurate and general hypotheses and second, the interpretation ability is highly important to increase our structural knowledge for the ADME data sets and the activity data set investigated in this work.