SAMFA: Simplifying Molecular Description for 3D-QSAR

In this paper we consider the following question: How much can we simplify molecular description without sacrificing too much quality of 3D-QSAR models. We compare the performance of the newly developed Simple Atom Mapping Following Alignment (SAMFA) descriptors with CoMFA using nine different data sets from the literature, by using three regression approaches (PLS, SVM, RandomForest), as implemented in R, and Monte Carlo cross-validation (MCCV) numerical experiments. The results indicate that SAMFA descriptors, despite their simplicity, perform surprisingly well when compared to the much more refined CoMFA descriptors. Moreover, their simplicity makes them readily interpretable and applicable to the difficult problem of inverse QSAR.