Machine learning determined models of inhibitory activities for fluorinated Cinchona alkaloids

A series of 25 fluorinated Cinchona alkaloids derivatives was theoretically investigated by calculation of their potential energy surfaces (PES). PES for all compounds were sampled by performing molecular dynamics simulations [1] and then decomposed by principal component analysis. Each PES was represented by three points in the newly determined reduced space. These points were used as independent variables for establishing activity/PES regression models whereas previously measured inhibitory activities towards human acetyl- and butyrylcholinesterase were used as dependent variables. Multivariate linear regression models were built by applying an extensive machine learning protocol where linear combinations of original variables as well as their higherorder polynomial terms were used. Leave-one-out cross- validation (LOO-CV) was used to validate obtained models [2, 3]. Optimal activity/PES models were selected based on the adjusted R 2, predicted R 2 and the LOO-CV mean squared error (Figure 1).