Exposing the Limitations of Molecular Machine Learning with Activity Cliffs
The title of this paper pulled in my attention immediately. The authors performed a comprehensive study of the ability of ML and DL to predict the larger drops in potency that can occur with small structural changes. Activity Cliffs are poorly predicted (perhaps unsurprisingly), and the authors describe some of the potential reasons behind this. For those wishing to perform an ML/DL study you would do well to follow their methodology. For those that are unfamiliar with statistics, this paper may come across as rather heavy. Then again – if that is the case, why are you a computational chemist?