Preparation of Training Sets

It seems obvious that an ideal training set must include all tested active and inactive compounds. However, in practice it is necessary to be very careful during the design of training set because a data set consisting of database chemical drawings and HTS assay measurements may be very misleading . 

In a well-designed training set the structural diversity must be as uniform as possible. It is very difficult to control such uniformity however, the presence of closely similar compounds series in the set could (and have to) be checked, to avoid degeneracy. 

In general, any ligand-based virtual screening method is based on direct or generalized similarity between the screened compound and compounds from the training set. Therefore, if such similarity is absent at all, no reasonable prediction of screened compound s properties can be made by using this training set. 

---