Development of a method to assess the informational content of structure-activity data bases

The ability to predict the biological or toxicological properties of yet untested chemicals based upon structure-activity relationships (SAR) is very dependent upon the size and chemical diversity of the "learning set" used to develop the SAR model. In the present study it is shown that for noncongeneric chemicals, systematically increasing the informational contents of "learning sets" by iteratively selecting chemicals based upon structural diversity increases the predictivity of the SAR model. This approach can now be used to generate learning sets with maximal informational content while keeping the number of chemicals that require testing at a minimum.