To build a SAR/QSAR model applicable to a wide variety of chemotypes the training set should be diverse, but exactly how diverse and what is the appropriate measure of diversity is still an unresolved problem. Some of the frequently used ways to measure diversity include: a) Bio-activity profile or distribution of physico-chemical properties b) Distribution or central moments of pairwise similarity c) Clustering-based approaches and dimension reduction.Bio-activity data and physico-chemical properties are not always available for all chemical structures of interest.  Methods relying on mean pairwise similarity make an a priori assumption that the underlying distribution is of a specific shape (i.e. Gaussian), while methods using either clustering or a distribution of properties make it difficult to quantify to compare two very different sets of molecules to claim that one is more diverse than the other. Therefore we explored a possibility of creating a diversity measure which satisfies the following criteria: d) it is trivially computed from the structure itself for any structure, e) it is applicable to any set of molecules without any a priory assumptions about its composition f) it is easy to use to compare two different sets.With the increasing availability of large datasets of molecules numbering in tens of millions of structures there is enough publicly accessible data to conduct an investigation into the diversity distribution over a set of sets of randomly selected compounds which has not been possible before.We present a new application to estimate the diversity of a chemical set of molecules.

New application to estimate the diversity of molecular databases

247th National Meeting (2014)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.