On the intrinsic dimensionality of chemical structure space

Abstract

An important expectation in chemistry and pharmacology is that similar chemical structures have similar properties and behavior. New industrial chemicals, pesticides, and therapeutics are often sublte modifications of "lead" structures with known chemical behavior. Chemical properties and recreation rates can be can be estimated from suitable homologs 1,2. Moreover, the safety of untested chemicals is often evaluated by comparing the chemical toanalogous structures for which toxicological data are available. Despite the widespread use of terms such as "homolog" and "analogs" in research, chemical similarity has evaded quantitative interpretation from a perspective where all chemicals are considered simultaneously. One reason is that chemical similarity is inherently a multivariate problem or, in other words, chemicals are simultaneously similar and different from many perspectives. We have approached chemical similarity by attempting to define a structure space in which all chemicals can be identified. Because there are so many potentially important variables, multivariate tools are necessary to reduce the dimensionality of this problem. When this is accomplished, we need to comprehend what this space means and what can be predicted from it . This paper is one of the first attempts to define chemical structure space for a large universe of chemicals.