While the entirety of 'Chemical Space' is huge (and assumed to contain between 10(63) and 10(200) 'small molecules'), distinct subsets of this space can nonetheless be defined according to certain... Show moreWhile the entirety of 'Chemical Space' is huge (and assumed to contain between 10(63) and 10(200) 'small molecules'), distinct subsets of this space can nonetheless be defined according to certain structural parameters. An example of such a subspace is the chemical space spanned by endogenous metabolites, defined as 'naturally occurring' products of an organisms' metabolism. In order to understand this part of chemical space in more detail, we analyzed the chemical space populated by human metabolites in two ways. Firstly, in order to understand metabolite space better, we performed Principal Component Analysis (PCA), hierarchical clustering and scaffold analysis of metabolites and non-metabolites in order to analyze which chemical features are characteristic for both classes of compounds. Here we found that heteroatom (both oxygen and nitrogen) content, as well as the presence of particular ring systems was able to distinguish both groups of compounds. Secondly, we established which molecular descriptors and classifiers are capable of distinguishing metabolites from non-metabolites, by assigning a 'metabolite-likeness' score. It was found that the combination of MDL Public Keys and Random Forest exhibited best overall classification performance with an AUC value of 99.13%, a specificity of 99.84% and a selectivity of 88.79%. This performance is slightly better than previous classifiers; and interestingly we found that drugs occupy two distinct areas of metabolite-likeness, the one being more 'synthetic' and the other being more 'metabolite-like'. Also, on a truly prospective dataset of 457 compounds, 95.84% correct classification was achieved. Overall, we are confident that we contributed to the tasks of classifying metabolites, as well as to understanding metabolite chemical space better. This knowledge can now be used in the development of new drugs that need to resemble metabolites, and in our work particularly for assessing the metabolite-likeness of candidate molecules during metabolite identification in the metabolomics field. Show less
