Recent technical and analytical progress in the field of metabolomics has lead to the identification of vast amounts of new compounds in living organisms. However, the integration of these compounds into the context of known metabolism remains difficult. To address this challenge of incomplete knowledge, we propose a computational approach that identifies novel hypothetical reactions between known metabolites, integrates experimentally measured molecular structures into existing metabolic networks, and finally predicts chemical compounds that are probable to exist in metabolism. The computational framework BNICE.ch is used to exploit the known biochemistry contained in the Kyoto Encyclopedia of Genes and Genomes (KEGG). We summarize the vastly diverse functionalities of enzymatic reactions in a few hundred expert-curated reaction rules, each generalizing multiple biochemical reactions. We then apply these rules to all metabolites known to KEGG in order to create a database of all the biochemically plausible reactions between compounds reported to occur in living organism. This extrapolation of the known metabolism results in a network of more than 130’000 known and novel reactions, each connecting two or more KEGG compounds. The generated information has been organized into an online database, the “Atlas of Biochemistry”, and is available under http://lcsb-databases.epfl.ch/atlas/.