Recently, genome-wide association studies (GWAS) on metabolic quantitative traits have proven valuable tools to uncover the genetically determined metabolic individuality in the general population 
. Interestingly, a great portion of the genetic loci that were found to significantly associate with levels of specific metabolites are within or in close proximity to metabolic enzymes or transporters with known disease or pharmaceutical relevance. Moreover, compared to GWAS with clinical endpoints the effect sizes of the genotypes are exceptionally high.
The number and type of the metabolic features that went into these GWAS was mainly defined by the metabolomics techniques used: Gieger et al. 
and Illig et al. 
used a targeted mass spectrometry (MS)-based approach giving access to the concentrations of 363 and 163 metabolites, respectively. Suhre et al. 
and Nicholson et al. 
applied untargeted nuclear magnetic resonance (NMR) based metabolomics techniques, yielding 59 metabolites that had been identified in the spectra prior to the GWAS and 579 manually selected peaks from the spectra, respectively. In Suhre et al. 
, 276 metabolites from an untargeted MS-based approach were analyzed.
While these previous GWAS focused on metabolic features with known identity, untargeted metabolomics approaches additionally provide quantifications of so-called “unknown metabolites”
. An unknown metabolite is a small molecule that can reproducibly be detected and quantified in a metabolomics experiment, but whose chemical identity has not been elucidated yet. In an experiment using liquid chromatography (LC) coupled to MS, such an unknown would be defined by a specific retention time, one or multiple masses (e.g. from adducts), and a characteristic fragmentation pattern of the primary ion(s). An unknown observed by NMR spectroscopy would correspond to a pattern in the chemical shifts. Unknowns may constitute previously undocumented small molecules, such as rare xenobiotics or secondary products of metabolism, or they may represent molecules from established pathways which could not be assigned using current libraries of MS fragmentation patterns 
or NMR reference spectra 
The impact of unknown metabolites for biomedical research has been shown in recent metabolomics-based discovery studies of novel biomarkers for diseases and various disease-causing conditions. This includes studies investigating altered metabolite levels in blood for insulin resistance 
, type 2 diabetes 
, and heart disorders 
. A considerable number of high-ranking hits reported in these biomarker studies represent unknown metabolites. As long as their chemical identities are not clarified the usability of unknown metabolites as functional biomarkers for further investigations and clinical applications is rather limited.
In mass-spectrometry-based metabolomics approaches, the assignment of chemical identity usually involves the interpretation and comparison of experiment-specific parameters, such as accurate masses, isotope distributions, fragmentation patterns, and chromatography retention times 
. Various computer-based methods have been developed to automate this process. For example, Rasche and colleagues 
elucidated structural information of unknown metabolites in a mass-spectrometry setup using a graph-theoretical approach. Their approach attempts to reconstruct the underlying fragmentation tree based on mass-spectra at varying collision energies. Other authors excluded false candidates for a given unknown by comparing observed and predicted chromatography retention times 
, or by the automatic determination of sum formulas from isotope distributions 
. Furthermore, Gipson et al. 
and Weber et al. 
integrated public metabolic pathway information with correlating peak pairs in order to facilitate metabolite identification. However, these methods might not be applicable for high-throughput metabolomics datasets that have been produced in a fee-for-service manner, since the mass spectra as such might not be readily available.
Approaching the problem from a conceptually different perspective, we here present a novel functional metabolomics method to predict the identities of unknown metabolites using a systems biological framework. By combining high-throughput genotyping data, metabolomics data, and literature-derived metabolic pathway information, we generate testable hypotheses on the metabolite identities based solely on the obtained metabolite quantifications (). No further experiment-specific data such as retention times, isotope patterns and fragmentation patterns are required for this analysis.
Data integration workflow for the systematic classification of unknown metabolites.
The concept of our approach is based on the following observations from our previous work on genome-wide association studies and Gaussian graphical modeling (GGM) with metabolomics: We showed that GWAS with metabolic traits can reveal functional relationships between genetic loci encoding metabolic enzymes and metabolite concentration levels in the blood 
. A genetic variant can alter, for instance, the expression levels of mRNAs or affect the properties of the respective enzymes through changes of the protein sequence (e.g. enzyme activity, substrate specificity). Moreover, we found that GGMs, which are based on partial correlation coefficients, can identify biochemically related metabolites from high-throughput metabolomics data alone 
. These observations suggest that if an unknown compound displays a similar statistical association with a genetic locus in a GWAS or a known metabolite in a GGM, then this may provide specific information of where it is located in the metabolic network. Based on this information we can then derive testable hypotheses on the biochemical identity of the unknown metabolite. This annotation idea parallels classical concepts from functional genomics, where, for instance, co-expression between RNA transcripts is used to predict the function of poorly characterized genes 
The manuscript is organized as follows: We first conduct a full genome-wide association study on 655,658 genotyped SNPs with concentrations of 225 unknown metabolites using fasting blood serum samples from a large German population cohort (n
. We thereby extend our previous work on known metabolites 
to a GWAS with hitherto unpublished unknown metabolic traits. We then compute a Gaussian graphical model including both known and unknown metabolites. In a third step, we integrate the results of the GWAS and GGM computations and combine them with metabolic pathway information from public databases to derive predictions for a total of 106 unknown metabolites. In order to validate the approach, we investigate six distinct cases, in which we derive specific identity predictions for a total of nine unknown metabolites, which we then confirm experimentally. Finally, we discuss the relevance of newly discovered genetic loci and unknown identity predictions in the context of existing disease biomarker discovery and pharmacogenomics studies.