Analytical analyses of highly complex mixtures, such as biofluids or liquid food products, often give rise to signals for unknown compounds, particularly for compounds at low concentration. Here we compare two conventional chemometric approaches for NMR spectral analysis ("spectral binning" and "high-resolution analysis") with a novel library-based method ("targeted profiling of unknowns", TPU). The three methods were applied to a proton NMR spectral data set of ultrafiltered mouse serum typical of those examined in metabolomics/metabonomics studies. The advantages of high-resolution analysis of typical NMR peaks have been well described previously, and as a result we examined low intensity unknowns peaks (LIUPs). A total of 25 LIUPs were assessed based on their significance to multivariate statistical analysis of the data set using the TPU method. The linearity of NMR signals at low incremental concentration changes (< 10 microM) was determined by titration of endogenously occurring metabolites into filtered mouse serum. Carbon-13 decoupling of the NMR spectra was used to ensure isotope-satellite peaks were eliminated. Four peaks were noted as significant to separation between arthritic and diseased animals. The conventional spectral methods were hampered by baseline noise or overlap with high concentration metabolites and were not able to identify these LIUPs reliably. In general, conventional methods, particularly high-resolution analysis, are recommended for peaks with moderate signal-to-noise. The TPU method is recommended for peaks with low signal-to-noise or when compression of spectral data with high fidelity is desirable, such as integration of NMR data into cross-platform studies.