An approach to automated frequency-domain feature extraction in nuclear magnetic resonance spectroscopy.

For the analysis of metabolite systems, nuclear magnetic resonance (NMR) spectroscopy has become an important quantitative monitoring technology. Automated quantitation methods are highly desired and mainly characterized by the tasks of model selection and parameter approximation. This paper proposes a promising automated two stage approach in the frequency-domain, in which signaling peaks are first identified and filtered from noise based on curvature properties of the spectrum, and then proportionally approximated based on the analytical solution of a Lorentz-function. Remarkably, in opposition to common least-squares approaches, the proposed approximation scheme does not rely on partial derivatives, and furthermore, the runtime is independent to the number of spectral datapoints. Simulations provide promising empirical evidence for successful peak selection and parameter approximation, with the results for the latter highly outperforming the Levenberg-Marquardt algorithm in terms of error minimization and robustness.

A bottom-up approach to the biclustering-problem.

The biclustering problem adresses the discovery of locally significant correlation within a data matrix and has recently become quite popular in the field of microarray data analysis. The preservation of a particularly defined degree of homogeneity between elements within a bicluster plays a key role in the search procedure. This work proposes a pairwise distance function related to the mean squared residue to introduce multiple enrichment algorithms. Based on a theoretical framework, the impact is demonstrated empirically by the enrichment of commonly available and also on artificially generated bicluster sets.