Localized Deconvolution

The localized deconvolution is a robust quantification method to process highly congested spectra. It is paralleizable and a Web service interface is provided to a map-reduce implementation, which makes it scalable and accessible to a wide variety of tools.

Our experiments have shown that the removal of adjacent, convoluting, and irrelevant signals results in significantly improved absolute and relative quantification, as demonstrated on realistic synthetic data. The performance metrics also demonstrate that including a buffer region does not improve overall accuracy, and allowing the baseline to be positive or negative results in the best accuracy. However, it was observed that specific spectral configurations did benefit from including a buffer region. Developing an algorithm to take advantage of the strengths of both methods is currently in process.

The advantages of our method were also observed on an experimental metabolomics data set of organ toxicity. Specifically, the within group scatter was reduced by localized deconvolution, resulting in an improved cross-validation score (Q2); however, this increase in accuracy leads to additional computing costs. Such issues can easily be overcome by parallelizing the process with map-reduce and making use of cheaply available cloud resources. Alternative methods to binning, such as targeted and direct quantification methods, could be improved by by using localized deconvolution to filter and remove obfuscating signals.

This software is available as part of our Metabolomics Analysis Toolkit, which can be downloaded at GitHub.