Fundamental analytical chemistry and signal processing

We design analytical platforms for high-throughput analysis and provide information on (in theory) all small organic compounds in samples.

Sample preparation should therefore be non-selective or with complementary selectivity such that the entire range of compound properties can be covered by few analytical platforms. This is a very challenging task as the exact composition of any field sample is unknown and complex.

We focus on the multidimensional chromatographic platforms GC×GC-TOFMS and LC×LC-high resolution-TOFMS that provide 3D data (e.g., rt1 × rt2 × m/z) for each sample. We have custom-built these platforms that represent the ideal combination for chemical fingerprinting due to their complementarity and large individual compound coverage. We aim to find column combinations that provide the most orthogonal separations (i.e., providing different relative retention times for the same compounds). The testing of novel phase chemistries such as the polar ionic liquid columns for GC is of special interest to develop GC×GC separations with high peak capacity. For LC and LC×LC separations, we combine the hydrophobic subtraction model and the linear solvent strength model with laboratory trials to predict and test combinations and gradients that provide the most orthogonal separations.

For each application, we construct custom-made signal processing procedures. We develop generic functions for import and cropping of different types of analytical data and develop new strategies for feature detection and pixel-based analysis of multidimensional data. We work on solving the problems that arise for multidimensional separation systems such as retention time shifts in multiple dimensions, complex baselines and the very large number of variables compared to samples.

We develop new scaling procedures using the analytical variation of each signal in replicate analyses to reduce the importance of non-chemical information and allow robust modelling of data with many more variables than samples. To do this, we also exploit the structure of data (sample × rt1 × rt2 × m/z) by employing multi-way models such as PARAFAC that can resolve and quantify even grossly overlapping peaks.