Melikerion

The integration of numerous sources of biochemical and clinical information sets a challenge to the medical scientists who are striving for a deeper understanding of human disease mechanisms. Classical reductionist approaches often overlook non-linearity and rely on univariate diagnostic criteria as targets of inference, which may prevent the detection of complex multi-dimensional interactions. Unsupervised analysis of the regularities within a dataset is the first step in transforming the measurements into usable knowledge, but estimating the statistical significance of the observed patterns is often difficult.

Melikerion is an implementation of the Kohonen self-organizing neural network algorithm, and was designed for unsupervised analysis of clinical materials in particular. A typical process flow – with emphasis on the biochemical profiles – begins with the preprocessing of the measurement data so that the variables become comparable in scale and mean value. Next, linear decomposition is applied to create an initial map layout of patients. The final layout is achieved after several iterations of a batch version of the Kohonen algorithm. Once the map is complete, the full dataset is visualized based on the biochemical profiles, with statistical significance estimates for the clinical variables.

One of the design goals was to make the method accessible to the scientific community, with minimal software production costs. We therefore made the first version on the Matlab programming environment, and utilized the SOM Toolbox as the computational core, but have now expanded compatibility to the open source equivalent Octave with a revised SOM core.

The most convenient way to try MeliKerion is via the online system, which will also produce a full complement of visualizations automatically. We hope that this way also those researchers who are not familiar with the Matlab/Octave environment will be able to benefit from our work.