A Mean Pattern Model for Integrative Study - Integrative Self-Organizing Map

integrating multiple experiments to explore genetic factors contributing to the commonality and the diversity among species, omics or platforms has drawn an increasing attention recently. The study is in fact a pattern discovery process and the accuracy varies using different approaches. Most focused on multivariate structure of data and over-looked the nature of biological data, i.e. they are replicated samples. It is well known that a well-designed experiment can significantly reduce the variance among the measurements of replicated samples. This indicates that the measurements (count, expression or flux) of each molecule such as a gene, a metabolite, or a protein from replicated samples can be considered as random samples of a Gaussian density whose mean value is the truth. When we experiment many molecules together, it is common that most of them correlate. Therefore, it is obvious to believe that the measurements of all molecules are random samples of a mixture of Gaussian densities. These mean values of these Gaussian densities can be estimated using a statistical model, which we refer to as a mean pattern model. We generalize the self-organizing map to implement this mean pattern model and call it as an integrative self-organizing map (iSOM). We compared this new approach with existing algorithms using simulated and real data. The result shows that iSOM works well.