Classification approaches have been developed, adopted, and applied to distinguish disease classes at the molecular level using microarray data. Recently, a novel class of hierarchical probabilistic models based on a kernel-imbedding technique has become one of the best classification tools for microarray data analysis. These models were first developed as kernel-imbedded Gaussian processes (KIGPs) for binary class classification problems using microarray gene expression data, then they were further improved for multiclass classification problems under a unifying Bayesian framework. Specifically, an adaptive algorithm with a cascading structure was designed to find appropriate featuring kernels, to discover potentially significant genes, and to make optimal disease (e.g., tumor/cancer) class predictions with associated Bayesian posterior probabilities. Simulation studies and applications to publish real data showed that KIGPs performed very close to the Bayesian bound and consistently outperformed or performed among the best of a lot of state-of-the-art methods. The most unique advantage of the KIGP approach is its ability to explore both the linear and the nonlinear underlying relationships between the target features of a given disease classification problem and the involved explanatory gene expression data. This line of research has shed light on the broader usability of the KIGP approach for the analysis of other high-throughput omics data and omics data collected in time series fashion, especially when linear model based methods fail to work.