Lattice based Clustering of Temporal Gene-Expression Matrices

Individuals show different cell classes when they are in the different stages of a disease, have different disease subtypes, or have different response to a treatment or environmental stress. It is important to identify the individuals ’ cell classes, for example, to decide which disease subtype they have or how they will respond to a certain drug. In a temporal gene-expression matrix (TGEM) each row represents a time series of expression values of a gene. TGEMs of the same cell class should show similar gene-expression patterns. However, given a set of TGEMs, it can be difficult to classify matrices by cell classes. In this paper, we develop a tool called LAB-STER (LAttice Based cluSTERing) to cluster gene-expression matrices by cell classes. Rather than treating each row or column as a vector, we create a Galois lattice for each matrix, which yields a natural distance function between gene expression matrices. Finally, we cluster based on these distances. A key advantage of our method is that it effectively handles missing values, which is a problem in gene expression data. We evaluated LABSTER on both simulation data and clinical data. The results show that LAB-STER has better clustering performance than several widely used vector-based clustering methods. A bootstrapping procedure is also proposed to further improve the performance of LABSTER. LAB-STER has the potential to be used on matrices containing data other than gene expression