Time serial microarray or RNA-Seq experiments were frequently used in searching specific genes or biomarkers associated with carcinogenesis, and subsequently in identifying novel cancer subtypes and implementing molecular cancer diagnosis (Chung et al., 2002; Liau and Whang, 2009). Thus, developing methods to analyze time serial gene expression profiling remains ever important and is always a challenge.
Popular statistical methods, such as the Student's t-test, clustering and analysis of variance analyses, etc., are widely used in analyzing time serial gene expression data, and are successful for finding disease-related genes or biomarkers. However, there still are many general problems remaining to be solved. Some major problems in mining gene expression data include high false positives, low sensitivity and low overlap of genes found by different methods. In general, the reason for drawbacks can come from a variety of sources, but two main aspects may be behind the problems. First, almost all of the methods mentioned above used to ignore the continual changes of a gene during the whole stages of disease; secondly, they take no account of the continuous stabilities or conservative properties of time serial gene expression. In term of this, two approaches, multiclass ordinal analyses and multi-category logit model algorithm (Pyon and Li, 2009), were presented and used in finding disease-related gene signatures and biomarkers. Both methods are based on a simple hypothesis that consistently increasing/decreasing expressed genes will show concordance/discordance with cancer progressions.
However, due to the complexity and heterogeneousness of gene expression profiling, no methods can promise a perfect solution in analyzing this kind of data.