Abstract
Tiling arrays of high-density oligonucleotide probes spanning the entire genome are powerful tools for the discovery of new genes. However, it is very difficult to determine the expression value of structurally unknown gene from noisy array signals. Here we introduce a statistical method that estimates the exon/intron structure of a structurally unknown gene by maximizing the likelihood of observed intensities and sequences based on a bi-directional hidden Markov model. The method was implemented as a program ARTADE, and was applied to the Arabidopsis thaliana whole-genome array data, and predicted the structures and expression values of 20,168 genes including novel 6,513 genes. Gene expression values predicted without reference to structural information highly correlated with those calculated in the conventional fashion and based on gene structures. This approach opens up new possibilities for combined analysis of structurally known and unknown genes and the discovery of new coding and non-coding transcripts.