Time-Frequency Feature Detection for Time-course Microarray Data

摘要

Gene clustering based on microarray data provides useful functional information to the working biologists. Many current gene-clustering algorithms rely on Euclidean-based distance metrics and fail to capture the time-dependent features of the data, usually corrupted by high levels of experimental noise. Here we propose an algorithm capable of dealing with the noise through a time-frequency approach and related measure of correlation between time-course expressions of different genes (trajectories). The approach makes use of fast multi-resolution feature classification algorithms and allows for the desired functional characteristics (such as phase delay, activation/repression etc.) to be enhanced and detected. We have applied our algorithm to time-course microarray data of Drosophila melanogaster (Arbeitman et al., Science, Sep 27, 2002, page 2270-2275). We examined various relations among homeodomain genes (referred to as group H) and regulators of homeodomain genes (group RH) as follows: After normalization, the trajectories were projected on to CosBell wavelet basis. The four genes in group RH form two clusters: three of them stayed close to each other, and the last one, CG8651 (trithorax), was singled out. The group H genes, forming four clusters, showed functional features that are more similar to trithorax than the other three. We further analyzed ten homeodomain genes that have good correlations with trithorax in the wavelet basis. Literature search showed that there are five genes thought to be in the downstream pathway of trithorax. Although only two of these five genes were in the dataset available to the algorithm, it was able to identify both of these. Our study suggests that time-frequency analysis provides a powerful tool for discovering the underlying regulatory networks when applied to time-course microarray data.