Informative or noninformative calls for gene expression: a latent variable approach.

Area of Science:

• Genomics and Bioinformatics
• Statistical Modeling in Biology

Background:

• Microarray technology generates vast amounts of data, necessitating methods to identify informative genes.
• Existing gene filtering techniques, such as FLUSH and I/NI calls, have limitations in handling complex data variability.
• Minimizing irrelevant genes is crucial for accurate differential gene expression analysis and sample classification.

Purpose of the Study:

• To propose and evaluate a flexible linear mixed model for identifying informative genes in microarray data.
• To compare the performance of the linear mixed model against existing methods, including FLUSH and I/NI calls.
• To introduce and assess additional gene filtering criteria, such as R2, intra-cluster correlation, and information criteria (AIC, BIC).

Main Methods:

• Development and application of a linear mixed model for probe-level gene filtering.
• Implementation of criteria including R2, intra-cluster correlation, likelihood ratio testing, Akaike Information Criteria (AIC), and Bayesian Information Criteria (BIC).
• Validation using the HGU-133A Spiked-in data set.

Main Results:

• The proposed linear mixed model demonstrates superior performance compared to the FLUSH method.
• The linear mixed model achieves performance comparable to the I/NI calls method.
• Different filtering criteria (conditional variance, R2, intra-cluster correlation, AIC, BIC) exhibit varying stringency levels.

Conclusions:

• The linear mixed model offers a robust and flexible approach for filtering noninformative genes in microarray datasets.
• This method enhances the reliability of gene expression analysis and classification derived from microarray data.
• The study provides a comprehensive evaluation of multiple gene filtering strategies, aiding in the selection of optimal methods.