LEMMA - Laplace approximated EM Microarray Analysis

Background

A common objective in microarray analysis is to identify genes that are differentially expressed between two treatment groups. Typically, the number of genes simultaneously assayed is large, while the number of samples for each gene is relatively small. In such typical cases, gene-wise tests are under-powered, resulting in few, or no discoveries. Bar, Booth, Schifano, and Wells, (2009) developed a statistical model that achieved highest power is simulations, while maintaining low false discovery rate. For more details about their random effect model, see: lemma.pdf. For a short summary, click here.

The Program

LEMMA is an R program that implements the RR model to analyze normalized microarray data. This version (1.2-1, 2009-09-03) supports two treatments and
  • three-way classification: null genes, for which statistically there is no difference in expression between the two treatment groups; nonnull group #1 - genes that are significantly more expressed in group 1 than in group 2; and nonnull group #2 - genes that are significantly more expressed in group 2 than in group 1; or
  • two-way classification (null and nonnull genes, as in the LEMMA paper)

    • In the near future it will support any number of treatments.
    The program runs on both Windows and Linux. For execution instructions, read the LEMMA documentation page.

    To download the R code, go to the CRAN contributed packages site, or go directly to the lemma package page.

    For questions, support, or feedback, contact the authors, Haim Bar (hyb2 at cornell dot edu) and Elizabeth Schifano (eds27 at cornell dot edu), (Cornell University, Department of Statistical Science , and the Department of Biological Statistics and Computational Biology)

    Examples
    The web page "LEMMA - Examples" contains two examples, with data, invocation instructions, and corresponding output. The first example is the ApoA1 dataset (Callow et al., 2000), and the second is a simulated dataset.

    A fully Bayesian alternative

    For a fully Bayesain alternative, see here