What is BAM?

BAM is a new statistical technique for detecting differentially expressing genes from microarray data. At the heart of the method is a special type of signal detection technique that squashes down false signal while leaving real signal alone. Standard methods look for genes by relying on elementary test statistics using data from one gene at a time (such as t-tests, or contrasts from ANOVA models). Gene lists are obtained by filtering statistics by attempting to control overall type-I error, or false discovery rate, or by filtering using a user specified cutoff level.

Patterns of Interest

BAM takes a fundamentally different approach and focuses instead on estimating the differential effect of a gene by synthesizing information across all genes simultaneously. Rather than filtering genes, estimated effects are mapped into pattern types. BAM's special signal detection mechanism shrinks to zero estimated effects for genes unlikely to be differentially expressing, and consequently patterns of interest become highly interpretable using graphical visualization tools.

Patterns of interest might be "hit-and-run" genes which affect a biologic system for only a certain amount of time, or genes involved throughout the process. See Colon Cancer Illustration.

What types of problems can BAM be used for?

BAM applies to multigroup experimental designs, such as data collected over different stages of a disease. Other applications include gene expression time profiling for time-course data, outlier detection, invariant set normalization, gene atlas transcriptome mappings, as well an many other problems.

BAMarray*

BAM is implemented by the user friendly Java software, BAMarray 3.0, which can be downloaded for free.

*Supported by NSF grants DMS-0405675 and DMS-0405072 and Case Western Reserve University/Cleveland Clinic CTSA grant: UL1 RR024989 from NCRR.

Key Methodological Features

• The underlying methodology for BAM has been rigorously studied theoretically (Ishwaran and Rao 2005a, 2005b).
• BAM has been shown to be robust to non-normality of gene expression measurements and to correlations between expression measurements on a given chip (Ishwaran and Rao 2003).
• BAM discovers more genes by optimally balancing the number of falsely detected and falsely non-detected genes. This is a property guaranteeing lower total gene misclassification (Ishwaran and Rao 2005a). This differs from methods which control type-I error rates or false detection rates. Controlling such rates tends to identify obviously changing genes but misses many subtle changes.
• BAM adaptively reduces correlations between estimated differential effects for a gene. This allows true gene expression patterns to be readily identified and reduces the number of implausible patterns (Ishwaran and Rao 2005a).