Many methods are available for the analysis of expression data from microarrays 123. Most software falls into one of two categories: programs which generate a list of significantly differentially regulated genes and programs which perform an analysis based on such a significant genes list (SGL). The first category includes statistical methods (e.g. SAM 4) and pattern-grouping methods (e.g. Cluster 5), the end result of either type of analysis being a list of genes which appear to be significantly differentially regulated. The second category of software includes a variety of approaches which search for relationships between the members of the SGL, such as methods which identify common regulatory motifs 6 or establish significance based on genetic position 7.

In browsing a SGL, lexical trends are frequently observed as genes involved in similar metabolic functions are often coordinately regulated. Identification of such lexical trends is critical for the generation of hypotheses for further experimentation, but the large size of a typical microarray dataset makes objective evaluation by eye difficult, if not impossible. Additionally, the genetic content of the organism, the style of the genome annotation, the features selected for representation on the microarray, and experimental design all contain biases which are ultimately reflected in the SGL, which further confounds objective analysis. A number of publications have addressed some of these issues by designing lexical analyses based on functional classifications 8, Medline MeSH terms 9, or Medline titles and abstracts 101112 that are associated with the genes in a SGL. Each of these approaches draw upon external information to overlay biological relationships onto the expression data, and the significance of these relationships can then be analyzed using standard statistical tools.

The described methods have all been used to identify previously undetected relationships of possible biological significance in published microarray datasets. However, each approach has drawbacks that limit their use as a generalized analysis tool for the average microarray user. Methods based on functional classifications require that the genes be categorized prior to the analysis. This is often not performed by the sequencing group and too laborious to be practical for most microarray end users. Medline-based methods cannot include genes with putative functions in their analyses, as there is no associated literature for such genes. As over 50% of the genes in most genomes are still of putative function, abstract-based approaches exclude a large proportion of the expression data from the analysis. All of the described approaches are also confounded by ambiguity and variability in the annotations used by different researchers (e.g. genes with multiple names). Most importantly, these methods require a significant degree of computer savvy to set up and execute, which makes use impractical for many microarray users. We developed a program, LACK, which performs a simple but flexible lexical analysis of microarray data that circumvents many of the described problems.

We have demonstrated one example of how we have employed LACK for statistical analysis of lexical trends. In analyzing our microarray data, we realized that few simple tools are available for assessing significance of observed trends. While there are a number of other lexical analysis methods available, they employ different goals (i.e. de novo identification of lexical trends) than LACK (significance analysis of specific lexical trends). The primary advantages of LACK over other lexical analysis methods for the typical microarray user are the simplicity of use and the flexibility in searchable concepts. The inputs are text files that are generated as part of a typical microarray analysis, and the output is provided as tab-delimited text; no external information is required apart from the microarray annotation. The inherent flexibility in user-specified search terms makes LACK potentially useful for many microarray users who are interested in determining whether or not specific trends are significant in their data, rather than performing a global search of all possible trends. In addition, this flexibility makes lexical analysis possible for datasets that are difficult to analyze by other lexical methods due to inconsistencies in annotation nomenclature, as the user can specify terms to encompass multiple naming conventions.

It is important to note that with this flexibility comes an additional point at which bias can be introduced: careless selection of search terms can lead to false determination of significance. For example, search terms should not be chosen only from the SGL itself, but rather from the full dataset. To demonstrate such an inappropriate example, we selected a search list based on metabolites found in the top 20 genes of the SGL described above. While there is no clear relationship between the search list terms (glucosamine, methionine sulfoxide, citrate, 2-aminoethylphosphonate, ethanolamine, UDP-glucose, cysteine), LACK indicates a high level of significance in these terms being over-represented in the SGL (p < 0.0005), illustrating the danger of choosing terms directly from the SGL.

One possible approach that avoids this problem is to select a wordlist prior to generation of the SGL. This has the obvious limitation that it cannot be implemented for exploratory microarray experiments, only for specific hypothesis testing. Another approach is to generate the word list in a blinded fashion; for example, a researcher who did not generate the SGL could select search terms from the full dataset. We have also made a tool available, ALACK (automated LACK), which selects single search terms which are over-represented in the SGL in an automated, unbiased fashion (although multiple search terms are not supported in ALACK). While the benefits of LACK far outweigh the potential abuses, it is clear that care must be taken to avoid over-interpretation of significance.

Analysis of the significance of trends in expression data is critical for the generation of further hypotheses, as variation in the arrays and in annotations can confound visual assessment of significance. We have described a method for assessing the statistical significance of lexical trends in microarray data. Our approach is similar to, but distinct from, other lexical analysis methods 3891011, in that our method does not incorporate external information. Rather, LACK is designed to analyze the SGL generated from another analysis within the context of the microarray from which the data was generated. We expect that LACK can serve as a useful adjunct to other microarray analysis methods, including other global lexical trend identification methods.

CK performed programming of LACK, microarray experiments, and preparation of the manuscript. SF provided funding and supervision for the experiments described herein. Both authors have read and approved the final manuscript.