Classifi_gofile – The GO file contains the results from the automated probe annotation with all Gene Ontology ID’s listed in the fourth column. These include the primary GO annotations along with the GO parentage associated with these primary annotations. This file is useful to identify which probes/genes are responsible for gene cluster classifications.

Classifi_outputfile – The output file contains all of the enumerated variables that were used in the hypergeometric distribution calculation and the probability results from the calculation. The data is divided by gene cluster and the GO ID’s are ranked from lowest p value to highest p value for each gene cluster. The column descriptions are as follows:

GO id – is the identification number given for the specific GO annotation.
g – total number of probes (for Affy these are actually probe sets) in the entire data set.
f – the number of occurrences of this particular ontology in the entire data set.
c – the number of probes in the gene cluster.
n – the number of occurrences of the particular ontology in the gene cluster.
expt – the expected number of occurrences of the particular ontology based on “g”, “f” and “c” given a Poisson distribution.
prob – the probability that the co-clustering of this ontology in this gene cluster would have occurred by chance (this is the important number).
GO type – the gene ontology type, either Molecular Function, Biological Process or Cellular Component (cellular location).
GO name – the ontology description linked with the GO ID in the first column.
clusterid – gene cluster ID assigned by user.

Classifi_topfile – The Top file is simply a list of the GO ID’s that gives the lowest p value in each of the gene clusters.

Recommended data set size - CLASSIFI does not work very well for small data sets. We have obtained good results using data sets that contain at least 1000 probes, with gene clusters containing an average of ~100 probes. But the larger the data set, the longer it will take to do the calculations. Data sets of 1000 – 3000 probes containing 10 – 30 gene clusters are recommended.

Establishing p value cutoff - Low probabilities are highly informative. For this kind of data, selecting a p value cutoff of 0.05 is not appropriate, since this will include a large number of false positive results. To determine an appropriate p value cutoff, one needs to either do a permutation analysis in which a series of randomized data sets are generated with a structure similar to the original data set and then analyzed using the same algorithm (time consuming), or apply a Bonnferoni correction to determine the appropriate cutoff (easy and fast). While the first approach is more accurate, using the Bonnferoni correction gives a reasonable estimate for most datasets and is simple to calculate. To estimate the significance cutoff, select an a level (e.g. 0.05 or 0.01) and divide it by the number of data points used in the calculations (equivalent to the number of rows in the data set).

This number provides an estimate for the p value cutoff that should be used for this data set.

Limitations - High probabilities do not necessarily rule out a significant co-clustering. Since the probabilities are determined by all four parameters, high probabilities can occur for a variety of reasons. For example, high p values are always obtained if the number of occurrences of a particular ontology in the entire data set (f) is small.

The bottom line is low p values are significant; high p values may or may not be significant.

Related GO ID’s in the Output file - Several GO id’s come from the same gene, and often are listed near each other in the table. For example, you may find that the first 5 GO ID’s listed for a given gene cluster are functionally related and have the exact same n values. This may be because all of these GO annotations are attached to the same set of genes. Remember that since both the primary GO annotations and the entire GO parentage is included, a group of genes will often have multiple GO ID’s in common.