Using mutual information to discover temporal patterns in gene expression data

Finding relations among gene expressions involves the definition of the similarity between experimental data. A simplest similarity measure is the Correlation Coefficient. It is able to identify linear dependences only; moreover, is sensitive to experimental errors. An alternative measure, the Shannon Mutual Information (MI), is free from the above mentioned weaknesses. However, the calculation of MI for continuous variables from the finite number of experimental points, N, involves an ambiguity arising when one divides the range of values of the continuous variable into boxes. Then the distribution of experimental points among the boxes (and, therefore, MI) depends on the box size. An algorithm for the calculation of MI for continuous variables is proposed. We find the optimum box sizes for a given N from the condition of minimum entropy variation with respect to the change of the box sizes. We have applied this technique to the gene expression dataset from Stanford, containing microarray data at 18 time points from yeast Saccharomyces cerevisiae cultures (Spellman et al.,). We calculated MI for all of the pairs of time points. The MI analysis allowed us to identify time patterns related to different biological processes in the cell.