Optimizing the inference of linkages from mRNA co-expression.

<p>(A) Examples of a functionally informative DNA microarray data set and a non-informative one. Each set is illustrated as a scatter plot showing the log likelihood of functional association for each successive bin of 1,000 gene pairs (circles) ranked by decreasing Pearson correlation coefficient between expression vectors derived from that array set. The set of microarray data measuring oxidative stress responses following Menadione treatment <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0000988#pone.0000988-Shapira1" target="_blank">[75]</a> (filled circles) does not show a significant relationship between co-expression and the likelihood of functional association. In contrast, the set of cell cycle time course experiments <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0000988#pone.0000988-Eisen1" target="_blank">[76]</a> (open circles) shows a strong relationship. The effect of filtering genes using the parameters M and R is illustrated in (B). A data set of genes changing expression during the diauxic shift <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0000988#pone.0000988-Brauer1" target="_blank">[77]</a> (open circles) shows a noisy relationship between co-expression and the likelihood of functional association, especially for gene pairs with the highest Pearson correlation coefficients. However, by introducing the two threshold parameters, the relationship improves (filled circles), in particular decreasing variance considerably and improving the corresponding regression model. (C) The divide-test-integrate strategy <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0000988#pone.0000988-Lee1" target="_blank">[1]</a> for inferring linkages, shown here calculated across all 500 microarray experiments (empty triangles) considerably outperforms analysis of the expression vectors constructed by concatenating the 500 experiments (filled circles). Precision is measured using reference linkages derived from MIPS functional annotation, masking the term “protein synthesis”, and recall is calculated for either reference linkages or total yeast genes (inset).