In our hands, use of the biological knowledge of important immune cell subsets and our definitions of their expression profiles increased the accuracy of deconvolution predictions. Another limitation of microarray deconvolution is the discrete nature of the component cell types in the basis set. The model assumes that cells do not exist in significant quantities in states intermediate between those that are purified and profiled for inclusion in the basis set. This issue has been addressed in one study: continuous variation in the expression levels of genes have been mapped to continuous cellular states in cell cycle experiments in yeast. Immune cells also exist at different points in a continuum of states of differentiation and activation. However, we did not consider these states here because of the complexity of modeling the large number of intermediates stemming from a group of eighteen basis cell types. We did observe that the residuals from fitting of immune cell basis matrices to the SLE blood in this study were small, indicating that the discrete model��s assumptions are valid. We assume that the residuals are due to a combination of technical noise and Ginsenoside-Ro incomplete sampling of expression profiles of leukocyte populations. The relative contributions of these two factors are not known, but we predict that fit could be further improved if the different states of immune cells were more fully represented. For example, it might be advantageous to capture varying degrees of activation by contrasting signatures of cells that are resting and activated in vitro. Also, the canonical activation of cells used in this study are widely believed to simulate in vivo activation reasonably well, but there are other forms of activation that could be performed and profiled on microarrays to better capture the spectrum of leukocyte populations in blood. Expression deconvolution may actually even have the Orbifloxacin potential to help identify these cells or states. In this study we selected genes that discriminated between the cell types we had chosen to assay; this step improves the performance of the method but likely implicitly excludes the best markers for cell types that were not chosen. If the approach was modified to include more genes, then the subset of those extra genes that fit poorly might be good markers for cell types or states that should have been profiled but were not.