These data include expression of genes in different activation and differentiation states that represent a spectrum of cell species present in blood, providing a basis set for microarray deconvolution of blood samples. Here we test fifteen cell subsets including several resting and activated dyads. Some are not Echinatin readily distinguishable based on surface markers alone. Moreover, it should be possible to distinguish even greater numbers of cell types by deconvolution. The expression signatures in blood samples from SLE patients show significant, specific differences from those of healthy controls. Some of these differences are changes in the abundance of specific leukocyte populations, suggesting that systematic large-scale characterization of the cellular composition of SLE patient blood would measure quantitative differences relevant to the disease pathophysiology. Here we use microarray deconvolution to explore immune cell subsets and activation states in SLE patient blood. First, we measure the accuracy of the method with a “truth” experiment where known proportions of immune cells are mixed, assayed on expression microarrays, and computationally separated. Next, we performed a proof of concept experiment by deconvolving white blood cell profiles into a modest number of immune cell subsets. We then use this validated method to derive immune cell signatures for a panel of eighteen major populations and states of white blood cells. Finally, we deconvolve expression profiles of blood samples from healthy donors and SLE patients into the proportions of these different white blood cell subsets and identify patterns in their dynamics related to disease and treatment. The process of deconvolving mixtures of cells was developed using a system of four transformed cell lines of immune origin: Raji, IM-9, Jurkat, and THP-1 cells. These cell lines provided the abundant sources of pure cells necessary to support experimental mixing of different types of cells in several different ratios. These cell lines are useful because they show similar but distinguishable expression profiles; their immune derivation is not important to the purpose of the experiment. We chose two B cell lines to gauge the ability of the assay to discriminate between cells that are very similar to each other. The algorithm was Tubeimoside-I trained and the performance limits of deconvolution were measured by creating various mixtures of cells, assaying the pure cells and the cell mixtures on expression microarrays, and using the expression data from the pure cells to deconvolve the expression data from the cell mixtures. Data for many probesets in a given expression microarray dataset are comprised of noise but little or no biological signal. Here we show that reducing the contribution of these noisedominant probesets to deconvolution improves performance, and we establish an approach for weighting probesets to define a highperformance basis matrix for performing deconvolution. Probesets were ranked by their degree of differential expression as described in the Methods section, and a thorough set of matrices comprised of different quantities of the most differentially-expressed probesets was tested in deconvolution by comparing the results of each matrix to the known mixture ratios. Both small and very large matrices performed poorly. The distribution of matrix size to the least squares fit to the data was continuous and exhibited a gently rounded optimum at 275 probesets. We observed that goodness of fit correlated very closely with how well conditioned each matrix was.