This web supplement contains additional data on the following topics:

A) Assignments of Uncharacterized Genes.

B) Immunohistochemistry of previously uncharacterized proteins whose mRNAs were highly enriched in the membrane-associated fraction confirms their membrane-association.

C) Theoretical and statistical evidence for confidence estimates.

D) Details of the subcellular fractionation and reproducibility of the method.

Where possible we have assigned localization information and estimated confidences to uncharacterized human and yeast genes whose mRNAs were highly enriched in the cytosolic or membrane-associated fractions. The confidence measurements were derived from the percentage of characterized mRNAs encoding membrane-associated proteins at a given Cy5/Cy3 ratio using a moving average algorithm (see Figure 4). The confidence assigned to each uncharacterized gene is the percentage of membrane-associated proteins encoded by the 151 (human) or 175 (yeast) characterized genes with Cy5/Cy3 ratios greater than or equal to that of the uncharacterized gene.

We have begun to examine the subcellular localization of some of the unknown gene products using antibodies raised against predicted peptides from these proteins. The results shown below are from five unknowns whose mRNAs were highly enriched in the membrane-associated fraction of Jurkat T cells and for which predicted coding sequences were available. The antibodies were produced through a collaboration with Research Genetics, Inc. As can be seen, uncharacterized mRNAs that were enriched in the membrane-associated fraction were enriched for those coding for membrane-associated proteins. Although this represents a small data set, the overall percentage of membrane-associated proteins (4 of 5 or 80%) agrees well with the 85% confidence interval from which these unknowns were chosen.

Protein 1: Protein showing Golgi-like staining.

Protein 2: Protein showing vesicular/reticular staining.

Protein 3: Protein showing nuclear membrane staining.

	Protein 4: Protein showing plasma membrane staining.
	This panel shows immunofluoresence data using the antibody against Protein #4. The bottom image was acquired on a confocal microscope.

Protein 5: Protein showing nuclear staining (excluded from nucleolus).

As discussed in the paper, the probability that an uncharacterized gene with a given Cy5/Cy3 ratio encodes a membrane or secreted protein can be estimated from the fraction of the characterized genes with similar fluorescence ratios that encode membrane-associated proteins. This assumes that the prior probability that any uncharacterized gene in the set encodes a membrane-associated or secreted protein is equivalent to the fraction of characterized proteins that have been assigned to this class, an assumption that is supported, at least for yeast, by the very similar frequency of computationally-predicted signal peptides and transmembrane domains in the characterized and uncharacterized gene products. Specifically, among known yeast genes, 9% contain signal peptides and 31% contain transmembrane domains; among unknown yeast genes, 12% contain signal peptides and 31% contain transmembrane domains. Thus, based on sequence, the unknown genes and known genes are not significantly different. If anything, current prediction algorithms suggest that there is a higher fraction of membrane-associated proteins in the unknown genes than in the known genes (12% vs. 9% with signal peptides).

Given this analysis, it is possible to test our method by using half of the known yeast genes ("known set") to construct the confidence calibration curve and then using the other half as a "test set" (i.e. as hypothetical unknowns.) To perform this analysis, yeast genes of characterized subcellular localization were randomly separated into two groups. A calibration curve was constructed using the "known set" as in Figure 4 of the paper (see below.) Using this curve, 83.3% of the genes in the "test set" that fall into the 85% confidence interval actually encode membrane-associated proteins. Together with the protein data presented above, we thus have multiple lines of evidence showing that unknowns which are highly enriched in the membrane fraction actually encode membrane-associated proteins.

Test Calibration Curve.

The following data present further details of the strength and accuracy of the fractionation of membrane-bound RNA from free RNA. As described in the text, tissue culture cells were hypotonically lysed, and membrane-bound RNA was separated from free RNA by equilibrium density centrifugation in a sucrose gradient. The gradient was harvested in 1.5 ml fractions and an OD260 profile was acquired in order to determine which fractions contained nucleic acids. Figure 1 shows data from a representative fractionation. The larger peak (left) represents free nucleic acids while the smaller peak (right) represents membrane-associated nucleic acids. Based on this profile it is straightforward to decide which fractions to pool for the membrane-associated and free categories.

Figure 1. OD260 measurements for a subcellular fractionation.

The reproducibility of the assay has been tested in a number of ways. First, we have found similarly strong separations between membrane-associated and free mRNAs in two different organisms (see manuscript Figure 4). Secondly, we have achieved remarkably parallel results in experiments from a number of different human cell lines, including lymphoid cells and fibroblasts. The calibration curves (as those in Figure 4a,b of the manuscript) for four different experiments are shown in Figure 2.

Figure 2. Sliding-window algorithm-generated callibration curves for 4 different fractionations of human cell lines. The experiments shown here were performed on Jurkat T cells, MOLT4 T cells, and MCF7 fibroblasts.

Thirdly, we have performed the fractionation on the same cell line multiple times. Figure 3 below shows the scatter plot of normalized Cy5/Cy3 ratios for genes that were well measured in two independent fractionations of the MOLT4 T cell line. The scatter plot shows good correlation between the experiments, especially for genes with large or small Cy5/Cy3 ratios. The R value for a linear regression fit of the data is 0.82. The experiments were done under different conditions (specifically flask size, culture density, and time on ice before lysis) and this likely accounts for many of the differences between the two experiments. Since only those mRNAs that are being translated at the time of fractionation will remain associated with the rough ER, the translational state of the cells is an important factor in determining which membrane-associated transcripts will be highly enriched in the membrane fraction. Of the 729 genes that were well measured in both experiments and that were enriched in the membrane fraction at the 80% confidence interval in Experiment #1, 566 (or 78%) were also enriched in the membrane fraction at the 80% confidence interval in experiment #2. Thus, the fractionation shows good reproducibility, even between slightly different conditions.

Figure 3. Scatter plot for two fractionations of MOLT4 T cells under two different conditions.

Finally, we have performed a strong test of the reproducibility of this method by comparing the results of applying this procedure to two different T-cell derived lines, the Jurkat cells reported in the manuscript, and Molt4 cells (data not reported in manuscript.) In both analyses, similarly efficient and clean separations were achieved. Despite considerable differences in their gene expression profiles, we found that 205 of the 343 genes with Cy5/Cy3 fluorescence ratios corresponding to greater than 85% secreted/membrane proteins in the reported analysis of Jurkat cells were similarly classified in the analysis of MOLT4 cells, while only 1 of these 343 genes was classified as cytosolic/nuclear with greater than 85% probability. Thus, the experimental procedure was not only reproducible in separate cell-fractionations and microarray hybridizations, but also in the analyses of different cell lines.