Nice problem, its solution would be helpful. I don't think there is any, at present. Most data are from tissues or from cell lines not from distinct cell populations within tissues. White blood cells make an exception. Since most genes are expressed in many cell types you would need some discretization of the expression levels, deciles or quintiles would probably be suitable. One could start with cell line data keeping in mind, that cell lines are far from what happens in the organism.

Have you thought about using the data in the GNF Expression Atlas? It has a mixture of data from tissues and cell-type specific expression, but has good coverage for immune cell types, if those are your targets. Alternatively, you could try to mine the Exon Array data from the ENCODE project for cell-line specific expression, provided their target cell lines overlap with your favorite targets.

We are looking into this problem as well. I think as Ulrich pointed out it will be challenging to find "hard" cell type markers (meaning not expressed in any other cell type). So the context would matter a great deal. Consider for example liver. You would have Kupfer cells, hepatocytes, smooth muscle cells, endothelial cells, maybe some immune cells, etc. If you can find or generate expression data on the same platform for all those cell types (not impossible, but hard) you can run ANOVA and select genes that might appear in just one of the cell types and not in the others. That would give you the set you look for, however it will be relevant only in the tissue context from which you select the available cell types.

Next, you can look at these paper:

http://www.nature.com/nmeth/journal/v8/n11/abs/nmeth.1710.html

http://web.cbio.uct.ac.za/~renaud/CRAN/web/CellMix/vignettes/CellMix_appnote.pdf

CellMix also exists as an R project:

https://r-forge.r-project.org/projects/cellmix/

Basically these methods would allow you to check what is the cell composition of your sample and maybe decompose the expression to that of the originating cell types.

Hope this helps.

Thanks. cellmix and similar approaches are interesting. Nonetheless I'd *love* to see a simple ontology based on significant expression / lack thereof. As for the tissue / cell type and context problem -- well, that is what ontology is for, one could for example have a node "Kupfer cell" connected to the node "liver tissue" (etc).

I'm on the verge of doing the work myself, I have a pretty good idea how to combine data (or, rather, information) from different platforms. No time, though.

Having just a few notions in biology, I wonder how many genes can actually be related to particular cell types in a significant way. I know no such ontology, but maybe our Gene Expression Atlas might be useful: http://www.ebi.ac.uk/gxa/

I don't think this resource exists; you would have to generate this type of GO descriptor de novo.

I think it is theoretically possible to arrive at a set of bar-coding gene expression patterns for a cell, however what is a pure cell to develop the bar code from is a more difficult question. Typically I limit this kind analysis to explainable tissues. if I am differentiated ESC then I would look for the three germ layer lineages, but no point looking of liver or lung cell signatures.

So first order is to create a set of "pure" cell expression profiles, then learn a signature of gene that discriminate them. Your false positives will come from not having enough pure cell type signature for comparison.

Submarainian present a paper using about 1500 gene to predict the expression patterns of the remaining 20K genes in the genome. He had an 80% accuracy. This is analogous to your approach. You just need to determine how many genes are required to give you a signature set. For every cell in the body from all developmental phases this is likely 1500, but for a more limited scope maybe 20-30, like early endoderm lineage specification. You would need to integrate it with MGI's expression data and data from Emage that maps developmental gene expression patterns in situ. Learning from just array data will only be as good as the array data and the sample annotation. The array data is likely better than the annotation.

Hi, I made some attempts to this a number of years ago. I made gene sets based on microarray data which I used for GSEA analysis. Examples attached. Maybe this is of help.

Here another gene set collection on tissue specific genes. The strange extension .gmx is due to its use in GSEA.