Does anyone have an idea on methods to deconvolute expression data? Often in gene expression studies a bulk sample is profiled, that actually contains multiple cell types, making it hard afterwards to know what actually caused the signal. Especially in cancer it is well known that stroma such as infiltrating immune cells play important roles in its development however it is often tedious and expensive to generate cell type specific expression data. I'm looking for ideas that are completely de novo or methods that start from known relative contributions of each cell type.
For the in-silco based methods, the paper we published (Shen-Orr et al. Nature Methods 2010) - thank you Steve for the reference!)., requires the relative cell-type subsets of each sample (though there are complementary methods to estimate the cell-freq themselves from the gene expression data itsefl too). The output is cell-type specific differential expression which effectively removes the very strong bias which comes merely from variations in cell-freq.
If this is of interest to anyone, we just released a Microsoft Excel Add-in for the R package as well. Available here:
http://shenorrlab.technion.ac.il/Data-and-Tools
Olivier,
If you could sample a wide range of samples with varying contributions of cell types, it should in principle be possible to find the transcripts most closely associated with the cancerous cells using artificial neural networks. cf Chapman et al 2011 in Molecular Ecology for an example of how this can be done for chemical contaminants
This is a very interesting and ambitious undertaking. We have used laser capture methods in the laboratory to isolate cell types prior to microarray analysis but that is not an "in silico" methodology and can be quite expensive. A simple approach would be to analyze the data using gene sets unique to a subset of the cell types in the tissue. Perhaps you could then speculate on the contributions of cell type " A" to the expression of common genes. I imagine we could create a relational network with some type of weighting for the contribution from the each cell type based on published data ? GEO may be helpful resource. Of course there are so many confounding factors! I will be interested to see how this thread develops.
If you have an estimate of the amount of tumor or stromal infiltrate, then use it as a covariate in the expression analysis. transformation of the covariate may be necessary. For example, log(%infiltrate). If you are using an ANOVA analysis of some factor, then this is equivalent to asking about the effect of that factor above and beyond the linear effects of the covariate.
Hello Olivier,
There are a couple of methods out to "deconvolute" gene expression data.
There is the classical PCA that you could apply or correspondence analysis and (weighted) spectral map analysis. All methods try to disentangle a "sample effect" and a "gene effect". The named methods are implemented in the R package "mpm" (http://cran.r-project.org/web/packages/mpm/index.html).
Further you can use independent component analysis (ICA), which could be very interesting for gene expression data, because it would detect independent sources of variation in the data, that could be interpreted as different pathways that are switched on or of in different samples. Implemented in the R package "fastICA" (http://cran.r-project.org/web/packages/fastICA/index.html).
In a project, where cell lines were treated with different compounds a biclustering approach was very successful. Biclustering would find a subset of genes that have a similar behaviour in a subset of samples. So some compounds would affect some genes. These could be genes of one pathway. FABIA is the best method for biclustering of gene expression data: http://www.bioconductor.org/packages/2.10/bioc/html/fabia.html
Regards,
Günter
There was a paper in the last few years from Tibshirani;
Cell type-specific gene expression differences in complex tissues.
Shen-Orr SS, Tibshirani R, Khatri P, Bodian DL, Staedtler F, Perry NM, Hastie T, Sarwal MM, Davis MM, Butte AJ.
Nat Methods. 2010 Apr;7(4):287-9. Epub 2010 Mar 7. which speaks to this from a computational viewpoint.
Understood, but what should be the start up data for such analysis.
For the in-silco based methods, the paper we published (Shen-Orr et al. Nature Methods 2010) - thank you Steve for the reference!)., requires the relative cell-type subsets of each sample (though there are complementary methods to estimate the cell-freq themselves from the gene expression data itsefl too). The output is cell-type specific differential expression which effectively removes the very strong bias which comes merely from variations in cell-freq.
If this is of interest to anyone, we just released a Microsoft Excel Add-in for the R package as well. Available here:
http://shenorrlab.technion.ac.il/Data-and-Tools
To answer your question unambiguously is to do single cell expression profiling using RT-qPCR and NGS after FACS or MACS.
Thank you all for very interesting suggestions. I also want to add that the Shen-Orr study also showed that for the majority of genes, adding cell types in a mix has linear effects. Meaning that the linear assumption for deconvolution is reasonable. However, their method should be applied in linear space, before log transformation as pointed out by Zhong & Liu.
@Toni-Ann yes this is an approach that I have seen where patterns of expression from individual cell types are used when the relative frequencies are unknown. I found this paper an excellent review on this topic: "Gene expression deconvolution in clinical samples" by Zhao & Simon, in genome medicine, 2011.
@Steven thanks for putting this paper in this thread, I was aware of this work, but did not want to bias this question from the start.
@Shai Shen-Orr, thanks for sharing the updates about csSAM, it's interesting to know that you are still working on this area and providing tools to implement it.
@Michael What you are suggesting is going to correct for infiltration of one subtype, but is it also possible to get cell type specific expression levels if you have more than 2 cell types in the mix.
@Günter thanks for suggesting these methods. I will check them out for sure. Is the second approach an unsupervised approach? So does it attempt to find unknown variations in the data? The FABIA method also looks great since I'm also interested in bi-clustering methods.
I'm not sure whether it will help you but here is the link for CTen (Cell type enrichment) analysis. http://www.ncbi.nlm.nih.gov/pubmed/22953731. Hope you find some...
Olivier,
All mentioned methods are unsupervised approaches. I attached a pdf with an R script that performs all the analyses on the famous Golub data set.
There are some papers from Tom Freeman using network analysis to deal with that kind of problem.
In http://www.ncbi.nlm.nih.gov/pubmed/20580463, http://www.ncbi.nlm.nih.gov/pubmed/21635249 and http://www.ncbi.nlm.nih.gov/pubmed/20211243 there are some work done.
Hope it helps
Edgardo
- Badges
- Science topic