Classify sample based on RNAseq data and micro-array data

[francisb]

I have a sample on which I will perform RNAseq. I would like to compare its transcriptome to the transcriptomes of several cell types and find the cell type it resembles most. The cell type data is, however, micro-array data.

Does anyone have an idea how I can correlate my RNAseq data with the micro-array data? How can I include biological replicates of my RNAseq experiment, and how many replicates would I need? (I do not need any statistical significance, I just need to pick the best match)

Thanks a lot in advance!

[khughitt]

Hi Francis,

Assuming there aren't any publicly available microarray datasets for your cell type of interest, you may be able to make your RNA-Seq data look more like microarray data by log-transforming it and applying voom.

Then, an easy place to start to look for relationships between the cell types would be to perform look at a PCA plot of all of the samples. In this case, the more closely related samples should group together in the plot.

HTH,
Keith

[mbblack]

Since you need to compare to microarray data, why are you not performing microarray experiments with your data? If you read through the literature you will find that actual relative expression or estimated fold change does not correlate well at all between the two technologies. Nor, at least much of the time, do the two technologies even detect the same sets of significantly differentially expressed genes (the overlap in genes can often be less than 50%).

If the intent is to try to identify the cell type by matching expression patterns, then I'd say you should at least be comparing directly similar data. Go ahead and do RNAseq with your samples if you want that data for other types of analysis, but don't try and compare RNAseq from an unknown cell type to microarray data of known cell types. I think any conclusions based on that sort of comparison will be highly criticized and far too ambiguous. Expression as measured by the two technologies is not really comparable across different cell types (or tissues or species, etc), not for pattern matching as you are proposing.

As for how many replicates. how many were used in the microarray experiments? What was the variance in expression observed across those (i.e. how variable are these cell types when it comes to gene expression)? Are these different cell strains, cell lines, or actual wild type cell samples?

[francisb]

Thanks for your replies!

@mbblack Good question. The plan was to do RNAseq because we want to do other analyses as well, and we prefer RNAseq for those. If micro-array is needed to address this question, then I will consider doing that as well.

@khughhitt Thanks for this suggestion! How should I include variability across replicates in this type of analysis?

[khughitt]

When using voom, create a factor for each of your cell types, and possibly another one if you have differing batches across the samples, e.g.:

R code

celltype = c("a", "a", "a", "b", "b", "b")
model = model.matrix(~0+celltype) 

# Or, if you have multiple batches, 
# model = model.matrix(~0+celltype+batch)

Next, you will want to perform some normalization (e.g. size factor or quantile normalization) on the counts, and call voom passing in the above model:

voom_result = voom(exprs(normed_counts), model, plot=TRUE)

# Your transformed counts
head(voom_result$E)

Just to be clear, I haven't actually made this kind of comparison yet, so please take everything with a grain of salt and spend some time reading up on Voom and the differences between microarray/RNA-Seq. I have used the above approach to transform RNA-Seq data for use with Microarray methods with good success, so it seems like a reasonable place to start for your goals.
Keith

Good luck!