I have 40 cancer cell lines that were treated with a specific drug, and drug response was quantified with an appropriate numeric score ranging form 0 to 1. I have gene expression data (untreated) for all 40 cell lines and I wanted to create a regressive randomforest model in order detect features (genes) associated with drug sensitivity. For feature selection I wanted to used rfe, Boruta or others. Basically, a classic scenario.
Here is my problem:
For all the cell lines, I have both RNA-Seq (Illumina) and Microarray (AFFY HGU133P2) gene expression data. RNASeq has been properly processed and normalized for library size in order to be used for RandomForest. (STAR+Cuffquant+Cuffnorm).
1 SCENARIO: R SQUARED ASSOCIATED TO SCORE
Separately, for both RNASeq and Micorarray, I measured the correlation between gene/probe expression and response score. I calculated the r squared and the r - associated p-value.
1883 genes have p60%+) between RNASeq and Microarray which is more than I expected, great.
2 SCENARIO: EXTRACTING FEATURES WITH RANDOMFOREST
When using Boruta for feature extraction, or other more relaxed feature selection methods, despite the very good overlap that I mentioned before, I end up with 2 different signatures that overlap by only 10% of the genes.
In an example, the gene with the highest variable importance in Microarray signature was validated to be associated to drug response in vitro using siRNAs. But this gene is not part of the RNASeq signature, even though it has a high R squared value (expression/score correlation) in both RNASeq and Microarray and was part of the "intersection" mentioned in point 1.
I would like to combine the information coming from both RNASeq and Microarray for selecting final genes in the signature. I was thinking about using only the overlapping genes from point 1 for randomforest model generation, but this way I would loose possibly important transcripts measured only by RNASeq and not microarray. The opposite might also be true in certain conditions.
How should I proceed? What would you do?
Apart from the random forest model you and do GO based analysis separately for both and calculate p valus for each GO class then compare to get common and unique set of enriched GO terms
My first thought is that the overlaps in the original correlations sound reasonable, but you didn't say how big the signatures are (how may genes). Given that more than 1000 genes are well-correlated with outcome, there may be a lot of stochasticity in which ones are selected as most important features, which plays out differently in the two independent sets. If your final selected feature sets are small, then maybe a 10% overlap is exactly what you would expect by selecting features at random from the well-correlated ones?
Just a few ideas:
One idea is to try to bring some biological insight to the party. Microarrays measure intensity values for all probes, even when the probes are detecting lowly expressed transcripts such as transcription factors, whereas RNA-Seq has a much stronger detection bias towards high;y-expressed transcripts. It might be that the array data provides better features at the low expression end than RNA-Seq, and vice versa, and you could check this out from your data to hand. Transcriptions factors may be important correlates of biological activity, as they orchestrate the cellular response, but they may have high variance in the RNA-Seq data.
Another idea might be to carry out feature selection separately on each set, and then use the RNA-Seq data as test data for the array features, and the array data as test data for the RNA, in order to re-evaluate each feature's importance independent of measurement technology.
If you are looking for features that would theoretically work independently of measurement technology, one idea might be to use each data set to generate a ranked list of feature importance, then than the mean of ranks as your criteria for inclusion of features.
Finally, is it possible to use all the data together in a single feature-selection exercise, and see which technology wins for which genes?
I'd be interested to hear how it turns out.
- Badges
- Science topic