correlation coefficient versus DEGs analysis: what's the best approach for low expressed genes?

Question

I have 6 folders. Each one contains 7 datasets of a specific type of cancer (RNA-Seq) and 7 datasets of normal tissue (healthy control for that type of cancer). A total of 84 datasets.

I want to investigate my gene of interest in each type of cancer.

I believe the best strategy is to merge the FPKM columns and perform correlation analyses of my gene of interest (GOI) versus all other genes in both cancer and normal tissues. As a result, I will have 14 correlation tables, one for each type of cancer and one for each control tissue. Then, I can investigate the pathways related to the genes most correlated with my GOI in each type of cancer and write my biological interpretation.

Am I in the right direction? Is this a good strategy to answer to my question?

I could also compare the control and cancer samples to investigate the significant DEGs (fold change > 1) and the pathways activated by these DEGs, but my gene of interest doesn't have a strong pattern of gene expression, so I'm afraid it won't be in the list of statistically significant DEGs, that's why I'm preferring the correlation analysis approach.

Any other idea?

The controls were collected from the same patient and processed at the same time. I have the gene counts and can use EdgeR, but what should I do if my GOI doesn't appear among the statistically significant DEGs (fold change > 1)? How can I investigate the role of one single gene in a RNA-Seq dataset if it is not differentially expressed? It can have a role in cancer, even if it is not an important DEG. That's why I thought correlation analyses could give me interesting information.

Answer 1

First of all, RNASeq is extremely sensitive to batch effects. Are these matched controls processed at the same time as the tumors? IF they are not, lots of your differences will be caused by batch effect, and not cancer.

I believe the best strategy is to merge the FPKM columns and perform correlation analyses of my gene of interest (GOI) versus all other genes in both cancer and normal tissues.

What you propose isn't publishable. Use the software that people have been using and validating for years. Use either DESeq2 or EdgeR, or limma if you don't have access to raw gene counts.

If the underlying counts are low, the correlation won't mean much, because your real values could be quite off. If your gene has 2 counts in one sample, and 3 counts in another, that does not mean it really correlates with a gene that has 500 reads in the one sample and 750 in the other. I don't know that you can trick your way around it; you need more reads, or more samples for more power.