How to prove correlation between gene expression and functions using omics data in bioinformatics?

Question

I am new to bioinformatics, coming from a background in mathematical biology. I am currently participating in a project where the Principal Investigator (PI), who is a biologist, has tasked me with proving a correlation between the expression and functions of two or more elements. Specifically, one of them is a kinase enzyme and the other is mTOR. The PI has asked me to use omics data or databases for this purpose.

As someone with limited experience in bioinformatics, I am struggling to understand how a biological pathway or correlation can be definitively proven, beyond basic statistical methods. Additionally, the ultimate goal of this research is to develop treatments using substances that our lab has discovered, which have shown promise in killing specific cancer cells in murine and cell lines (although our data on this is limited).

My question goes as followes- is it possible to demonstrate a connection between these elements using only databases and omics methods? Why might a dynamic modeling approach, such as ordinary differential equations (ODEs) not be enough?

I am feeling quite confused and would greatly appreciate any guidance or direction on how to approach this problem effectively.

Thank you.

Answer 1

In general terms data mining is needed, then supervised learning and finally ODE, because you are performing DE against enzyme kinetics and thats doable, but you'll need a lot of biochemical knowledge.

I would first take the 'omics' data sets and perform a PCA analysis on it and see where your mTOR and kinase genes are located. PCA is known as unsupervised learning and can be accessed by Seurat or better using Scanpy packages. That will help establish the overall behaviour of your key enzymes pre- and post- treatment ... against the control.

Basically you want:

• Data sets of cancer mice - pre-treatement
• Data set of cancer mice under treatment (drugs targeting kinases)
• Healthy controls

See what unsupervised learning looks like between the 3 groups and where your enzymes are. The PCA plots then compared for the 3 groups and you want to see the targets you are interested in changing groups, especially the treatment group shifting between the control and the diseased unsupervised plots. This would provide a basis for exploring supervised learning (outlined below).

---

Supervised learning What you will do is look at gene expression and train it against your phenotype of interest (cancer and recovery from cancer). The training 'target' is the drug the mice are being treated with for each given data set, obviously they'll be a "untreated" and "control" categories. If your model trains correctly >0.85 to 0.9 accuracy ... post-back and I'll explain how to extract the answer. This approach will extract the biochemical pathway for how the cancer treatment works, but requires more work than described here and that provides the framework for the DE.

Again, a lot of serious knowledge of biochemistry is needed to deploy ODE, unless you've got a brilliant a priori - then you could dive straight in. To do that you'd need to know the exact mechanism of how these enzymes are involved in cancer and then parse that data from your data sets (totally different approach to what is described above). The problem with this approach is if someone knows both the maths and genomics (like a reviewer or examiner) and you've made an error, that will that will be a problem. If you've been through an orthodox approach (above), its a lot stronger justification for the ODE, thus difficult to criticise.

Note The key is what data set types comprise 'omics', I'm assuming this is RNA-based (thats normal), but it might not be.