How to account for cohort stratification?

[vanshikabidhan]

Hello! Thanks for developing the tool! I went through your example scripts, and I felt that this wasn't really clarified in the script.

So, I am doing a case-control analysis whereby I run mvNMF on separate inputs for my cases and controls. Using the original input and the exposures from a de-novo run, I identified the clusters stratifying my data. At the same time, I do not find any outliers in my data. How should I account for cohort stratification while interpreting the signatures? Should I run the mvNMF separately for the stratified cohorts?
I also observed that in the latter case, if I re-run a de-novo discovery on my stratified data (let's say all the samples from cluster 0), and re-run the pre-processing script again (using these new exposures and stratified input), the function further stratified the data into more clusters. Is this expected? Could you please share some insights?

Feel free to correct me if I got something wrong.

Thanks a lot!
Vanshika

[Hu-JIN]

Hi, thanks for the question and apologies for the late response.

You are getting to an important yet subtle question in signature discovery related to when to or when not to stratify the cohort to subsets. One the one hand, stratification can sometimes help increase the sensitivity of discovering signatures that are only present (at appreciable levels) in a subset of the cohort (see the example in Extended Figure 6 of our paper). On the other hand, stratification certainly reduces the number of samples for signature discovery (# of samples in subsets < # of samples in total cohort), which could hurt us in terms of power. So whether to do it and to what level we do it really depend on the samples you are working with. It usually helps comparing the de novo discovery results from the total cohort with those from the stratified subsets. A judgement can be made based on whether new signatures are discovered from the subsets.

Hope this helps! Let me know if this still does not clear things up. I'd happy to provide further assistance given more info on the data.