3
$\begingroup$

I am looking for advice on how to best approach a problem I am faced with.

I have a dataset of numerous degradative biomarkers, clinical information and various other measures (from clinical trials). I would like to see how different biomarkers relate to each other, and effect their levels (high/low etc).

My goal is to gain an understanding of how these biomarker levels fluctuate in different disease profiles - and understand if there is a relationship between each of them (co/independent). The limitation is that I do not have access to genetic or environmental data so this will be a "random effect". There are around 1000 marker samples and 10's of biomarkers mixed between clinical and biochemical. There are also numerous time points to be incorporated.

I have thought about implementing something similar to a protein-protein interaction network, although this will of course not involve physical interactions, but more pathway interactions for further hypothesis driven research.

I wanted to firstly: validate if this project actually makes sense and secondly: ask for some advice around how to best implement this.

I have experience implementing neural/deep nets, directed and undirected networks (Bayes) in python, but not sure exactly how to go about this.

Thanks!

Improve this question
$\endgroup$
4
  • $\begingroup$ First of all it is important to know very well what kind of data you have and the limitations and advantages it has. It is also equally important to know your goal: which is your goal? Find the best biomarker? I was starting to answer, but then I thought that asking this won't damage my answer. Also which is the number of samples for each data type? $\endgroup$
    – llrs
    Jul 13, 2017 at 19:28
  • $\begingroup$ Yes agreed! My goal is to gain an understanding of how these biomarker levels fluctuate in different disease profiles - and understand if there is a relationship between each of them (co/independent). The limitation is that I do not have access to genetic or environmental data so this will be a "random effect". Its not necessarily "best" biomarker as different diseases will have different marker profiles :) there are around 1000 marker samples and 10s of biomarkers mixed between clinical and biochemical. There are also numerous time points to be incorporated. $\endgroup$
    – JB1
    Jul 14, 2017 at 7:38
  • $\begingroup$ reddit.com/r/bioinformatics/comments/6n0x2s/… Someone alse answered here :) $\endgroup$
    – JB1
    Jul 14, 2017 at 7:39
  • 1
    $\begingroup$ It would be fine if you edit the post to include the information from the previous comment and from reddit in your post. It will improve the question and make it self contained $\endgroup$
    – llrs
    Jul 14, 2017 at 10:32

1 Answer 1

Reset to default
2
$\begingroup$

You can start by looking at the correlation (if data is non-parametric) of each variable with each variable of each data point you have. One thing that might change is the correlation between variables along the disease.

Next you can determine which relationships are constant along the disease progression/clinical trial. Identify also the ones that don't show a constant relationship. To do so you can use WGCNA (although with just 10 variables is too low to expect a scale free network you can use to correlate the variables and group them by correlation profile)

Then the question is which variables are more important, those who kept the same relationship or those which change? Use the eigengenes of those variables (separately by those that have a constant relationship and those which don't) to estimate the dependency with the parametric variables. Or you simply can do an ANOVA of all the variables with each non-parametric variable.

Improve this answer
$\endgroup$
6
  • $\begingroup$ Thanks for your help, I am perhaps slightly further along the road/ am looking at something slightly different. I have looked at correlations between all variables be they biomarkers or end points etc. What (for me) is important to look at are co-dependencies of biomarkers across the whole system - all be it only 10 markers. I am here in the process of literature research :) arxiv.org/pdf/1501.04731.pdf $\endgroup$
    – JB1
    Jul 14, 2017 at 11:22
  • $\begingroup$ If you read carefully my answer. correlations are the start of the process to find the relationship between markers and profiles $\endgroup$
    – llrs
    Jul 14, 2017 at 11:27
  • $\begingroup$ Ok i see what you mean. I will have to look at WGCNA to see if it is extendable outside of genetic data. I am happy to implement this myself. I primarily wanted to make sure what I was proposing wasnt stupid (have poele done thins before - yes ) and what others experience was. Thanks for the advice! $\endgroup$
    – JB1
    Jul 14, 2017 at 13:01
  • $\begingroup$ Yes, it is extendable outside genetic data $\endgroup$
    – llrs
    Jul 14, 2017 at 13:01
  • $\begingroup$ @JB1 Perhaps if you wait a bit before selecting my answer as the best one you will get more answers (I doubt it, but is in general recommended). BTW if you find other methods let me know I am also interested in this area and will be doing a similar project soon $\endgroup$
    – llrs
    Jul 14, 2017 at 14:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.